在线低秩稀疏表示的鲁棒视觉跟踪

发布时间：2018-05-10 23:04

本文选题：视觉跟踪 + 低秩表示　；参考：《工程科学与技术》2017年04期

【摘要】：基于L_1最小化的鲁棒视觉跟踪算法(L_1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L_1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L_(1,1)范数正则化约束,对琐碎模板系数实施L_(1,1)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。
[Abstract]:The robust visual tracking algorithm based on L1 minimization uses image grayscale feature to describe the target, neglecting the structure information between templates, and modeling the appearance change of the target is not accurate enough, which leads to the low tracking accuracy. In order to balance the tracking speed and tracking effect, L1 tracking algorithm uses 12 脳 15 image blocks with low resolution, so it is difficult to obtain enough information to represent the target. Aiming at the deficiency of LST1 tracking algorithm, this paper presents a visual tracking algorithm based on online low rank sparse representation. Firstly, the algorithm makes full use of the ability of principal component analysis (PCA) basis vector to represent the appearance change of the target and considers the occlusion phenomenon. The PCA basis vector template is used to describe the appearance change of the target, and the trivial template is used to deal with the abnormal noise such as occlusion. The candidate target is expressed as a linear combination of PCA base template and trivial template. Secondly, in the optimization of the target representation model, the low rank constraint and the norm regularization constraint are applied to the PCA base template coefficient, and the L / L / 1) norm constraint is applied to the trivial template coefficient. An inexact augmented Lagrangian multiplier (IALM) method is used to solve the representation coefficient. Then, in the framework of particle filter, the reconstruction error and sparse error of the unoccluded part of the target are used to establish the observation model to track the target. Finally, in order to overcome the problem of model drift, occlusion detection and update mechanism is used to update the template. In the experiment of testing 8 groups of video sequences, the resolution of the image block is set to 32 脳 32. Compared with the four existing tracking algorithms, the proposed algorithm achieves the highest average overlap rate of 0.78 and the lowest average center error of 4.05. Experimental results show that the proposed tracking algorithm has good tracking accuracy and robustness.
【作者单位】：南京航空航天大学航天学院;
【基金】：国家自然科学基金资助项目(61401200;61201365) 江苏省普通高校研究生科研创新计划资助项目(SJLX15_0138)
【分类号】：TP391.41

【参考文献】