核相关滤波融合稀疏表示的目标跟踪算法  

Object tracking algorithm based on kernelized correlation filters and sparse representation

作　　者：邱泽敏 QIU Ze-min(School of Information Science,Xinhua College of Sun Yat-Sen University,Guangzhou 510520,China)

机构地区：[1]中山大学新华学院信息科学学院,广东广州510520

出　　处：《计算机工程与设计》2022年第3期757-763,共7页Computer Engineering and Design

基　　金：2018年度广东省普通高校重点科研平台和科研项目基金项目(2018KTSCX316);2019年中山大学新华学院校级科研基金项目(2019KYYB03);2020年广州市科技创新发展专项资金项目基础与应用基础研究基金项目(202002030273);2020中山大学新华学院校级重点学科基金项目(2020XZD02)。

摘　　要：为解决稀疏表示跟踪速度慢和核相关滤波算法难以处理遮挡、尺度变化的问题,提出一种核相关滤波结合稀疏表示的视频跟踪算法。用核相关滤波快速定位目标的最大可能位置,密集采样后采用基于局部图像块的稀疏表示和对齐池化构建外观模型,更好表征目标的局部特征以应对遮挡问题。为处理跟踪漂移,采取基于滤波和稀疏编码的探测策略实现模型更新。OTB100上的实验结果表明,该算法在处理遮挡、旋转和快速运动上与其它算法相比达到最好的跟踪效果。An algorithm combining sparse representation and correlation filter was proposed,to solve the low tracking speed of sparse representation and the difficulty in processing occlusion and scale changes in kernelized correlation filters.The kernelized correlation filters were used to locate the maximum possible position of the target rapidly,the appearance model was constructed using the sparse representation with local image block and aligning pooling after dense sampling,which represented the local characteristics of the target and dealt with the occlusion problem preferably.To deal with tracking drift,the model was updated after detection based on filtering and sparse coding.Experimental results on OTB100 show that the proposed algorithm achieves the best tracking effects compared with other algorithms in processing occlusion,rotation,and fast motion.

关 键 词：核相关滤波 稀疏表示 外观模型 对齐池化 目标跟踪 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]