基于目标分块多特征核稀疏表示的视觉跟踪

Robust Fragments-Based Tracking with Multi-Feature Joint Kernel Sparse Representation

大多数现有的基于稀疏表示的跟踪器仅采用单个目标特征来描述感兴趣的目标,因而在处理各种复杂视频时不可避免会出现跟踪不稳定的情况.针对这个问题,提出一种基于多特征联合稀疏表示的粒子滤波跟踪算法.该算法的主要思想是对随时间不断更新的字典模板和抽样粒子的局部块依据其位置进行分类,用字典中所有类别块对抽样粒子的局部块进行稀疏表示,而仅用与字典中具有相同类别的局部块及表示系数进行重构,根据重构误差构建似然函数以确定最佳粒子(候选目标),实现对目标的精确跟踪.该方法不仅实现了局部块的结构稀疏性,而且充分考虑了粒子之间的依赖关系,提高了跟踪精度.将算法进一步推广到采用基于核的多种特征描述,经混合范数约束并利用KAPG(kernelizable accelerated proximal gradient)方法求解联合特征的稀疏系数.定性和定量的实验结果均表明该算法在目标发生遮挡、旋转、尺度变化、快速运动、光照变化等各种复杂情况下,依然可以准确地跟踪目标.

Most existing sparse representation based trackers only use a single feature to describe the objects of interest and tend to be unstable when processing challenging videos.To address this issue,we propose a particle filter tracker based on multiple feature joint sparse representation.The main idea of our algorithm is to partition each particle region into multiple overlapped image fragments.Eevery local fragment of candidates is sparsely represented as a linear combination of all the atoms of dictionary template that is updated dynamically and is merely reconstructed by the local fragments of dictionary template located at the same position.The weights of particles are determined by their reconstruction errors to realize the particle filter tracking.Our method simultaneously enforces the structural sparsity and considers the interactions among particles by using mixed norms regularization.We further extend the sparse representation module of our tracker to a multiple kernel joint sparse representation module which is efficiently solved by using a kernelizable accelerated proximal gradient（KAPG） method.Both qualitative and quantitative evaluations demonstrate that the proposed algorithm is competitive to the state-of-the-art trackers on challenging benchmark video sequences with occlusion,rotation,shifting and illumination changes.