The gray value features of images were used to describe targets in robust visual tracking via L1 minimization (L1 tracking algorithm).However,the ignorance of structural information between the templates and inaccurate modeling of target appearance changes leaded to a low tracking accuracy.What's more,since the low resolution image block with 12× 15 was adopted to balance the tracking speed and tracking effect,it was difficult to obtain sufficient information to represent the target.To solve the problem of the L1 tracking algorithm,a visual tracking method based on online low-rank sparse representation was proposed.Firstly,in order to take advantage of the ability to represent target appearance changes of principal component analysis (PCA) basis vectors,the PCA basis vector template was used to describe target appearance changes.To take occlusion into account,the occlusion and other abnormal noise were modeled with the trivial template.Consequently,the candidate target was represented as a linear combination of PCA basis templates and trivial templates.Secondly,for the optimization problem of object representation model,a low-rank and L1,1 regularization constraint were added on the PCA basis template coefficients and a L1,1 regularization constraint on the trivial template coefficients.Then the inexact augmented lagrange multiplier(IALM) method was adopted to solve the representation coefficients.Thirdly,visual tracking was achieved by taking the reconstruction error of the unoccluded part and sparse error term as an observation model in particle filter framework.Finally,to avoid model drifting,an occlusion detection updating mechanism was proposed to update templates.In the experiments of 8 video image sequences with 32×32 image block resolution,compared with four state-of-the-art trackers,the proposed algorithm had achieved the highest average overlap rate of 0.78 and the lowest average center error of 4.05.Accordingly,the experimental results demonstrated that the proposed algorithm had better tracking accuracy and robustness.%基于L1最小化的鲁棒视觉跟踪算法(L1跟踪算法)使用图像灰度值特征描述目标,忽略了模板间的结构信息,对目标外观变化的建模不够准确,导致跟踪准确度较低.而且L1跟踪算法为了平衡跟踪速度和跟踪效果而采用分辨率较低的12×15图像块,难以获取足够的信息来表征目标.针对L1跟踪算法的不足,该文提出一种在线低秩稀疏表示的视觉跟踪算法.首先,该算法充分利用主成分分析(PCA)基向量对目标外观变化的表示能力并考虑目标遮挡现象,以PCA基向量模板描述目标外观变化,以琐碎模板处理遮挡等异常噪声,从而将候选目标表示为PCA基模板和琐碎模板的线性组合.其次在目标表示模型的优化问题中,对PCA基模板系数进行低秩约束和L1,1范数正则化约束,对琐碎模板系数实施L1,1范数约束,并采用非精确增广拉格朗日乘子(IALM)方法求解表示系数.然后在粒子滤波框架下,用目标未被遮挡部分的重建误差和稀疏误差项建立观测模型跟踪目标.最后为了克服模型漂移问题,采用遮挡检测更新机制进行模板更新.在对8组视频图像序列进行测试的实验中,图像块分辨率设定为32×32,与4个现有的跟踪算法相比,该算法取得了最高的平均重叠率0.78和最低的平均中心误差4.05.实验结果表明,该文提出的跟踪算法具有较好的跟踪准确性和鲁棒性.
展开▼
