应对遮挡问题对DeepSORT进行轨迹拟合优化  

Trajectory fitting optimization for DeepSORT to cope with occlusion problem

作　　者：李伟[1] 颜旒 LI Wei;YAN Liu(School of Mechanical Electrical and Vehicle Engineering,Chongqing Jiaotong University,Chongqing 400074,China)

机构地区：[1]重庆交通大学机电与车辆工程学院,重庆400074

出　　处：《现代电子技术》2023年第10期173-180,共8页Modern Electronics Technique

摘　　要：检测跟踪范式是多目标跟踪的主要研究方向,也是自动驾驶汽车的主要应用框架。完善的检测跟踪范式可以在提高多目标跟踪精度的同时有效降低跟踪框的失真率。然而在现有的先进方法中,这种范式通常会遇到多目标重叠后的ID交换问题,严重影响跟踪精度和轨迹判断。为解决该问题,文中基于经典的DeepSORT算法提出改进方案。首先,在卡尔曼滤波器中添加跟踪框进行置信度的预测和更新,并按降序对更新后的置信度进行排列,在后续匹配中优先匹配预测置信度更高的跟踪目标;然后,比较预测置信度和检测置信度之间的差异,识别出置信度突变的目标,以进行跟踪轨迹的优化和剪枝;最后,使用余弦相似度和交并比(IoU)识别重叠目标,并对重叠目标中置信度最高的目标消失后的轨迹进行多项式轨迹拟合,以纠正错误的ID,完成精确匹配。实证分析结果表明,相比于目前多种先进算法,所提方法的ID交换次数为172,优于其他算法,验证了该算法的可行性和正确性。Tracking⁃by⁃detection is now a major research direction of multi⁃object tracking(MOT),which is also the primary application framework for self⁃driving automobiles.A comprehensive tracking⁃by⁃detection can reduce tracking box distortion while increasing MOT accuracy.However,in existing advanced methods,it often encounters ID exchange issues after multiple overlapping targets,which seriously affects tracking accuracy and trajectory judgment.An improved scheme is proposed based on clissical DeepSORT algorithm to slove above problems.A tracking box is added to the Kalman filter for the confidence prediction and updation,and the updated confidence is arranged in descending order.In subsequent matching,priority is given to matching tracking targets with higher prediction confidence.The difference between predictive confidence and detection confidence is compared to identify targets with sudden changes in confidence,so as to optimize and prune the tracking trajectory.The cosine similarity and inersection over union(IoU)are used to identify overlapping targets,and perform polynomial trajectory fitting on the trajectory of the target with the highest confidence among overlapping targets after disappearance to correct incorrect IDs and achieve accurate matching.The empirical analysis results show that in comparison with various advanced algorithms at present,the proposed method has an ID exchange frequency of 172,which is superior to other algorithms,verifying the feasibility and correctness of the algorithm.

关 键 词：遮挡目标检测 多目标重叠 DeepSORT算法 轨迹拟合算法 卡尔曼滤波 置信度 算法对比 

分 类 号：TN820.4-34[电子电信—信息与通信工程] TP311[自动化与计算机技术—计算机软件与理论]