GMGCN: Group-Masked Graph Convolution Network for Pedestrian Trajectory Prediction

Accurately predicting pedestrian movements can enhance decision-making in autonomous vehicles and logistics robots. However, the complex social relationships within and outside pedestrian groups make this task challenging. To address these challenges, we propose a group-masked graph convolution network (GMGCN) for pedestrian trajectory prediction. We construct a social graph and then utilize the self-attention and time-frequency domain convolution module to obtain an interaction matrix. To model in-group and out-group interactions, we generate in-group and out-group mask matrices based on the pedestrian feature similarity by setting appropriate thresholds. We then fuse the interaction matrix with out-group and in-group mask matrices, extract interaction features via graph convolution, and predict future trajectories with temporal convolution networks. Experiments on multiple real-world trajectory datasets demonstrate that our method achieves lower prediction errors than existing methods and effectively models group interactions.