第16卷第4期 智能系统学报 Vol.16 No.4 2021年7月 CAAI Transactions on Intelligent Systems Jul.2021 D0:10.11992/tis.202005027 网络出版地址：https:/ns.cnki.net/kcms/detail/23.1538.TP.20210412.1132.004html 相关滤波的运动目标抗遮挡再跟踪技术 戴煜彤，陈志国，傅毅 (1.江南大学人工智能与计算机学院，江苏无锡214122：2.无锡环境科学与工程研究中心，江苏无锡 214153) 摘要：针对相关滤波在抗遮挡方面效果不佳的问题，本文在ECO_HC(efficient convolution operators handcraft)的基础上提出了一种多特征融合的抗遮挡相关滤波算法。在相关滤波算法的框架下，将目标 ULBP(uniform local binary pattern)纹理特征和目标HOG(histogram of oriented gridients)特征进行线性加权融合； 在模型建立与更新阶段通过高斯掩码函数缓解循环移位造成的边界效应：通过计算目标最大响应值的峰值均 值比来判断目标状态，并将卡尔曼算法作为目标被遮挡后重定位策略。实验结果显示，在16个视频序列上，该 文算法的平均精确度达到87.3%.成功率达到76.5%.相比基线算法.分别提升了27.7%和23.7%。 关键词：目标跟踪：相关滤波；特征融合；ULBP:高斯掩码；参数峰值均值比；卡尔曼预测：抗遮挡 中图分类号：TP391.41文献标志码：A文章编号：1673-4785(2021)04-0630-11 中文引用格式：戴煜形，陈志国，傅毅.相关滤波的运动目标抗遮挡再跟踪技术J.智能系统学报，2021,16(4)：630-640. 英文引用格式：DAI Yutong,.CHEN Zhiguo,FUYi.Anti-occlusion retracking technology for a moving target based on correlation filtering J CAAI transactions on intelligent systems,2021,16(4):630-640. Anti-occlusion retracking technology for a moving target based on correlation filtering DAI Yutong',CHEN Zhiguo',FU Yi2 (1.School of Artificial Intelligence and Computer,Jiangnan University,Wuxi 214122,China;2.Wuxi Research Center of Environ mental Science and Engineering,Wuxi 214153,China) Abstract:To address the poor anti-occlusion effect of correlation filtering,this paper proposes an anti-occlusion correla- tion filtering algorithm by means of multifeature fusion based on efficient convolution operators handcraft.First,based on the framework of correlation filtering,a method of linearly weighted fusion is adopted to deal with the target uni- form local binary pattern texture feature and the target histogram of oriented gradients feature.Second,the Gaussian mask function is used during the model establishment and update phase to ease the boundary effect caused by cyclic shift.Lastly,the target state is judged by calculating the peak-to-average ratio of the target maximum response value, and the Kalman algorithm is utilized as the relocation strategy after the target is blocked.Experimental results show that the average accuracy of the proposed algorithm reaches 87.3%,and the success rate reaches 76.5%on 16 test sequences, which are 27.7%and 23.7%higher than those of the baseline algorithm,respectively. Keywords:object tracking;correlation filter;multi-feature fusion;ULBP;Gaussian mask;peak-to-average ratio;Kal- man prediction;anti-occlusion 目标跟踪山近年来因其横跨视频监控、无人 收稿日期：2010-05-21.网络出版日期：2021-04-12. 驾驶、无人飞行器、医学图像分析、空中预警等诸 基金项目：江苏省高等学校自然科学研究面上项目 (17KJB520039):江苏省“333高层次人才培养工程科 多领域而迅速成为计算机视觉研究的热点之一。 研项目”(BRA2018147):江苏省高校“青蓝工程”项 目(2020年). 日标跟踪的主流方法目前正由生成类方法逐渐转 通信作者：陈志国.E-mail:427533@qq.com 向判别式方法，其中基于相关滤波的目标跟踪算DOI: 10.11992/tis.202005027 网络出版地址: https://kns.cnki.net/kcms/detail/23.1538.TP.20210412.1132.004.html 相关滤波的运动目标抗遮挡再跟踪技术 戴煜彤1 ，陈志国1 ，傅毅2 （1. 