基于视觉跟踪的实时视频人脸识别

(厦门大学软件学院,福建 厦门 361005)

视觉跟踪; 人脸识别; 监控视频

Real-time Face Recognition in Videos Based on Visual Tracking
REN Zihan,YANG Shuangyuan*

(Software School,Xiamen University,Xiamen 361005,China)

visual tracking; face recognition; surveillance video

DOI: 10.6043/j.issn.0438-0479.201709014

备注

目前基于深度学习的人脸识别方法准确率高,但是模型复杂,识别速度慢.为了实现监控视频中人脸的实时识别,提出了一种基于视觉跟踪的实时视频人脸识别(RFRV-VT)方法.首先将监控视频的帧序列分组,每一组中分为人脸识别帧和人脸跟踪帧; 然后在人脸识别帧中使用基于深度学习的人脸检测和人脸特征提取方法,在人脸跟踪帧中使用基于核相关滤波(KCF)的视觉跟踪方法以加快识别速度.将该方法应用于数据集YouTube Faces(YTF)上进行测试,实验结果显示该算法在监控视频中具有实时性和较高的识别准确性(99.60%).

At present,face recognition methods based on deep learning yield high accuracies,but their complex models recognize face slowly.To achieve the real-time face recognition in surveillance videos,we propose a real-time face recognition method in videos based on visual tracking(RFRV-VT).Firstly,this algorithm divides the surveillance video frame sequence into several groups,and each group contains face recognition frames and face tracking frames.Then,face detection method and face feature extraction method based on deep learning are used in the face recognition frame,and visual tracking method based on kernelized correlation filters(KCF)is used to speed up the recognition in the face tracking frame.This method is applied to the YouTube Faces(YTF)dataset for testing.Experimental results show that the proposed algorithm exhibits real-time performances and high recognition accuracies in videos(99.60%).