摘要
行人重识别旨在不同时间、不同摄像头拍摄范围中检索特定目标行人,在实际应用场景中,可能会存在行人被严重遮挡的图像,不仅不利于行人检测,还会消耗大量的时间.行人姿态检测可以通过定位行人关键点位置判断行人是否存在遮挡,因此,本研究提出在重识别检测之前,对行人姿态进行分析,提出一种基于AlphaPose的重识别行人姿态评价方法.首先,利用AlphaPose进行姿态检测,得到行人各个关键点的置信度;然后,利用各个关键点的置信度得到各个行人的姿态评分;最后,根据姿态评分结果筛选出多个测试集进行验证分析.利用torchreid框架在数据集DukeMTMC-reID及Market1501进行实验,实验结果表明,与初始测试集相比,筛选后的测试集检测效率明显提高,且mAP和rank-n值也有所提高.
Person re-identification aims to retrieve specific target person in different time and camera shooting range.In practical application scenarios,there may be redundant images lacking person information or images where persons are severely occluded,which is not only detrimental to person detection,but also consumes a lot of time.Person posture detection can determine whether there is occlusion by locating the key position of persons.Therefore,we proposed to analyze the person pose before the rerecognition detection,then proposed a person re-identification pose evaluation method based on AlphaPose.Firstly,AlphaPose was used for pose detection to obtain the confidence degree of each key point of person.Then,the attitude score of each person was obtained by the confidence of each key point.Finally,several test sets were selected according to the attitude score results for verification and analysis.In this paper,torchreid framework is used to conduct experiments in DukeMTMC-reID and Market1501 data sets.The experimental results showed that compared with the initial test set,the detection efficiency of the filtered test set was significantly improved,and the mAP and rank-n values were also improved.
作者
刘立名
马传香
LIU Liming;MA Chuanxiang(School of Computer Science and Information Engineering,Hubei University,Wuhan 430062,China;The Key Research Institute of Humanities and Social Sciences in Hubei Province(Research Center of Information Management for Performance Evaluation),Wuhan 430062,China)
出处
《湖北大学学报（自然科学版）》
CAS
2023年第5期702-711,共10页
Journal of Hubei University：Natural Science
基金
湖北省技术创新专项(重大项目)(2019ACA144)资助。
作者简介
刘立名(1999-),女,硕士生;通信作者:马传香,博士,教授,研究方向为数据挖掘、信息安全、计算机视觉,E-mail:mcx 838@hubu.edu.cn.
