Research On Real-Time Video-Based Multiple Faces Detection,Tracking And Optimization Methods

With the increasing demand of information security,access control,border security,video surveillance,and other applications,video-based face recognition technology attracts more and more people's attention and is widely used in a variety of social lives and works.Video-based face recognition system sets face collection,detection,tracking and recognition in one and is able to automatically recognize the faces in the video.Since the operating target of video-based face recognition system is the video sequence,how to pick out the faces from a large number of detection and tracking faces for recognition will directly affect the performance of the video-based face recognition system.Therefore,this paper constructs a real-time multiple faces detection,tracking and optimization system based video.Before face recognition,the system detects and tracks the faces in the video firstly,and then picks out everybody's relatively optimal face for subsequent face recognition during tracking.Therefore,the system consists of three parts which are multi-face detection,tracking and optimization.During face detection,the method synthetically uses image grayscale and color information which is based on skin color segmentation and Adaboost face detection algorithm.Firstly,it preselects face candidate regions according to the image color information.Then,the Adaboost algorithm is used to detect human faces in face candidate regions.So that we can effectively solve the problem that Adaboost algorithm is prone to mistakenly detect faces in complex background.Meanwhile,aiming at the problem of real-time monitoring people in and out of the video and changing numbers without detecting faces frame by frame,this paper presents a method that uses background subtraction on image edge to determine whether someone comes in or out,and then starts face detection program correspondingly.During face tracking,Camshift algorithm is used to track faces within a limited search range.To solve the problem that Camshift algorithm is prone to lose the tracking objects in the practical application of complex situations,the CamShift based on the Kalman filter are put forward.It uses the tracking result of Camshift algorithm as the observation value of the Kalman filter to constantly update facial motion model.When it is determined that some complex cases occur,such as multiple faces occlusion,faces overlap and similar background color interference,Kalman filter is applied to correct the tracking results of CamShift algorithm.This method improves the accuracy and robustness of the tracking method.During face optimization,there are a variety of face poses in video and its changes will greatly influence the performance of face recognition.In addition,since people in the video are in motion,it is inevitable that there will be small and blurred faces which also badly influence the performance of face recognition.Therefore,an optimization module for screening face size,definition and posture is introduced.The module will screen numerous faces of one person in size,definition and posture layer by layer to provide a face with appropriate size,high definition and relatively best posture for the subsequent recognition.Finally a real-time system based above algorithms for multiple faces detection,tracking and optimization is developed by using OpenCV and MFC under Visual C++ 6.0 SDE.This system can accurately detect and track multiple faces in the video and save a relatively optimal face for everyone by comparing face size,definition and posture.