Three-dimensional multi-target tracking of point cloud from bird's-eye view
Shuzhao Hou, Zhenxin Wang, Xueyan Li, Junfeng Song · 2021 International Conference on Electronic Information Engineering and Computer Science (EIECS) · 2021
Currently, the most popular approach for multitarget tracking is tracking-by-detection. Existing algorithms cannot balance the quality and speed of three-dimensional (3D) multi-target online point cloud tracking. We propose a low-complexity high-quality 3D multi-target online tracking algorithm to track a 3D point cloud from a bird's-eye view. First, we use a detector to obtain a bird's-eye view of a 3D point cloud and object information. Then, a 3D multi-target online tracking scheme is constructed to obtain the target trajectory. The scheme is mainly divided into state estimation and target association. State estimation uses a 3D Kalman filter algorithm Target association is divided into two steps: cascade matching and intersection over union (IoU) matching. The apparent feature used by the cascade matching algorithm is determined by using the unsupervised algorithm variational autoencoder on the candidate target in a bird's-eye view. The IoU matching algorithm uses the 3D spatial features of the candidate target. We demonstrate the effectiveness of our algorithms on the KITTI dataset. The values of multiple object tracking accuracy (MOTA), multiple object tracking precision (MOTP), and speed are 66.02, 85.81, and 43 frames per second, respectively, indicating the effectiveness of our algorithm in tracking quality and speed.