基于Struck的在线学习和相似度匹配的双重更新跟踪算法  

Struck-based double update tracking algorithm for online learning and similarity matching

作　　者：袁振 王传江 张腾飞[1,2] YUAN Zhen;WANG Chuanjiang;ZHANG Tengfei(College of Electrical Engineering and Automation,Shandong University of Science and Technology,Qingdao,Shandong 266590,China;Shandong Provincial Key Laboratory of Robotics and Intelligent Technology,Qingdao,Shandong 266590,China;School of Control Science and Engineering,Shandong University,Jinan,Shandong 250061,China)

机构地区：[1]山东科技大学电气与自动化工程学院,山东青岛266590 [2]山东省机器人与智能技术重点实验室,山东青岛266590 [3]山东大学控制科学与工程学院,山东济南250061

出　　处：《山东科技大学学报（自然科学版）》2019年第6期74-80,共7页Journal of Shandong University of Science and Technology(Natural Science)

基　　金：国家自然科学基金青年基金项目（61803235）;山东省重点研发计划项目（2016GSF201197）;山东省高等学校科技计划项目（J16LB11）

摘　　要：针对Struck算法在遇到完全遮挡后难以恢复目标的跟踪问题,提出了利用双重更新策略对目标进行跟踪的算法。首先标定首帧中目标的所在位置,提取目标特征作为初始模板。其次,设计相似函数判别当前帧目标区域与初始模板的相似度,超过阈值的区域选为正样本加入到在线学习的过程。最后,当目标遇到完全遮挡时,通过遍历搜索的方式寻找与目标相似的图像块,选择超过阈值中相似度最高的图像块作为目标继续跟踪。实验结果表明,改进后的算法可以更好地解决因遮挡或背景相近等复杂条件引起的目标跟踪丢失的问题。In view of the difficulty of Struck algorithm in recovering the tracking target in the case of complete occlusion,an algorithm for tracking the target by using the double update strategy was proposed in this paper.First,the position of the target in the first frame was calibrated,and the target feature was extracted as the initial template.Secondly,the similarity function was designed to discriminate the similarity between the current frame target area and the initial template.The area above the threshold was selected as the positive sample to join the online learning process.Finally,when the target encountered full occlusion,the image block with similar target was searched by traversing the search.The image block that had the highest similarity among the thresholds was selected as the target to continue tracking.The experimental results show that the improved algorithm can better solve the problem of target tracking loss caused by complex conditions such as occlusion or similar background.

关 键 词：在线学习 相似度匹配 双重更新 结构化内核 视觉目标跟踪 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]