© 2018, Society for Imaging Science and Technology. This paper proposes a guitar fingering assessing system based on CNN (Convolutional Neural Network) hand pose estimation and SVR (Support Vector Regression) evaluation. To spur our progress, first, a CNN architecture is proposed to estimate temporal 3D position of 16 joints of hand; then, based on a DCT (Discrete Cosine Transform) feature and SVR, fingering of guitarist is scored to interpret how well guitarist played. We also release a new dataset for professional guitar playing analysis with significant advantage in total number of video, professional judgement by expert of guitarist, accurate annotation for hand pose and score of guitar performance. Experiments using videos containing multiple persons' guitar plays under different conditions demonstrate that the proposed method outperforms the current state-of-art with (1) low mean error (Euclid distance of 6,1 mm) and high computation efficiency for hand pose estimation; (2) high rank correlation (0.68) for assessing the fingering (C major scale and symmetrical excise) of guitarists.

© 2018, Society for Imaging Science and Technology. Towards the actualization of an air bag system for the UAV's crash, this paper proposes a method for estimating the UAV's crash site from the video sequence acquired by the camera attached to the UAV. The crash point can be considered to correspond to the divergence point of the optical flows. In the accumulator, the cells at which the optical flows (straight lines) pass through are incremented by one. After performing this process for all the optical flows, the cell with the largest vote is obtained as the crash point (divergence point) in the image plane. Experiments using a hand held camera show that the accuracy of estimating the crash site is increased as the camera approaches the target plane. Overall, the experimental results are promising.

© 2018 Society for Imaging Science and Technology. Towards the actualization of an automatic guitar teaching system that can supervise guitar players, this paper proposes an algorithm for accurately and robustly tracking the 3D position of the fretboard from the video of guitar plays. First, we detect the SIFT features within the guitar fretboard and then match the detected points using KD-tree searching based matching algorithm frame by frame to track the whole fretboard. However, during the guitar plays, due to movements of the guitar neck or occlusions caused by guitar players' fingers, the feature points on the fretboard cannot always be matched accurately even though applying traditional RANSAC homography. Therefore, by using our modified RANSAC algorithm to filter out the matching error of the feature points, perspective transformation matrix is obtained between the correctly matched feature points detected at the first and other frames. Consequently, the guitar neck is tracked correctly based on the perspective transformation matrix. Experiments show promising results such as high accuracy: the total mean tracking error of only 4.17 mm and variance of 1.5 for the four tracked corners of the fretboard. This indicates the proposed method outperforms related tracking works including state-of-art Fully-convolutional Network.

© 2018, Society for Imaging Science and Technology. Towards the establishment of the preventive medical care for the cerebral aneurysm, this paper proposes an SVM based method for building a discrimination function that classifies the presence or absence of the cerebral aneurysm using the cerebral blood vessel's shape features obtained from medical images such as MR images. Using the discrimination function, this paper explores how much each feature affects the onset of the cerebral aneurysm. This paper deals with the internal carotid artery (ICA). The blood vessel (ICA)'s shape features are extracted from medical images of 18 persons without cerebral aneurysm and 13 patients with a cerebral aneurysm. From the medical image, the cross sections and centerline of the ICA are obtained. The cross sections are divided into nine sections along the centerline. Shape features such as the cross sectional area, its circularity, curvature, torsion, length of the centerline and branch angles are obtained in each section; as a total, 113 features including the mean and variance of some features in each section are used for building the SVM. As a result of conducting the experiments, the accuracy for discriminating the presence/absence of the aneurysm by the SVM is 90.3%. In the obtained discrimination function, the coefficient values of the function can be considered how much the features affect the onset of the aneurysm. The features that could significantly cause the onset of the cerebral aneurysm are clarified, and the reasons why these features are significant are discussed.

© 2018, Society for Imaging Science and Technology. To contribute to the actualization of the care worker assistance robot, this paper proposes a method for detecting whether the care receiver is chewing from the video sequence acquired by the camera that observes that receiver. The proposed method detects the receiver's face and areas for both cheeks and chin. After applying some normalization to the areas, chewing detection that uses a variable-intensity template is performed, where the template consists of shape models, interest points and intensity distribution model. A likelihood based on the variableintensity template is computed so that the receiver is judged whether the receiver is chewing. Experiments using seven subjects are conducted. As a result, the accuracy of chewing detection by the proposed method is 83%, which is quite promising.

To alleviate the recent shortage problem of nurses, the actualization of RSN (Robotic Scrub Nurse) that can autonomously judge the current step of the surgery and pass the surgical instruments needed for the next step to surgeons is desired. The authors developed a computer vision based algorithm that can early-recognize only two steps of suture surgery. Based on the past work, this paper explores the effectiveness of utilizing temporal order of the six steps in suture surgery for the early-recognition. Our early-recognition algorithm consists of two modules: start point detection and hand action early-recognition. Segments of the test video that start from each quasi-start point are compared with the training data, and their probabilities are calculated. According to the calculated probabilities, hand actions could be early-recognized. To improve the early-recognition accuracy, temporal order information could be useful. This paper checks confusions of three steps' early recognition results, and if necessary, early-recognizes again after eliminating the wrong result, while for the other three steps, temporal order information is not utilized. Experimental results show our early-recognition method that utilizes the temporal order information achieves better performances.

This paper explores the effectiveness of applying a deep learning based method to segment the amniotic fluid and fetal tissues in fetal ultrasound (US) images. The deeply learned model firstly encodes the input image into down scaled feature maps by convolution and pooling structures, then up-scale the feature maps to confidence maps by corresponded un-pooling and convolution layers. Additional convolution layers with 1×1 sized kernels are adopted to enhance the feature representations, which could be used to further improve the discriminative learning of our model. We effectively update the weights of the network by fine-tuning on part of the layers from a pre-trained model. By conducting experiments using clinical data, the feasibility of our proposed approach is compared and discussed. The result proves that this work achieves satisfied results for segmentation of specific anatomical structures from US images.

This paper proposes a temporal Grouping and pattern analysis-based algorithm that could track the fingertips of guitarists during their guitar playing towards the actualization of the automatic guitar fingering recognition system. First a machine learning-based Bayesian Pixel Classifier is used to segment the hand area on the test data. Then, the probability map of fingertip is generated on the segmentation results by counting the voting numbers of the Template Matching and Reversed Hough Transform. Furthermore, a temporal Grouping algorithm, which is a geometry analysis for consecutive frames, is applied to removal noise and group the same fingertips (index finger, middle finger, ring finger, little finger). Then, a data association algorithm is utilized to associate 4 tracked fingers (index finger, middle finger, ring finger, little finger) with their correspondent tracked results frame by frame. Finally, particles are distributed only between the associated fingertip candidates to track the fingertips of guitarist effectively. The experimental result demonstrates that this fingertip tracking algorithm is robust enough for tracking fingertips (1) without any constrains such us color marker
(2) under the complex contexts, such us complicated background, different illumination conditions, (3) with the high tracking accuracy (mean error 3.36 pixels for four fingertips).

This paper proposes a temporal Grouping and pattern analysis-based algorithm that could track the fingertips of guitarists during their guitar playing towards the actualization of the automatic guitar fingering recognition system. First a machine learning-based Bayesian Pixel Classifier is used to segment the hand area on the test data. Then, the probability map of fingertip is generated on the segmentation results by counting the voting numbers of the Template Matching and Reversed Hough Transform. Furthermore, a temporal Grouping algorithm, which is a geometry analysis for consecutive frames, is applied to removal noise and group the same fingertips (index finger, middle finger, ring finger, little finger). Then, a data association algorithm is utilized to associate 4 tracked fingers (index finger, middle finger, ring finger, little finger) with their correspondent tracked results frame by frame. Finally, particles are distributed only between the associated fingertip candidates to track the fingertips of guitarist effectively. The experimental result demonstrates that this fingertip tracking algorithm is robust enough for tracking fingertips (1) without any constrains such us color marker
(2) under the complex contexts, such us complicated background, different illumination conditions, (3) with the high tracking accuracy (mean error 3.36 pixels for four fingertips).

Towards the actualization of an autonomous guitar teaching system, this paper proposes the following two video analysis based methods: (1) pressed chord recognition and (2) fingertip tracking. For (1), an algorithm that can extract finger contours and chord changes is proposed so that the chords pressed by the guitar player are recognized. For (2), an algorithm that can track the fingertips by continuously monitoring the appearance and disappearance of the regions of fingertip candidates is proposed. Experimental results demonstrate that the proposed two modules are robust enough under complex contexts such as complicated background and different illumination conditions. Promising results were obtained for accurate tracking of fingertips and for accurate recognition of pressed chords.

The purpose of abstract painters is to let viewers get the various images and abstract images. However, viewers who do not have enough knowledge of art, cannot easily get abstract images. The authors have proposed a music generation system that utilizes viewers’ gazes. It can be expected that the authors’ music generation system can prompt the viewer of abstract paintings to imagine abstract images, which the painter intended to express. This paper explores whether the authors’ music generation system can enhance abstract imaginations of persons who see abstract paintings, by subjective tests. Experiments using 19 subjects and eight abstract paintings were conducted for the two cases in which the subjects see the abstract paintings without hearing any music and while hearing the viewers’ gaze based music generated by the authors’ system. Experimental results imply that “hearing gaze based music” could enhance the viewers’ abstract imagination.

The purpose of abstract painters is to let viewers get the various images and abstract images. However, viewers who do not have enough knowledge of art, cannot easily get abstract images. The authors have proposed a music generation system that utilizes viewers’ gazes. It can be expected that the authors’ music generation system can prompt the viewer of abstract paintings to imagine abstract images, which the painter intended to express. This paper explores whether the authors’ music generation system can enhance abstract imaginations of persons who see abstract paintings, by subjective tests. Experiments using 19 subjects and eight abstract paintings were conducted for the two cases in which the subjects see the abstract paintings without hearing any music and while hearing the viewers’ gaze based music generated by the authors’ system. Experimental results imply that “hearing gaze based music” could enhance the viewers’ abstract imagination.

This paper proposes a method that combines Temporal Modified-RANSAC(TMR) with a fixation-based segmentation algorithm for reconstructing the 3D structure of moving and still objects that are tracked in video and depth image sequences acquired by moving Kinect© and/or range finders First, we compute 3D optical flow of feature points. Second, TMR classifies all flows into consistent 3D flow sets for the background and each moving object
simultaneously, the rotation matrix and the translation vector for each 3D flow set are computed. Then, an improved fixation-based segmentation method segments each object’s area. Finally, dense 3D models for the background and each moving object are constructed along with each object’s rotation matrix and translation vector in each frame. Experiments using multiple moving objects in color and depth image sequences acquired by Kinect(c) demonstrate the effectiveness of our proposed method.

Toward the actualization of an automatic navigation system for fetoscopic tracheal occlusion (FETO) surgery, this paper proposes a 3D ultrasound (US) calibration-based approach that can locate the fetal facial surface, oral cavity, and airways by a registration between a 3D fetal model and 3D US images. The proposed approach consists of an offline process and online process. The offline process first reconstructs the 3D fetal model with the anatomies of the oral cavity and airways. Then, a point-based 3D US calibration system based on real-time 3D US images, an electromagnetic (EM) tracking device, and a novel cones' phantom, computes the matrix that transforms the 3D US image space into the world coordinate system. In the online process, by scanning the mother's body with a 3D US probe, 3D US images containing the fetus are obtained. The fetal facial surface extracted from the 3D US images is registered to the 3D fetal model using an ICP-based (iterative closest point) algorithm and the calibration matrices, so that the fetal facial surface as well as the oral cavity and airways are located. The results indicate that the 3D US calibration system achieves an FRE (fiducial registration error) of 1.49 +/- 0.44 mm and a TRE (target registration error) of 1.81 +/- 0.56 mm by using 24 fiducial points from two US volumes. A mean TRE of 1.55 +/- 0.46 mm is also achieved for measuring location accuracy of the 3D fetal facial surface extracted from 3D US images by 14 target markers, and mean location errors of 2.51 +/- 0.47 mm and 3.04 +/- 0.59 mm are achieved for indirectly measuring location accuracy of the pharynx and the entrance of the trachea, respectively, which satisfy the requirement of the FETO surgery.

Because of the shortage of nurses in the world, the realization of a robotic nurse that can support surgeries autonomously is very important. More specifically, the robotic nurse should be able to autonomously recognize different situations of surgeries so that the robotic nurse can pass necessary surgical tools to the medical doctors in a timely manner. This paper proposes and explores methods that can classify suture and tying actions during suture operations from the video sequence that observes the surgery scene that includes the surgeon's hands.
First, the proposed method uses skin pixel detection and foreground extraction to detect the hand area. Then, interest points are randomly chosen from the hand area so that their 3D SIFT descriptors are computed. A word vocabulary is built by applying hierarchical K-means to these descriptors, and the words' frequency histogram, which corresponds to the feature space, is computed Finally, to classify the actions, either SVM (Support Vector Machine), Nearest Neighbor rule (NN) for the feature space or a method that combines "sliding window" with NN is performed.
We collect 53 suture videos and 53 tying videos to build the training set and to test the proposed method experimentally It turns out that the NN gives higher than 90% accuracies, which are better recognition than SVM. Negative actions, which are different from either suture or tying action, are recognized with quite good accuracies, while "Sliding window" did not show significant improvements for suture and tying and cannot recognize negative actions.

While the Optical Music Recognition (OMR) of printed and handwritten music scores in modern standard notation has been broadly studied, this is not the case for early music manuscripts. This is mainly due to the high variability in the sources introduced by their severe physical degradation, the lack of notation standards and, in the case of the scanned versions, by non-homogenous image-acquisition protocols. The volume of early musical manuscripts available is considerable, and therefore we believe that computational methods can be extremely useful in helping to preserve, share and analyse this information. This paper presents an approach to recognizing handwritten square musical notation in degraded western plainchant manuscripts from the XIVth to XVIth centuries. We propose the use of image processing techniques that behave robustly under high data variability and which do not require strong hypotheses regarding the condition of the sources. The main differences from traditional OMR approaches are our avoidance of the staff line removal stage and the use of grey-level images to perform primitive segmentation and feature extraction. We used 136 images from the Digital Scriptorium repository (DS, 2007), from which we were able to extract over 90% of the staves and over 88% of all symbols present. For symbol classification, we used gradient-based features and SVM classifiers, obtaining over 90% precision and recall over eight basic symbol classes.

Because of the shortage of nurses in the world, the realization of a robotic nurse that can support surgeries autonomously is very important. More specifically, the robotic nurse should be able to autonomously recognize different situations of surgeries so that the robotic nurse can pass necessary surgical tools to the medical doctors in a timely manner. This paper proposes and explores methods that can classify suture and tying actions during suture operations from the video sequence that observes the surgery scene that includes the surgeon's hands. First, the proposed method uses skin pixel detection and foreground extraction to detect the hand area. Then, interest points are randomly chosen from the hand area so that their 3D SIFT descriptors are computed. A word vocabulary is built by applying hierarchical K-means to these descriptors, and the words frequency histogram, which corresponds to the feature space, is computed. Finally, to classify the actions, either SVM (Support Vector Machine), Nearest Neighbor rule (NN) for the feature space or a method that combines sliding window with NN is performed. We collect 53 suture videos and 53 tying videos to build the training set and to test the proposed method experimentally. It turns out that the NN gives higher than 90% accuracies, which are better recognition than SVM. Negative actions, which are different from either suture or tying action, are recognized with quite good accuracies, while Sliding window did not show significant improvements for suture and tying and cannot recognize negative actions. © 2014 SPIE.

This paper explores the effectiveness of prompting abstract paintings' viewers' inspiration and imagination by the authors' gaze based music generation system. The authors' music generation system detects the viewer's gaze by a gaze detection equipment. At each of the gaze staying positions in the painting, the color of that point is converted to the sound so that as the gaze moves, music that consists of the converted time series sounds is generated. Experiments using six subjects and six abstract paintings were conducted for the three cases in which the subjects see the abstract paintings without hearing any music, while hearing pre-selected music and while hearing the viewers' gaze based music generated by the authors' system. The experimental results imply that "hearing gaze based music" could stimulate the viewers' inspiration and imagination best, "hearing pre-selected music" second best, and "without music" third best. © Springer International Publishing Switzerland 2014.

