Face Recognition System Of ATM Video Based On SIFT Algorithm

In recent years, pattern recognition techniques are widely used in condition of fulldevelopment of computer technology. As the main technique of pattern recognition techniques,face recognition technology is already used in many aspects of life. Face recognitiontechnology usually refers to identity verification and some other information, face images aremostly got through video surveillance and identity card photos in daily life. Automatic tellermachines are used more and more wide, at the same time, the frequency of crimes on lawlessautomatic teller machines by person are increasingly high, such as illegal withdrawals andillegal steal bank card password. How to effective monitor the withdraw process and protectthe remittee, has become a hot issue in contemporary research.Since the structural characteristics from video of Automatic teller machines are obvious,it is made up of series orders of withdraw money. Video entry on structural time series isrelatively fixed, remittees normally Individual appear, the contents of each frame imagechange little, and are small motions, moreover,a remittee to ATM machines is often thebeginning of a withdrawal. In this paper, the face recognition process is divided into thefollowing processes:Firstly, compressing the video accessed in uniform format, decomposing the compressedvideo into a single image, removing the background image frame, storing each of theremaining frames chronologically in.Then preprocessing each frame image in group andstoring as pending group;Afterwards，face detecting each frame image in processed group b[n] in accordance withthe use of pre-trained color threshold based on skin color algorithm, then expanding thedetected image in order to broaden the color area, storing as pending group,treating the imagein with binary processing and level accumulation projection, approximately plottinghistogram curve, all the video image are divided into positive face image class, left profileclass and right profile class; After larges number of video training, ATM video is generally35seconds, in accordance with the minimum avi compression format of15frames/seccompressed, there are almost500frames, selecting10frames of positive face image canrepresent the video. Calculating the ratio of detected skin color area in positive face images and the whole images, reordering the sequence of images in each class according to the ratio,picking up first five frames as key frames, then take five frame images as one group(the lastgroup less than five frames is considered to be noise), separately counting the squaredeviation of image sequence number and ratio, adding up two square deviations, calculateminimum group as another five key frames. Picking the most obvious peak value ofprojection histogram of right and left profiles as secondary key frame to be identifiedmatching.Finally, matching the10positive face key frames selected upon local feature matchingalgorithm and Face Database and storing the result, then using RANSAC algorithm removingthe error matched images.