Detection and Tracking Dynamic Vehicles for Autonomous Driving Based on 2-D Point Scans

Processing dynamic vehicle detection and tracking tasks based on three-dimensional (3-D) point clouds is computationally overconsuming, which hinders its practical realization on embedded devices. To release the processing burden, we attempt to transform 3-D LiDAR data to 360 FOCUS points (FS points) before this task. Then, a novel detecting and tracking dynamic vehicle approach is designed based on the obtained 2-D scan. The approach classifies the FS points into static points and unstatic points by generating a Dempster–Shafer (D–S) evidence theory-based static obstacle map (SOM). At first, the SOM is predicted and updated by interacting with the model-free dynamic vehicle clustering and tracking module (MFCTM). Then, the unstatic points determined in SOM are clustered in the MFCTM, and each obtained cluster is initialized as a new track or assigned to the existing tracks. Finally, the dynamic states are calculated. The computational burden performance of the proposed method is compared with that of the other state-of-the-art methods. Meanwhile, it is also applied to three simulation scenarios and three field tests, and its validity is demonstrated through comparison with the results of the L-shape-based tracking method.