Enhancing Multi-Object Tracking Using Roadside Camera Based on Trajectory Prediction in BEV

With the continuous development of vehicle-infrastructure collaborative autonomous driving, the importance of roadside tracking is becoming increasingly prominent. However, tracking on the image plane is difficult to solve the occlusion problem in roadside environments. Therefore, a camera multi-object tracking method based on BEV plane trajectory prediction enhancement is proposed. Specifically, the paper adopts geometric methods for depth estimation to improve the generalization of roadside camera tracking, and effectively solves the occlusion problem through trajectory prediction in the BEV plane. We conducted tests on the CROP dataset and the experiments showed that our method is effective, with an average improvement of 4.04 on the MOTA metric.