Intention-based and Risk-Aware Trajectory Prediction for Autonomous Driving in Complex Traffic Scenarios
Wen Wei1, Jiankun Wang1,2
[1] Shenzhen Key Laboratory of Robotics Perception and Intelligence, SUSTech, Shenzhen, China.
[2] Jiaxing Research Institute, SUSTech, Jiaxing, China.
Abstract: Accurately predicting the trajectory of surrounding vehicles is a critical challenge for autonomous vehicles. In complex traffic scenarios, there are two significant issues with the current autonomous driving system: the cognitive uncertainty of prediction and the lack of risk awareness, which limit the further development of autonomous driving. To address this challenge, we introduce a novel trajectory prediction model that incorporates insights and principles from driving behavior, ethical decision-making, and risk assessment. Based on joint prediction, our model consists of interaction, intention, and risk assessment modules. The dynamic variation of interaction between vehicles can be comprehensively captured at each timestamp in the interaction module. Based on interaction information, our model considers primary intentions for vehicles to enhance the diversity of trajectory generation. The optimization of predicted trajectories follows the advanced risk-aware decision-making principles. Experimental results are evaluated on the DeepAccident dataset; our approach shows its remarkable prediction performance on normal and accident scenarios and outperforms the state-of-the-art algorithms by at least 28.9\% and 26.5\%, respectively. The proposed model improves the proficiency and adaptability of trajectory prediction in complex traffic scenarios.
Pipeline:
Overview of our model
The Interaction Module processes historical trajectories and map information through the map encoder and trajectory encoder. The Relative Position Encoder further refines the positional relationships between agents. Subsequently, the interaction relationship between agents is modeled. In the Intention Module, the longitudinal and lateral intention probabilities are calculated and fused to obtain the agent's motion intention. Finally, the scene risk value is calculated through the Risk Assessment Module and used to guide the trajectory optimization. By collaborating with different modules, the entire system can effectively predict and evaluate the risks in multi-vehicle interactions to ensure the proficiency and adaptability of driving in complex scenarios.
Qualitative Results:
Illustration of prediction results in different traffic scenes. Top row: scenario exists for one right-turn operation. Bottom row: scenario exists collision probability. Compared with baselines, our model performs better in both scenarios, not only predicting accurately but also considering potential risks in the future.