Research On Personnel Recognition In Video Surveillance Based On Deep Learning

The use of biometric features,such as the use of faces,fingerprints,and irises,for personnel recognition is more secure than digital systems such as passwords.However,the recognition of personnel in a single mode is not effective in some cases,so the integration of multiple models of biometrics for human recognition has become a research hotspot.In this paper,based on the requirements of personnel recognition under video surveillance,the information of attitude characteristics and target tracking is integrated on the basis of face recognition,and two innovative schemes for personnel recognition are designed.In this paper,the CASIA Webface face data set published by the Institute of Automation of Chinese Academy of Sciences is used in the research of face recognition.First,the camera image is obtained by using OpenCV,and face detection and face alignment are implemented based on MTCNN(Multi-Task Cascaded Convolutional Neural Networks).Secondly,the face recognition model is built on the basis of FaceNet,and the cosine distance is selected on the method of calculating the feature distance.Finally,the two parts are combined to realize a real-time face recognition system,and the accuracy rate can reach 100% on the local face data set.Generally,in the video surveillance,the face information that can be effectively recognized cannot be captured all the time,so the person cannot be continuously identified,resulting in poor recognition effect.The existing target tracking system is capable of continuously tracking the same person but without the ability to recognize people.Therefore,this paper combines target tracking and face recognition.After capturing the face information that can be effectively recognized,the face recognition and target tracking feedback coordinate information on the screen is used to transmit the face recognition result to the target tracking.Personnel information,to achieve tracking and recognition,to achieve the effect of continuously identifying specific people in video surveillance.Faces captured in video surveillance have low resolution in long-distance situations,affecting face feature extraction,resulting in inaccurate face recognition results,or even unrecognizable.The attitude feature information is easy to collect,and although the definition is low,the gesture recognition still has a certain recognition accuracy.Therefore,the improvement of the GaitSet algorithm for gesture recognition is improved by 6%.Combining face recognition and gesture recognition,the face recognition and face feature are calculated respectively,and the highest confidence is taken as the recognition result.Multi-modal recognition based on face and attitude is realized.Based on the above research,this paper developed a system VSR(Video Surveillance Recognition)for monitoring human recognition in video.The system implements the above two schemes of tracking recognition and multi-modal recognition.In addition,due to the current sensitivity of the public to the privacy of the face,the system does not need to specifically collect a single face image,and can use the web camera to implicitly collect face information and posture information at a long distance,and store it as a local database to recognize people in video surveillance.