基于YOLOv5s与扩展卡尔曼滤波的人体跟踪器设计  

A Human Tracker Based on YOLOv5s and the Extended Kalman Filter

作　　者：徐振宇[1] 蔡敏雅 陈子钰 秦晋 XU Zhenyu;CAI Minya;CHEN Ziyu;QIN Jin(College of Information Science&Technology,Zhejiang Shuren University,Hangzhou Zhejiang 310015,China;Institute of New Electron Device and Application,Hangzhou Dianzi University,Hangzhou Zhejiang 310018,China;Hangzhou Pioneer Technolegy,Hangzhou Zhejiang 310018,China)

机构地区：[1]浙江树人学院信息科技学院,浙江杭州310015 [2]杭州电子科技大学新型电子器件与应用研究所,浙江杭州310018 [3]杭州派尼澳电子科技有限公司,浙江杭州310018

出　　处：《传感技术学报》2024年第5期841-848,共8页Chinese Journal of Sensors and Actuators

基　　金：浙江树人学院省属高校基本科研业务费专项资金项目(2022XZ015)。

摘　　要：YOLOv5s在COCO数据集上的预训练权重可以检测出人体目标,然而对于特殊的人体姿态无法识别。提出了一种基于YOLOv5s与扩展卡尔曼滤波的跟踪器,当YOLOv5s能检测出人体目标时,对扩展卡尔曼滤波进行初始化,当YOLOv5s无法检测时,由扩展卡尔曼滤波进行跟踪,在先验结果创建候选区域,使用差异值哈希匹配出最优候选区域作为观测值,从而更新目标位置,实现人体检测的连续性。实验结果表明,在跟踪精度上所提出的跟踪器与“真实边界框”的重叠率为50.65%,中心位置误差为51.78像素,在实时性上优于KCF和TLD跟踪器,帧率比KCF跟踪器快了26 frame/s。YOLOv5s’s pre-training weight on the COCO dataset can detect human targets,while failing to recognize specific human poses.A human tracker based on YOLOv5s and the extended Kalman filter is proposed.When YOLOv5s can detect the human object,the extended Kalman filter is initialized.When YOLOv5s cannot detect,the extended Kalman filter is used to track,and the candidate region is created in the prior results.The optimal candidate region is matched by using the different hash algorithm as the observation value,to update the target position and realize the continuity of human detection.The experimental results show that in terms of tracking accuracy,the overlap rate between the proposed tracker and the“real boundary box”is 50.65%,and the center position error is 51.78 pixels.The real-time performance is better than that of the KCF and TLD trackers,and the frame rate is 26 frames per second faster than that of the KCF tracker.

关 键 词：YOLOv5s 扩展卡尔曼滤波 目标跟踪 差异值哈希 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]