专注学习时空关系的相关滤波跟踪算法

Correlation Lilter Tracking Algorithm Focusing on Learning Spatio-Temporal Relationship

传统相关滤波跟踪算法试图引入预定义的正则项,如抑制背景学习或限制相关滤波器的学习率来提高算法的鲁棒性,但在复杂场景下还是容易发生目标跟踪丢失,因为传统相关滤波跟踪算法没有关注相邻两帧之间的信息变化。针对以上问题,本文提出专注学习时空关系的相关滤波跟踪算法,引入相邻两帧的响应图变化作为空间正则项权值的参考权重,而当前帧的响应图的振荡程度确定时间正则项权值,最后本文通过交替方向乘子法(ADMM)迭代优化本文的损失函数。通过在OTB-50、OTB-100和OTB-2013三个基准数据集上进行了实验,验证了本文算法在复杂场景下更具有鲁棒性。

Traditional correlation filter tracking algorithms try to introduce predefined regularization terms,such as restraining background learning or limiting the learning rate of correlation filter,to improve the robustness of the algorithm.However,in complex scenes,target tracking loss is easy to occur,because the traditional correlation filter tracking algorithm does not pay attention to the information changes between two adjacent frames.In order to solve the above problems,this paper proposes a correlation filter tracking algorithm focusing on learning the spatial-temporal relationship.The change of the response graph of two adjacent frames is introduced as the reference weight of the weight value of the spatial regularization term,and the oscillation degree of the response graph of the current frame determines the weight value of the temporal regularization term.Experiments on OTB-50,OTB-100 and OTB-2013 benchmark datasets show that the proposed algorithm is more robust in complex scenes.