Trajectory Prediction Based on Planning Method Considering Collision Risk

Ya Wu, Jing Hou, Guang Chen, Alois C. Knoll

Abstract: Anticipating the trajectory of Autonomous Vehicles (AV) plays an important role in improving its driving safety. With the rapid development of learning-based method in recent years, the long short-term memory (LSTM) network for sequential data has achieved great success in trajectory forecasting. However, the previous LSTM only considered forward time cues and did not reason on motion intent of rational agents. In this paper, we use planning-based methods follow a sense-reason-predict scheme in which agents reason about intentions and possible ways to the goal. In addition, the collision risk is considered, and the most appropriate future trajectory will be selected with the current state of the agent. We have compared our method against two baselines in highD dataset. Our experimental results show that the planning-based method improves prediction accuracy compared with the baselines.