Abstract: In order to address local occlusion and deformation of target tracking, an inverse sparse visual target tracking strategy with local cosine similarity training weights is proposed in this paper. According to the core idea of particle filter in parameter space, based on the framework of inverse sparse, the inverse sparse representation coefficients of candidates are obtained by using the candidates to reconstruct template, and the best candidate which is more similar with template is selected according to the feature of coefficient distribution. In order to overcome the influence of occlusion, a new objective function is introduced to improve the discrimination of the local block in template：calculate the local cosine similarity difference between the positive samples and negative samples and the template, and then the function of quadratic optimization is used to update weights with ability of discrimination. In order to avoid invalidation of template updating, this paper updates template based on the information of weights selectively. Experimental results demonstrate that the proposed algorithm is robust on challenging benchmark video sequences with complicated conditions including occlusion and so on, has its own advantages compared with other existing algorithms.

Key words: visual tracking, particle filter, inverse sparse representation, local cosine similarity

摘要： 针对目标跟踪的遮挡与局部形变，提出局部余弦相似度训练权重的逆稀疏视觉目标跟踪策略。借鉴参数空间的粒子滤波的核心思想，以逆稀疏表示为理论框架，用候选目标重构模板获得候选目标的稀疏表示系数，依据表示系数分布特征筛选出最佳候选目标。为克服遮挡影响，引入新的目标函数构建模板的局部块判别能力：计算正负样本与模板之间的局部余弦相似度差值，利用二次优化，更新具有判别能力的权重。依据权重信息综合进行有选择的模板更新，避免模板更新的无效性。多组实验结果表明，该算法在部分遮挡等复杂环境下，仍然可以准确地跟踪目标，相比已有算法具有自己的优势。

关键词: 目标跟踪, 粒子滤波, 逆稀疏表示, 局部余弦相似度