Toward the actualization of an automatic navigation system for fetoscopic tracheal occlusion (FETO) surgery, this paper proposes a 3D ultrasound (US) calibration-based approach that can locate the fetal facial surface, oral cavity, and airways by a registration between a 3D fetal model and 3D US images. The proposed approach consists of an offline process and online process. The offline process first reconstructs the 3D fetal model with the anatomies of the oral cavity and airways. Then, a point-based 3D US calibration system based on real-time 3D US images, an electromagnetic (EM) tracking device, and a novel cones' phantom, computes the matrix that transforms the 3D US image space into the world coordinate system. In the online process, by scanning the mother's body with a 3D US probe, 3D US images containing the fetus are obtained. The fetal facial surface extracted from the 3D US images is registered to the 3D fetal model using an ICP-based (iterative closest point) algorithm and the calibration matrices, so that the fetal facial surface as well as the oral cavity and airways are located. The results indicate that the 3D US calibration system achieves an FRE (fiducial registration error) of 1.49±0.44mm and a TRE (target registration error) of 1.81±0.56mm by using 24 fiducial points from two US volumes. A mean TRE of 1.55±0.46 mm is also achieved for measuring location accuracy of the 3D fetal facial surface extracted from 3D US images by 14 target markers, and mean location errors of 2.51±0.47 mm and 3.04±0.59 mm are achieved for indirectly measuring location accuracy of the pharynx and the entrance of the trachea, respectively, which satisfy the requirement of the FETO surgery. © 2014 The Institute of Electronics, Information and Communication Engineers.

Since long time ago, speech recognition has been researched, though it does not work well in noisy places such as in the car or in the train. In addition, people with hearing-impaired or difficulties in hearing cannot receive benefits from speech recognition. To recognize the speech automatically, visual information is also important. People understand speeches from not only audio information, but also visual information such as temporal changes in the lip shape. A vision based speech recognition method could work well in noisy places, and could be useful also for people with hearing disabilities.
In this paper, we propose an automatic lip-reading method for recognizing the speech by using multimodal visual information without using any audio information such as speech recognition. First, the ASM (Active Shape Model) is used to track and detect the face and lip in a video sequence. Second, the shape, optical flow and spatial frequencies of the lip features are extracted from the lip detected by ASM. Next, the extracted multimodal features are ordered chronologically so that Support Vector Machine is performed in order to learn and classify the spoken words. Experiments for classifying several words show promising results of this proposed method.

Recently, a minimally invasive surgery (MIS) called fetoscopic tracheal occlusion (FETO) was developed to treat severe congenital diaphragmatic hernia (CDH) via fetoscopy, by which a detachable balloon is placed into the fetal trachea for preventing pulmonary hypoplasia through increasing the pressure of the chest cavity. This surgery is so dangerous that a supporting system for navigating surgeries is deemed necessary. In this paper, to guide a surgical tool to be inserted into the fetal trachea, an automatic approach is proposed to detect and track the fetal face and mouth via fetoscopic video sequencing. More specifically, the AdaBoost algorithm is utilized as a classifier to detect the fetal face based on Haar-like features, which calculate the difference between the sums of the pixel intensities in each adjacent region at a specific location in a detection window. Then, the CamShift algorithm based on an iterative search in a color histogram is applied to track the fetal face, and the fetal mouth is fitted by an ellipse detected via an improved iterative randomized Hough transform approach. The experimental results demonstrate that the proposed automatic approach can accurately detect and track the fetal face and mouth in real-time in a fetoscopic video sequence, as well as provide an effective and timely feedback to the robot control system of the surgical tool for FETO surgeries.

A new approach is proposed for estimating illuminant colors from color images under an unknown scene illuminant. The approach is based on a combination of a gray-world-assumption-based illuminant color estimation method and a method using color gamuts. The former method, which is one we had previously proposed, improved on the original method that hypothesizes that the average of all the object colors in a scene is achromatic. Since the original method estimates scene illuminant colors by calculating the average of all the image pixel values, its estimations are incorrect when certain image colors are dominant. Our previous method improves on it by choosing several colors on the basis of an opponent-color property, which is that the average color of opponent colors is achromatic, instead of using all colors. However, it cannot estimate illuminant colors when there are only a few image colors or when the image colors are unevenly distributed in local areas in the color space. The approach we propose in this paper combines our previous method and one using high chroma and low chroma gamuts, which makes it possible to find colors that satisfy the gray world assumption. High chroma gamuts are used for adding appropriate colors to the original image and low chroma gamuts are used for narrowing down illuminant color possibilities. Experimental results obtained using actual images show that even if the image colors are localized in a certain area in the color space, the illuminant colors are accurately estimated, with smaller estimation error average than that generated in the conventional method. © 2013 SPIE.

In fetoscopic tracheal occlusion (FETO) surgery, the localization of a flexible tool has an important role in easing high demands for surgeons. In this paper, a point-based 3D ultrasound (US) calibration system is proposed for localizing the surgical tool, based on real-time 3D US images, an electromagnetic (EM) tracking device, and a novel cones' phantom. A 3D US probe attached with an EM sensor is used to acquire the 3D US images of the surgical tool
meanwhile, an EM transmitter is fixed at a certain point in the environment as the origin of the world coordinate system. From these 3D US images, the skeleton of the surgical tool is extracted, and 15 labeled points along the surgical tool are then distinguished, whose coordinates in the world coordinate system are calculated by the proposed 3D US calibration system. The results demonstrate that our proposed approach achieves high localization accuracy, i.e. the average TRE of 1.99 ± 0.49 mm and the maximum TRE of 2.61 ± 0.49 mm, which satisfy the requirement of the FETO surgery: less than 3.85 mm. © Springer-Verlag 2013.

With the wide clinical application of 3D ultrasound (US) imaging, automatic location of fetal facial features from US volumes for navigating fetoscopic tracheal occlusion (FETO) surgeries becomes possible, which plays an important role in reducing surgical risk. In this paper, we propose a feature-based method to automatically detect 3D fetal face and accurately locate key facial features without any priori knowledge or training data. The candidates of the key facial features, such as the nose, eyes, nose upper bridge and upper lip are detected by analyzing the mean and Gaussian curvatures of the facial surface. Each feature is gradually identified from the candidates by a boosting traversal scheme based on the spatial relations between each feature. In experiments, all key feature points are detected for each case, and thus a detection success rate of 100% is achieved by using 72 3D US images from a test database of 6 fetal faces in the frontal view and any pose within 15 degrees from the frontal view, and the location error 3.18 +/- 0.91 inn of the detected upper lip for all test data is obtained, which can be tolerated by the FETO surgery. Moreover, this system has a high efficiency and can detect all key facial features in about 625 ms on a quadcore 2.60 GHz computer.

An investigation into encouraging higher user self-disclosure during text-based Computer Mediated Communication (CMC) system known as the Prompter Communication system (PCS), proposed by Yonemura et al. in previous studies, in which 1 to 3 black circular prompter images ("•") are embedded onto the background of a text-field canvas as stimuli and as an embryonic basis for conversation, have been investigated and evaluated. The effects on supporting the projection or reflection of generated mental representations onto messages with the "•" prompter image as part of the representations within the message contents have also been examined. During the investigation, the concepts and features of the Geneplore cognitive creative thinking process model, together with self-projections, were integrated into the proposed system. Based on these concepts and features, the PCS aims to encourage users to express their mental representations by projecting or reflecting them onto the messages in combination with the provided "•" prompter image as part of the message contents. Communication experiments using the prototype PCS system have been conducted, where a Normal Communication System, NCS which uses only a plain text-field canvas without the "•" prompter image, is compared with the PCS. The experimental results indicate that communication conducted using the PCS results in an increased number of messages with the "•" prompter image(s) involved as part of the message contents than as background images. Further analysis into the self-disclosure level of the message contents and the relationships between the degree of self-disclosure and the representations of the "•" prompter image(s) within the messages have indicated that the degrees of self-disclosure increases when the "•" prompter image(s) are used to represent the participant him/herself, or the participant's thoughts and emotions, and decreases when the "•" prompter image(s) are used to represent animals or objects, or when the "•" prompter image(s) presented within the messages have been ignored. Investigations into the differences in self-disclosure level between message contents from both the PCS and the NCS have shown that using the PCS results in a generally higher degree in the average frequency of the appearance of high self-disclosure information than when using the NCS. During the conversation, these messages contain information with a high level of self-disclosure, which, consequently, may further lead to the development of intimate and deep interpersonal relationships.

In fetoscopic tracheal occlusion (FETO) surgery, the localization of a flexible tool has an important role in easing high demands for surgeons. In this paper, a point-based 3D ultrasound (US) calibration system is proposed for localizing the surgical tool, based on real-time 3D US images, an electromagnetic (EM) tracking device, and a novel cones' phantom. A 3D US probe attached with an EM sensor is used to acquire the 3D US images of the surgical tool
meanwhile, an EM transmitter is fixed at a certain point in the environment as the origin of the world coordinate system. From these 3D US images, the skeleton of the surgical tool is extracted, and 15 labeled points along the surgical tool are then distinguished, whose coordinates in the world coordinate system are calculated by the proposed 3D US calibration system. The results demonstrate that our proposed approach achieves high localization accuracy, i.e. the average TRE of 1.99 ± 0.49 mm and the maximum TRE of 2.61 ± 0.49 mm, which satisfy the requirement of the FETO surgery: less than 3.85 mm. © Springer-Verlag 2013.

A new approach is proposed for estimating illuminant colors from color images under an unknown scene illuminant. The approach is based on a combination of a gray-world-assumption-based illuminant color estimation method and a method using color gamuts. The former method, which is one we had previously proposed, improved on the original method that hypothesizes that the average of all the object colors in a scene is achromatic. Since the original method estimates scene illuminant colors by calculating the average of all the image pixel values, its estimations are incorrect when certain image colors are dominant. Our previous method improves on it by choosing several colors on the basis of an opponent-color property, which is that the average color of opponent colors is achromatic, instead of using all colors. However, it cannot estimate illuminant colors when there are only a few image colors or when the image colors are unevenly distributed in local areas in the color space. The approach we propose in this paper combines our previous method and one using high chroma and low chroma gamuts, which makes it possible to find colors that satisfy the gray world assumption. High chroma gamuts are used for adding appropriate colors to the original image and low chroma gamuts are used for narrowing down illuminant color possibilities. Experimental results obtained using actual images show that even if the image colors are localized in a certain area in the color space, the illuminant colors are accurately estimated, with smaller estimation error average than that generated in the conventional method. © 2013 SPIE.

Recently, a minimally invasive surgery (MIS) called fetoscopic tracheal occlusion (FETO) was developed to treat severe congenital diaphragmatic hernia (CDH) via fetoscopy, by which a detachable balloon is placed into the fetal trachea for preventing pulmonary hypoplasia through increasing the pressure of the chest cavity. This surgery is so dangerous that a supporting system for navigating surgeries is deemed necessary. In this paper, to guide a surgical tool to be inserted into the fetal trachea, an automatic approach is proposed to detect and track the fetal face and mouth via fetoscopic video sequencing. More specifically, the AdaBoost algorithm is utilized as a classifier to detect the fetal face based on Haar-like features, which calculate the difference between the sums of the pixel intensities in each adjacent region at a specific location in a detection window. Then, the CamShift algorithm based on an iterative search in a color histogram is applied to track the fetal face, and the fetal mouth is fitted by an ellipse detected via an improved iterative randomized Hough transform approach. The experimental results demonstrate that the proposed automatic approach can accurately detect and track the fetal face and mouth in real-time in a fetoscopic video sequence, as well as provide an effective and timely feedback to the robot control system of the surgical tool for FETO surgeries.

With the wide clinical application of 3D ultrasound (US) imaging, automatic location of fetal facial features from US volumes for navigating fetoscopic tracheal occlusion (FETO) surgeries becomes possible, which plays an important role in reducing surgical risk. In this paper, we propose a feature-based method to automatically detect 3D fetal face and accurately locate key facial features without any priori knowledge or training data. The candidates of the key facial features, such as the nose, eyes, nose upper bridge and upper lip are detected by analyzing the mean and Gaussian curvatures of the facial surface. Each feature is gradually identified from the candidates by a boosting traversal scheme based on the spatial relations between each feature. In experiments, all key feature points are detected for each case, and thus a detection success rate of 100% is achieved by using 72 3D US images from a test database of 6 fetal faces in the frontal view and any pose within 15 degrees from the frontal view, and the location error 3.18 +/- 0.91 inn of the detected upper lip for all test data is obtained, which can be tolerated by the FETO surgery. Moreover, this system has a high efficiency and can detect all key facial features in about 625 ms on a quadcore 2.60 GHz computer.

This paper proposes a method for reconstructing 3D detailed structures of internal organs such as gastric wall from endoscopic video sequences. The proposed method consists of the four major steps: Feature-point-based 3D reconstruction, 3D point cloud stitching, dense point cloud creation and Poisson surface reconstruction. Before the first step, we partition one video sequence into groups, where each group consists of two successive frames (image pairs), and each pair in each group contains one overlapping part, which is used as a stitching region. Fist, the 3D point cloud of each group is reconstructed by utilizing structure from motion (SFM). Secondly, a scheme based on SIFT features registers and stitches the obtained 3D point clouds, by estimating the transformation matrix of the overlapping part between different groups with high accuracy and efficiency. Thirdly, we select the most robust SIFT feature points as the seed points, and then obtain the dense point cloud from sparse point cloud via a depth testing method presented by Furukawa. Finally, by utilizing Poisson surface reconstruction, polygonal patches for the internal organs are obtained. Experimental results demonstrate that the proposed method achieves a high accuracy and efficiency for 3D reconstruction of gastric surface from an endoscopic video sequence.

This paper proposes a method for tracking and recognizing the white line marked in the surface of the road from the video sequence acquired by the camera attached to a walking human, towards the actualization of an automatic navigation system for the visually handicapped. Our proposed method consists of two main modules: (1) Particle Filter based module for tracking the white line, and (2) CLAFIC Method based module for classifying whether the tracked object is the white line. In (1), each particle is a rectangle, and is described by its centroid's coordinates and its orientation. The likelihood of a particle is computed based on the number of white pixels in the rectangle. In (2), in order to obtain the ranges (to be used for the recognition) for the white line's length and width, Principal Component Analysis is applied to the covariance matrix obtained from valid sample particles. At each frame, PCA is applied to the covariance matrix constructed from particles with high likelihood, and if the obtained length and width are within the above-mentioned ranges, it is recognized as the white line. Experimental results using real video sequences show the validity of the proposed method.

This paper improves the authors' conventional method for reconstructing the 3D structure of moving and still objects that are tracked in the video and/or depth image sequences acquired by moving cameras and/or range finder. The authors proposed a Temporal Modified-RANSAC based method [1] that (1) can discriminate each moving object from the still background in color image and depth image sequences acquired by moving stereo cameras or moving range finder, (2) can compute the stereo cameras' egomotion, (3) can compute the motion of each moving object, and (4) can reconstruct the 3D structure of each moving object and the background. However, the TMR based method has the following two problems concerning the 3D reconstruction: lack of accuracy of segmenting into each object's region and sparse 3D reconstructed points in each object's region. To solve these problems of our conventional method, this paper proposes a new 3D segmentation method that utilizes Graph-cut, which is frequently used for segmentation tasks. First, the proposed method tracks feature points in the color and depth image sequences so that 3D optical flows of the feature points in every N frames are obtained. Then, TMR classifies all the obtained 3D optical flows into regions (3D flow set) for the background and each moving object; simultaneously, the rotation matrix and the translation vector for each 3D flow set are computed. Next, Graph-Cut using the energy function that consists of color probability, structure probability and a-priori probability is performed so that pixels in each frame are segmented into object regions and the background region. Finally, 3D point clouds are obtained from the segmentation result image and depth image, and then the point clouds are merged using the rotation and translation from the N-th frame prior to the current frame so that 3D models for the background and each moving object are constructed with dense 3D point data.

Human gesture as a natural interface plays an utmost important role for achieving intelligent Human Computer Interaction (HCI). Human gestures include different components of visual actions such as motion of hands, facial expression, and torso, to convey meaning. So far, in the field of gesture recognition, most previous works have focused on the manual component of gestures. In this paper, we present an appearance-based multimodal gesture recognition framework, which combines the different groups of features such as facial expression features and hand motion features which are extracted from image frames captured by a single web camera. We refer 12 classes of human gestures with facial expression including neutral, negative and positive meanings from American Sign Languages (ASL). We combine the features in two levels by employing two fusion strategies. At the feature level, an early feature combination can be performed by concatenating and weighting different feature groups, and PLS is used to choose the most discriminative elements by projecting the feature on a discriminative expression space. The second strategy is applied on decision level. Weighted decisions from single modalities are fused in a later stage. A condensation-based algorithm is adopted for classification. We collected a data set with three to seven recording sessions and conducted experiments with the combination techniques. Experimental results showed that facial analysis improve hand gesture recognition, decision level fusion performs better than feature level fusion.

