L_2范数正则化鲁棒编码视觉跟踪

Robust Coding via L_2-norm Regularization for Visual Tracking

针对基于稀疏表示的视觉跟踪计算效率低和易于产生"模型漂移"的不足,该文提出一种基于L2范数正则化鲁棒编码的视觉跟踪方法。该方法利用L2范数正则化鲁棒编码求解候选目标的编码系数,以粒子滤波为框架,利用候选目标的加权重建误差建立似然模型跟踪目标。为了适应目标的变化并克服"模型漂移"问题,利用L2范数正则化鲁棒编码估计当前目标的加权矩阵用于遮挡检测,根据遮挡检测结果实现模型更新。对提出的跟踪方法进行实验的结果表明:与现有跟踪方法相比,该方法具有较优的跟踪性能。

Sparse representation based visual trackers are very computationally inefficient and prone to model drifting. To deal with these issues, a novel visual tracking method is proposed based on L2-norm regularized robust coding. The proposed method solves the coding coefficient of candidate objects via robust coding based on L2-norm regularization, and it achieves visual tracking by taking weighted reconstruction errors of the candidate object as observation likelihood in particle filter framework. In addition, to adapt the changes of object appearance and avoid model drifting, an occlusion detection method for template update is proposed by investigating the weight matrix of current object estimated with L2-norm regularized robust coding. The experimental results on several challenging sequences show that the proposed method has better performance than that of the state-of-the-art tracker.