Research On Face Tracking And Identification System Based On Deep Learning

With the rapid development of deep learning,computer vision combined with powerful computing capabilities is applied in various fields.In intelligent surveillance systems,the combination of face identification and tracking technology with deep learning has always been a very important and cutting-edge computer vision research field.This paper designs a set of intelligent surveillance system with automatic identification,automatic tracking and automatic identification functions,combined with micro-computing equipment.The system is divided into three aspects:face detection,face feature extraction and face tracking.Research and improvement are carried out based on the scene of processing video surveillance sequence images.In terms of face detection,based on the MTCNN algorithm,considering that there is a depth relationship between the intersecting detections in the surveillance video frame,the nonmaximum suppression algorithm is combined with the detection area,and the face detection with relatively intact facial features is selected.The detection is used as the filtering result,and at the same time,for the problems of face motion blur,posture change,illumination change,etc.,through regression detection combined with the information of the previous frame of the sequence image,the detection rate is improved by the way of partially reducing the confidence.The experimental results on the self-built data set prove that the improved algorithm has increased the detection rate by about 5%-8%compared to the original.In terms of face feature extraction,this paper compares and analyzes the optimization objectives and gradients of other loss functions,proposes a loss function with angle as the optimization objective.Analyzing and comparing the loss function gradient graph,it is concluded that the loss function in this paper has a larger gradient in the similarity within the class.Combine the loss function and other loss functions with the Resnet50 backbone network,train and test on public data sets and select accuracy evaluation indicators.The loss function in this paper increases by 0.5%-2%compared with other loss functions on AgeDB after adding margins,which proves that the loss function in this paper has stronger convergence within the class.In terms of face tracking,multiple object tracking DeepSORT algorithm framework uses face detection and face feature extraction to achieve face tracking.This paper speeds up the process of judging that the tracking target leaves the monitoring screen by introducing the division of monitoring areas,using the location-related characteristics of regression detection in the face detection algorithm to improve the matching process and improve the matching efficiency.The intelligent surveillance system is divided into a face detection module and a trackingidentification module,which are deployed on the micro computing device Raspberry Pi and the personal computer,and transmit data in the form of client server.The face detection,face tracking,and face identification functions of the system were tested respectively,and they all achieved the expected goals.