Research On Visual Privacy Protection Technology In Cooperative Surveillance System

With the continuous advancement of policies such as intelligent cities and safe cities,also the rapid development of the security monitoring system,the trend of large-scale surveillance networking has become increasingly apparent.The monitoring system is becoming more sophisticated.Video surveillance also improves regional defense capacity with the perfect combination of security and the Internet.However,because a large capacity of surveillance videos are stored,transmitted,and accessed through the Internet,people who are monitored have to worry about their privacy being exposed and illegally exploited,so the original conflict between need of security and privacy protection breaks out.The achievement of an acceptable balance between protection and security monitoring becomes a challenge.Currently,most existed visual privacy protection methods are based on the perspective of video users to protect visual privacy,with video encryption as the core,and with hierarchical authorization to provide access.The method proposed in this thesis focus on the visual privacy protection mechanism in region with fixed personnel such as communities.For fixed personnel with security attributes,their visual information will be absolutely protected from the source image,and non-fixed personnel will be monitored.The security attribute of fixed personnel directly determine that their visual information will be strictly protected,rather than the identity of the video users.Specifically,this method will selectively track and monitor pedestrians and use face recognition technology to identify the pedestrians whose visual information need to be protected.At the same time,privacy replacement method based on dynamic background models is used to protect visual privacy.The major work is as follows:First,aiming at the problem that the original Vibe algorithm is used to detect foreground with too much computation in high-definition video,we proposed an improved Vibe method to model for background quickly in high-definition surveillance video.The improved method abandons the idea of creating a sample set for each pixel in the original method and conducts a large scale reduction of background model.A moderately uniform sampling is taken for each frame to create a Vibe background model(1).The images to be detected are sampled in the same way,the foreground region is detected by the proposed Vibe method,and the false detection results are subtracted by taking contour and area of each target as the main reference information.Finally,the coordinate of moving targets in sampling image are mapped to the high-definition image according to the linear relationship between the source image and the sampled image.This method decreases the calculations cost of foreground detection multiplicatively.In addition,the result of moving object detection is further applied to the background model(2)updating in the background replacement method,so that the background model becomes flexible.Even if light,texture,and other information are slowly transitioned with time,the foreground and background are kept harmonize state at all times,and the visual integrity of surveillance videos maintains well after privacy protection processing is completed.Secondly,a collaborative surveillance visual privacy protection method is proposed to deal with the confliction of security monitoring and visual privacy protection in area with relative fixed people.Because the camera is fixed without shaking and rotation,the method establishes a dynamic background model(1)for foreground detection,also the model(1)adapts to changes in the light intensity during the day.We adopts the improved Vibe foreground detection method to extract the foreground region in high-definition surveillance video.Besides Pedestrian detection is used to separate moving targets with mild adhesions in foreground,which significantly reduces the cost of computation in pedestrian detection.For each candidate target,a multi-region local sensitive histogram is used to represent moving target,the data association algorithm and the Kalman filter are used to match the tracking targets and the candidate targets,and then the face recognition technology is used to identify whether the region of target includes privacy(the region corresponds to fixed person)or not.Finally,the result of foreground detection is used to update the dynamic background model(2)for privacy replacement,and we use the background model(2)to replace the visual privacy information in current frame.Our experiment shows that the technology of collaborative surveillance visual privacy protection method can accurately protect the visual privacy of the fixed staff in the normal motion scene and effectively monitor non-fixed personnel.