融入深度特征的多模板相关滤波跟踪算法

针对现实场景中跟踪算法因背景杂乱、遮挡、尺度变化、目标形变等情况易导致跟踪失败的问题,提出融入深度特征的多模板相关滤波跟踪算法.首先对图像或图像区域分别提取深度特征和Color Name特征,经过核相关滤波器学习得到不同的模板;然后采用核相关滤波跟踪算法获得2个特征下的响应集合,并对所得到的集合进行加权融合得到最终的目标位置;最后使用贝叶斯统计通过最大化后验的方式估计最佳目标尺度,同时更新核相关滤波器参数,以实现自适应尺度的目标跟踪.在OTB2013和OTB2015这2个基准数据库上进行实验,并与当前6种优秀的算法进行比较,结果表明该算法性能最优,在2个数据集上的成功率OP(AUE)较KCF算法分别提升10.7%和12.4%.

Multi-template Correlation Filter Tracking Based on Deep Feature

In view of the problem that tracking algorithm may easily fail in realistic scene due to background clutter, occlusion, scale change, target deformation, etc., a multi-template correlation filter tracking algorithm fusing deep feature is proposed. Firstly, the deep feature and Color Name feature of images or image regions are extracted separately. The parameters of multiple templates are learned by kernel correlation filters. Then the kernel correlation filter tracking algorithm is used to obtain the response sets under the two kinds of features. The final target position is obtained by weighted fusion of the response in response sets. Finally, in order to achieve the adaptive scale target tracking, Bayesian statistics is utilized to estimate the optimal target scale and update parameters of kernelized correlation filter simultaneously by maximizing the posterior. Comparative experiments are conducted on the OTB2013 and OTB2015 benchmark databases, involving comparisons of 6 excellent tracking algorithms in the current. The results show that the proposed algorithm has the best performance. The success rate OP(AUE) on the two databases exceeds the KCF algorithm by 10.7% and 12.4% respectively.