Disentangled Multi-Relational Graph Convolutional Network for Pedestrian Trajectory Prediction

Pedestrian trajectory prediction is one of the important tasks required for autonomous navigation and social robots in human environments. Previous studies focused on estimating forces among individual pedestrians. However, they did not consider groups pedestrians, which results over-collision avoidance problems. To address this problem, we present a Disentangled Multi-Relational Graph Convolutional Network (DMRGCN) socially entangled pedestrian prediction. We first introduce novel disentangled multi-scale aggregation to better represent interactions, pedestrians weighted graph. For aggregation, construct multi-relational graphs based distances relative displacements In step, propose global temporal alleviate accumulated errors changing their directions. Finally, apply DropEdge into our DMRGCN avoid over-fitting issue relatively small datasets. Through effective incorporation three parts within an end-to-end framework, achieves state-of-the-art performances variety challenging benchmarks.