Generative Pedestrian Trajectory Prediction with Graph Representation

Abstract

Multi-agent trajectory prediction is one of the core modules of unmanned driving and intelligent robots. The traditional method is difficult to measure the relationship between multiple agents, and the modeling ability is rigid. Nowadays, most of the methods make less use of geographic information, and social relationship modeling is not sufficient. Our model uses Graph Neural Network (GNN) to measure social relationships to improve its usability. The model is built by the basic Conditional Auto-encoder (CVAE) framework, using Gaussian Mixture Model to mix multiple Gaussian distributions and the possibility of obtaining more potential space through dynamic integration to make the model achieve better results. Our model has achieved excellent results on the Stanford Drone Dataset. The evaluation by average displacement error (ADE) and final displacement error (FDE) metrics has exceeded the majority of existing models.