Human gesture as a natural interface plays an utmost important role for achieving intelligent Human Computer Interaction (HCI). Human gestures include different components of visual actions such as motion of hands, facial expression, and torso, to convey meaning. So far, in the field of gesture recognition, most previous works have focused on the manual component of gestures. In this paper, we present an appearance-based multimodal gesture recognition framework, which combines the different groups of features such as facial expression features and hand motion features which are extracted from image frames captured by a single web camera. We refer 12 classes of human gestures with facial expression including neutral, negative and positive meanings from American Sign Languages (ASL). We combine the features in two levels by employing two fusion strategies. At the feature level, an early feature combination can be performed by concatenating and weighting different feature groups, and PLS is used to choose the most discriminative elements by projecting the feature on a discriminative expression space. The second strategy is applied on decision level. Weighted decisions from single modalities are fused in a later stage. A condensation-based algorithm is adopted for classification. We collected a data set with three to seven recording sessions and conducted experiments with the combination techniques. Experimental results showed that facial analysis improve hand gesture recognition, decision level fusion performs better than feature level fusion.

The characteristics of line drawing video communication are described in this paper. In the experiment, an original video and a line drawing video were compared. Subjects performed the same task using both video systems and their performances were assessed in terms of the number of utterances in the conversation and the subject's switching pause. The subject's conversation strategy was extracted based on gaze measurement. In the case of the subujects who often gaze at a partner's eyes during conversation, the switching pause for the line drawing video was significantly longer than for the original video. In both video formats, subjects who does not gaze at a partner's eyes had more switching pauses and utterances than the subjects who do.

This paper proposes a gray world assumption based method for estimating an illuminant color from an image by hue categorization. The gray world assumption hypothesizes that the average color of all the objects in a scene is gray. However, it is difficult to estimate an illuminant color correctly if the colors of the objects in a scene are dominated by certain colors. To solve this problem, our method uses the opponent color properties that the average of a pair of opponent colors is gray. Thus our method roughly categorizes the colors derived from the image based on hue and selects them one by one from the hue categories until selected colors satisfy the gray world assumption. In our experiments, we used three kinds of illuminants (i.e., CIE standard illuminants A and D-65, and a fluorescent light) and two kinds of data sets. One data set satisfies the gray world assumption, and the other does not. Experiment results show that estimated illuminants are closer to the correct ones than those obtained with the conventional method and the estimation error for both using CIE standard illuminants A and D-65 by our method are within the barely noticeable difference in human color perception.

While Optical Music Recognition (OMR) of modern printed and handwritten documents is considered a solved problem, with many commercial systems available today, the OMR of ancient musical manuscripts still remains an open problem. In this paper we present a system for the OMR of degraded western plainchant manuscripts in square notation from the XIV to XVI centuries. The system has two main blocks, the first one deals with symbol extraction and recognition, while the second one acts as an error detection stage for the first block outputs. For symbol extraction we use widely known image-processing techniques, such as Sobel filtering and Hough Transform, and SVM for classification. The error detection stage is implemented with a hidden Markov model (HMM), which takes advantage of a priori knowledge for this specific kind of music.

The use of gesture as a natural interface plays an utmost important role for achieving intelligent Human Computer Interaction (HCI). Human gestures include different components of visual actions such as motion of hands, facial expression, and torso, to convey meaning. So far, in the field of gesture recognition, most previous works have focused on the manual component of gestures. In this paper, we present an appearance-based multimodal gesture recognition framework, which combines the different groups of features such as facial expression features and hand motion features which are extracted from image frames captured by a single web camera. We refer 12 classes of human gestures with facial expression including neutral, negative and positive meanings from American Sign Languages (ASL). We combine the features in two levels by employing two fusion strategies. At the feature level, an early feature combination can be performed by concatenating and weighting different feature groups, and LDA is used to choose the most discriminative elements by projecting the feature on a discriminative expression space. The second strategy is applied on decision level. Weighted decisions from single modalities are fused in a later stage. A condensation-based algorithm is adopted for classification. We collected a data set with three to seven recording sessions and conducted experiments with the combination techniques. Experimental results showed that facial analysis improve hand gesture recognition, decision level fusion performs better than feature level fusion.

We propose a Tensor Decomposition based algorithm that recognizes the observed action performed by an unknown person and unknown viewpoint not included in the database. Our previous research aimed motion recognition from one single viewpoint. In this paper, we extend our approach for human motion recognition from an arbitrary viewpoint. To achieve this issue, we set tensor database which are multi-dimensional vectors with dimensions corresponding to human models, viewpoint angles, and action classes. The value of a tensor for a given combination of human silhouette model, viewpoint angle, and action class is the series of mesh feature vectors calculated each frame sequence. To recognize human motion, the actions of one of the persons in the tensor are replaced by the synthesized actions. Then, the core tensor for the replaced tensor is computed. This process is repeated for each combination of action, person, and viewpoint. For each iteration, the difference between the replaced and original core tensors is computed. The assumption that gives the minimal difference is the action recognition result. The recognition results show the validity of our proposed method, the method is experimentally compared with Nearest Neighbor rule. Our proposed method is very stable as each action was recognized with over 75% accuracy.

This paper proposes a Temporal Modified-RANSAC based method that can discriminate each moving object from the still background in the stereo video sequences acquired by moving stereo cameras, can compute the stereo cameras' egomotion, and can reconstruct the 3D structure of each moving object and the background. We compute 3D optical flows from the depth map and results of tracking feature points. We define "3D flow region" as a set of connected pixels whose 3D optical flows have a common rotation matrix and translation vector. Our Temporal Modified-RANSAC segments the detected 3D optical flows into 3D flow regions and computes the rotation matrix and translation vector for each 3D flow region. As opposed to the conventional Modified-RANSAC for only two frames, The Temporal Modified-RANSAC can handle temporal images with arbitrary length by performing the Modified-RANSAC to the set of a 3D flow region that classified in the latest frame and new 3D optical flows detected in the current frame iteratively. Finally, the 3D points computed from the depth map in all the frames are registered using each 3D flow region's matrix to the initial positions in the initial frame so that the 3D structures of the moving objects and still background are reconstructed. Experiments using multiple moving objects and real stereo sequences demonstrate promising results of our proposed method.

This paper proposes a GPU based method that can visualize voxelized surface data with fine and complicated features has high rendering quality at interactive frame rates and provides low memory consumption The surface data is compressed using run length encoding (RLE) for each level of detail (LOD) Then the loop for the rendering process is performed on the GPU for the position of the viewpoint at each time instant The scene is raycasted in planes where each plane is perpendicular to the horizontal plane in the world coordinate system and passes through the viewpoint For each plane one ray is cast to rasterize all RLE elements intersecting this plane starting from the viewpoint and ranging up to the maximum view distance This rasteritization process projects each RLE element passing the occlusion test onto the screen at a LOD that decreases with the distance of the RLE element from the viewpoint Finally the smoothing of voxels in screen space and full screen anti aliasing is performed To provide lighting calculations without storing the normal vector inside the RLE data structure our algorithm recovers the normal vectors from the rendered scene s depth butter After the viewpoint changes the same process is re executed for the new viewpoint Experiments using different scenes have shown that the proposed algorithm is faster than the equivalent CPU implementation and other related methods Our experiments further prove that this method is memory efficient and achieves high quality results.

This paper proposes a method for detecting ellipses from an image despite (1) multiple colors within the ellipses, (2) partially occluded ellipses' boundaries, (3) noisy. locally deformed boundaries of ellipses, (4) presence of multiple objects other than the ellipses in the image, and (5) combinations of (1) through (4). After boundary curves are obtained by edge detection, by utilizing the first-order difference curves of the edge orientation of each pixel in the boundary curves, a segment-reconnect method obtains boundary clusters. Then, a modified RANSAC detects ellipses by choosing five pixels randomly from the boundary clusters, where overlapped ellipses are merged. Experimental results using synthesized images and real images demonstrate the effectiveness of the proposed method together with comparison with the Randomized Hough Transform, a well-known conventional method.

Recognizing hand gestures from the video sequence acquired by a dynamic camera could be a useful interface between humans and mobile robots. We develop a state based approach to extract and recognize hand gestures from moving camera images. We improved Human-Following Local Coordinate (HFLC) System, a very simple and stable method for extracting hand motion trajectories, which is obtained from the located human face, body part and hand blob changing factor. Condensation algorithm and PCA-based algorithm was performed to recognize extracted hand trajectories. In last research, this Condensation Algorithm based method only applied for one person's hand gestures. In this paper, we propose a principal component analysis (PCA) based approach to improve the recognition accuracy. For further improvement, temporal changes in the observed hand area changing factor are utilized as new image features to be stored in the database after being analyzed by PCA. Every hand gesture trajectory in the database is classified into either one hand gesture categories, two hand gesture categories, or temporal changes in hand blob changes. We demonstrate the effectiveness of the proposed method by conducting experiments on 45 kinds of sign language based Japanese and American Sign Language gestures obtained from 5 people. Our experimental recognition results show better performance is obtained by PCA based approach than the Condensation algorithm based method.

In this paper, the end-of-life products (ELPs) collecting and recycling system is studied under the application of extended producer responsibility (EPR) programs. Economic performance is evaluated based on the decomposition and analysis of major cost and revenues components during the recycling process. We identified two problematic scenarios where EPR programs cannot run smoothly. The according prosperous incentive management strategies are discussed as the corresponding solution.

This paper proposes a method that identifies and tracks a walking human across discontinuous fields of views of multiple cameras for the purpose of video surveillance. A typical video surveillance system has multiple cameras, but there are several spaces within the surveillance area that are not within any of the camera's field of view. Also, there are discontinuities between the fields of views of adjacent cameras. In such a system, humans need to be tracked across discontinuous fields of views of multiple cameras. Our proposed model addresses this issue using the concepts of gait pattern, gait model, and motion signature. Each human's gait pattern is constructed and stored in a database. This gait pattern spans a tensor space that consists of three dimensions: person, image feature, and spatio-temporal data. A human's gait model can be constructed from the gait pattern using the "tensor decomposition based approach" described in this paper. When human(s) appears in one of the camera's field of a view (which is often discontinuous from the other camera's field of views), the human's motion signature is calculated and compared to each person in the database's gait model. The person with the gait model that is most similar to the motion signature is identified as same person. After the person is identified, the person is tracked within the field of view of the camera using the mean-shift algorithm based on color parameters. We conducted two experiments; the first experiment was identifying and tracking humans in a single video sequence, and experiments, the percentage of subjects that were correctly identified and tracked was better than that of two currently widely-used methods, PCA and nearest-neighbor. In the second experiment was the same as the first experiment but consisted of multiple-cameras with discontinuous views. The second experiment (human tracking across discontinuous images), shows the potential validity of the proposed method in a typical surveillance system.

This paper proposes a modified RANSAC based method that can detect multiple independently moving objects from the video sequences acquired by active stereo cameras, and can estimate the egomotion of the stereo cameras. We first compute 3D optical flows from consecutive frames, where dense 3D optical flows, which are needed for the subsequent egomotion estimation based on the modified RANSAC, are obtained by the process proposed in this paper. From the dense flows, three 3D optical flows are randomly selected for computing an egomotion (model). The validity of that egomotion model is checked using the weight obtained by calculating a consistency between that model and each of the dense flows. If it turns out that the estimated egomotion model is valid, the label that identifies this model is assigned to the inliers, which are consistent with this model, and then the labeled inliers are deleted. The random choice of three flows and its subsequent processes are repeated till one of the termination con itions is met. Using the inliers present in the static background, the real egomotion is computed. Experiments using synthesized and real stereo sequences demonstrate the effectiveness of the proposed method. © 2010, The Institute of Image Electronics Engineers of Japan. All rights reserved.

In this paper, we propose an improved Kernel-based Fuzzy C-means Algorithm (iKFCM) with spatial information to reduce the effect of noise for brain MR image segmentation. We use k-nearest neighbour model and a neighbourhood controlling factor by estimating image contextual constraints to optimize the objective function of conventional KFCM method. Conventional KFCM algorithms classify each pixel in image only by its own gray value, but the proposed method classifies by the gray values of its neighbourhood system. For this reason, the proposed iKFCM has a strong robustness for image noise in image segmentation. In experiments, some synthetic grayscale images and simulated brain MR images are used to assess the performance of iKFCM in comparison with other fuzzy clustering methods. The experimental results show that the proposed iKFCM method achieves a better segmentation performance than other fuzzy clustering methods. © 2010 IEEE.

This paper explores the feasibility of using machined vision based cash register system in a university cafeteria and automatically determine price of the food items that are contained on a cafeteria food tray. The idea is to take an image of the food tray, identify all the food items on the tray, and display the total food cost. A fully automated version of the system is envisioned to use a standard biometric technology for customer identification and for automatic debit process of the food cost from the customer's pre-registered account. The analysis method uses shape and size analysis to identify food items that are served on distinct type dishes and color analysis for those items that are served in similar type dishes. The paper explains prototype of a developed system and experimental results for 10 types of meals served at Ritsumeikan Asia Pacific University Cafeteria to justify the validity of this approach. ©2010 IEEE.

In this paper, the administration-driven WEEE management system in Europe and the market-driven end-of-life automobile management system in the US are studied first in this paper. Successful experiences and deficiencies of each system are investigated and compared. Then the potentials for integrating these two systems under each single-instrument-dominant scenario are identified followed by the discussion on key factors for soundly building an integrated system under various scenarios in general. Based on these achievements, the WEEE management system in China is taken as an example to demonstrate how to practically set up an integrated system in such a specific conditions, so that the application of EPR principle could be better promoted in a more economic efficient and ecological effective manner. © 2010 IEEE.

We develop a framework based approach to extract and recognize hand gestures from the video sequence acquired by a dynamic camera, which could be a useful interface between humans and mobile robots. We use Human-Following Local Coordinate (HFLC) System, a very simple and stable method for extracting hand motion trajectories, which is obtained from the located human face and body part. Hand trajectory motion models (HTMM) are constructed by HFLC and hand blob changing factor. In this paper, we apply a principal component analysis (PCA) based approach to improve the recognition accuracy. For further improvement, temporal changes in the observed hand area changing factor are utilized as new image features to be stored in the database after being analyzed by PCA. Each HTMM in the database is classified into gesture categories, or temporal changes in hand blob changes. We demonstrate the effectiveness of the proposed method by conducting experiments on 51 kinds of sign language based Japanese and American Sign Language gestures obtained from 7 people. Our experimental recognition results show better performance is obtained by PCA based approach than the Condensation algorithm based method.

This paper proposes a system that automatically deforms the original video images into line drawing expression transmitted in both directions so as to achieve two goals: easing the concerns of users privacy and ensuring good enough support by visual channel. We examine the effect of line drawing video for support system on the effectiveness and impression of privacy on a collaboration between a novice user and an operator. As a result, it became clear that concern of the user on privacy or security are greatly reduced by using a line drawing image, and there was no significant difference on efficiency with an original image and a line drawing image. © 2010, The Institute of Image Electronics Engineers of Japan. All rights reserved.

This paper describes about the media which touch off informal text communication by black spot stimulus. A black spot stimulus is shown as a background image of the text field in advance of text input. The purpose of black spot stimulus presentation is to touch off conceiving topic based on the cognitive process model of creative thinking. The prototype system based on this media concept was developed, and the communication experiment was conducted. It turned out as a result of the experiment that these media touched off the idea about making topic strongly in the text communication. Moreover, it was implied that the topic touched off deepens the mutual understanding between speakers. © 2010, The Institute of Image Electronics Engineers of Japan. All rights reserved.

In this paper, we propose a novel ellipse detection method which is based on a modified RANSAC, with automatic sampling guidance from the edge orientation difference curve. Hough Transform family is one of the most popular and methods for shape detection, but the Standard Hough Transform loses its computation efficiency if the dimension of the parameter space gets high. Randomized Hough Transform, an improved version of Standard Hough Transform has difficulty in detecting shapes from complicated, cluttered scenes because of its random sampling process. As a pre-process for random selection of five pixels to be used to build the ellipse's equation, we propose a two-step algorithm: (1) region segmentation and contour detection by mean shift algorithm (2) contour splitting based on the edge orientation difference curve obtained from the contour of each region. In each contour segment obtained by step (2), 5 pixels are randomly selected and the modified RANSAC is applied to the 5 pixels so that an accurate ellipse model is obtained. Experimental result show that the proposed method can achieve high accuracies and low computation cost in detecting multiple ellipses from an image.

