相关滤波追踪算法中的测地线更新策略

Geodesic updating strategy in tracking with correlation filter

在追踪视频序列中目标物体常常会出现形变、光照变化、旋转等情况,为了适应目标物体发生的这些变化,得到更加鲁棒的追踪效果,需要对求得的滤波器进行更新.经典相关滤波追踪算法中采用的更新策略是对新旧两个滤波器进行简单的加权,得到具有前一帧目标信息的新滤波器.但这种仅使用加权的更新策略忽略了核函数矩阵间可能存在的几何结构,限制了追踪效果的进一步提高.针对滤波器更新问题,提出了基于正定矩阵流形测地线的更新方法.由于自相关核函数矩阵是对称正定矩阵,所以首先在流形上对自相关核函数矩阵进行更新,进而得到相关滤波器的更新策略.在公开数据集OTB-50上,将提出的算法应用于使用Histogram of Oriented Gradient(HOG)特征和Convolutional Neural Networks(CNNs)特征的相关滤波追踪算法中,并与基线算法和其他先进的算法进行了对比实验,实验结果表明,新的更新策略提高了原始算法的精确率与成功率,达到了更好的追踪结果.

In the tracking video sequence,the target object often appears deformation,light change,rotation and other conditions.In order to adapt to these changes of the target object and obtain more robust tracking effect,it is necessary to update the filter.The updating strategy of classical correlation filters algorithm simply summing over the filters of current frame and previous frame in the Fourier domain.However,this strategy ignores the possible geometric structure that may exist between the kernel function matrices,which allows a limitation in the tracing performance.To solve the problem of filter updating,a novel updating method based on positive definite matrix manifold geodesic is proposed.The proposed approach works by exploiting the special structure of the autocorrelation kernel matrix which belongs to the positive definite(PD)matrices manifold.The autocorrelation kernel matrix is first updated on the manifold,and then the updated filter is obtained.We apply our proposed algorithm to the correlation filter tracking algorithm using Histogram of Oriented Gradient(HOG)features and Convolutional Neural Networks(CNNs)features,compared with baseline algorithms and other state-of-the-art algorithms on the public data set OTB-50.The experimental results show that the new update strategy can improve the precision rate and the success rate of the baseline algorithm,achieved better tracking results.