基于局部敏感核稀疏表示的视频跟踪

摘要

In order to solve the problem of lack of discriminability in thel1-norm constraint sparse representation, visual tracking via locality-sensitive kernel sparse representation is proposed. To improve the linear discriminable power, the candidates’ Scale-Invariant Feature Transform (SIFT) is mapped into high dimension kernel space using the Gaussian kernel function. The locality-sensitive kernel sparse representation is acquired in the kernel space. The candidates’ representation are obtained after multi-scale maximum pooling. Finally, the candidates’ representation is put into the classifier and the candidate with the biggest Support Vector Machines (SVMs) score is recognized as the target. And the experiments demonstrate that the robustness of the proposed algorithm is improved due to the use of the data locality under the kernel sparse representation.%为了解决l1范数约束下的稀疏表示判别信息不足的问题，该文提出基于局部敏感核稀疏表示的视频目标跟踪算法。为了提高目标的线性可分性，首先将候选目标的SIFT特征通过高斯核函数映射到高维核空间，然后在高维核空间中求解局部敏感约束下的核稀疏表示，将核稀疏表示经过多尺度最大值池化得到候选目标的表示，最后将候选目标的表示代入在线的SVMs，选择分类器得分最大的候选目标作为目标的跟踪位置。实验结果表明，由于利用了核稀疏表示下数据的局部性信息，使得算法的鲁棒性得到一定程度的提高。