This paper proposes a Tensor Decomposition Based method that can recognize an unknown person's action from a video sequence, where the unknown person is not included in the database (tensor) used for the recognition. The tensor consists of persons, actions and time-series image features. For the observed unknown person's action, one of the actions stored in the tensor is assumed. Using the motion signature obtained from the assumption, the unknown person's actions are synthesized. The actions of one of the persons in the tensor are replaced by the synthesized actions. Then, the core tensor for the replaced tensor is computed. This process is repeated for the actions and persons. For each iteration, the difference between the replaced and original core tensors is computed. The assumption that gives the minimal difference is the action recognition result. For the time-series image features to be stored in the tensor and to be extracted from the observed video sequence, the human body silhouette's contour shape based feature is used. To show the validity of our proposed method, our proposed method is experimentally compared with Nearest Neighbor rule and Principal Component analysis based method. Experiments using 33 persons' seven kinds of action show that our proposed method achieves better recognition accuracies for the seven actions than the other methods.

To achieve environments in which humans and mobile robots co-exist, technologies for recognizing hand gestures from the video sequence acquired by a dynamic camera could be useful for human-to-robot interface systems. Most of conventional hand gesture technologies deal with only still camera images. This paper proposes a very simple and stable method for extracting hand motion trajectories based on the Human-Following Local Coordinate System (HFLC System), which is obtained from the located human face and both hands. Then, we apply Condensation Algorithm to the extracted hand trajectories so that the hand motion is recognized. We demonstrate the effectiveness of the proposed method by conducting experiments on 35 kinds of sign language based hand gestures. © 2009 SPIE-IS&amp
T.

Research into creative thinking-support tools and communication is commonly focused on how to develop and share ideas between participants or with others. In this paper, we proposes a creative thinking support method that utilizes randomly generated visual prompter (black circle) image patterns (VP-patterns) and free hand-drawing and writing functions. Concepts and ideas of the research have been explained together with the development of the systems (CSP1 and CSP2). Experiments have been Conducted in order to evaluate the potentials and effectiveness of the system. From the results, a tendency towards inspiring creative ideas by participants has been observed.

hi this paper, an efficient approach to segmentation of different independent motion areas from a moving platform is described. This approach is implemented on a stereo vision system, depth information could be computed by matching feature points between stereo images. For two consecutive frames, ego-motion is estimated from the optical flows, including depth information belonging to the background, which has a larger space distribution comparing to those of independent moving objects. In order to distinguish different motion areas, we proposed a modified version of RANSAC mechanism, which could handle the problem of multiple model extraction in a noisy environment.

This paper proposes a new method for enhancing the contrast of color images based on Wavelet Transform and human visual system. The RGB (red, green, and blue) values of each pixel in a color image are converted to HSV (hue, saturation and value) values. To the V (luminance value) components of the color image, Wavelet Transform is applied so that the V components are decomposed into the approximate components and detail components. The obtained coefficients of the approximate components are converted by a grey-level contrast enhancement technique based on human visual system. Then, inverse Wavelet transform is performed for the converted coefficients so that the enhanced V values are obtained. The S components are enhanced by histogram equalization. The H components are not changed, because changes in the H components could degrade the color balance between the HSV components. The enhanced S and V together with H are converted back to RGB values. The effectiveness of the proposed method is demonstrated experimentally.

This paper proposes an efficient and effective computation method for simulating articulated 3D elastic objects. The articulated elastic object consists of bones, elastic object primitives and joining surfaces. After transforming the bones by inverse kinematics, the elastic object primitives are deformed basically by the BEM (Boundary Element Method). To reduce the computation cost and memory usage and to achieve robust deformations, this paper improves the BEM based module of the conventional method: more specifically, a new method for computing the displacements of the vertices in the joining surfaces. Experimental results show that the proposed model can be simulated 10 times as fast as the conventional model or faster, and that real-time performances such as 60fps can be achieved even for models with 20,000 vertices. High quality simulations of different deformations including expansions and contractions are achieved using character type models.

We present a novel skinned skeletal animation system based on spline-aligned deformations for providing high quality and fully designable deformations in real-time. Our ambition is to allow artists the easy creation of abstract, pose-dependent deformation behaviors that might directly be assigned to a large variety of target objects simultaneously To achieve this goal, we introduce the usage of deformation styles and demonstrate their applicability by our animation system. We therefore enhance spline-skinned skeletal animation with two sweep-based free-form-deformation (FFD) variants. The two FFD variants are pose-dependent, driven by three textures and three curves, which can be designed by the artist. As the three textures are similar to height-maps, their creation is very intuitive. Once designed, the deformation styles can be directly applied to any number of targets for imitating material behaviors of cloth, metal or even muscles. Our GPU based implementation shows promising results for real-time usage, as about 30 Million vertices per second can be animated. The basic spline-skinning even reaches more than twice the speed and gets close to the performance of skeletal subspace deformation (SSD). Furthermore, our method can easily be combined along with other existing deformation techniques as pose space deformation or SSD.

This paper proposes a system that automatically deforms the images transmitted in both directions so as to achieve two goals: easing the concerns of users and ensuring well enough support. We examine here the effect of visual information quality on the effectiveness and impression of collaboration between a novice user and an operator.

Research into thinking-support tools is commonly focused on how to develop and share ideas between participants or with others. In this paper, we propose and develop a communication system that stimulates the thought processes and inspires the creative ideas of participants by using a visual "circle" prompt within the framework of a communication pallet. Experiments have been conducted into methods of stimulating the thought process and inspiring ideas during conversation and the results have been analyzed. From the results, a tendency towards inspiring creative ideas by participants has been observed.

Research into thinking-support tools is commonly focused on how to develop and share ideas between participants or with others. In this paper, we propose and develop a communication system that stimulates the thought processes and inspires the creative ideas of participants by using a visual "circle" prompt within the framework of a communication pallet. Experiments have been conducted into methods of stimulating the thought process and inspiring ideas during conversation and the results have been analyzed. From the results, a tendency towards inspiring creative ideas by participants has been observed.

Presented is a novel skeletal animation system for providing high quality geometric deformations in real-time. Each bone of the skeleton is therefore represented by a spline, rather than using conventional matrix rotation. In our approach, each vertex of the animated character can be influenced by a maximum of three spline-curves, which is sufficient for skinned animation. One spline is parameterized by three control points and a possible twist. As opposed to conventional Bezier curves does our arc-spline rely on trigonometric functions for providing better curvatures. The optimized implementation using the OpenGL-shading language shows very promising results for real-time character animation, as even about 1 Million vertices were able to be transformed at interactive 43 frames per second on a GeForce 7800 GTX graphics card.

We have been researching the relationship between human motion and emotion. In this paper, our purpose is to extract motion features specific to each emotion. We propose a new approach for motion data analysis, which applies the higher order Singular Value Decomposition(HOSVD) direct to the motion data and the wavelet analysis to the synthesized data with SVD. The HOSVD models the mapping between persons and emotions. The model can synthesize a complete data acting with each emotion for a given new person. The wavelet analysis extracts each motion feature from the synthesized data for each emotion. Some experimental results using motion capture data for "gait" action and 6 emotions - "angry, joy, sad and so on" show that our method can synthesize novel gait motions for a person by using the extracted motion elements and can extract some features specific to each emotion.

In this paper we describe a mixed reality-supported interactive viewing enhancement museum display system: Visual-dimension Interact System (VIS). With a transparent interactive inter-face, the museum visitor is able to see, manipulate, and interact with the physical exhibit and its virtual information, which are overlapped on one other. Furthermore, this system provides the possibility for visitor to experience the creation process in an environment as close as possible to the real process. This has the function of assisting the viewer in understanding the exhibit and most importantly, gaining a so-to-speak hands-on experience of the creation process itself leading to a deeper understanding of it.

This paper studies how audiences should be expressed in a Cyber-theater, in which remotely located persons can direct plays as directors, perform as performers and/or see the performances as audiences through a networked virtual environment. It is noted that the audience effect has been widely acknowledged in the real-world theater: that is, the audience reaction has a significant effect on the acting of player and performance of the play itself. However, only a few works relevant to audiences in the Cyber Theater can be seen. This paper studies whether the audience effect exists also in the Cyber-theater. By constructing a system in which two actors are displayed a remotely located audience's avatar in which the audience can display his/her emotional actions, we clarified that interactions between the actors and audience are effective.

This paper applies an algorithm, based on Tensor Decomposition, to a new synthesis application: by using sampled motions of people of different ages under different emotional states, new motions for other people are synthesized Human motion is the composite consequence of multiple elements, including the action performed and a motion signature that captures the distinctive pattern of movement of a particular individual. By performing decomposition, based on N-mode SVD (singular value decomposition), the algorithm analyzes motion data spanning multiple subjects performing different actions to extract these motion elements. The analysis yields a generative motion model that can synthesize new motions in the distinctive styles of these individuals. The effectiveness of applying the tensor decomposition approach to our purpose was confirmed by synthesizing novel walking motions for a person by using the extracted signature.

A method for enabling efficient facial data transmission over the communication network and reconstructing realistic 3D face images is proposed. A conventional method which combines the advantages of the model-based and image-based approaches to reconstruct 3D facial images with high efficiency and precision on a standalone environment is extensively used. In our approach, the face image recognition part and the 3D image reconstruction part, which were conventionally carried out on the same PC, are separated and implemented on two network-connected PCs. Also, by using JPEG compression, the size of image data transmitted over the network is reduced. Our method was implemented on network-connected PCs, and its effectiveness was confirmed. Subjective tests showed that image data could be reduced to around 6% of its original size. © 2005, The Institute of Image Electronics Engineers of Japan. All rights reserved.

This paper proposes a method of tracking a human object by using nonsynchronous multiple-viewpoint images. The proposed method tracks human forms efficiently by using a Kalman filter to integrate observed information which is obtained nonsynchronously from multiple viewpoints. The experimental system is composed of multiple observation nodes, which operate nonsynchronously to process the multiviewpoint images, a tracking node, which tracks the human figure, and the discovering node, which finds the human figure. The image features are matched to the tracking model in the observation node based on the prediction of the observed value which is sent from the tracking node. The image features matched to the model are sent to the tracking node, and the tracking model is updated. The image features which are not matched are sent to the discovering node to find a new human figure. With the proposed approach, it is possible to construct a large-scale tracking system while reducing the deterioration of processing efficiency and the redundancy among observations that occur in synchronous systems. The effectiveness of the proposed method is demonstrated by an experiment using real images. (C) 2004 Wiley Periodicals, Inc.

In this paper we describe a mixed reality supported interactive viewing enhancement museum display system. With a transparent interactive interface, the museum visitor is able to see, manipulate, and interact with the physical exhibit and its virtual information, which are overlapped on one other. Furthermore, this system provides the possibility for the visitor to experience the creation process in an environment as close as possible to the real process. This has the function of assisting the viewer in understanding the exhibit and most importantly, gaining a so-to-speak hands-on experience of the creation process itself leading to a deeper understanding of it.

We have been studying about people's visual impression and image features for texture images in order to clarify the human subjective interpretation mechanism for images[1]. In corresponding image features of human impressions for the images, we found that the impressions for material were bottle-necked. We have studied a new analysis method which gives the impression for material from texture images. Especially, we mainly focused on the properties of visual targets which people can feel tactile sense. In this paper, we propose a new texture analysis method which is based on frequency analysis with 3D texture which is designed for photorealistic rendering. We found that our new method can estimate not only the surface roughness but also the surface softness.

The authors have developed a binocular-type display system that allows digital archives of cultural assets to be viewed in their actual environment. The system is designed for installation in locations where such cultural assets were originally present. The viewer sees buildings and other heritage items as they existed historically by looking through the binoculars. Images of the cultural assets are reproduced by stereoscopic 3D CG in cyberspace, and the images are superimposed on actual images in real-time. This system consists of stereoscopic CCD cameras that capture a stereo view of the landscape and LCDs for presentation to the viewer. Virtual cameras. used to render CG images from digital archives. move in synchrony with the actual cameras, so the relative position of the CG images and the landscape on which they are superimposed is always fixed. The system has manual controls for digital zoom. Furthermore, the transparency of the CG images can be altered by the viewer. As a case study for the effectiveness of this system, the authors chose the Heijyoukyou ruins in Nara, Japan. The authors evaluate the sense of immersion, stereoscopic effect, and usability of the system.

A Dynamic Programming (DP) based algorithm that achieves real-time, high quality stereo-matching is presented. A special coarse to fine approach as well as the MMX based assembler implementation mainly contribute to the reached computation speed. A novel vertical smoothing approach by re-integrating paths inside the DP matching allows the avoidance of noisy horizontal strokes, so that high quality stereo-matching is achieved. The current implementation capable of running at about 30 FPS on an 2.2GHz PC, which is sufficient for the utilization in real-time applications.

This is a position paper reporting an on-going collaboration project between SUNY Binghamton, USA, and Waseda University, Japan, on multimodal information retrieval through exploiting the cognitive synergy across the different modalities of the information to facilitate an effective retrieval. Specifically we focus on image retrieval in the applications where imagery data appear along with collateral text. It is noted that these applications are ubiquitous. We have proposed the Synergistic Indexing Scheme (SIS) to explicitly exploit the synergy between the information of imagery and text modalities. Since the synergy we have exploited between the information of imagery and text modalities is subjective and depends on specific cognitive context, we call this type of synergy as cognitive synergy. We have reported part of the empirical evaluation and are in the process to fully implement the SIS prototype for an extensive evaluation.

As a step towards subjective image retrieval, this paper reports an on-going collaboration project between Waseda University and SUNY Binghamton, on relating texture images to haptic impressions. To grasp the surface height variations, texture images are taken under different illuminations and viewing conditions. Our method applies a new frequency analysis method to the texture images. We evaluate the performances of our feature and other typical conventional features by checking whether texture images are correctly classified into "soft" or "hard" by the SVM (support vector machine) method, where the training data for the SVM are collected by subjective tests. Experimental results show that our texture feature can classify "soft" or "hard" better than the other features.

This paper deals with a method that analyzes a botanical tree's behaviors in real space by a computer vision approach so as to reproduce the analyzed behaviors in virtual space. Instead of applying unstable local tracking to the tree in a video sequence, we estimate the direction and strength of the wind that shakes the tree by a learning based method that classifies the input video sequence into one of the stored winds with different directions and strengths. In the learning phase, sample video sequences are used for constructing the Eigenspace and Fisherspace, which is obtained from Fisher discriminant analysis. In the classification phase, the input video sequence is compared with each of the stored sample sequences so that the direction and strength of the wind are estimated. An interpolation method improves the estimation accuracy. Experimental results demonstrate the effectiveness of the proposed method.

A method for constructing 3D human facial images which maintains the chararcteristics of video input image is proposed. First, a personalized 3D polygonal facial model is constructed from two orthogonal facial images. Video input frames are modified, and the vertices composing the 3D polygons are altered according to the contraction rule of facial muscles. Moreover, the modified input frame is texture mapped directly onto to the forementioned model to improve the quality of the reconstructed image. This method was implemented onto an ordinary PC, and an experiment using five persons' face data showed the effectiveness of the method both quantitatively and subjectively.

This paper suggested an image processing method using a multiple-colored suit for estimating complicated postures of a human body. One of the major problems on the posture estimation when using the silhouette image analysis is the overlap of the body parts' silhouettes. In order to solve this problem, the human wears the multiple-colored suit so that each body part has a different color from the other body parts. Experimental results show the effectiveness of the proposed method. It has confirmed that the complicated overlapped postures can be analyzed through the contour of the human's silhouette and the color information from the body segments that are taken using three directional cameras which are located at the front, left, and right side of the subject.

This paper proposes a computer vision based method that recognizes interactions between human body and object. In two successive frames in a video sequence, our function based contour analysis method checks whether the silhouettes of the human body and object overlap. If they do not overlap, our method judges that the human body and object does not interact. If the two silhouettes overlap, our method checks whether the motion vectors obtained from the areas for the human body and object coincide. If they coincide, our method judges that the human body and object are interacting. Some experimental results show the effectiveness of the proposed method. © 2003 by Springer Science+Business Media New York.

This paper proposes a non-contact type method for estimating human body postures. One of the major problems on the posture estimation using the silhouette image analysis is the overlapping of the body parts' silhouettes. In order to solve this problem, this paper proposes a method for estimating the posture of a human wearing a multiple-colored suit based on color information processing. By analyzing the contour of the human's silhouette, the method judges whether feature points are occluded by another body parts. If the occlusions occurs, color region segmentation is performed in order to know which region is frontal. The feature point in the frontal region is located in the skeleton of the region. Experimental results show the effectiveness of the proposed method.

This paper proposes anew efficient method that can animate botanical trees in 3D realistically. In this paper, a 3D botanical tree model consists of a set of branch segments, to which leaf models are attached. To reduce the amount of computation, instead of calculating the motions of all the branch segments, only the representative segment in each branch is numerically analyzed. The numerical analysis is constrained to a 2D plane so that 3D numerical analysis need not be performed. Concerning the leaf model, a set of four leaves is systematically attached to each branch segment. Experimental results clarify the conditions for real-time, realistic animations of dynamic behaviors of trees.

