Key Technical Research Of People Observation Under Dynamic Background In Video Surveillance

With the rapid development of computer science, intensive study ofimage processing and more attention to city safety, intelligent surveillanceare widely used in our daily life to ensure the public security. Owing to theimportant role of people in society, observing and monitoring people’sbehavior in video surveillance is more significant.The surveillance video is always too long and too boring to locatepeople quickly. It is necessary to organize video information and supervisepeople by intelligent system, which can not only save a great deal ofmanpower, resources and time, but also achieve more accuracy and moreobjective result. Based on this demands, a system named “peopleobservation and video analysis” was proposed in this thesis to help to dovideo management. The system provides two main functions, one is to dopeople observation, getting the attributes of people and extracting videoabstraction. Another one is video searching based on people observation.In this thesis, we carry out several theoretical and experimental studiesbased on the “people observation and video analysis” system.Considering the drawback of conventional foreground detectingalgorithms, a stereo local binary pattern based on appearance and motion(SLBP-AM) descriptor is proposed for background modeling and objectsdetection. The algorithm regards the motion of pixels as dynamic texture inellipsoidal domain, and combines texture histograms in the XY, XT, YT planes in the ellipsoid as the new descriptor for background subtraction.Experiment results show that the new proposed method can not only berobust to slight disturbance, but also adapt quickly to the large-scale andsudden changes. Besides, by reducing the dimensionality of featurehistograms and the background models, the time performance ofSLBP-AM has been widely improved.For system design part, a new structure is proposed in this thesis. Thethesis divides the people observation function into three modules:foreground detecting, people detecting and feature extraction. Each moduleis organized independently with specified interfaces, offering the followingadvantages:1) More kinds of video abstracts can be extracted by systemmodularization.2) Algorithms adopted in each module can be easilyupdated.3) The system can be extended to other fields.