Research On Video Outlier Mining Based On Markov Random Field

Video surveillance is a key means to reinforce security in some important public places suchas school, railway stations, banks, airports etc, and traditional surveillance systems depend entirelyon the humans for analysis and decision making. But limited by human physiological character,people can only monitor a limited number of surveillance screens for a limited duration of time.Therefore, there is an increasing demand for automatic methods for analyzing the vast quantitiesof data generated continuously by Video surveillance systems. Video exception mining is aninterdisciplinary study for video normal behavior recognition and anomaly detection, and it denotesthe synergy of data mining, image processing, computer vision, patter detection, and artificialintelligence, recent years have been widely used in Video surveillance systems. Broadly speaking,deals with the extraction of implicit knowledge or other patterns not explicitly store in thelow-level, pixel representation video frames is the fundamental task of video exception mining.Starting from the traditional video image processing and exception mining techniques, focuson the detection of abnormal behaviors in the monitored scene video, this paper studies into thekey algorithms of intelligent surveillance systems. The major work and innovation of this paperare as follows.(1) Proposed an object detection scheme using Markov Random Field and a method instationary camera situation. It is showed the detection results of this algorithm are robust and notsensitive to noise in picture frame.(2) Introduced a new paradigm for abnormal behavior based on Markov chain model, thismethod models the local density of object feature vector, hence reduce the dimensionality of thefeature vector and leading to simple and elegant criterion for behavior classification. The proposedalgorithm is tested on the detection of pedestrian under the influence of alcohol, and the resultsshow that it has an overall accuracy of80%on the dataset.(3) Concerning the violence abnormal behavior between people, a method which couldindentify abnormal behavior more accurately from video was proposed. This pattern was based onthe simulation of All matters have both potential and kinetic energies, adopted the feature ofneighborhood related and the motion feature of adjacent frame in Markov random field,reconstructed the energy function, then added all the pixel energy to calculate the whole energyvalue, which could be analyzed by using the energy curve. Finally the effectiveness of thisapproach has been proved by comparing with other similar methods.