A method for constructing 3D human facial images which maintains the chararcteristics of video input image is proposed. First, a personalized 3D polygonal facial model is constructed from two orthogonal facial images. Video input frames are modified, and the movements of facial feature points are tracked and measured based on the the contraction rule of facial muscles. The vertices of the 3D polygons are altered accordingly, and the modified input frame is texture mapped directly onto to the forementioned model. This method was implemented using an ordinary CCD camera and computer, and an experiment using one person's data showed that realistic facial images could be reconstructed with high speed and precision.

In this paper, we present a novel representation of the human face for estimating the orientation of the human head in a two dimensional intensity image. The method combines the use of the much familiar eigenvalue based dissimilarity measure with image based rendering. There are two main components of the algorith described here: The offline hierarchical image database generation and organization, and the online pose estimation stage. The synthetic images of the subject's face are automatically generated offline, for a large set of pose parameter values, using an affine coordinate based image reprojection technique. The resulting database is formally called as the IBR (or image based rendered) database. This is followed by the hierarchical organization of the database, which is driven by the eigenvalue based dissimilarity measure between any two synthetic image pair. This hierarchically organized database is a detailed, yet structured, representation of the subject's face. During the pose estimation of a subject in an image, the eigenvalue based measure is invoked again to search the synthetic (IBR) image closest to the real image. This approach provides a relatively easy first step to narrow down the search space for complex feature detection and tracking algorithms in potential applications like virtual reality and video-teleconferencing applications. © 2001 Pattern Recognition Society. Published by Elsevier Science Ltd. All rights reserved.

An efficient method for constructing facial images for use in telecommunication applications is proposed. This method uses a simple 3D feature model, which consists of polygons, which describe the shape of the face, and elastic linear springs, which simulate the natural movements of facial muscles. This method requires only two orthogonal facial images, and could easily be implemented on a relatively low-spec PC. Experimental results showed good results that various facial expressions could be synthesized and displayed from arbitrary directions.

This paper describes the author's recent activities related to computer vision based methods for analyzing human behaviors: more specifically, posture estimation and recognizing interactions between a human body and object. For estimating postures in 3D from multiple camera images, the authors developed a heuristic based method and non-heuristic method. The heuristic based method heuristically analyzes the contour of a human silhouette so that significant points of a human body can be located in each image. The non-heuristic method utilizes a function for analyzing contours without using heuristic rules. Recognizing the interactions exploits the function based contour analysis and motion vector based analysis so that the system can judge whether the human body interacts with the object.

Traditionally most camera-based position estimation systems use only a few points to calibrate cameras. In this paper, we investigate a novel and alternate approach for 3D position estimation by using a larger number of points arranged in a 3D grid. We present an implementation of the active-space indexing mechanism which uses three cameras. Given the corresponding points in camera images, a precise estimation of the position can be obtained. The active-space indexing method can be also used as a spatial filter to eliminate the large number of possible corresponding pairs from consideration. This capability, unique only to the active-space indexing method, provides a tractable algorithm to the otherwise intractable situation. (C) 2001 Academic Press.

This paper overviews our research activities on computer vision based non-verbal information analysis that can be applied to virtual communication environments and human computer interactions. In virtual communication environments, a user's facial expressions and body motions are estimated by computer vision approaches, and the estimated non-verbal information is reproduced in the user's avatar. For human computer interfaces, hand gestures are recognized as pre-defined commands by analyzing multiple camera images that observe the hand. In addition, facial expressions and body gestures are recognized from a time-sequential images by HMM (Hidden Markov Models).

This paper proposes a new real-time method of estimating human postures in three dimensions from trinocular images. The proposed method extracts feature points of the human body by analyzing contours of human silhouettes. The feature points are extracted by using the subtraction images when self-occlusions occur in the silhouette images. Dynamic compensation is carried out with a Kalman filter so that all feature points are tracked. The 3D coordinates of the feature points are reconstructed by considering the geometrical relationship between the three cameras. Experimental results confirm both the feasibility and the effectiveness of the proposed method.

This paper has two parts. In the first part of the paper, we note the property that under the para perspective camera projection model of a camera, the set of 2D images produced by a 3D point can be optimally represented by two lines in the affine space (alpha - beta space). The slope of these two lines are same, and we observe that this constraint is exactly the same as the epipolar line constraint. Using this constraint, the equation of the epipolar line can be derived. In the second part of the paper, we use the "same slope" property of the lines in the alpha - beta space to derive the affine structure of the human face. The input to the algorithm is not limited to an image sequence of a human head under rigid motion. It can be snapshots of the human face taken by the same or different cameras, over different periods of time. Since the depth variation of the human face is not very large, we use the para perspective camera projection model. Using this property, we reformulate the (human) face structure reconstruction problem in terms of the much familiar multiple baseline stereo matching problem. Apart from the face modeling aspect. we also show how we use the results for reprojecting human faces in identification tasks.

Head and face detection and eye tracking in realtime are the first steps for head gesture recognition and/or face expression recognition for a human-computer interaction interface. We propose a two-step approach for eye tracking in video streams.
First, we detect or track a point between the eyes. For this task, we apply a special filter rye proposed before. Once we detect the point between the eyes, it is fairly easy to locate the eyes, which are the two small darkest parts on each side of this point. Because defecting the point between the eves is easier and more stable than directly detecting the eyes, the system can robustly track the eyes
We implement the system on a SGI O2,workstation. The video image size is 320x240 pixels. The system processes images at seven frames per second in the detection mode, and 13 frames per second in the tracking mode without arty special hardware.

This paper proposes a skin color information based face location algorithm that can locate the human face in a complex background including non-face skin color regions, such as hands. Our algorithm is divided into three steps. The first step is to use color components of the input image to segment the skin colors from the complex background. The second step is to detect a curve, which is called the division curve, between the hair region and face region. The last step is to use the division curve to estimate the face location. Our algorithm can be applied to some gesture analysis and understanding systems that combine the gestures of the head and hands.

For face detection and tracking in a video sequence, many systems are now applying a technique to extract skin-color areas effaces. In these systems, the skin-color model should be built beforehand. However skin-color changes when the lighting condition varies when an input video camera is changed, when the white-balance of the camera is readjusted, etc.
In this paper we describe a method to build a skin-color model for operation while a face detection system is running. When it has no skin-color model, it uses the differences between adjacent frames to extract moving objects. Once the system recognizes that the moving object is a face, it analyses the color histograms of the moving urea and extract skin-color distribution
We applied this method to our face detection and tracking system. Experiments showed good results.

In this paper, we propose and adaptive human tracking system with non-synchronous multiple observations. Our system consists of three types of processes, discovering node for detecting newly appeared person, tracking node for tracking each target person, and observation node for processing one viewpoint (camera) images. We have multiple observation nodes and each node works fully independently. The tracking node integrates observed information based on reliability evaluation. Both observation conditions (number of cameras), relative distance between a human and cameras, extent of occlusion, etc.) and human motion states (walking, standing, sitting) are considered in the evaluation. Matching between tracking models and observed image features are performed in each observation node based on the position, size and color similarities of each 2D image. Due to the non-synchronous property, this system is highly scalable for increasing the detection area and number of observing nodes. Experimental results for some indoor scenes are also described.

This paper proposes a new real-time method for estimating human postures in 3D from trinocular images. In this method, an upper body orientation detection and a heuristic contour analysis are performed on the human silhouettes extracted from the trinocular images so that representative points such as the top of the head can be located. The major joint positions are estimated based on a genetic algorithm-based learning procedure. 3D coordinates of the representative points and joints are then obtained from the two views by evaluating the appropriateness of the three views. The proposed method implemented on a personal computer runs in real-time. Experimental results show high estimation accuracies and the effectiveness of the view selection process. © 2000 IEEE.

Among head gestures, nodding and head-shaking are very common and used often. Thus the detection of such gestures is basic to a visual understanding of human responses. However it is difficult to detect them in real-time, because nodding and head-shaking are fairly small and fast head movements. We propose an approach for detecting nodding and head-shaking in real time from a single color video stream by directly detecting and tracking a point between the eyes, or what we call the "between-eyes". Along a circle of a certain radius centered at the "between-eyes", the pixel value has two cycles of bright parts (forehead and nose bridge) and dark parts (eyes and brows). The output of the proposed circle-frequency filter has a local maximum at these characteristic points. To distinguish the true "between-eyes" from similar characteristic points in other face parts, we do a confirmation with eye detection. Once the "between-eyes" is detected, a small area around it is copied as a template and the system enters the tracking mode.combining with the circle-frequency filtering and the template, the tracking is done not by searching around but by selecting candidates using the template
the template is then updated. Due to this special tracking algorithm, the system can track the "between-eyes" stably and accurately. It runs at 13 frames/s rate without special hardware. By analyzing the movement of the point, we can detect nodding and head-shaking. Some experimental results are shown. © 2000 IEEE.

This paper proposes a real-time human body posture estimation method using ANNs. The network is composed of three ANNs and a decision logic unit. The ANNs' input is the result of a function analysis on a human silhouette's contour extracted from camera images and the ANNs' output indicates the feature points' positions on the contour. The decision logic unit synthesizes each of the ANNs' output vectors and then the 2D coordinates of the human body's feature points are calculated. The proposed method is implemented on a personal computer and runs in real-time (17-20 frames/sec). Experimental results confirm both the feasibility and the effectiveness of the proposed method for estimating human body postures.

This paper proposes a new real-time method of estimating human postures in 3D from trinocular images. The proposed method extracts feature points Of the human body by applying a type of function analysis to contours of human silhouettes. To overcome self-occlusion problems, dynamic compensation is carried out using the Kalman filter and all feature points are tracked. The 3D coordinates of the feature points are reconstructed by considering the geometrical relationship between the three cameras. Experimental results confirm both the feasibility and the effectiveness of the proposed method, and an application example of the 3D human body posture estimation to a motion recognition system is presented.

This paper proposes a new communication method, called "Virtual Me", that enables a user to simultaneously participate in multiple Events such as meetings or lectures as a virtual-metamorphosed avatar controlled by the user and/or as a computer-controlled autonomous agents. By interpreting non-verbal data generated by the other participants at each event, the system determines when to switch between using virtual metamorphosis or employing agents. We demonstrate the effectiveness of this method in a prototype of the Virtual Me system. It is our hope that this technology may be in the interest not only of time efficiency but also of enriched human communication.

Indexing techniques usually are used in model-based object recognition and ray tracing algorithms. In this paper we present a new method for estimating the three-dimensional position of a subject (resp. object) in a circumscribed space based on an indexing method. We construct two and three-dimensional indices of a space, which are used to estimate the three-dimensional position by an interpolation technique. There are two processes in estimating the three-dimensional position of a subject (resp. object): preprocessing and three-dimensional position estimation. We have implemented this idea using stereo camera, and tested by using two different sizes of a grid pattern. Promising results for preprocessing and 3D position estimation are presented. Moreover, we show that this approach can also be extended for multiple cameras.

In this paper, a new method was proposed for implementing 3-D facial image models designed to allow faithful reconstruction of facial images in virtual space teleconferencing by using 3-D measurement while detecting various facial expressions. In the method proposed, first many dots are painted on the face. Then, for a variety of facial expressions (eight expressions in this study) selected with regard to the action of the major facial muscles, 3-D displacement vectors for the normal (neutral) face are measured at each dot and recorded in the facial image plane as reference vectors. When the facial image is reconstructed, 2-D displacement vectors are detected by tracking markers that were previously placed on the face and are represented as the sum of two enclosing reference vectors. Based on these data, the vertices of a 3-D wireframe face model (WFM) are moved appropriately, and the facial expression is reconstructed. (C) 1999 Scripta Technica.

The virtual metamorphosis system lets people change their forms into any other form in a virtual scene. To realize these changes, a computer Vision system estimates facial expressions and body postures and reproduces them in a computer graphics avatar in real time. We introduce three systems in order of their development: the Virtual Kabuki system, Networked Theater, and "Shall We Dance?".

A method for extracting facial motion parameters is proposed. The method consists of three steps. First, the feature points of the face, selected automatically in the first frame, are tracked in successive frames. Then, the feature points are connected with Delaunay triangulation so that the motion of each point relative to the surrounding points can be computed. Finally, muscle motions are estimated based on motions of the feature points placed near each muscle. The experiments showed that the proposed method can extract facial motion parameters accurately. In addition, the facial motion parameters are used to render a facial animation sequence.

This paper presents a hybrid estimation method of human body postures from CCD camera images. In the hybrid estimation method, the feature points of the human body (top of the head, tips of the hands, and feet, and elbow joints) are obtained from the results of heuristic contour analyses of human silhouettes or those of a time subtraction image depending on the reliability of the silhouette information. A dynamic compensation is then carried out by tracking all feature points using the AR model in order to obtain their optimal position and to overcome self-occlusion problems. The AR model's parameters are estimated through on-line processing by the Kalman filter. The proposed method is implemented on a personal computer and the process runs in real-time. Experimental results show high estimation accuracy and the feasibility of the proposed method.

This paper proposes a new real-time method for estimating human postures in 3D from trinocular images. In this method, an upper body orientation detection and a heuristic contour analysis are performed on the human silhouettes extracted from the trinocular images so that representative points such as the top of the head can be located. The major joint positions are estimated based on a genetic algorithm based learning procedure. 3D coordinates of the representative points and joints are then obtained from the two views by evaluating the appropriateness of the three views. The proposed method implemented on a personal computer runs in real-time (30 frames/second). Experimental results show high estimation accuracies and the effectiveness of the view selection process.

In this paper, we describe a visual surveillance system for evaluating the audience's reaction in meeting sessions. The system, which works in real-time, can recognize and evaluate the reaction of the audience. It is mainly composed of three subsystems. The first subsystem is a face detection and head motion segmentation system which is used to detect the face from complex background and segment the head motion into different units, with each unit including different information about the audience's reaction. The second subsystem is a gesture and pose recognition system which can recognize the gestures and the poses of human head. The third subsystem is an evaluation system which is used to evaluate the reaction of the audience by using the recognition results in the second system. Our system has been tested for determining the concentration of the audience in meeting sessions and the experimental results are good.

This paper proposes a new real-time method for reproducing facial expressions in 3D face models realistically based on anatomy for artists. To reproduce facial expressions in a face model, the detected expressions need to be converted to the data for deforming the face model. In the proposed method, an artist who has learned anatomy for artists creates arbitrary facial expressions in the 3D face model by mixing the reference expressions chosen by the artist so that the synthesized expressions realistically represent the respective expressions displayed by real persons. The parameters obtained by this manual operations are used to construct the equations that convert the expression features obtained by the detection module to the displacement vectors of the vertices of the face model. During human communications through face models, the equations are used to reproduce the detected expressions in real-time. The effectiveness and robustness of the proposed method were demonstrated by experimental results and demonstration systems.

In this paper we present a unified theoretical framework for novel scene synthesis, merging real and virtual worlds, and view stitching. To start with, we have a set of real images from weakly calibrated cameras, for which we compute the dense point match correspondences. For applications like novel view synthesis, one may first solve the 3D scene reconstruction problem, followed by a view rendering process. However, errors in 3D scene reconstruction usually gets reflected in the quality of the new scene generated, so we seek. a more direct method. In this paper, we use the knowledge of dense point matches and their affine coordinate values to estimate the corresponding affine coordinate values in the new scene. Our technique of reprojection is extended for other applications like merging real and synthetic worlds, and view stitching. (C) 1998 Elsevier Science B.V. All rights reserved.

This paper describes the vutual metamorphosis system, whtch was originally proposed by ATR Media Integratton &amp
Communicattons Research Laboratories, Kyoto, Japan. In the vntual metamorphosis system, people can change their forms mto any other form in a vutual scene. To realize. this changes, facial expressions and body postures are estimated by computer vision and reproduced m a computer graphics avatar m real-ume. We developed computer vision based technologies for automatic face trackmg, estimatmg deformations of factal components,and estimatmg body postures in real-time. We also developed real-time facial expresston reproduction method based on "Anatomy for artists". Usmg these technologies, prototyping systems for virtual metamorpho-SIS were constructed. This paper introduces two systems in order of their development: the Virtual Kabuki system and the "Shall We Dance" system.

