Spatial and Temporal Based Pedestrian Trajectory Prediction

Abstract

Pedestrian trajectory prediction is a critical task that aims at predicting the future movement trajectory of pedestrians over the next several seconds. Recent pedestrian trajectory prediction methods assume that the number of pedestrians and the relationship between pedestrians is constant or based on the distance, and heavily rely on contextual information and homography matrix. We propose a dynamic graph structure to describe the pedestrians, using self-attention along both the spatial and temporal dimensions. During each time step, spatial self-attention captures how pedestrians interact. The temporal self-attention predicts how pedestrians will move in the future by giving weight to the representation of prior data. Thus, we can accurately predict the pedestrian's future movements by considering their past behavior and interactions. Experiment results show that our model achieves great performance on ETH/UCY datasets.