基于多模态模板的抗遮挡Staple跟踪算法  

Anti-occlusion Staple Tracking Algorithm Based on Multimodal Template

作　　者：黄育明 戴奕婧 李丽惠 何博 严嘉怡 陈振雕 陈颖频 HUANG Yuming;DAI Yijing;LI Lihui;HE Bo;YAN Jiayi;CHEN Zhendiao;CHEN Yingpin(Minnan Normal University School of Physics and Information Engineering,Zhangzhou 363000,China;Electric and Information School,Zhangzhou Institute of Technology,Zhangzhou 363000,China;University of Electronic Science and Technology of China,School of Mathematical Sciences,Chengchu 611731,China)

机构地区：[1]闽南师范大学物理与信息工程学院,福建漳州363000 [2]漳州职业技术学院电子信息学院,福建漳州363000 [3]电子科技大学数学科学学院,四川成都611731

出　　处：《探测与控制学报》2023年第2期99-108,114,共11页Journal of Detection & Control

基　　金：福建省自然科学基金项目(2020J05169);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(JAT210278,JAT190393,JAT190382);闽南师范大学高级别科研项目(GJ19019);大学生创新创业训练计划项目(202210402009,S202210402038,S202210402025)。

摘　　要：Staple算法将方向梯度直方图与颜色直方图两种视觉特征进行结合,互相弥补彼此的不足;然而Staple未考虑上下文信息且未对每帧最大响应样本进行置信度评估,在遮挡场景下鲁棒性较差。为解决上述问题,提出一种基于多模态模板池的抗遮挡Staple跟踪算法。首先,引入上下文感知框架与Staple进行融合,有效解决了运动模糊、背景复杂等挑战下跟踪效果受限的问题;然后,使用金字塔尺度搜索策略,提高原算法在尺度变化场景下的跟踪鲁棒性;最后,建立历史多模态目标池,提高对可靠样本的筛选能力,有效解决了遮挡挑战下跟踪漂移的问题。将提出方法在OTB数据集上进行测试并与其他先进跟踪算法比对,实验结果表明,该算法性能优于比对算法。The Staple algorithm combines the two visual features of the directional gradient histogram and the color histogram to make up for each other s shortcomings.However,Staple does not consider context information and does not perform confidence evaluation on the sample with maximum response per frame,which is less robust in the occlusion scene.To solve the above problems,an anti-occlusion Staple tracking algorithm based on multimodal template pooling was proposed in this paper.First of all,the context-aware framework and Staple were introduced for fusion,which effectively solved the problem of limited tracking effect under the challenges of motion blur and complex background.Then,the pyramid scale search strategy was used to improve the tracking robustness in scale-changing scenarios;Finally,a historical multi-modal target pool was established,which improved the screening ability of reliable samples and effectively solves the problem of tracking drift under occlusion challenges.The proposed method was tested on the OTB dataset and compared with other advanced tracking algorithms,the experimental results showed that the performance of the proposed method outperformed the comparison algorithm.

关 键 词：目标跟踪 互补特征 上下文感知模型 多模态模板 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]