A new method for extracting human silhouettes from color images in real-time regardless of changes in lighting conditions and backgrounds is proposed. In the method, RGB values of each pixel in color images acquired by a CCD camera are converted to YIQ values. Before extracting human silhouettes, the mean YIQ values of each pixel and the threshold values for the Y, I, Q components are obtained in advance from the image sequence acquired by the camera that observed the background. Based on the mean values and the threshold values, each pixel is judged whether it belongs to the human silhouette or not. The proposed method was used with the virtual metamorphosis system, in which anyone can change his or her form to any other form, and the robustness and real-time-ness were confirmed. © 1998 IEEE.

A method that can be used for spotting segments that display facial expression is proposed. The motion of the face is modeled by HMM in such a way that each state corresponds to the conditions of facial muscles, e.g., relaxed, contracting, apex and relaxing. The probability assigned to each state is updated iteratively as the feature vector is obtained from image processing. A spotted segment is placed into a certain category when the probability of that category exceeds a threshold value. Experiments show that the segments for the six basic expressions can be spotted accurately in near real time. © 1998 IEEE.

In this paper we present a new camera projection model, which is intermediate between the affine camera mode and the pin hole projection model. It is modeled as a perspective projection of 3D points into an arbitrary plane, followed by an affine transform of these projected points. We observe that the reprojection of a point into a navel image can be achieved uniquely provided that we have located a set of Jive reference points over four images (of which three are input images, and the fourth is the novel image). Also, the reprojection, theory; does not assume that the input images are captured from cameras with identical internal calibration parameters. Thus, we apply our technique two different domain: (a) generation of novel images from a stereo pair (b) generation of virtual walkthrough sequence with a monocular image sequence as input.

We have implemented a new algorithm, called the geometric-imprints method for extracting key-features of a given contour. Series of points on the curve are used to find a geometric-imprint set of the contour. The algorithm is based upon the assumption that an articulated human and other figures have mostly cylindrical body-parts. The algorithm recursively partitions the contour based upon the cylindrical nature of human-silhouettes, and a starting point. Similar topological contours are expected to produce similar geometric-imprints. Thus the correspondence problem is somewhat simplified. Results of our implementation are presented.

In this paper we present an algorithm to estimate the human face structure. The input to the algorithm is not limited to an image sequence of a human head under rigid motion. It can be snapshots of the human face taken by the same or different cameras, over different periods of time. Since the depth variation of the human face is not very large, we use the affine camera projection model. Under this assumption, it can be shown that the set of 2D images produced by a 3D point feature of a rigid object can be optimally represented by two lines in the affine space. Using this property, we reformulate the (human) face structure reconstruction problem in terms of the much familiar multiple baseline stereo matching problem [8]. Apart from the face modeling aspect, we also show how we use the results for reprojecting human faces in identification tasks.

At ATR Media Integration & Communications Research Laboratories, our research group is aiming at realizing virtual communication environments in which the users at remote sites can feel that they are virtually co-located. This paper presents the ongoing research in two subareas chosen to address this problem: (a) Generating human images in virtual scenes (b) Act image based approach for generation of (background) scenes. In the virtual metamorphosis system, 3D models of characters are created in advance. For metamorphosis, the facial expressions and body posture of a person are detected in real-time fr om the face images and monocular thermal images of the person, respectively. The detected expressions and body movements are reproduced by deforming the 3D models. Demonstrations using SD models of a Kabuki actor and a dinosaur show good performance. for novel view generation, we use as affine coordinate based re-projection scheme, which is also extended to merging real and synthetic objects.

This paper introduces a new real-lime method to estimate the posture of a human from thermal images acquired by an infrared camera regardless of the background and lighting conditions. Distance transformation is performed for the human body area extracted from the thresholded thermal image for the calculation of the center of gravity. After the orientation of the zipper half of the body is obtained by calculating the moment of inertia, significant points such as the top of the head, the tips oft he hands and foot are heuristically located. In addition, the elbow and knee positions are estimated from the detected (significant) points using a genetic algorithm based learning procedure.
The experimental results demonstrate the robustness of the proposed algorithm and real-time (faster than 20 frames per second) performance.

We propose a novel method of extracting a moving object region from each frame in a series of images regardless of complex, changing background using statistical knowledge about the target. In vision systems for 'real worlds' like a human motion tracker, a priori knowledge about the target and environment is often limited (e.g., only the approximate size of the target is known) and is insufficient for extracting the target motion directly. In our approach, information about both target object and environment is extracted with a small amount of given knowledge about the target object. Pixel value (color, intensity, etc.) distributions for both the target object and background region are adaptively estimated from the input image sequence based on the knowledge. Then, the probability of each pixel being associated with the target object is calculated. The target motion can be extracted from the calculated stochastic image. We confirmed the stability of this approach through experiments.

In this paper we present an algorithm to generate new views of a human face, starting with at least two other views of the face. In a typical face recognition system, the task of comparison becomes easier if the faces have similar orientation with respect to the camera. The affine coordinate based reprojection algorithm presented in the paper enables us to do that. Dense point matches between the two input faces of the same individual are computed using an affine coordinate based reprojection framework. This is followed by the reprojection of one of these to faces to the target face once the user has matched four feature points across two input face images and the target face image.

In this paper we present an algorithm to generate new views of a scene, starting from stereo images. Errors in 3D scene reconstruction usually get reflected in the quality of the new scene generated, so we seek a direct method for reprojection. In this paper, we use the knowledge of dense point matches and their affine coordinate values to estimate the corresponding affine coordinate values in the new scene.

This paper proposes a new real-time method for estimating human body postures from thermal images acquired by an infrared camera, regardless of the background and lighting conditions. Distance transformation is performed for the human body area extracted from the thresholded thermal image, in order to calculate the center of gravity. After the orientation of the upper half of the body is obtained by calculating the moment of inertia, significant points such as the top of the head and the ends of the hands and feet are heuristically located. In addition, the elbow and knee positions are estimated from the detected (significant) points, using a genetic-algorithm-based learning procedure. The experimental results demonstrate the robustness of the proposed algorithm and real-time performance (faster than 20 frames per second).

A new method for estimating human posture from multiple images using a genetic algorithm is proposed. In the proposed algorithm, the posture parameters to be estimated are assigned to the genes of an individual in the population. For each individual, its fitness evaluates to what extent the human multiple images synthesized by deforming a 3D human model according to the values of the genes are registered to the real human multiple images. Genetic operations such as natural selection, crossover and mutation are performed so that individuals in the next generation are generated. After a certain number of repetitions of these processes, the estimated parameter values are obtained from the individual with the best fitness. Experiments using multiple synthesized images show promising results for estimating 17 joint angle values for each degree of freedom of the joints and also the three translational and three rotational degrees of freedom.

A new method to detect deformations of facial parts from a face image regardless of changes in the position and orientation of a face using the genetic algorithm is proposed. Facial expression parameters that are used to deform and position a 3D face model are assigned to the genes of an individual in a population. The face model is deformed and positioned according to the gene values of each individual and is observed by a virtual camera, and a face image is synthesized. The fitness which evaluates to what extent the real and synthesized face images are similar to each other is calculated. After this process is repeated for sufficient generations, the parameter estimation is obtained from the genes of the individual with the best fitness. Experimental results demonstrate the effectiveness of the method. © 1996 IEEE.

This paper considers the problem of extracting and tracking complex contours without user interaction. We assume that a complex contour consists of contour segments whose spatial coordinates and intensity gradient vary smoothly in the direction normal to themselves. In our algorithm, digital curves that could correspond to contour segments are extracted by connecting edge pixels using a B-spline based contour segment model. The extracted curves trace the contour segments at the next frame by using the active contour model technique. Experimental results show even occluded contours can be tracked automatically. © 1996 IEEE.

This paper describes the Virtual Kabuki Theater the authors have recently developed. In the Virtual Kabuki Theater, people at different locations can be Kabuki actors in a Kabuki scene. In our system, Kabuki actors' SD models are created in advance. Facial expressions of a person are detected an real-tame in the face images from the small camera fixed to the helmet worn by the person. Body movements of the person are estimated in real-time from the thermal images acquired by the infrared camera that observes the person. The detected expressions and body movements are reproduced in. the Kabuki actor's model. Our implementation shows good performance. The Virtual Kabuki Theater is a first step towards human metamorphosis systems, in which anyone can change (metamorphose) his/her form into any other characters.

In this paper, we present a method of obtaining the approximate transformation parameter values as a starting point in estimating the pose of rigid 3D free form objects using multiple 2D images. We back project the edge silhouettes in the images, and obtain the approximate volume in the 3D space containing the object. Next, for a point selected in the volume, we hypothesize a set of points within the 3D CAD model of the object it can possibly correspond to, using the spatial extent function introduced in this paper. This is repeated for three arbitrarily chosen point in the volume. The hypothesized (match point) lists of these three points are next used to derive the pose parameter by enforcing the conditions of rigidity. Our initial experiments demonstrate the potential of this idea, and the pose parameters estimated using this method can be refined using the standard methods available in the literature.

This paper describes realistic, real-time 3-D reproduction of human body movements based on detection data from a small number of sensors attached to a participant for virtual space teleconferencing. A 3-D human model is created by a skeletal link model. WFM (Wire Frame Model), body segments for FFD (Free Form Deformations), and IFP (Implicit Function Primitives) for body parts. In this model WFM provides fast rendering ability and FFD and IFP give natural deformation representation. A reconstruction algorithm that can deal with human movements in teleconference situations is proposed
sensed data are used to reconstruct whole body movements by utilizing motion constraints. Experiments, in which four magnetic sensors for detecting the six pose parameters are attached to the upper half of a human body, show the effectiveness of the proposed 3-D human model and reconstruction method, in which human movements are reproduced at a speed of 8 frames/sec.

Real-time reproduction of 3D human images is realized by an experimental system recently built as a prototype for virtual space teleconferencing, in which participants at different sites can feel as if they are colocated and can work cooperatively, At each sending and receiving site of the teleconferencing system, a 3D model of each participant is constructed from a wire frame model mapped by color texture and is rendered on a 3D display, In the current experimental system, real-time detection of facial features at the sending site is achieved by visually tracking tape marks pasted to the participant's face. Movements of the head, body, hands, and fingers are detected in real time using magnetic sensors and data gloves. At the receiving site, the detected motion parameters are used to move nodes in the wire frame model to reproduce the movements of the participants at each sending site. Realistic facial expressions are reproduced through the-use of simple motion rules applied to the tape mark tracking information, Through experimental evaluation, the optimal number of nodes for best quality has been obtained. Reproduction of facial expressions and synthesis of arm movements are examined, The reproduction speed using the optimal human model is approximately 6 frames/s. Examples of cooperative work by participants using the experimental system illustrate the effectiveness of virtual space teleconferencing. (C) 1995 Academic Press, Inc.

An effective algorithm for character recognition in scene images is studied. Scene images are segmented into regions by an image segmentation method based on adaptive thresholding. Character candidate regions are detected by observing gray-level differences between adjacent regions. To ensure extraction of multisegment characters as well as single-segment characters, character pattern candidates are obtained by associating the detected regions according to their positions and gray levels. A character recognition process selects patterns with high similarities by calculating the similarities between character pattern candidates and the standard patterns in a dictionary and then comparing the similarities to the thresholds. A relaxational approach to determine character patterns updates the similarities by evaluating the interactions between categories of patterns, and finally character patterns and their recognition results are obtained. Highly promising experimental results have been obtained using the method on 100 images involving characters of different sizes and formats under uncontrolled lighting.

This paper explores the validity of embedding the image processing based extraction of 10 image features that could describe customers' shopping behaviors to a Bayesian Network based classification method that can recognize the four categories: combinations of with/without hesitation and bought/not bought. By constructing a system that consists of the three modules for tracking customers, extracting the image features and classifying the customers' behaviors, recognition tests using 48 video data were conducted. As a result of the experiments, a recognition accuracy of higher than 89% is achieved, which is promising.

This paper explores the validity of embedding the image processing based extraction of 10 image features that could describe customers' shopping behaviors to a Bayesian Network based classification method that can recognize the four categories: combinations of with/without hesitation and bought/not bought. By constructing a system that consists of the three modules for tracking customers, extracting the image features and classifying the customers' behaviors, recognition tests using 48 video data were conducted. As a result of the experiments, a recognition accuracy of higher than 89% is achieved, which is promising.

This paper studies an automatic method that generates portraits by utilizing a neural network trained by facial features obtained by applying Active Shape Model to original face images and by portraits drawn by an illustrator. The authors' conventional research works had remaining issues such as low subjective similarities of generated portraits, validity of the nose, and absence of subjective evaluation of generated portraits. To increase the subjective similarity, this paper proposes a new portrait depicting tool. We obtain promising feeling that subjective similarity can be improved by the new depicting tool. In addition, we conduct experiments on subjective discrimination of portraits. It turns out that the new tool gives higher subjective discrimination rate than our conventional tool, and that the case of the new tool and presence of the nose gives the highest discrimination rate.

This paper proposes a method that can discriminate feature points of each moving object from the still background by utilizing supervoxel and Multi-label Graph-Cut. The ordinal clustering method based on multi label graph-cut has a problem that can not classify small movement object. To solve this problem, our method utilize initial cluster based on supervoxel and construct multi-labeled-graph by the unit of group that belongs same supervoxel. This improvement enable accurate clustering feature points even if the amount of movement is small. Experiments using multiple moving objects and real stereo sequences demonstrate the effectiveness of our proposed method.

Towards the realization of a robotic nurse that can support surgeries autonomously by recognizing surgical situations only using video informations, this paper proposes an improved method by using sectioned-SIFT and sliding window based neural network that can recognize surgeon's hand actions: suture and tying. Hand area is detected by using color information and then the video sequence is partitioned into sections. Sectioned-SIFT descriptors are computed in each section and built a word vocabulary. Histogram feature of the action is spliced by using word's frequency in each section. Finally, sliding window and neural network is used to recognize the significant actions: suture and tying. The proposed method has achieved the 100% recognition rate for manually extracted actions and 90% recognition rate for whole surgery video sequences.

Towards the realization of a method that can recognize the psychological state that causes the purchasing behaviors of consumers, this paper proposes a method for recognizing customers' behaviors based on analyses of the customers' facial expressions. First, in each frame of the video sequence, facial feature points are extracted by ASM (Active Shape Model), and features are obtained from feature points in the eye-brows and mouth. Then, each feature is assigned to a node of Bayesian Network. On recognition, by inputting a video sequence, the Bayesian Network outputs the behavior recognition result. The validity of the proposed method is presented experimentally.

In this paper, we aim to an automatic tongue diagnosis system in traditional Chinese medicine. First, we propose a new method to extract the tongue area automatically from digital photos using the active contour model based on the watershed segmentation. Second, we classify the following symptoms: the teeth-printed tongue, which could be diagnosed by the contour of tongue, the tongue coating, the fissured tongue, which cloud be diagnosed by the texture feature of tongue, as well as the swelling tongue or thin tongue, which could be diagnosed by the area of tongue. Finally, we confirm the effectiveness of the proposed method by experiments.

We study a system that can automatically generate portraits from face images using multi-layered neural networks, for which image features of the face detected by ASM are the inputs, the portrait drawn by an illustrator are teaching signals, and back-propagation algorithms are performed for learning. This system aims at outputting portraits that do not depend on deformations of the line drawing traced from the face image, and captures the drawing style of an individual artist. The number and locations of feature points detected by ASM could affect quality of generated portraits. This paper reports new results that was not clear in our laboratory conventional research. We study the optimal number and locations of the detected facial feature points in generating portraits.

In the paper, we present a lip-reading method that can recognize speech by using only visual features. Lip-reading can work well in noisy places such as in the car or in the train. In addition people with hearing-impaired or difficulties in hearing can be benefited. First, the Active Shape Model (ASM) is applied to track and detect the face and lip in a video sequence. Second, three visual features, the shape, optical flow and Discreet cosine transformation of the lip are obtained from the lip area detected by ASM. The extracted features are ordered chronologically so that Support Vector Machine (SVM) is performed so as to learn and classify the spoken words. Hierarchical SVMs are used to recognize the words. Each visual feature is trained by the respective middle-layer SVM, and those outputs of SVM's are integrated by the final SVM. Experimental results show that the integration of these features improves the recognition accuracy.

Structure-from-motion (SFM) is a significant issue of 3D reconstruction for many applications. However, some image collector like endoscopy only provides a narrow view angle and 2D structures of the internal organs. In order to expand the view angle and describe the details of 3D structures, we propose an approach to reconstruct 3D surface based on image stitching from a video sequence. There are four major steps of the proposed method: 3D reconstruction, 3D point cloud stitching, dense point cloud generation and Poisson surface reconstruction. Experimental results demonstrate that the proposed method achieves a satisfied accuracy and efficiency for 3D reconstruction, and promise a large view angle.

