Research On Face Recognition Algorithm In Intelligent Video Surveillance

With the improvement of people's safety awareness,video surveillance technology has been widely used in present society,and has developed to where there is a security precaution,there must be a video surveillance.The traditional video surveillance system only provides the post-event video data for query,but can't do pre-warning.And because the video data are large,the monitor can't monitor hundreds of thousands of video images,which lead to the lost of pre-warning function.Therefore,intelligent video surveillance technology has been vigorously developed and applied.Intelligent video surveillance systems use the powerful data processing ability of computers to analysize massive video data,filter out irrelevant information,provide critical information,and give alarms based on the preset rules.As most of the intelligent video surveillance objects are humans,the function of intelligent video surveillance systems is mainly to detect,track,process and identify the human abnormal activities and biometrics.In view of the fact that face recognition technology has many advantages such as non-contact,simple operation,intuitive results and high reliability,the application of face recognition technology on intelligent video monitoring has become a hotspot.Because of the affection of device and other disturbances,image resolution obtained from video surveillance is often very low.The low resolution face image is not conducive to the subsequent face recognition.In this paper,the super resolution reconstruction of video face images is firstly carried out,and the reconstructed high resolution face images are used for subsequent recognition.The content of this paper consists of the following aspects:(1)The methods of moving object detection,face detection and image preprocessing are studied.Firstly,background subtraction and frame difference method are used to detect moving targets in video sequences.Then face detection is performed using the trained face classifier in OpenCV source code.Subsequently,histogram equalization algorithm,mean filtering algorithm,median filtering algorithm,geometric normalization and so on are used for face image preprocessing.(2)Image observation models,image quality assessment methods,common image super resolution reconstruction algorithms and face super resolution reconstruction algorithm based on neighborhood embedding are studied.In order to overcome the shortcomings of the existing super-resolution reconstruction algorithms based on neighborhood embedding,a new face super resolution reconstruction method jointing local constraint neighbor embedding and adaptive neighborhood selection is proposed.And the experiments are carried out on CAS-PEAL-R1 face database.Compared with the traditional face super resolution reconstruction algorithm via neighborhood,the new proposed algorithm increased by 0.39 dB and 0.02 in PSNR and SSIM respectively.(3)Feature extraction methods used in face recognition are studied,especially the feature extraction algorithm based on patterns of oriented edge magnitudes.Considering that facial features extracted by the patterns of oriented edge magnitudes have high dimensionality and complex computing,a noval face recognition algorithm based on the patterns of oriented edge magnitudes and supervised locality preserving projections are proposed.This algorithm first extracts facial features by POEM operator,and then project the high-dimensional feature data to the sample space obtained by SLPP algorithm for dimension reduction.Finally,the proposed algorithm classifies test samples by the nearest neighbor method.Experimental results on CAS-PEAL-R1 face database(including the expression,background,accessory,age,distance test set)indicate that the average recognition rate of the new algorithm increases by 22% than the POEM+LPP algorithm,and increases by 2% than the POEM+PCA algorithm.(4)A video based face recognition system is set up which can achieve the functions of face detection,preprocessing,super resolution reconstruction and recognition.The system has been tested by entering recorded video sequences,and the results show that the system can achieve the recognition of faces in videos.The correct recognition rate of samples in the gallery set is 90%,and the rejection rate of samples outside the gallery set is 100%.