Robust Object Tracking Method Based on Multi-Level Association Matching

Target tracking is an important component of intelligent vehicle perception systems, which has outstanding significance for the safety and efficiency of intelligent vehicle driving. With the continuous improvement of technologies such as computer vision and deep learning, detection based tracking has gradually become the mainstream target tracking framework in the field of intelligent vehicles, and target detection performance is the key factor determining its tracking performance. Although remarkable progress has been made in current 3D object detection networks, a single network still struggles to provide stable detection for distant and occluded targets. Besides, traditional tracking methods are based on single-stage association matching, which can easily lead to identity jumps and target loss in case of missed detections, resulting in poor overall stability of the tracking algorithm. To solve the above problem, a hierarchical association matching method using a dual object detection network is proposed, which compensates for the poor performance of multi-modal 3D detection on distant targets due to the sparsity of point clouds through a lightweight 2D detection network. In addition, to alleviate the problem of missing detection caused by inadequate confidence of occluded targets, a strategy of implementing hierarchical correlation based on the confidence level of detection boxes is also proposed. The above proposed methods are deployed in an unscented Kalman filter based on the constant turn rate and velocity model, and the sort method is used for trajectory lifecycle management to construct a complete target tracking system. The experiments of the nuScenes dataset and real vehicle data shows that our method can achieve more robust multi-target tracking performance.