CAPHA: A Novel Context-Aware Behavior Prediction System of Heterogeneous Agents for Autonomous Vehicles

To safely and successfully navigate complex and dense driving scenes, automated vehicles need to develop situational awareness, which requires understanding the current behavior and states of their surrounding traffic participants as well as reasoning about their future intent and motions. One considerable challenge is to estimate the distribution of future motions of surrounding agents considering the uncertainty in driving behavior. Most recent works model this uncertainty through multimodality and forecasting diverse modes of future motions. However, these methods may assign a high probability to several unlikely predictions with infeasible outcomes that are neither realistic nor scene-compliant. To address this issue, we propose conditioning future behavior of agents on grid-based plans modeled as a Markov decision process. We design a novel Context-Aware behavior Prediction system for Heterogeneous Agents (CAPHA) based on inverse reinforcement learning. More specifically, CAPHA works based on mapping the scene and kinematic features into rewards and learning a policy through a value iteration network which is further sampled and fed to stacked transformers and the behavior decoder to generate multiple plausible motions. We evaluate the performance of CAPHA on two large-scale motion forecasting datasets with heterogeneous agents, and the experiments show our model's high performance and efficiency. The results indicate that CAPHA outperforms several state-of-the-art models by generating diverse and scene-compliant motions.