基于深度信息的车辆识别和跟踪方法

Vehicle Detecting and Tracking Method Based on Range Information

根据激光雷达检测车辆目标的特点,提出了一种基于距离的自适应聚类方法,距离阈值可根据目标与本车的相对距离与方位自动调整,提高了聚类的准确性.采用一种基于多特征融合的确定性多目标关联方法,以目标的运动特征为代价方程的主要约束条件,同时考虑了目标的外形特征,可提高关联的准确性.针对现有的确定性目标关联方法只能对给定数目的目标进行跟踪的缺陷,应用一种改进的目标关联方法和跟踪器管理策略,根据实际道路情况动态地增加和删除跟踪器,实现了对暂时遮挡或者漏检的目标保持跟踪的连贯性.通过实验验证了本文识别和跟踪方法的有效性.

An adaptive point-distance-based clustering method was developed based on the traits of the range information obtained by laser scanner. The point distance threshold could be automatically adjusted in this method according to both the distance and orientation of the objective car to the subjective one. Thereby the clustering accuracy was improved. A deterministic data association method, which took the motion property as the main constraints and the figuration as minor ones for cost function, was applied for the object tracking in order to improve the accuracy of correspondence. Considering the fact that the number of objects had to be given prior to tracking for the present deterministic data association methods, a modified data association method and tracker management strategy were developed. The trackers could be dynamically added and deleted according to the real road scene and the tracking process for the temporarily occluded and undetected objects could be maintained for a set period in this tracking method. The validity of the detection and tracking algorithm was verified by experiments.