Scenario-Transferable Semantic Graph Reasoning for Interaction-Aware Probabilistic Prediction

Accurately predicting the possible behaviors of traffic participants is an essential capability for autonomous vehicles. Since vehicles need to navigate in dynamically changing environments, they are expected make accurate predictions regardless where and what driving circumstances encountered. Several methodologies have been proposed solve prediction problems under different situations. These works usually combine agent trajectories with either color-coded or vectorized high definition (HD) map as input representations encode this information behavior tasks. However, not all relevant scene forecasting such irrelevant may be even distracting certain Therefore, paper, we propose a novel generic representation various environments by taking advantage semantics domain knowledge. Using enables situations modeled uniform way applying knowledge filters out unrelated elements target vehicle’s future behaviors. We then general semantic framework effectively utilize these formulating them into spatial-temporal graphs reasoning internal relations among graphs. theoretically empirically validate highly interactive complex scenarios, demonstrating that our method only achieves state-of-the-art performance, but also processes desirable zero-shot transferability.