人脸和步态特征注意力融合的身份识别方法  

Human Face and Gait Feature Attention Fusion Based Identity Recognition Method

作　　者：沈澍 张文昊 王汝传 沙超 丁浩[3] SHEN Shu;ZHANG Wenhao;WANG Ruchuan;SHA Chao;DING Hao(School of Computer Science,Nanjing University of Posts and Telecommunications,Nanjing 210023,China;Jiangsu High Technology Research Key Laboratory for Wireless Sensor Networks,Nanjing 210003,China;Jiangsu Police Institute,Nanjing 210031,China)

机构地区：[1]南京邮电大学计算机学院、软件学院、网络空间安全学院,南京210023 [2]江苏省无线传感网高技术重点实验室,南京210023 [3]江苏警官学院,南京210031

出　　处：《小型微型计算机系统》2024年第7期1695-1701,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金面上项目(62272244)资助;江苏省高等学校自然科学研究重大项目(22KJA520010)资助;江苏省研究生科研与实践创新计划项目(KYCX22_0973,SJCX23_0273,SJCX22_0266)资助;浙江大学CAD&CG国家重点实验室开放课题项目(A2118)资助;痕迹检验鉴定技术公安部重点实验室(中国刑事警察学院)开放课题项目(HJKF201915)资助.

摘　　要：真实的身份认证场景往往存在面部遮挡和远距离等难点,给人脸识别等传统识别方法带来挑战.步态识别等新型识别方法助力身份认证.步态识别适用于面部遮挡场景,且远距离时优于人脸识别.为了发挥人脸识别和步态识别在远距离遮挡下的互补作用,本文提出了一种基于人脸和步态多模态融合的身份识别方法.该方法包括面向低分辨率和有遮挡场景的人脸识别模块、基于轻量化模型GaitLight的多视角步态识别模块、融合人脸和步态特征的注意力融合模块.人脸和步态融合数据集上的实验结果表明,提出的多模态方法在面部无遮挡和面部遮挡条件下,识别率均高于单模态方法和现有的多模态方法.两种条件下识别率分别达到98.5%和98.4%,高于人脸识别算法1.2%和7.1%.多模态识别方法既能满足日常识别需求,也适用于远距离遮挡下的身份识别,识别性能优于目前应用的人脸识别方案.Real authentication scenarios often have difficulties such as facial occlusion and long distance,which brings challenges to traditional recognition methods such as face recognition.New recognition methods such as gait recognition help identity authentication.Gait recognition is suitable for face occlusion scenarios and is superior to face recognition at long distances.To play the complementary role of face and gait recognition under long-distance occlusion,this paper proposes an identity recognition method based on the multi-modal fusion of face and gait.The method includes a face recognition module for low-resolution and occluded scenes,a multi-view gait recognition module based on the lightweight model GaitLight,and an attention fusion module that integrates face and gait features.Experimental results on the face and gait fusion dataset show that the recognition rate of the proposed multi-modal method is higher than that of single-modal methods and existing multi-modal methods under conditions of no facial occlusion and mask occlusion.The recognition rates under two conditions are 98.5%and 98.4%,respectively,1.2%and 7.1%higher than face recognition.The multi-modal recognition method not only meets the needs of daily recognition,but also is suitable for identification under long-distance occlusion,and the recognition performance is superior to the current face recognition scheme.

关 键 词：人脸识别 步态识别 注意力机制 多模态融合 身份识别 视频视觉转换器 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]