第14卷第1期 智能系统学报 Vol.14 No.I 2019年1月 CAAI Transactions on Intelligent Systems Jan.2019 D0:10.11992/tis.201806038 网络出版地址：http:/kns.cnki.net/kcms/detail/23.1538.tp.20180928.1338.004.html 一种自适应模板更新的判别式KCF跟踪方法 宁欣2，李卫军2，田伟娟2，徐驰2，徐健 (1.中国科学院半导体研究所高速电路与神经网络实验室，北京100083,2.威富集团形象认知计算联合实验 室，北京100083,3.中国科学院大学微电子学院，北京100029) 摘要：为了解决单目标跟踪算法中存在的目标旋转、遮挡和快速运动等挑战，提出了一种基于自适应更新策 略的判别式核相关滤波器(kernelized correlation filter,KCF)目标跟踪新框架。构建了外观判别式模型，实现跟 踪质量有效性的评估。构造了新的自适应模板更新策略，能够有效区分目标跟踪异常时当前目标是否发生了 旋转。提出了一种结合目标检测的跟踪新构架，能够进一步有效判别快速运动和遮挡状态。同时，针对上述 3种挑战，分别采用模板更新、目标运动位移最小化以及目标检测算法实现目标跟踪框的恢复，保证了跟踪的 有效性和长期性。实验分别采用2种传统手动特征HOG和CN(color names)验证提出的框架鲁棒性，结果证明 了提出的目标跟踪新方法在速度和精度方面的优越性能。 关键词：目标跟踪；目标检测；高速核相关滤波算法：模板更新；卷积神经网络 中图分类号：TP391文献标志码：A文章编号：1673-4785(2019)01-0121-06 中文引用格式：宁欣，李卫军，田伟娟，等.一种自适应模板更新的判别式KC℉跟踪方法J川.智能系统学报，2019,14(1)： 121-126. 英文引用格式：NING Xin,LI Weijun,TIAN Weijuan,etal.Adaptive template update of discriminant KCF for visual tracking J CAAI transactions on intelligent systems,2019,14(1):121-126. Adaptive template update of discriminant KCF for visual tracking NING Xin,LI Weijun2,TIAN Weijuan2,XU Chi,XU Jian' (1.Laboratory of Artificial Neural Networks and High-speed Circuits,Institute of Semiconductors,Chinese Academy of Sciences, Beijing 100083,China;2.Image Cognitive Computing Joint Lab,Wave Group,Beijing 100083,China;3.School of Microelectron- ics,University of Chinese Academy of Sciences,Beijing 100029,China) Abstract:To solve the challenges of in-plane/out-of-plane rotation(IPR/OPR),fast motion(FM),and occlusion(OCC), a new robust visual tracking framework of discriminant kernelized correlation filter(KCF)based on adaptive template update strategy is presented in this paper.Specifically,the proposed discriminant models were first used to determine the tracking validity and then a new adaptive template update strategy was introduced to effectively distinguish whether or not the object has rotated when the object tracking was abnormal.Furthermore,a new visual tracking framework com- bining object test is presented,which could further effectively distinguish FM and OCC.Meanwhile,to overcome the above-mentioned challenges,three measures were taken to recover the object tracking frame:template updating,object movement displacement minimization,and use of an object detection algorithm ensuring validity and long-term visual tracking.We implemented two versions of the proposed tracker with representations from two conventional hand-actu- ated features,histogram of oriented gradient(HOG),and color names(CN)to validate the strong compatibility of the al- gorithm.Experimental results demonstrated the state-of-the-art performance in tracking accuracy and speed for pro- cessing the cases of IPR/OPR,FM,and OCC. Keywords:visual tracking;object detection;high-speed kernelized correlation filters;template update;convolution neural network 收稿日期：2018-06-22.网络出版日期：2018-09-30 近年来，计算机视觉作为一门新兴学科发展 基金项目：国家自然科学基金项目(61572458) 通信作者：李卫军.E-mail:wjli@semi.ac.cn 十分迅速，目标跟踪作为视频监督、分析和理解DOI: 10.11992/tis.201806038 网络出版地址: http://kns.cnki.net/kcms/detail/23.1538.tp.20180928.1338.004.html 一种自适应模板更新的判别式 KCF 跟踪方法 宁欣1,2，李卫军1,2,3，田伟娟2 ，徐驰2 ，徐健1 （1. 