采用局部-全局区域重检测机制的无人机长期跟踪算法

A Long-Term Unmanned Aerial Vehicle Tracking Algorithm Using Local-Global Region Redetection Mechanism

为解决基础跟踪器面对遮挡和移出视野等长期跟踪场景时易出现跟踪失败等问题,提出了一种基于局部-全局区域重检测的无人机长期跟踪算法。设计了基础滤波器,将高置信度样本与其结合,并融入自适应时空正则化,解决了滤波器退化问题,提高了模型鲁棒性以及复杂场景下的性能;优化了滤波器更新策略,通过评价跟踪结果进行自适应更新;设计快速尺度滤波器,解决了跟踪过程中的尺度变化问题;设计了局部-全局区域重检测机制,跟踪失败时启动重检测器恢复跟踪目标,先完成局部区域重检测,若恢复跟踪失败,再利用全局区域重检测器继续恢复目标跟踪状态。实验结果表明:所提算法在UAV20L数据集上的精确度和准确率分别可达0.724和0.621,与基于时空正则化相关滤波器的跟踪算法(STRCF)相比分别提升了25.9%和20.6%,与同类主流算法相比,跟踪效果得到提升,证明了算法的有效性。

In order to solve the problem that the basic tracker is prone to tracking failure in long-term tracking scenarios such as occlusion and out of view,a long-term tracking algorithm for unmanned aerial vehicle(UAV)based on local-global region redetection is proposed.The basic filter is designed,and the high-confidence samples are combined with it,and the adaptive spatio-temporal regularization is integrated to solve the filter degradation problem and improve the robustness of the model and its performance in complex scenarios.The filter update strategy is optimized,and the adaptive update is performed by evaluating the tracking results.A fast scale filter is designed to solve the problem of scale change in the tracking process.A local-global region redetection mechanism is designed.When the tracking fails,the re-detector is started to recover the tracking target and the local region re-detection is completed first.If the tracking recovery fails,the global region re-detector is used to continue to recover the target tracking state.The experimental results show that the precision and accuracy of the proposed algorithm on the UAV20L data set can reach 0.724 and 0.621 respectively,representing improvement of 25.9%and 20.6%respectively compared with the STRCF algorithm.Compared with the similar mainstream algorithms,the tracking effect of the algorithm is improved,which proves its effectiveness.