结合置信度评估与再检测的目标长时跟踪

Long-term object tracking combined with confidence evaluation and re-detection

针对目标长时跟踪中遮挡和出视野等因素导致的跟踪准确率降低甚至跟踪失败问题,提出一种结合置信度评估与再检测的目标长时跟踪算法。根据相关滤波响应的平均峰值相关能量与颜色得分判断跟踪结果的可靠性;通过引入DIOU约束,有效限制候选位置的数量,增加惩罚项计算候选框得分,得到最优再检测结果;评估再检测结果的可靠性决定是否替换跟踪位置。在OTB-2015数据集上与先进相关滤波算法进行对比,该算法精确度与成功率分别为0.855和0.798。在目标遮挡和出视野等复杂环境下,具有较强的鲁棒性和实时性。

For the problem of reduced tracking accuracy and even tracking failure caused by factors such as occlusion and out-of-view in long-term target tracking,a long-term target tracking algorithm combining confidence evaluation and re-detection was proposed.The reliability of tracking result was judged by average peak-to-correlation energy of correlation filtering response and color score.By introducing DIOU constraints,the number of candidate positions was limited effectively,and a penalty item was used to calculate candidate score,so as to obtain the optimal re-detection result.The reliability of re-detection results was evaluated to decide whether to replace the tracking position.The accuracy and success rate of the proposed algorithm are 0.855 and 0.798 in comparison with the advanced correlation filtering algorithms on OTB-2015 benchmark datasets.It has strong robustness and real-time performance in complex environments such as target occlusion and out-of-view situations.