基于改进DeepSORT的视觉/激光雷达目标跟踪与定位方法CSCD

通过结合目标跟踪与相对定位,在对多帧检测目标进行关联与分析的同时,可以获取其三维信息。但当目标外观特征变换较大时,传统目标跟踪算法较易发生漏匹配或身份变换,而仅依靠对齐点云的相对定位算法较易出现定位失效的情况。针对以上问题,提出了一种基于改进DeepSORT的目标跟踪与定位方法在原始DeepSORT算法中加入基于位置约束的匹配,解决了因外观改变导致的漏匹配问题;在获取跟踪信息的基础上,设计了基于目标运动模型的相对定位方法,解决了图像中目标较小时相对定位不连续且定位精度较低的问题。试验结果表明,与传统DeepSORT算法相比,多目标跟踪准确度提高了5.9%;与仅依靠对齐点云的相对定位算法相比,定位精度提高了62.4%。

Vision/Lidar Object Tracking and Localization Method Based on Improved DeepSORT

By combining object tracking and relative positioning, three-dimensional information can be obtained while correlating and analyzing multi-frame detection objects. However, when the object appearance feature transforms largely, the traditional object tracking algorithm is prone to missing match or identity transformation, and the relative positioning algorithm that only relies on aligning point clouds is prone to localization failure. To address the above problems, an object tracking and localization method based on improved DeepSORT is proposed. The matching based on position constraints is added to the original DeepSORT algorithm to solve the problem of missing match caused by appearance changes; on the basis of obtaining tracking information, the relative positioning method based on target motion model is proposed, which solves the problem of discontinuous relative positioning and low positioning accuracy of small image objects under sparse lidar point cloud. The experimental results show that compared with the traditional DeepSORT algorithm, the multi-object tracking accuracy is improved by 5.9%, and the positioning accuracy is improved by 62.4% compared with the relative positioning algorithm that only relies on aligned point clouds.