核相关滤波与孪生网络相结合的目标跟踪算法

Traget Tracking Algorithm Combining Kernel Correlation Filter and Siamese Network

针对核相关滤波目标跟踪算法中对局部上下文区域图像提取的HOG特征图在复杂环境下不能保证目标跟踪的精度问题,提出了一种核相关滤波与孪生网络相结合的目标跟踪算法.首先在首帧输入图像中提取HOG特征图并建立相关滤波器模板,同时提取经过孪生网络的目标区域图像特征图;然后若后续帧输入图像帧数不为5的倍数则提取仿射变换HOG特征图,否则提取经过孪生网络的搜索区域图像特征图;最后根据遮挡处理的结果自适应获取目标位置并更新模型和最终相关滤波器模板.仿真实验结果表明本文算法在保证目标跟踪精度的前提下具有满足实时跟踪要求的跟踪速率.

Aiming at the problem that the HOG feature map extracted from the local context area image in the kernel correlation filter target tracking algorithm can’t guarantee the accuracy of traget tracking in complex environments.A target tracking algorithm combining kernel correlation filter and siamese network is proposed.Firstly,extracting the HOG feature map to establish a correlation filter template in the first frame input image,and simultaneously extracting target area image feature map through siamese network.Then if the number of the input image frame in subsequent frames is not a multiple of five,extracting the affine transformation HOG feature map,otherwise extracting serach area image feature map through siamese network.Finally,according to the result of occlusion process to adaptively obtain the target position and update model and the final correlation filter template.Simulation results show that the algorithm in this paper has a tracking rate that meets the requirements of real-time tracking on the premise of ensuring traget tracking accuracy.