I conducted two complementary studies on generalized, deployment-oriented pedestrian trajectory prediction by integrating social scene information and motion priors. The first study, training-free prediction via segmentation-guided path planning, derives a walkable-region representation from scene segmentation and converts it into a grid or cost map. Future trajectories are generated through classical path planning while enforcing feasibility constraints from the environment and simple kinematic consistency from short observations, such as heading and speed trends. This eliminates the need for model training or dataset-specific fine-tuning, which is advantageous under data scarcity, domain shift, or rapid deployment constraints.The second study, Kinematic Temporal VAE for generalized pedestrian prediction, proposes a lightweight temporal generative model that explicitly encodes kinematic and temporal structure. Given a short observed trajectory, the model learns a compact latent representation and produces multi-modal future trajectories that preserve motion realism, for example smoothness and physically consistent progression. The design emphasizes efficient inference and robust generalization across heterogeneous scenes and observation conditions, and it supports uncertainty-aware prediction by sampling diverse plausible futures.Overall, these works bridge classical feasibility priors and probabilistic sequence generation. The segmentation-guided planning approach provides strong environmental validity without training, while the kinematic temporal VAE captures uncertainty and behavioral diversity with an efficient learned model. Together, they offer a practical toolkit for pedestrian prediction in autonomous driving and mobile robotics, improving reliability across varying social scene contexts and deployment settings.

    Pedestrian trajectory prediction is a key challenge in AI-driven traffic scenarios due to two main constraints: (1) pedestrians move unpredictably without strict traffic rules, and (2) in-vehicle systems have limited computing power, making traditional methods less feasible. Previous approaches often include redundant information, causing feature imbalance and increasing the risk of overfitting.    To address these challenges, we propose a lightweight conditional variational autoencoder model with post-processing (L-CVAE-P) for pedestrian trajectory prediction. Our method efficiently integrates multidimensional features, enhancing the model’s effectiveness for real-world deployment. We evaluate L-CVAE-P on two public datasets, where it achieves state-of-the-art performance while maintaining computational efficiency. Experimental results confirm that our approach successfully optimizes pedestrian trajectory prediction for practical applications.    Our approach employs a Conditional Variational Autoencoder (CVAE) to encode temporal kinematic data into a standard Gaussian distribution, generating diverse predictions that deviate from it. To enhance spatio-temporal feature interactions, we introduce a spatial feature filtering algorithm based on scenario information. Finally, a hybrid attention-based and temporal processing network refines the output.    This framework maintains the diversity of pedestrian trajectories while ensuring reasonable and safe constraints in an intelligent and efficient manner. Unlike previous methods that directly process high-dimensional spatio-temporal features, L-CVAE-P first prioritizes time-series features and then implicitly integrates scenario-based spatial information, thereby avoiding overfitting and reducing model redundancy. This structured approach ensures a balanced representation of temporal and spatial features without excessive dimensional transformations.    Compared to previous work, L-CVAE-P offers a more balanced processing of temporal and spatial features while minimizing redundancy, making it a computationally efficient solution for pedestrian trajectory prediction in autonomous systems.Contributions:(1) Multi-dimensional Information BalancingInstead of directly integrating spatial and temporal features, our method prioritizes temporal processing and implicitly incorporates spatial information for spatio-temporal interactions.This reduces the model’s dependency on scenario data, preventing overfitting and ensuring a more robust feature representation.(2) High-Diversity Generative Framework with Post-processingThe proposed model preserves output diversity while a post-processing module enhances model robustness, making it more effective for real-world applications.(3) Novel Statistical Evaluation for Practical UseA data distribution statistical experiment is conducted to objectively assess real-world performance.By analyzing outliers and extreme values, the model’s effectiveness in practical deployment is demonstrated.“A Specialized Variational Autoencoder for Cost-Efficient Pedestrian Trajectory Prediction”The paper has been accepted by IEEJ and will be published in IEEJ Vol. 20, No. 8, TEEE Section C.