人体结构化特征与核相关滤波器算法融合的目标跟踪方法  

Target tracking method based on fusion of human structural features and kernelized correlation filter algorithm

作　　者：马敬奇 钟震宇 雷欢 吴亮生 MA Jingqi;ZHONG Zhenyu;LEI Huan;WU Liangsheng(Guangdong Key Laboratory of Modern Control Technology(Guangdong Institute of Intelligent Manufacturing),Guangzhou Guangdong 510070,China)

机构地区：[1]广东省现代控制技术重点实验室(广东省智能制造研究所),广州510070

出　　处：《计算机应用》2020年第S01期56-60,共5页journal of Computer Applications

基　　金：广东省重点领域研发计划项目(2018B010108006);广东省科学院建设国内一流研究机构行动专项(2019GDASYL‑0105066)。

摘　　要：针对视频图像行人跟踪过程中因遮挡、大尺度姿态变化而导致的目标丢失问题,提出了人体结构化特征(依据人体结构划分的局部区域特征)与核相关滤波器(KCF)算法融合的跟踪方法。首先,基于人体骨架关键点的图像位置,提取目标肩部、胸部、腿部等局部区域的图像特征,形成与人体姿态相关联的结构化特征模板;然后,建立平均峰相关能量(APCE)和KCF响应函数峰值结合的目标丢失判据,目标丢失后,利用行人骨架关键点判断待匹配人体姿态异常和局部区域遮挡情况,矫正异常姿态人体局部图像,提取待匹配行人有效的区域结构化特征;最后,通过人体结构化特征匹配算法,计算待匹配人体与跟踪目标的相似度,重新定位目标,重定位后恢复KCF对目标的跟踪。实验结果表明,在遮挡、大尺度姿态变化时算法能够快速重新定位目标,精度和成功率可达到79.3%和68.2%,与典型跟踪算法KCF、Struck、TLD、DLT相比性能更好,具有更强的鲁棒性。Aiming at the problem of target loss caused by occlusion and large-scale posture change in the process of pedestrian tracking in video,a tracking method based on the structured features of human(local features divided according to human structure)and Kernelized Correlation Filter(KCF)algorithm was proposed. Firstly,based on the positions of the skeleton key points,the features of the local areas such as the shoulder,chest and leg of the target were extracted to form the structural feature template associated with the human posture. Secondly,the target loss criterion combined with the Average Peak to Correlation Energy(APCE)and the peak of KCF response function was established. When the target lost,the key points of the pedestrian skeleton were used to judge the abnormal posture of the human body to be matched and the occlusion of the local areas,so as to correct the partial image of the human body with abnormal posture,and extract the effective structured features of the pedestrian to be matched. Thirdly,through the human body structured feature matching algorithm,the similarity between the human body to be matched and the tracked target was calculated,the target was relocated,and the tracking of the target by KCF was recovered. The proposed method can solve the problem that the target cannot be tracked continuously due to occlusion and scale change. The experimental results show that the algorithm can quickly relocate human target in occlusion and large-scale attitude,and the accuracy and recall are 79. 3% and 68. 2% respectively. The proposed method outperforms the traditional tracking methods,such as KCF,Structured output tracking with kernels(Struck),Tracking Learning Detection(TLD),Deep Learning Tracker(DLT),and is more robust.

关 键 词：骨架 关键点 核相关滤波器 行人目标跟踪 遮挡 姿态 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]