江南大学 人工智能与计算机学院，江苏 无锡 214122; 2. 无锡环境科学与工程研究中心，江苏 无锡 214153） 摘 要 ：针对相关滤波在抗遮挡方面效果不佳的问题，本文在 ECO_HC(efficient convolution operators handcraft) 的基础上提出了一种多特征融合的抗遮挡相关滤波算法。在相关滤波算法的框架下，将目标 ULBP(uniform local binary pattern) 纹理特征和目标 HOG(histogram of oriented gridients) 特征进行线性加权融合； 在模型建立与更新阶段通过高斯掩码函数缓解循环移位造成的边界效应；通过计算目标最大响应值的峰值均 值比来判断目标状态，并将卡尔曼算法作为目标被遮挡后重定位策略。实验结果显示，在 16 个视频序列上，该 文算法的平均精确度达到 87.3%，成功率达到 76.5%，相比基线算法，分别提升了 27.7% 和 23.7%。 关键词：目标跟踪；相关滤波；特征融合；ULBP；高斯掩码；参数峰值均值比；卡尔曼预测；抗遮挡 中图分类号：TP391.41 文献标志码：A 文章编号：1673−4785(2021)04−0630−11 中文引用格式：戴煜彤, 陈志国, 傅毅. 相关滤波的运动目标抗遮挡再跟踪技术 [J]. 智能系统学报, 2021, 16(4): 630–640. 英文引用格式：DAI Yutong, CHEN Zhiguo, FU Yi. Anti-occlusion retracking technology for a moving target based on correlation filtering[J]. CAAI transactions on intelligent systems, 2021, 16(4): 630–640. Anti-occlusion retracking technology for a moving target based on correlation filtering DAI Yutong1 ，CHEN Zhiguo1 ，FU Yi2 (1. School of Artificial Intelligence and Computer, Jiangnan University, Wuxi 214122, China; 2. Wuxi Research Center of Environ￾mental Science and Engineering, Wuxi 214153, China) Abstract: To address the poor anti-occlusion effect of correlation filtering, this paper proposes an anti-occlusion correla￾tion filtering algorithm by means of multifeature fusion based on efficient convolution operators handcraft. First, based on the framework of correlation filtering, a method of linearly weighted fusion is adopted to deal with the target uni￾form local binary pattern texture feature and the target histogram of oriented gradients feature. Second, the Gaussian mask function is used during the model establishment and update phase to ease the boundary effect caused by cyclic shift. Lastly, the target state is judged by calculating the peak-to-average ratio of the target maximum response value, and the Kalman algorithm is utilized as the relocation strategy after the target is blocked. Experimental results show that the average accuracy of the proposed algorithm reaches 87.3%, and the success rate reaches 76.5% on 16 test sequences, which are 27.7% and 23.7% higher than those of the baseline algorithm, respectively. Keywords: object tracking; correlation filter; multi-feature fusion; ULBP; Gaussian mask; peak-to-average ratio; Kal￾man prediction; anti-occlusion 目标跟踪[1] 近年来因其横跨视频监控、无人 驾驶、无人飞行器、医学图像分析、空中预警等诸 多领域而迅速成为计算机视觉研究的热点之一。 目标跟踪的主流方法目前正由生成类方法逐渐转 向判别式方法，其中基于相关滤波的目标跟踪算 收稿日期：2010−05−21. 网络出版日期：2021−04−12. 基金项目：江苏省高等学校自然科学研究面上项 目 (17KJB520039)；江苏省“333 高层次人才培养工程科 研项目”(BRA2018147)；江苏省高校“青蓝工程”项 目 (2020 年). 通信作者：陈志国. E-mail：427533@qq.com. 第 16 卷第 4 期 智 能 系 统 学 报 Vol.16 No.4 2021 年 7 月 CAAI Transactions on Intelligent Systems Jul. 2021