A Specialized Variational Autoencoder for Cost-Efficient Pedestrian Trajectory Prediction

Abstract

The prediction of pedestrian trajectories represents a crucial and widely discussed topic in the field of AI-driven traffic scenarios. The prediction of pedestrian trajectories is constrained by two factors. First, pedestrians do not have the same traffic rule constraints as vehicles. Second, the computational power of in-vehicle systems is limited. This renders the application of traditional methods challenging. Previous methods have been observed to utilize redundant information, which can result in feature imbalance and the potential for model overfitting. In light of these limitations, we propose a lightweight conditional variational autoencoder model with post-process (L-CVAE-P) for pedestrian prediction scenarios. The L-CVAE-P focuses on the efficient interaction of multidimensional features to achieve a comprehensive enhancement of the model for real-world use. The model is tested on two public datasets and achieved state-of-the-art performance, while maintaining efficiency. The experimental results demonstrate that our work has developed and optimized a pedestrian trajectory prediction model for practical applications. © 2025 The Author(s). IEEJ Transactions on Electrical and Electronic Engineering published by Institute of Electrical Engineers of Japan and Wiley Periodicals LLC.