Trajectory Prediction for Surrounding Traffic Participants via Local Perception and Attentive Map Encoding.

Predicting the trajectory of the surrounding traffic participants is a pivotal factor in autonomous driving. The prevalent methods in trajectory prediction hinge upon offline pre-stored high-precision maps to incorporate driving environment data. However, updating high-precision map in real time is a huge challenge in real-world scenarios. Based on the observation, we propose a novel local perception-based trajectory prediction method, which utilizes a bird’s eye view (BEV) local perception map to replace the demand for high-precision map. Our BEV local perception map is built by leveraging sensor perception information, including semantic segmentation, 3D object detection and lidar point cloud data. Then, an attentive map encoding (AME) module is designed to extract the traffic environment features from the local perception map. Finally, an encoder-decoder structure based on Transformer is applied to achieve the trajectory prediction. Comprehensive experiments are conducted on the nuScenes dataset, demonstrating that our method can maintain prediction accuracy while avoiding dependence on high-precision maps.