Surveillance Attendance And Stranger Verification Method And System Implementation Based On Face Recognition

Surveillance video cameras in today's society are ubiquitous,however most surveillance video systems only have simple video recording capabilities,which could be the evidence of a criminal case happened instead of advance alarm.Because the monitored pedestrians are often in an unconstrained state,and face recognition and stranger verification under this influence of factors such as face deflection and illumination changes still have great challenges.This paper focuses on the face recognition and stranger verification of face multi-angle deflection problem in real scenes.According to the logic of motion detection,face detection,face evaluation,face synthesis,face recognition and stranger verification,a series of researches are carried out,and a real-time low-resource video surveillance system of attendance and stranger verification based on face recognition is developed in general computer(CPU I5-7400-3.0Ghz,No GPU).The main work of this paper includes:1.A real-time face detection algorithm based on SSD-ResNet10 is improved and used,which can achieve 28 FPS on 1080 P video coordinating with motion detection algorithm in single thread.The average accuracy and average recall rate of this method on FDDB face detection database are respectively 99.8% and 87.5%.2.A frontal face evaluation method based on face deflection angle prediction is proposed,whose R2(coefficient of determination)achieves 0.9723.The face detected in the video is scored under frontal face evaluation method,and the face image with high confidence of frontal face evaluation is selected to face recognition.Hence,the interference of deflection and occlusion in face recognition can be excluded.3.Frontal face synthesis method from multiple pose-variant faces with CGAN is proposed,which can synthesize frontal faces when the frontal face evaluation method doesn't find frontal faces for recognition.The face recognition accuracy rate based on the frontal face synthesis algorithm increases from the original 80.0% to 93.2%,closing to the recognition accuracy rate 98.8% of frontal face with manual selection.4.A face expression feature extraction model based on Triplet-Loss and MobileNetV2 is proposed.The face image feature is extracted by this model,and we use template matching strategy to recognize single frame face image in registrant set.In video face recognition,the exponential weighted average method and the frontal face evaluation are combined to be the single-frame image recognition score,which can complete the face recognition in the video motion event.The recognition accuracy rate of registrants reaches 96.13%,the false acceptance rate of registrants is 1.35% and the false alarm rate is 2.52% in the real scenes surveillance video dataset we built.5.A stranger discriminator based on GMM-UBM is proposed.The stranger model(UBM)is trained through the external massive data set,and then the registrant feature is used to learn the second-level judgment model(GMM)of the registrant under the stranger model.Stranger discrimination is achieved through cascade judgment and empirical threshold strategy.According to this method,the stranger detection rate reaches 88.75%.6.The Faiss vector search matching framework is introduced to large-scale fast registrant feature search matching recognition calculation,which can realize face feature vector search matching of thousands of people within 30 milliseconds to meet the actual environmental requirements.7.A low-resource real-time attendance and stranger verification system based on face recognition through surveillance video is built to achieve multi-target motion tracking,face detection,frontal face evaluation,face synthesis,registrant recognition and stranger verification.The cross-platform surveillance video system management website is developed,which can satisfy practical functions such as video samples viewing and surveying,registrant management,system setting management,and surveillance video live broadcast.