IMGCN: interpretable masked graph convolution network for pedestrian trajectory prediction

Abstract

Pedestrian trajectory prediction holds significant research value in various fields, such as autonomous driving, autonomous service robots, and human flow monitoring. Two key challenges in pedestrian trajectory prediction are the modeling of pedestrian social interactions and movement factors. Previous methods have not utilized interpretable information to explore complex situations when modeling social interactions. These methods also focus too much on temporal interactions at each moment when modeling movement factors and are therefore susceptible to slight motion changes. To solve the above problems, we propose an Interpretable Masked Graph Convolution Network (IMGCN) for pedestrian trajectory prediction. The IMGCN utilizes interpretable information such as the pedestrian view area, distance, and motion direction to intelligently mask interaction features, resulting in more precise modeling of social interaction and movement factors. Specifically, we design a spatial and a temporal branch to model pedestrians' social interaction and movement factors, respectively. Within the spatial branch, the view-distance mask module masks pedestrian social interaction by determining whether the pedestrian is within a certain distance and view area to achieve more accurate interaction modeling. In the temporal branch, the motion offset mask module masks pedestrian temporal interaction according to the offset degree of their motion direction to achieve accurate modeling of movement factors. Ultimately, the 2D Gaussian distribution parameters of future trajectory points are predicted by the temporal convolution networks for multi-modal trajectory prediction. On the ETH, UCY and SDD datasets, our proposed method outperforms the baseline models in terms of average displacement error and final displacement error. The code is publicly available at https://github.com/Chenwangxing/IMGCN_master.