欢迎访问《图学学报》 分享到：
图学学报

图学学报 ›› 2022, Vol. 43 ›› Issue (1): 60-69.DOI: 10.11996/JG.j.2095-302X.2022010060

• 图像处理与计算机视觉 • 上一篇    下一篇

局部模板更新逆向联合稀疏表示目标跟踪算法

  

  1. 1. 闽南师范大学计算机学院，福建 漳州 363000；   2. 闽南师范大学数据科学与智能应用福建省高校重点实验室，福建 漳州 363000；   3. 闽南师范大学物理与信息工程学院，福建 漳州 363000
  • 出版日期:2022-02-28 发布日期:2022-02-16
  • 基金资助:
    闽南师范大学校长基金项目(KJ19019)；福建省中青年教师科研教育项目(JAT190378，JAT190393，JAT190382)；闽南师范大学高级别 项目(GJ19019)；福建省大学生创新创业训练计划(202010402016，202110402012)；福建省自然科学基金项目(2020J01816)

Object tracking of reverse joint sparse representation with local template update 

  1. 1. School of Computer, Minnan Normal University, Zhangzhou Fujian 363000, China;   2. Key Laboratory of Data Science and Intelligence Application, Minnan Normal University, Zhangzhou Fujian 363000, China;  3. School of Physics and Information Engineering, Minnan Normal University, Zhangzhou Fujian 363000, China
  • Online:2022-02-28 Published:2022-02-16
  • Supported by:
    Principal Fund of Minnan Normal University (KJ19019); Young and Middle-aged Teachers Research and Education Project of Fujian Province (JAT190378, JAT190393, JAT190382); High-level Project of Minnan Normal University (GJ19019); Fujian University Studentsʹ Innovation and Entrepreneurship Training Plan (202010402016, 202110402012); Natural Science Foundation Project of Fujian Province (2020J01816) 

PDF (PC)

201

摘要： 逆向联合稀疏表示算法可充分利用跟踪过程中的时间相似性和空间连续性，但由于遮挡、光照 变化等的影响，易出现跟踪漂移。为解决上述问题，提出一种基于局部模板更新逆向联合稀疏表示目标跟踪算 法，其通过逆向局部重构目标模板集完成逆向联合稀疏表示。首先，在首帧初始化目标模板集，利用粒子滤波 获取候选图像，并对其分块处理，构建逆向联合稀疏编码模型；然后，利用交替方向乘子法求解出稀疏编码系 数，并通过 2 步评分机制获取最优候选图像；最后，根据相似性得分判断当前帧是否存在局部遮挡，若无遮挡， 则局部更新目标模板集以减少跟踪漂移现象。实验结果表明，本文算法的跟踪精度和成功率在 OTB-2013 数据 集上分别达到了 85.4%和 62.8%，在 OTB100 数据集上分别达到了 76.8%和 68.6%，速度达到每秒 5.76 帧，能 有效提高鲁棒性，减少跟踪漂移。

关键词: 目标跟踪, 逆向联合稀疏表示, 时-空信息, 局部模板更新, 交替方向乘子法

Abstract: The reverse joint sparse representation algorithm can make full use of the temporal similarity and spatial continuity in the tracking process. However, tracking drift can be easily incurred under the influence of occlusion and illumination change. Aiming at this problem, we proposed the reverse joint sparse representation tracker (RJST). It can accomplish the reverse joint sparse representation through the reversely local reconstruction of the object template set. Firstly, the object template set was initialized in the first frame, and the candidate images were generated by particle filtering. They were partitioned into blocks, and the reverse joint sparse representation model was constructed. Then, the sparse coding matrix was solved using the alternating direction method of multipliers. The optimal candidate image was acquired by the two-step scoring mechanism. Finally, whether the current object had local occlusion was evaluated according to the similarity score. If there was no occlusion, the object template set was locally updated to eliminate the tracking drift. Experimental results show that the precision and success rate of RJST reached 85.4% and 62.8% on the OTB-2013 benchmark, and 76.8% and 68.6% on the OTB100 benchmark, respectively, and that the speed was 5.76 frames per second, which can effectively boost robustness and eliminate tracking drift. 

Key words: object tracking, reverse joint sparse representation, spatio-temporal information, local template update; alternating direction method of multipliers

中图分类号: