Pedestrian Trajectory Prediction Based on Tree Method using Graph Neural Networks

Abstract

Pedestrian trajectory prediction in real-world scenarios is a challenging task for several computer vision applications, such as autonomous driving, video surveillance, and robotic systems. This is not a trivial task due to the numerous potential trajectories. In this article, it provides a tree-based approach to handle this multimodal prediction challenge. The tree is designed based on the observed data and is also used to predict future trajectories. In particular, an individual's potential future trajectory is represented by the tree's root-to-leaf route. Compared to previous approaches that use implicit latent variables to describe possible future paths, the movement behaviors may be directly represented by the path in the tree (e.g., go straight and then turn left), and thus offer more socially suitable trajectories. The experimental results on the ETH, UCY, and Stanford Drone datasets show that this approach can exceed the performance of the state-of-the-art approaches. The solution is more efficient and compact, with a smaller model size and a higher accuracy, and delivers better results with reference to average displacement error (ADE) and final displacement error (FDE) metrics.