To achieve a comment system that can automatically estimate emotions of a person who watches web news, this paper proposes a computer vision based method that can recognize facial expressions generated in response to Web news. During the learning session, first, ASM (Active Shape Model) is applied to the face so that some feature points are located. From the located feature points, some features such as distances between feature points are obtained. The features collected from sample data are used for nodes of Bayesian Networks. During facial expression estimation session, the features that are computed from the input facial video sequence by the same procedure as the learning session are input to the Bayesian Networks so that the recognition result is obtained as the facial expression category that maximizes the a-posteriori probability. Experimental results demonstrate the effectiveness of the proposed method.

To achieve a comment system that can automatically estimate emotions of a person who watches web news, this paper proposes a computer vision based method that can recognize facial expressions generated in response to Web news. During the learning session, first, ASM (Active Shape Model) is applied to the face so that some feature points are located. From the located feature points, some features such as distances between feature points are obtained. The features collected from sample data are used for nodes of Bayesian Networks. During facial expression estimation session, the features that are computed from the input facial video sequence by the same procedure as the learning session are input to the Bayesian Networks so that the recognition result is obtained as the facial expression category that maximizes the a-posteriori probability. Experimental results demonstrate the effectiveness of the proposed method.

Structure-from-motion (SFM) is a significant issue of 3D reconstruction for many applications. However, some image collector like endoscopy only provides a narrow view angle and 2D structures of the internal organs. In order to expand the view angle and describe the details of 3D structures, we propose an approach to reconstruct 3D surface based on image stitching from a video sequence. There are four major steps of the proposed method: 3D reconstruction, 3D point cloud stitching, dense point cloud generation and Poisson surface reconstruction. Experimental results demonstrate that the proposed method achieves a satisfied accuracy and efficiency for 3D reconstruction, and promise a large view angle.

Recently reading news through web news media is becoming popular. Most of web news is delivered together with a comment system, which asks the user to rate items such as "the news let you think", and "boring". However, many users do not like manual operations for rating; therefore, not many users actually input the rates. One solution for this issue is to utilize results of recognizing facial expressions from the video sequence acquired by the camera that observes the user. So far, there are very many works on recognizing six fundamental expressions such as sad, surprise and happy, for example, by HMM (Hidden Markov Models). However, not many works dealt with recognizing expressions that could appear when the user watches web news; main difficulty in recognizing these expressions could be caused by the fact that these expressions are spontaneously generated. It is difficult to recognize the spontaneous expressions, which come with tiny movements hard to be detected. This paper proposes a method that aims at recognizing spontaneous expressions.

This paper proposes a modified Active Shape Model (ASM), which is featured by the following two modules: (1)A flexible method that automatically generates training set and does not require time-consuming manual operations, which are required by the conventional ASM; (2)Two-step fitting that consists of the initial global (rough)fitting for the entire face and the subsequent local (precise)fitting for each facial component, where both fittings are based on ASM. Experiments demonstrate that more accurate location results were obtained by the proposed method than the conventional ASM.

This paper proposes a method for estimating the light source direction from RGB image and Depth image (K-D image), which can be acquired by Kinect sensor etc. The proposed method utilizes the property that the 3D reconstructed result, which is obtained by converting the RGB image to gray-level image and then converting that gray-level image to depth image (D-D image) by Khan et al. 's method based on "Deep is dark paradigm", is deformed to the light source direction. The value difference at each pixel of the D-D image and K-D image is obtained, and that pixel's surface normal is computed from the K-D image. In the 2D space defined by the two angle parameters that represent the surface normal's direction, each pixel's difiFerence is added to the cell that corresponds to that pixel's normal's direction. After this process is performed for all the pixels, the light source direction is determined based on the peak in the parameter space. Experiments using some diflferent objects under different light source directions were performed. Promising results were obtained.

For supermarkets and stores, marketing surveillance is a very important issue: in particular, customers' hesitations on purchase, which cannot be recorded by POS systems, are extremely important. This paper studies how to classify the following four categories: "purchased without hesitation", "purchased with hesitation", "not purchased without hesitation", and "not purchased with hesitation". Since these behaviors tend to vary depending on persons and situations, Bayesian network is considered to be a useful method for the classification. By manually extracting image features from video sequences that observe a shop's comer, Bayesian network based classification was explored. Consequently, promising experimental results were obtained.

In recent years, care of aged people's turnovers is heavy burdens for helpers; in particular, automatic judgment of turnovers is desired. This paper proposes a method that can judge whether aged people being cared turned over by utilizing a depth image sequence acquired by Kinect sensor that observes the aged people in beds. Specifically, image features are extracted from aged people's bodies in the depth image sequences. and are used for classification by Support Vector Machine. As the image features, this paper uses time series data of "xyz differences between left and right shoulders and between the left and right edges of the waist". and "average depth in each block obtained by partitioning the bed area into w by h blocks" for exploring classification accuracies. Experimental results show that the former and latter features achieve 99.5% and 89.5% classification accuracies, respectively. The fonner gives a better accuracy, but its applicability is limited. while the latter can be applied for more general cases.

In this paper, we propose an approach that captures a distant person's face image with high resolution by zooming the camera. Like an ordinary security system, cameras observe down from upper positions. The proposed approach extracts human skin color areas and zooms to the estimated position of the face, even if the resolution of the initial face image is so low that the face cannot be recognized. The captured face image is converted to a DCT (Discrete Cosine Transform)based feature vector, and then SVM (Support Vector Machine) judges if the person is a stranger. The proposed approach shares image information obtained by multiple cameras, and limits the number of zooming camera to one based on 3D position of the moving human in the continuous frames. This could achieve that both large area surveillance and face recognition can be performed simultaneously. Experimental results show validity of the proposed approach.

With the wide clinical application of 3D ultrasound (US) imaging, automatic location of fetal facial features from us volumes for navigating fetoscopic tracheal occlusion (FETO) surgeries becomes possible, which plays an important role in reducing surgical risk. In this paper, we propose a feature-based method to automatically detect 3D fetal face and accurately locate key facial features without any priori knowledge or training data. The candidates of the key facial features, such as the nose, eyes, nose upper bridge and upper lip are detected by analyzing the mean and Gaussian curvatures of the facial surface. Each feature is gradually identified from the candidates by a boosted cascade filtering scheme based on the spatial relations between each feature. In experiments, an identification rate of 100% is achieved by using 72 3D US images from a test database of 6 fetal faces in the frontal view and any pose within 15° from the frontal view, and the location error 3.18±0.91mm of the detected upper lip for all test data is obtained, which can be tolerated by the FETO surgery.

Reconstructing the dynamical behavior of a moving non-rigid object such as smoke, water and botanical tree is a tough task because it is very difficult to track the behavior of such objects due to their complicated structure and behavior. This paper studies on tracking the dynamical behaviors of a botanical tree by particle filter based methods. The proposed method uses particle filter, which is also known as Sequential Monte Carlo method, to estimate the position of the target leaf (leaf to be tracked) in each frame of the video sequences. The motion model predicts the position of the target leaf by generating different hypothesis of the target leaf's positions. The observation model corrects the prediction by evaluating each "particle" based on template matching score between the leaf model (template) and the target leaf. To increase the tracking accuracy and decrease the computation cost, a leaf is only searched for in its region of interest (ROI). The ROI position and size are updated by one frame delay method.

For a surgical navigation system with swept-volume 3D ultrasound images, it is necessary to estimate the extracted positions of ultrasound voxels relative to a tracking reference by a 3D ultrasoud calibration. In this paper, we propose a cone based 3D ultrasound calibration system by using an electromagnetic tracking device. The proposed phantom is not expensive, easy to construct, easy to scan, and has a simple experiment setup, while the identification of the tips on each cone from a single ultrasound volume is a manual process, the tracking position of each tip is measured by a pen probe, and the calibration matrix is estimated by SVD (Singular Value Decomposition) method. The precision of the system is experimentally evaluated. It turns out that the maximum distance root mean square (RMS) error is 3.32 mm by using 12 cones, which satisfies the requirement of our fetal surgical navigation system: less than 3.85 mm.

This paper analyzes how our Prompter Communication System (PCS), which has single or several black circular images "●" embedded in the background of the text-field canvas to be used for tele-communication between two users, affects telel-communication users' self-disclosures. To the analysis, this research applies the concepts and features of psychological projection techniques and tests. According to these concepts and features, PCS is expected to enhance and encourage the users' self-disclosure, which is essential to developing intimate and deep interpersonal relationship. Communication experiments by using the prototype system, PCS, are conducted, where a Normal Communication System (NCS), which uses only plain text-field canvas without black circles, is compared with PCS. Two subjects are required to make tele-communication using PCS or NCS. PCS/NCS user writes texts to the canvas with/without randomly generated black circles and sends it (message) to the other user. After tele-communication experiments using PCS and NCS, to each message, for coders give a self-disclosure level, which modifies Daibo et al.'s self-disclosure levels. Consequently, it turns out that PCS gives significantly higher frequencies for the high self-disclosure level, which corresponds to the users' inner thoughts and emotions, than NCS.

For a surgical navigation system with swept-volume 3D ultrasound images, it is necessary to estimate the extracted positions of ultrasound voxels relative to a tracking reference by a 3D ultrasoud calibration. In this paper, we propose a cone based 3D ultrasound calibration system by using an electromagnetic tracking device. The proposed phantom is not expensive, easy to construct, easy to scan, and has a simple experiment setup, while the identification of the tips on each cone from a single ultrasound volume is a manual process, the tracking position of each tip is measured by a pen probe, and the calibration matrix is estimated by SVD (Singular Value Decomposition) method. The precision of the system is experimentally evaluated. It turns out that the maximum distance root mean square (RMS) error is 3.32 mm by using 12 cones, which satisfies the requirement of our fetal surgical navigation system: less than 3.85 mm.

Reconstructing the dynamical behavior of a moving non-rigid object such as smoke, water and botanical tree is a tough task because it is very difficult to track the behavior of such objects due to their complicated structure and behavior. This paper studies on tracking the dynamical behaviors of a botanical tree by particle filter based methods. The proposed method uses particle filter, which is also known as Sequential Monte Carlo method, to estimate the position of the target leaf (leaf to be tracked) in each frame of the video sequences. The motion model predicts the position of the target leaf by generating different hypothesis of the target leaf's positions. The observation model corrects the prediction by evaluating each "particle" based on template matching score between the leaf model (template) and the target leaf. To increase the tracking accuracy and decrease the computation cost, a leaf is only searched for in its region of interest (ROI). The ROI position and size are updated by one frame delay method.

This paper analyzes how our Prompter Communication System (PCS), which has single or several black circular images "●" embedded in the background of the text-field canvas to be used for tele-communication between two users, affects telel-communication users' self-disclosures. To the analysis, this research applies the concepts and features of psychological projection techniques and tests. According to these concepts and features, PCS is expected to enhance and encourage the users' self-disclosure, which is essential to developing intimate and deep interpersonal relationship. Communication experiments by using the prototype system, PCS, are conducted, where a Normal Communication System (NCS), which uses only plain text-field canvas without black circles, is compared with PCS. Two subjects are required to make tele-communication using PCS or NCS. PCS/NCS user writes texts to the canvas with/without randomly generated black circles and sends it (message) to the other user. After tele-communication experiments using PCS and NCS, to each message, for coders give a self-disclosure level, which modifies Daibo et al.'s self-disclosure levels. Consequently, it turns out that PCS gives significantly higher frequencies for the high self-disclosure level, which corresponds to the users' inner thoughts and emotions, than NCS.

An ordinary treatment for Congential Diaphragmatic Hernia is Fetal Endoscopic Tracheal Occlusion (FTO), by which a balloon is placed in the trachea so that the lung can be restored in a natural fashion. However, this treatment required very dangerous surgery; thereby, a supporting system for the surgery is demanded. In this paper, to achieve a system that can automatically navigate a tool to be inserted into the fetal trachea, a method that can detect the mouth of the fetus from the endoscopic video sequence is explored. More specifically, a method that combines ellipse detection by randomized Hough transform with CamShift is proposed. Experimentally, the validity of the proposed method is confirmed.

Viewing the process of drawing caricatured portrait by artist myself introspectively, and it to be a computer program called 'Automatic caricatured portrait generate systems'.

In this paper, we present a new face makeup analysis and synthesis method for a rapid prototyping of various facial makeup styles. By taking an image pair of a person with and without makeup, we first measure the effect of the given makeup. Then, the extracted makeup is represented by using a finite mixture model. Based on this representation, we provide several operations to easily synthesize a variety of new face makeup effects such as transferring, transplanting, exaggeration, and attenuation. We also demonstrate that multiple makeup styles can be effectively blended to generate a new style in our framework.

This paper studies a system that can automatically generate portraits from face images using multi-layered neural networks, for which image features of the face are the inputs, the portrait drawn by an illustrator are teaching signals, and back-propagation algorithms are performed for learning. This system aims at outputting portraits that do not depend on deformations of the line drawing traced from the face image, and reflect human feeling. First, this paper explores the neural networks' structures and relationships between the learning's parameter values and the output results. Second, similarity between the system's outputs and portraits drawn by an illustrator is evaluated by subjective tests, and then, using the evaluation results, the relationship between physical errors (physical quantity) and subjective scores (psychological quantity) is studied.

This paper proposes a method that allows music composition based on abstract paintings. Our approach utilizes gaze behaviors of viewers who see abstract paintings, where temporal changes in gaze positions are utilized as temporal information for music. In this paper, we measured the viewers' gaze behaviors when they see different abstract paintings, and found the following things: (1) durations in which the fix one's gaze on some position are classified into "average length" group and "extremely long" group, and (2) the number of gaze's stay in each element depends on the element's size and distance from the center of the painting. By analyzing mapping between color and sound based on synesthesia, the element at which the gaze stays is converted to sound so that a new music composition method for generating tone clusters is shown.

In this paper, we propose an improved iterative randomized Hough transform (IRHT) method to detect fetal head automatically in ultrasound images. With the update of region of interest (ROI) in the IRHT method, the noise pixels are gradually excluded from the region of interest during iteration process, and the estimation becomes progressively close to the target. In order to enhance the efficiency and stability of this algorithm, we consider introducing the number (N) of pixels on the detected ellipses, and select the ellipse with the maximal number of the pixels on the ellipse as the result for each iteration, which is selected from the top-M peaks in the accumulators of the whole detected ellipse samples. The experiments on fetal ultrasound images demonstrate that the proposed method achieves more robust and accurate results, and has a better performance for fetal head detection than the IRHT method.

The purpose of this study was to examine patterns of facial movements of "posed" (intended) facial expressions and "elicited" (unintended) emotional responding by feature point tracking. We videotaped participants' facial movements of intended and unintended facial expressions of emotion by a high speed camera which allowed us to analyze facial movements very closely in image sequences. The experiment consisted of two parts. First, the participants task was to produce the six expressions (anger, disgust, fear, happiness, sadness, and surprise). In the second part of the experiment, participants were shown a set of film stimuli that elicited each of emotional states (amusemet, anger, disgust, fear, sadness, and surprise). We recorded the participants' facial expressions to the film stimuli.

We propose a method to track multiple human motion using multiple-camera-based non-synchronous observations. Our method employed Kalman-filtering algorithm to integrate non-synchronous observations from multiple viewpoints effectively. Our experimental system consists of three types of processes, discovering node for detecting newly appeared person, tracking node for tracking each target person, and observation node for processing each viewpoint images. In observation node, occlusion occurred among multiple persons cause serious problem to the matching process between the observed feature and tracked model. In the proposed method, not only position but also human features (height and cloth color) are employed for the matching process to enhance the tracking accuracy. Experimental results with real scene images support our system effectiveness.

We propose a technique to track user's eye and mouth by using two Pan-Tilt-Zoom controllable cameras. Mouth and eye zones are detected by combine the binary images from the camera. The zoom, rotation direction, and capturing rate of the cameras are automatically changed depends on the results of the captured images. By studying the characteristics of the extracted binary images, we are able to track lip and eye movements such as Lip Reading and blink. Experiments using the technique have shown satisfactory result.

We propose a method to detect human motion information using multiple viewpoint system. In vision-based human tracking, self-occlusions and human-human occlusions are a part of the more significant problems. We employ multiple-viewpoint system to reduce these problems. Our system can retrieve both position and orientation of human bodies, face images and hand positions. Experimental results with five cameras are also described.

We propose a hand shape recognition system using multiple-viewpoint images. Most conventional hand shape recognition system did not concern about self-occlusion problem due to the hand rotation and a user needs to pay attention to his hand's direction to avoid the problem. We employ multiple-viewpoint images to estimate the pose of a human hand. After a pose estimation, a "best view" for the hand shape recognition is selected based on the estimation result. Hand shape recognition is performed based on the shape representation using P-type Fourier descriptor that is not affected with image scaling and translation. Based on the shape recognition, we developed a system where a user co create virtual graphical scenes interactively. In the system, a user can change the virtual objects' positions, sizes, colors, etc. with hand gestures. This system can be used as a user interface device, replacing glove-type devices and overcoming most of the disadvantages of contact-type devices.

