基于多子块联合估计的相关滤波跟踪

针对遮挡情况下相关滤波算法跟踪精度下降的问题，提出了一种基于多子块联合估计的核相关滤波跟踪方法。首先依据初始帧跟踪框的几何特征对目标自适应分块，并采用KCF方法对各子块独立跟踪得到联合置信图；然后以上帧目标的位置及尺度作为先验信息对搜索区域采样，同时将样本框中置信图的权值密度作为观测值，利用粒子滤波算法实现候选目标的最优估计；最后对置信度较低的子块反向投影至上帧图像进行遮挡检测，防止模板错误更新。定性和定量实验结果表明，该方法与原始KCF算法相比跟踪精度提升约10%，具有良好的抗遮挡性，并对目标尺度变化具有一定的估计能力。

Correlation filter object tracking based joint multi-blocks estimation

In order to solve the problem that the poor tracking accuracy of correlation filter algorithm under occlusion condition, this papre proposed a kernelized correlation filter tracking method using joint multiple blocks estimation. Firstly, it divided the target into several blocks adaptively according to the geometric features of the initial frame tracking box, and each block using the KCF method for tracking independently to get a combined confidence map. Then it took the location and scale of previous frame target as priori information sampling the search area, meanwhile, the weight density of the confidence map in the sample box is used as the observation value, achieve optimal estimation the candidate target using particle filter algorithm. Finally, blocks with lower confidence levels back-project to previous frame for occlusion detection to prevent template update mistakenly. The qualitative and quantitative experiment results shows that compared with the original KCF algorithm, the tracking accuracy of the proposed method improves about 10% and it is robust to occlusion and scale change in some degree.