全局块与局部块协作的相关滤波目标跟踪算法  

Correlation Filter Object Tracking Algorithm Based on Global and Local Block Cooperation

作　　者：喻露 胡剑锋 姚磊岳[1,2] YU Lu;HU Jian-feng;YAO Lei-yue(School of Information Engineering,Nanchang University,Nanchang 330031,China;Center of Collaboration and Innovation,Jiangxi University of Technology,Nanchang 330098,China)

机构地区：[1]南昌大学信息工程学院,南昌330031 [2]江西科技学院协同创新中心,南昌330098

出　　处：《计算机科学》2020年第6期157-163,共7页Computer Science

基　　金：国家自然科学基金(61762045);江西省科技厅项目(20171BAB202031);江西省科技厅科技攻关项目(20171BBE50060);江西省博士后援助项目(2017KY33);江西省教育厅项目(GJJ161143,GJJ151146);江西省科技厅科技计划专项重点研发项目(20181BBE50018)。

摘　　要：针对传统相关滤波跟踪器在目标尺度变化和部分遮挡时效果不佳等问题,基于KCF提出了一种全局块与局部块协作的分块跟踪算法。该算法首先根据目标的外观特征,对跟踪目标进行水平或垂直分块,并分别训练两个局部滤波器和一个全局滤波器;然后在跟踪过程中使用局部滤波器对局部块进行跟踪,并根据局部块的跟踪结果对全局块的中心点位置进行初始预测。最后通过全局滤波器确定目标的最终位置,并将相关更新和尺度参数反馈给局部滤波器,以更新全局滤波器和局部滤波器。此外,不同于KCF使用单一的HOG特征,该算法合并了CN特征,改善了HOG特征对目标形变和运动模糊的表达能力。另外,为解决目标部分遮挡导致的模型漂移问题,提出了一种基于有效局部块来指导模型更新的方法,并给出了有效局部块的评价标准。同时,该算法通过分析前后两帧局部块之间的距离变化对目标的尺度进行估计,解决了因目标尺度变化带来的跟踪失败问题。实验在包含100个视频序列的公共数据集OTB-100上进行,在评价指标上,以AUC得分为主,DP和OP为辅,对算法的性能进行评估。实验结果表明:所提出的算法能有效应对尺度变化和部分遮挡的问题,AUC得分在KCF的基础上提升了10%,总体性能也比KCF的其他4个改进算法更优,处理速度达到32 fps。Traditional correlation filter trackers are not effective in dealing with the problem that caused by target scale changing and partial occlusion.Aiming at sloving this problem,a block tracking algorithm based on KCF was proposed in this paper.In the first step,tracking object is divided horizontally or vertically according to its apperance feature.Then,in the tracking process,local filter is used to track local block,and center point position of the global block can be predicted by the tracked result of the local blocks.At last,the final position of the target is determined by the global filter.The relevant information renewal and scale parameters are fed back to the local filters to update both global and local filter.In addition,different from KCF,which only uses HOG feature,CN feature is imported in the proposed algorithm to enhace the ability of traget deformation tracing and motion blurring tracing.Moreover,in order to solve model drift problem caused by partial occlusion,a method based on effective local block is raised to guide model updating.Criteria of evaluating effective local block also defined.Furthermore,the scale of the target can be effectively estimated by analyzing the distance between local blocks,which solves tracking failure problem caused by target scale changing.The algorithm is evaluated on the public dataset OTB-100,which contains 100 video samples.The results show that the proposed algorithm performs quite well in the situation of scale changing and partial occlusion.Compared with KCF,the accuracy of the proposed algorithm is improved by 10%,and the overall performance is better than other four KCF based algorithms.The processing speed of the algorithm reaches 32 fps.

关 键 词：目标跟踪 相关滤波 分块 尺度变化 部分遮挡 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]