L-Shape-Model-Based Vehicle Tracking With Joint Kinematic and Geometric Estimation Using Lidar

In this article, the problem of tracking vehicles using lidar sensors mounted on an ego-vehicle is addressed. Due to line-of-sight limitations, the back (or front) of a vehicle as seen by the lidar on the ego-vehicle behind (or ahead of) it is often modeled as L-shaped. In this article, an L-shape-based vehicle tracking algorithm with joint kinematic and geometric estimation is presented. By feeding back tracking results to L-shape fitting, an L-shape detection method that is robust to outliers is proposed. In the L-shape tracker currently available in the literature, the kinematic and geometric states of the L-shape model are separately estimated and maintained. However, the kinematic and geometric states are not independent since the orientation of a vehicle influences its velocity. Also, the dependence between the kinematic and the geometric states is caused by anchor-point (the closest point on the vehicle being tracked) switching, which is required during changes in the relative position between vehicles. To address this limitation, the proposed L-shape tracker exploits this dependence and estimates the kinematic and geometric states jointly. The proposed L-shape-model-based tracking algorithm is evaluated and compared with the original algorithm using the real traffic data from the KITTI datasets. The results demonstrate the superiority of the proposed algorithm over the original algorithm in terms of L-shape detection and tracking accuracies.