The Research Of Abnormal Behavior Detection Based On ORB Interesting Points And Its Application Technology

Abnormal behavior detection is the first and very important step of behavior analysis and understanding based on video. With the increasing amount of object in the scenes, it is very easy to fail to track the object because of occlusion and overlapping. The traditional behavior detection and tracking object-based approaches can achieve good results but is limit in the dense complex scenes. In order to overcome these difficulties, the entire analysis approaches based on particle advection and interesting points tracking provide new research methods, which can analysis the features of images and videos in the macro and micro aspects.In this paper, we take the behavior of individual and crowd in the video sequences as research subjects, around this goal to explore and improve the methods which are based on interesting points. It takes the ORB interesting points which have stable characteristics of image as entry point of the and extracts the basic features of ORB interest points of image in spatial and time domain to describe the behavior features of the moving objects. The normal behavior benchmarking can be constructed with the analysis of optical flow and location of interesting points and the rectification of different scenes. The abnormal behavior can be detected when its scale of feature is higher then the normal behavior benchmarking line. The main conclusions are as follows:(1) In view of abnormal individual behavior detection, we treat the walking and the running as normal and abnormal behavior respectively in the video sequences and analysis the difference of interesting points features in spatial and time domain between them. A new abnormal individual behavior detection method that can describes the difference of individual behavior is proposed.The features of the method included the median of ratio between width and height in spatial domain and the mean optical flow in the continue video sequences with statistics.(2) Depending on the applications, the abnormal crowd behavior can be classified into two categories: the local and the global abnormal crowd behavior. In view of the global abnormal crowd behavior, we use the average kinetic energy of the ORB interesting points to measure the intensity of the movement in each frame and build the normal behavior benchmarking with two thresholds according to the location where the behavior appears, then we propose a method which uses the K-means approach to cluster the location of interesting points to select the normal behavior benchmarking adaptively. In view of the local abnormal crowd behavior, we statistics and analysis the optical flows of the interesting points in different location to build the normal behavior benchmarking line. We can detective and locate the abnormal crowd behavior when its optical flows is high then the line. Experimental results show that the method based on ORB interesting points has high performances and can well deal with the problem of occlusion and the variation of the scale about the objects and the accuracy of the detection are very close to the state-of-the-art algorithms.