Heterogeneous graph social pooling for interaction-aware vehicle trajectory prediction

Abstract

Predicting the trajectories of neighboring vehicles is vital for self-driving cars in intricate real-world driving. The challenge lies in accounting for diverse influences on a vehicle's movement, travel needs, neighboring vehicles, and a local map. To address these factors comprehensively, we have developed a framework with a Heterogeneous Graph Social (HGS) pooling approach. The framework represents vehicles and infrastructures in a single graph, with vehicle nodes holding historical dynamics information and infrastructure nodes containing spatial features from map images. HGS captures vehicle–infrastructure interactions in urban driving. Unlike existing methods that are restricted to a fixed vehicle count and highway settings, HGS can accommodate variable interactions and road layouts. By merging all features, our approach predicts the target vehicle's future path. Experiments on real-world data confirm HGS's superiority, boasting quicker training and inference, affirming its feasibility, effectiveness, and efficiency.