Vehicle Motion Prediction Based on Selective Perception and Weak Lane Guidance

Trajectory prediction is crucial for ensuring the safety and reliability of autonomous driving systems. Accurately predicting the future trajectory of a vehicle can aid drivers or automated driving systems in developing reasonable movement plans and avoiding potential collisions. In real-world scenarios, drivers selectively observe areas that influence their driving decisions based on real-time traffic scenarios, adapting to changing traffic situations to ensure safety. To simulate this behavior, we propose a selective perception module that utilizes historical trajectories and lane centerlines to extract vital information from the changing traffic scene. Furthermore, vehicles tend to follow lane guidance to some extent, we use a normalized flow decoder and small-scale guided loss to induce the distribution of predicted trajectories, slightly increasing the probability of lane centerlines in the predictions as a weak lane guidance. We conduct ablation studies using the public dataset nuScenes and comparative experiments with state-of-the-art methods, showcasing the predictive diversity of our approach while maintaining high accuracy.