Research Of Human Abnormal Detection Based On Video Sequence

Intelligent video surveillance is a cutting edge in area of computer vision, and ithas an extremely important practical significance and broad application prospects.Intelligent video surveillance can do work all the time, with automated, high detectionrate, real-time alarm and many other advantages. Because of its advantages, more andmore researchers pay close attention to intelligent video surveillance, especially thehigh cost of human is increasing.People are concerned about the abnormal behavior in video surveillance in a realscene. When the abnormal behavior is identified in time, video surveillance informsthe occurring risk behaviors and types to the monitoring person in the monitoringareas, so that related people can prevent more dangerous activities. So how to detectand identify abnormal behaviors are the main contents of this paper.A novel fast3D non-local means denoising algorithm for video sequences ispresented in this paper. According to the characteristics of the video sequences, a3Ddenoising algorithm based on the non-local means algorithm is introduced. Twoimprovements are proposed for the video sequences denoising algorithm based on the3D non-local means algorithm. The optical flow feature vector is applied to revise theresult of the moving object parts in different frames. Meanwhile, a video accelerationscheme utilizing inter-frame correlation has been proposed, which makes the processof video sequences denoising faster.The detection of moving object is the basic of tracking, recognition and behavioranalysis. The accuracy of all rely the effect of the target extract. There are three mainmethods, inter-frame difference, background subtraction, and optical flow. Theinter-frame difference is so simple, the algorithm principle need the difference of thedifferent frames to detect the moving object, that the extraction result susceptible toenvironmental interference and the target to be extracted is incomplete. Optical flowmethod using the optical flow field changes in the environment to extract movingtarget, and can be used in moving camera, but the result can easily be affected byillumination change. The algorithm of optical flow is so complexity that most ofsystem cannot be processing real-time. Background subtraction method is themainstream method. It uses the difference of the current frame image with theestablishment of the background model extracted moving target. So the key is to createand update of the background model. An improved dual background modeling, which includes an adaptive running average background model and HSV background model,was proposed to indicate the variation of background pixels in order to increase therobustness against illumination changes and environmental disturbances. The modelcan reliably extract the motion area. Foreground was obtained from video sequencesby background subtraction.A human abnormal behavior detecting approach was proposed based on opticalflow features in the motion area in order to meet the needs of intelligent videosurveillance. The motion area was labeled as several regions of interest, and the opticalflow features in each labeled region were obtained using the Lucas-Kanade algorithm.Amplitude based weighted unit energy derived from the optical flow features wasdefined to measure the anomaly of human activity. Experiments were conducted onvarious videos indoor and outdoor, and the results were presented to verify theeffectiveness of the proposed scheme.This paper will introduce some classical algorithm of video denoising andmoving object extraction, such as3D non-local means denoising algorithm, RunningAverage algorithm, Gauss Mixture Model algorithm, and Coodbook Model algorithm.Each chapter will give a basic introduction of used method. The method of abnormalbehavior detection based on weighted energy of optical flow will be described in detailin chapter4.