相关滤波目标跟踪算法鲁棒性提升研究

Research on improving the robustness of correlation filtering target tracking algorithm

摘要:

摘要：ECO算法虽然在跟踪效果上有了很大的提升，但是它在复杂场景中表现不好甚至会丢失目标，即鲁棒性不高。对此，文章探讨了怎样在ECO算法中利用深度特征处理深层语义的能力和浅层特征处理纹理颜色信息的能力来提升算法的鲁棒性，同时对比了深度特征和浅层特征在目标跟踪的的不同作用，并因此提出了改进方法，首先在深度网络上选择了具有更深层次的ResNet-101网络；其次修改了适宜此网络的参数。算法在OTB-2015进行的实验也取得了比较良好的结果，在低分辨率，背景杂波，光照变化，尺度变化四个挑战因素的成功率分别领先基准算法ECO为0.135，0.034，0.031，0.024。

Abstract:

Absrtact: Although ECO algorithm has greatly improved the tracking effect, it does not perform well in complex scenes and even loses the target, that is, its robustness is not high. In this regard, this paper discusses how to improve the robustness of ECO algorithm by using the ability of deep feature to process deep semantics and the ability of shallow feature to process texture color information. At the same time, it compares the different functions of depth feature and shallow feature in target tracking, and puts forward an improved method, Firstly, resnet-101 network with deeper level is selected on the deep network; Secondly, the parameters suitable for this network are modified. The experiment of the algorithm in otb-2015 also achieved good results,The success rates of low resolution, background clutter, illumination change and scale change are 0.135, 0.034, 0.031 and 0.024 respectively.