联合YOLOv4检测的候选框选择和目标跟踪方法

Candidate box selection and object tracking method jointing YOLOv4 detection

当前的目标跟踪算法主流是基于检测的跟踪(DBT),所以检测的质量对跟踪的性能影响很大,同时在跟踪过程中易受环境干扰、光照变化、目标尺度和类别的影响,针对以上目标跟踪存在的问题,提出一种联合深度学习神经网络YOLOv4检测算法和Kalman滤波的目标跟踪算法。首先利用目标检测器对目标进行分类和边界框提取,跟踪器用于在跟踪轨迹中收集候选数据;其次,提出一种对象选择器,用来选择检测和跟踪轨迹中的最优候选框;最后,将最优候选框和跟踪轨迹利用ReID进行数据关联判断是否对跟踪轨迹进行更新。实验结果表明,联合检测的目标跟踪方法与其他几种已经成型算法对比跟踪精度达到84.9%,跟踪成功率为82.2%。同时该方法在面对环境变化、类别变化、光照强度、遮挡等复杂情况下仍然具有很好的鲁棒性。

The current object tracking algorithm is mainly based on the detection framework,namely detection⁃based tracking(DBT),so the detection quality has a great impact on the tracking performance of the algorithm.At the same time,the object tracking algorithm is susceptible to environment interference,illumination change,object scale and category in the tracking process.In view of the above problems of object tracking,a new object tracking algorithm combining deep learning neural network YOLOv4 detection algorithm and Kalman filtering is proposed.The object detector is used to classify the objects and extract the boundary boxes.The tracker is used to collect candidate data in the tracking trajectory.An object selector is designed to select the optimal candidate box in the detection and tracking trajectory.The optimal candidate box and tracking trajectory are subjected to data association by ReID(re⁃identification)to determine whether to update the tracking trajectory.The experimental results show that,in comparison with the other established algorithms,the precision plots of OPE of the proposed method is 84.9%and its success plots of OPE is 82.2%.In addition,this method still has good robustness in the complex situations,such as changes of environment,category,illumination intensity and occlusion.