筆者らは、非装着型のヒューマンインタフェースへの応用を目指して、画像処理による手振り認識の研究を進めており、多数カメラを利用することで安定な手の位置・姿勢推定が行なえることを示した。推定された手の位置・姿勢は、手形状認識の簡素化および安定化のための視点選択に利用できる。今回、推定姿勢に基づいて正面像を選択することで、フーリエ記述子による形状認識を安定化できることを確認したので報告する。

We describe a system to estimate position and orientation of a human body. In such a system, deformation and occulusion of the human body often cause difficulties in the estimation. Our system employed multiple-camera based method to avoid these difficulty. Skeleton images detected from multiple cameras are utilized for stable estimation of both position and orientation. Furthermore, according to the estimation result, the best camera for the body side detection is selected. We confirmed the stability of those through experiments.

We describe a system to estimate position and orientation of a human body. In such a system, deformation and occulusion of the human body often cause difficulties in the estimation. Our system employed multiple-camera based method to avoid these difficulty. Skeleton images detected from multiple cameras are utilized for stable estimation of both posltion and orientation. Furthermore, according to the estimation result, the best camera for the body side detection is selected. We confirmed the stability of those through experiments.

This report proposes a new real-time method that estimates the posture of a human from a thermal image acquired by an infrared camera regardless of background and lighting conditions. Distance transformation is performed for the human body area extracted from the thresholded thermal image so that the center of gravity of body is calculated. After the orientation of the upper half of the body is obtained from calculating the moment of inertia, significant points such as the top of the head, the tips of the hands and foot are heuristically located. In addition, the elbow and knee positions are estimated from the siginificant points using a genetic algorithm based learning procedure. This method does not require to be put any type of device on a body and it can be applied for arbitrary person.

筆者らは既に、熱画像を用いた非接触な人物の姿勢推定手法を提案している。しかし、筆者らの従来法では、上半身の左右への大幅な傾きや、足の広範な動作に対応できず、また単眼のため3次元情報が得られないなどの課題があった。本報告ではより多くの姿勢に対応できるよう、従来の単眼用アルゴリズムを改良する。即ち単眼赤外線カメラから入力される熱画像から獲得される人物領域と、その輪郭の情報に基づいてヒューリスティックに頭頂・手先・足先の各位置を実時間で検出する手法、および遺伝的アルゴリズム (GA) を利用して肘および膝の位置を推定する手法を提案する。さらに、ステレオ視による3次元位置の獲得について検討する。

We′ve been working on the human-machine interface using facial e xpression animation.But the model based facial expression synthesis method that we proposed are not satisfaction.Because this model deformation rule is constructed based on the two dimensional measurement of human face.In this paper,we propose the three dimensional measurement method for facial surface movement and decide the new deformation rule of facial model.In this method, the three dimensional information is reconstructed from the front and side view image.Furthermore we reconsider the quantify of Action Units in FACS and the interpolation of feature points from the measurement result.

ギター演奏者の左手の動きをカメラで撮像することにより獲得される動画像を解析し、運指の良否を自動的に判定する方法の検討を行った。ギター演奏中の手と指の動画像の訓練データを収集する。各指の関節の位置の時系列データに対するDCT(離散コサイン変換)特徴を求める。一方、それぞれの訓練データにおける指の動きの良さについてのスコアを、ギターの教師が与える。スコアを目的変数、各関節のDCT特徴を説明変数として、訓練データに対してサポートベクトル回帰(support vector regression)を行い、回帰式を得る。未知データのDCT特徴をこの回帰式に代入し、スコアが正確に求まるかどうかを実験的に検討し、有効な結果を得た。

近年の医者不足のため、胎児検診の自動化が望まれている。本研究では、超音波画像処理技術とロボット制御技術を組み合わせることにより胎児検診の自動化を目指す。ロボットアームの先端に取り付けられた超音波プローブを、妊婦の腹部表面を走査することにより獲得される超音波画像群から、胎児の頭部等を抽出し、頭囲等を計測するアルゴリズムを明らかにするための検討を行う。本研究では、深層学習を用いて、妊婦の子宮を抽出し、胎児と羊水の範囲を判断する方法を提案し、有効な実験結果を得た。また、超音波画像中の胎児の頭部を楕円で近似して検出する新たな方法を提案し、有効性を実験により検証した。

　手術における看護師不足への対策として看護師ロボットの実現が望まれている。我々は手術者を支援する看護師ロボットの実現を目指し、手術野の動画像を分析し、人間の看護師と同様に、自動的に手術の進行や状況を把握する方法の研究を行っている。既に我々は、傷の縫合手術を構成する６つの段階のうち２つの段階（手術者の手の動きに対応）を早期認識(early-recognition)する手法を提案し、有効性の見通しを得ている。本期間は、早期認識の対象を、傷の縫合手術における６つの動作に拡張する検討を行った。我々が既に開発した２つの段階の早期認識法を単純に適用する場合と、６つの動作（段階）の順序関係を利用する場合とを比較検討した。その結果、６つ段階のうちいくつかは順序関係を利用することが有効であるという実験結果が得られた。

　手術における看護師不足への対策として看護師ロボットの実現が望まれている。本研究では手術者を支援する看護師ロボットの実現を目指し、人間の看護師と同様に、手術野の動画像を分析し、自動的に手術の進行や状況を把握する方法の研究を行っている。既に提案した傷の縫合手術における動作認識手法の認識対象を二つの動作から五つの動作に拡張した[1]。さらに、認識手法の動作が終了する前に予測が可能な方法を提案し、検討した。Siding Windowを利用して連続動画像における動作開始点を抽出し、Ryooらの予測法、動的計画法、ベイズ推定等に基づき、動作の予測を行う[2~4]。本提案手法の有効性を実験により検証した。ただし、前述の予測法において用いる3D-SIFTに基づく特徴量の計算時間について、まだ改善する必要がある。

（１）動的動画像中の動物体の抽出・追跡 　動的RGBD動画像から、移動物体を検出する方法を検討した．３次元特徴点抽出法とSuperVoxel法による領域分割法とを組み合わせ、移動物体と背景に対応する領域を得る方法を提案した。実験の結果，本提案法の有効性の見通しを得た。 （２）動画像処理に基づく手術の状況解析 　研究では、カメラにより手術野を撮像して得られる画像から、手術者の手の動作を解析することにより、縫合手術における「縫合」と「結紮」を認識する方法を提案した。即ち、新たな画像特徴と認識法の組み合わせを提案した。さらに、手術動作の完了前に、その動作を確率的に予測する方法を提案した。これらの手法の有効性を実験的に検証した。

　本研究では、胎児の先天的横隔膜ヘルニアの手術による治療法であるFETO (fetoscopic tracheal occlusion: 胎児鏡を利用する気管の閉塞)に有効と考えられる可撓性のある細い手術具を自律的に胎児の気道にナビゲートし、必要な施術を行う手術法の実現に資するいくつかの医用画像処理技術について検討した。また、内視鏡カメラが装着された手術具を内臓の内部に挿入しながら撮像される動画像を処理し、内臓の内部を３次元復元する方法を検討した。具体的には以下の通りである。（１）楕円の検出　超音波画像中の胎児の頭部に対応する楕円を検出するために、従来の反復的ランダム化ハフ変換を改良した準ランダムな方法を提案し、検討を行った。楕円の一部が遮蔽されている場合でも、従来法より高精度かつ高効率に楕円検出が行えることを実験的に示した。（２）超音波画像からの胎児の顔の３次元検出　３次元超音波画像中において胎児の顔の特徴点（鼻の頂点、両目の目尻、鼻の柱および上唇）間の幾何学的な関係に基づき、胎児の顔を検出するアルゴリズムを提案し、検討を行った。即ち、３次元超音波画像はノイズを多く含むので、３次元画像を構成する２次元のスライス画像ごとに（１）の楕円検出処理を施して胎児の頭部を抽出し、これに基づき顔の輪郭を求め、全てのスライス画像で求まった顔輪郭を統合して顔表面の３次元画像を得る。次に、このようにして得られた３次元顔画像の局所的な曲率と幾何学的な関係を利用して、目、鼻、口等の顔パーツを検出する。胎児ファントムを用いた実験により高い精度で顔の表面と顔パーツが検出できることを示した。（３）内視鏡動画像を用いる内臓内部の３次元復元　胃等の臓器の内部は、皺のような構造が随所に見られるため、内視鏡を動かしながら獲得される動画像からStructure from Motion (SFM)の原理で３次元復元をするための画像特徴点は豊富に存在すると言える。しかし、同じような皺構造の繰り返しが多数見られるため、対応付けが困難であるという課題もある。そこで本研究では、画像の縫い合わせ（stitching）を利用して、動画像におけるフレーム間の対応付けを安定に行う方法を提案した。即ち、SFMにより得られる３次元特徴点集合の縫い合わせを用いて密な３次元復元結果が得られるようにする。実験の結果、正確かつ効率的に３次元復元が行えることが確認できた。

　本研究では、移動するKinectセンサにより獲得されるRGB-D(距離)動画像から抽出されたsuper-voxelに基づき、移動物体の追跡を行うとともに、その移動物体の密な三次元構造を復元する方法を検討する。さらに、ビデオ画像から抽出された特徴をベイジアンネットに入力し、人物の無意識な行動を認識する方法を検討する。具体的には以下の通りである。（１）移動RGBD動画像からの移動物体追跡に関する検討移動する Kinect 等のRGBD カメラで複数の移動物体を撮像したシーンの特徴点を背景と各移動物体にクラスタリングする方法には，RANSAC を繰り返し適用する方法等，様々な方法が存在する．Delong らは，マルチラベルのGraphCut を特徴点の分類に用いた手法を提案した。しかし、Delong らの手法では、初期クラスタをランダムに特徴点を３点選択していたため、３点が複数の物体にまたがる場合、処理の精度が低下する問題があった。　そこで本研究では、以下の処理を提案する。(1) RGBD画像をsuper-voxelに分割し、ランダムな３つの特徴点の選択を１つのsuper-voxelに限定して行い、初期クラスタを作成する。(2) 初期クラスタから回転・並進情報を推定する。(3) 各特徴点に対して、推定した回転・並進情報を施すことにより得られる点の位置と、その特徴点におけるオプティカルフローの終点との誤差を計算し、super-voxel内の特徴点の誤差の合計をデータ項とする。(4) Super-voxel間のドロネー図を作成し，隣接するsuper-voxelを結ぶエッジの本数を平滑化項として利用する。(5) グラフカットを行い，ラベルを統合する。グラフカットのエネルギー式は，単一のノード（super-voxel）のみで決まる項であるデータ項と，隣接ノードとの相関によって決まる平滑化項と、ラベルの数とラベル毎のエネルギーの積となる統合処理項の3 つの項で構成される。ラベルの統合の結果によるデータ項・平滑化項のエネルギーの増加よりも、ラベル数の減少によるラベル数の項のエネルギーの減少が大きい場合は統合処理が進む。その結果、ラベルの数が未知の場合でも、適切なラベルの数を自動的に選択することが可能である。　実験を行い、本提案手法の有効性を示す結果を得た。（２）人物の無意識的な行動の認識法の検討　店舗内の顧客の購買行動、特に、迷って／迷わずに、購入したのか／購入しなかったのか、を認識可能な手法を検討した。このような購買行動には定型パターンは乏しいので、ベイジアンネットワークを用いて認識を行う方法を従来から検討している。ただ、従来の検討では、ベイジアンネットワークの各ノードの２１個の画像特徴量は、手動で取得していた。良好な認識結果を得ていたものの、手動特徴抽出は実用的とは言えない。そこで、本期間は、画像特徴量を自動的に得る方法を検討した。Felzenszwalb らが提案したDeformable Part Model(DPM)等を利用して、人物が様々な姿勢をとっても追跡が可能な手法を提案し、前述の２１個の特徴のうち１０個が自動的に計算可能となった。認識実験の結果、若干まだ認識性能が手動特徴の場合より劣るが、有効性の見通しを得たと言える。

　本研究では、以下に示す（１）～（４）の課題についてそれぞれ検討を行った。（１）人物全身像の姿勢推定　動作している人物の時々刻々の姿勢を、静止カメラ画像から推定するためには、人体パーツを追跡することが有効である。ここで、パーツが別のパーツを隠すオクルージョンの対処法の確立が重要である。そこで、人体パーツごとに異なった色に着色された衣服を装着した人物を撮像した動画像中で、パーティクルフィルターを用いて人体パーツを追跡する方法を検討した。追跡対象のパーツにオクルージョンが発生したか否かを、パーティクルの挙動により区別できる見通しが得られた。（２）未知人物の行動認識　人物、行動、時系列データから構成されるテンソルをデータベースとして予め構築しておき、テンソル分解を利用して未知人物の行動を認識する方法の検討を進め、我々の方法が関連手法より高い認識率を与える結果が得られた。一方、本手法を利用して、視野が重ならない複数カメラの視野間を移動する人物の対応付けを行うことにより追跡を行う手法を新たに提案し、検討を開始し、有効性の見通しを得た。（３）移動カメラ画像からの移動物体の抽出　パンティルトカメラにより獲得される動画像からカメラの動きを推定し、その動画像中にＣＧ画像を正確に配置できるようにする方法を検討した。特に、動画像中に動く物体を含む場合、その動物体に対応する点からカメラの動きを推定しようとすると、推定精度の低下を招くので、動画像中の動く物体から推定されたか否かを検証する方法を提案し、有効性の見通しを得た。（４）移動カメラ画像からの人物の動き認識　移動ロボットに搭載したカメラにより獲得される動画像から、人物の手の動作を認識する方法の検討を進めた。カメラの動きを考慮に入れて手の軌跡を推定し、これを用いて従来検討したCondensation Algorithmに加え、主成分分析を用いる方法により手話動作の認識の性能評価を行い、主成分分析の方が高い認識率を与える結果を得た。

　現実には距離を隔てた人物同士の仮想的シーンを介したコミュニケーション環境実現のためには、現実の空間における人物の像を画像処理により自動的に認識し、仮想的シーン中の３次元人物モデルにおいて実時間で再現する、人物像の認識・生成法の実現が必要である。本研究課題では、人物像の認識・生成法の研究を立ち上げるため、以下の項目について、環境整備、方式検討および予備的検討を行ったので、その概要を述べる。（１）人物と物体のインタラクションの認識法の検討　人物のふるまいを画像処理により認識する従来の方式のほとんどは、人物が単独で存在する状況のみを扱っていた。これに対して、現実の空間では、人物はシーンや物体とのインタラクションを日常的に行う。従って、人間にとって自然なマンマシンインタフェースシステムや、自動監視システムを実現するためには、このような人物と物体とのインタラクションを自動的に認識する方式の開発が必要である。そこで、人物のシルエット像の輪郭解析法、移動物体の追跡法、人物と物体が接する場所周辺の濃淡画像解析を統合的に用いる手法を提案し、現在基本的なプログラムを開発中である。（２）３次元顔モデルの生成と表情再現法の検討　従来の３次元顔モデルには、解剖学的な構造を忠実に再現したものはほとんど見られなかった。特に、頭蓋骨、表情筋、脂肪組織、皮膚組織の構造のモデリングは、表情再現のために重要と考えられる。また、表情筋の動作に伴う、皮膚表面の形状変化の再現も重要である。そこで、CT画像データを用いて前述のような顔の構造をモデリングするとともに、実際の人物の表情変化を実現するための表情筋の動的特性パラメータの決定を実写顔画像から行う方法を提案し、現在環境を整備中である。（３）表情再現のリアリティ向上方式の検討　実際の人物の表情を画像処理により推定し、その結果を通信回線を用いて受信側の３次元顔モデルに送信し、表情を再現する方式では、３次元顔モデルにおいて再現される表情のリアリティに課題が残っていた。そこで、実人物の顔表面のテクスチャを顔画像から実時間で獲得し、これを３次元顔モデルにマッピングする方式を提案した。ここで、皮膚表面の光学的特性は個人毎に異なるので、これを効率的に推定する手法を開発する必要がある。また、実人物のいる環境の照明条件を、３次元顔モデルにテクスチャをマッピングする際に反映する必要がある。さらに、３次元顔モデルの形状変化や位置・姿勢変化にも対応する必要がある。現在、これらの検討を可能とする実験環境の整備を進めるとともに、アルゴリズムの詳細を検討中である。