一种面向复杂场景的人脸识别与目标跟踪算法设计  

Design of face recognition and object tracking algorithm for complex scenes

作　　者：李昆仑[1] 熊婷[1] LI Kunlun;XIONG Ting(College of Science and Technology,Nanchang University,Jiujiang 332020,China)

机构地区：[1]南昌大学科学技术学院,江西九江332020

出　　处：《现代电子技术》2024年第24期167-171,共5页Modern Electronics Technique

基　　金：江西省高等学校教学改革研究项目(JXJG-22-30-7);江西省科技厅项目(GJJ2203914)。

摘　　要：为解决在复杂场景下传统算法对人脸的识别和跟踪精度低的问题,文中提出一种面向复杂场景的人脸识别与目标跟踪算法FaceNet。该算法由人脸识别与人脸跟踪两部分组成。在人脸识别方面,基于YOLOv8提出一种改进的人脸识别算法FaceD-YOLOv8,提高了识别准确率。再以DeepSort为基础,提出一种改进人脸跟踪算法FaceT-DeepSort,实现对人脸目标连续准确的跟踪。在公开数据集上进行的对比实验结果表明,与主流的传统方法相比,所提FaceD-YOLOv8算法人脸识别的mAP值提高3.5%,FaceT-DeepSort算法的人脸目标跟踪精度(TP)值提高9.1%。证明所提方法具有良好的综合性能,能够满足工程应用的需要。In order to solve the problem of low recognition and tracking accuracy of traditional algorithms for faces in complex scenes,a face recognition and target tracking algorithm FaceNet for complex scenes is proposed,which is composed of the face recognition and face tracking.In terms of facial recognition,an improved face recognition algorithm FaceD-YOLOv8 is proposed based on YOLOv8 to improve the recognition accuracy.Based on DeepSort,an improved face tracking algorithm FaceT-DeepSort is proposed,which can realize continuous and accurate tracking of face targets.The comparative experimental results conducted on public datasets show that,in comparison with mainstream traditional methods,the proposed FaceD-YOLOv8 algorithm can increase the face recognition accuracy mAP by 3.5%,and the FaceT-DeepSort algorithm can improve the face target tracking precision(TP)by 9.1%,proving that the proposed method has good comprehensive performance and can meet the needs of engineering applications.

关 键 词：人脸识别 人脸跟踪 复杂场景 YOLOv8 DeepSort GIoU 

分 类 号：TN911.73-34[电子电信—通信与信息系统]