Abstract: The Sparse Representation Classification（SRC） has been applied in many machine vision applications including object tracking. Despite its popularity, most existing SRC methods have low tracking accuracy. To solve these problems, this paper proposes an improved version of SRC by combining Discriminative Analysis（DA） principles and sparse decomposition. The reason why DA can improve SRC has been given. It also proposes a NormalHedge based method to estimate the location and size of the target. Furthermore, it proposes a time-loop based method to update target templates. To efficiently cope with sudden disappearances and sudden tracking failures, it proposes a cooperative tracking framework which can resume tracking process after sudden tracking failure occurs. Two types of?experiments have been performed. The first type is to evaluate the proposed methods. The second type is to compare the proposed method with other classical tracking methods on six carefully selected videos. Experiment results show the proposed algorithm and framework yield robust performances under multiple challenging conditions. Limitations of the proposed method are also mentioned.

Key words: object tracking, template updating, sparse representation, discriminative analysis, classifier

摘要： 针对当前基于稀疏分类的目标跟踪算法跟踪精度较低等问题，结合判别分析思想，提出改进型稀疏跟踪算法。采用基于在线学习的标准对冲算法估算目标的位置以及面积，并详细介绍了标准对冲算法原理。对于在跟踪过程中目标外形改变的问题，提出了基于时序循环的模板更新方法。对目标暂时消失或被完全遮挡时会产生跟踪失败的问题，创造性地提出了基于稀疏分类器网格SCG的合作跟踪框架。进行了两类实验，第一类实验验证了该算法的有效性。第二类实验在大量公共图像序列的基础上对该算法及其他图像跟踪算法进行测试比较。实验结果证明，该算法适用于复杂背景下的跟踪任务，在跟踪失败后能自动恢复跟踪，在目标被部分遮挡、长期遮挡或目标与背景有相似特征模式的情况下都能保持较高的跟踪精度。

关键词: 目标跟踪, 模板更新, 稀疏表示, 判别分析, 分类器