样本分块稀疏表示判决式目标跟踪  

Sample blocking sparse representation discriminative target tracker

作　　者：侯跃恩[1] 李伟光[2] Hou Yue＇en;Li Weiguang(School of Computer,Jiaying University,Meizhou Guangdong 514000,China;College of Mechanical Engineering,South China University of Technology,Guangzhou 510640,China)

机构地区：[1]嘉应学院计算机学院,广东梅州514000 [2]华南理工大学机械工程学院,广州510640

出　　处：《计算机应用研究》2018年第8期2510-2513,2531,共5页Application Research of Computers

基　　金：国家"863"计划资助项目(2015AA043005);国家高等教育教学改革重点项目(JYJG20170109)

摘　　要：为了提高目标跟踪算法的鲁棒性和准确性,提出了一种粒子滤波框架下的样本分块稀疏表示判决式跟踪算法。算法在首帧提取目标模板和背景模板,并将这些模板进行分块,构建模板字典。然后将候选目标进行分块处理,并使用模板字典稀疏重构候选目标分块,从而获得候选目标的稀疏系数和残差,进而构建一款贝叶斯分类器。分类器的输入为候选目标稀疏系数和残差中提取的相似度信息,输出为候选目标与真实目标的相似度。分类器通过跟踪过程中获得的正负样本进行训练,使之能够适应目标和背景的变化。最后,将所提算法在八组具有挑战性的视频中进行测试,平均跟踪误差为5.9个像素,跟踪成功率为89%。与选取的三种先进的算法比较,所提算法具有更高的鲁棒性和准确性。For improving the robustness and accuracy of tacking algorithm, this paper proposed a sample blocking sparse representation discriminative tracking algorithm, which was under the particle filter framework. Firstly, the algorithm sampled target and background templates in the first frame, and divided each template into several patches, and built a template dictionary. Secondly, each candidate target was divided into several patches. It sparsely rebuilt each patch by the template dictionary, and obtained the coefficients and residual errors of candidate targets. Thirdly, it constructed a Bayes classifier. The classifier’ s input was similar information extracted from coefficients and residual errors while the output was the likelihood between the candidate target and the real target. For adapting the changes of the target and the background, it trained the classifier by positive and negative samples, which were obtained during the tracking process. It tested the proposed tracker in 8 challenging video sequences. The average tracking error is 8.9 pixels, and the average success rate is 89%. Compared with 3 state-of-the-art trackers, the proposed tracker is more robust and accurate.

关 键 词：粒子滤波 样本分块 稀疏表示 分类器 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]