中国科学院半导体研究所 高速电路与神经网络实验室，北京 100083; 2. 威富集团 形象认知计算联合实验 室，北京 100083; 3. 中国科学院大学 微电子学院，北京 100029） 摘 要：为了解决单目标跟踪算法中存在的目标旋转、遮挡和快速运动等挑战，提出了一种基于自适应更新策 略的判别式核相关滤波器 (kernelized correlation filter，KCF) 目标跟踪新框架。构建了外观判别式模型，实现跟 踪质量有效性的评估。构造了新的自适应模板更新策略，能够有效区分目标跟踪异常时当前目标是否发生了 旋转。提出了一种结合目标检测的跟踪新构架，能够进一步有效判别快速运动和遮挡状态。同时，针对上述 3 种挑战，分别采用模板更新、目标运动位移最小化以及目标检测算法实现目标跟踪框的恢复，保证了跟踪的 有效性和长期性。实验分别采用 2 种传统手动特征 HOG 和 CN(color names) 验证提出的框架鲁棒性，结果证明 了提出的目标跟踪新方法在速度和精度方面的优越性能。 关键词：目标跟踪；目标检测；高速核相关滤波算法；模板更新；卷积神经网络 中图分类号：TP391 文献标志码：A 文章编号：1673−4785(2019)01−0121−06 中文引用格式：宁欣, 李卫军, 田伟娟, 等. 一种自适应模板更新的判别式 KCF 跟踪方法[J]. 智能系统学报, 2019, 14(1): 121–126. 英文引用格式：NING Xin, LI Weijun, TIAN Weijuan, et al. Adaptive template update of discriminant KCF for visual tracking[J]. CAAI transactions on intelligent systems, 2019, 14(1): 121–126. Adaptive template update of discriminant KCF for visual tracking NING Xin1,2 ，LI Weijun1,2,3 ，TIAN Weijuan2 ，XU Chi2 ，XU Jian1 (1. Laboratory of Artificial Neural Networks and High-speed Circuits, Institute of Semiconductors, Chinese Academy of Sciences, Beijing 100083, China; 2. Image Cognitive Computing Joint Lab, Wave Group, Beijing 100083, China; 3. School of Microelectron￾ics, University of Chinese Academy of Sciences, Beijing 100029, China) Abstract: To solve the challenges of in-plane/out-of-plane rotation (IPR/OPR), fast motion (FM), and occlusion (OCC), a new robust visual tracking framework of discriminant kernelized correlation filter (KCF) based on adaptive template update strategy is presented in this paper. Specifically, the proposed discriminant models were first used to determine the tracking validity and then a new adaptive template update strategy was introduced to effectively distinguish whether or not the object has rotated when the object tracking was abnormal. Furthermore, a new visual tracking framework com￾bining object test is presented, which could further effectively distinguish FM and OCC. Meanwhile, to overcome the above-mentioned challenges, three measures were taken to recover the object tracking frame: template updating, object movement displacement minimization, and use of an object detection algorithm ensuring validity and long-term visual tracking. We implemented two versions of the proposed tracker with representations from two conventional hand-actu￾ated features, histogram of oriented gradient (HOG), and color names (CN) to validate the strong compatibility of the al￾gorithm. Experimental results demonstrated the state-of-the-art performance in tracking accuracy and speed for pro￾cessing the cases of IPR/OPR, FM, and OCC. Keywords: visual tracking; object detection; high-speed kernelized correlation filters; template update; convolution neural network 近年来，计算机视觉作为一门新兴学科发展 十分迅速，目标跟踪作为视频监督、分析和理解 收稿日期：2018−06−22. 网络出版日期：2018−09−30. 基金项目：国家自然科学基金项目 (61572458). 通信作者：李卫军. E-mail：wjli@semi.ac.cn. 第 14 卷第 1 期 智 能 系 统 学 报 Vol.14 No.1 2019 年 1 月 CAAI Transactions on Intelligent Systems Jan. 2019