协同结构稀疏重构的判别性视觉跟踪

摘要

基于稀疏表示的表观似然模型在目标跟踪领域具有广泛的应用,但是这种单一产生式目标表观模型并未考虑完整的判别性结构信息,容易受复杂背景的干扰.为了缓解由该问题造成的目标跟踪漂移,提出了一种目标表观字典和背景字典协同结构稀疏重构优化的视觉跟踪方法.通过构建一个有判别力的基于稀疏表示的表观似然模型,实现了对目标表观模型更为准确的描述.通过合理选择约束候选目标区域和候选背景区域的稀疏系数,在表观似然模型中引入判别式信息,以进一步揭示候选目标区域的潜在相关性和候选背景区域的结构关系,从而更加准确地学习候选目标区域的表观模型.大量有挑战性的视频序列上的实验结果验证了算法在复杂背景下跟踪的鲁棒性,与其他相关算法的对比实验也体现了该算法的优越性.%Though the appearance likelihood model based on sparse representation has been widely applied in visual tracking,the single generation object representation model can easily be interrupted by background clutter due to not considering the full discriminative structural information.In order to alleviate the drift problem of the visual tracking, this paper presented a novel tracking method based on collaborative sparse reconstruction of object appearance dictiona-ry and background dictionary.This paper achieved a more accurate description of the target appearance model by con-structing a discriminative appearance likelihood model based on sparse representation.Then,it embedded discriminative information into the appearance likelihood model by a reasonable method of selecting the sparse coefficients of candidate target region and candidate background region.By that way,it can explore the potential correlation of candidate target region and the structure relation of candidate background region,so as to learn the appearance model of candidate target area more accurately.Many experimental results in challenging sequence verify the robustness of this method.The pro-posed tracker outperforms excellent performance in comparison with other state-of-the-art trackers.