A Simultaneous Object Detection and Tracking Framework Based on Point Cloud
Zhaohui Xiang · 2022
Achieving accurate and effective detection and multi-object tracking plays a very important role in the field of automatic driving. Accurate detection of pedestrians, cars, cyclists, and other objects can improve the performance of the downstream tasks, such as obstacle avoidance and path planning. Based on the object detection results, multi-object tracking can identify and track moving targets in a period of time. The traditional tracking-by-detection methods can not handle the occlusion situation or long-term tracking problem well. In this paper, we propose a simultaneous detection and tracking method based on spatial temporal map, which has superior performance of trajectory connectivity and tracking accuracy in a long period of time. In the detection module, we use RANSAC algorithm to extract and segment the plane for each frame of lidar data, and then, cluster and classify different objects using the Euclidean Clustering algorithm. In the tracking module, we propose a two-frame and multi-frame switchable architecture for online tracking. In two-frame tracking, we use Kalman filter to update the object location measurements to match the same object in the current frame. For long-term and multi-frame tracking, the trajectories of objects are directly represented on spatial temporal map, and we fit polylines in the map to find the best data association. Experiment results on the KITTI dataset show that our 3D object detection and tracking framework is elegant and complete, the tracking module is robust, widely applicable, and low in cost.