Research On Crowd Abnormal Behaviors Recognition In Video Surveillance Scene

Intelligent video surveillance is a popular application domain in the fieldof computer vision in recent years. Based on the traditional video monitoring, it usesthe related technologies, such as computer vision, image processing and patternrecognition, to detect target and behavior in monitoring scene. It also analyzes anddescribes the interested target. In a sense, it changes the characteristic of traditionalvideo monitoring that simply relying on the artificial experience and realizesintelligent monitoring in the true sense. In intelligent video surveillance, the analysisof the abnormal human behavior which is produced by the real urgent need is takenmore and more attention.Based on the detailed review and analysis of the abnormal human behaviorrecognition and the relevant information, this paper implements in-depth research forsome relative specific problems, such as the crowd moving region detection, motionfeature extraction and the crowd behavior classification. This main research contentof this paper is as follows:1. We introduce some common algorithms for moving region detection,including background subtraction division and interframe difference method.Based on these two algorithms, we propose a improved target area crowd detectionmethod. In the extraction process of frame differential method, we add the edgedirection scanning to extract more reliable prospect; in the background differencemethod of Gaussian mixture modeling, we used HSV color space differencediscriminant to determine the prospect area. Finally, we achieve accurate detection ofthe crowd movement area through accumulating them by the presupposed weights.This ensures the reliability of the follow-up study.2. We research and design a kind of feature extraction and description strategy:applying the speed and direction information calculated from the pixels in movementarea to the space-time cube to ensure it is suitable for feature extraction in the sensewith strong optional. We combine the space-time cube characteristics with neuralnetwork models with the competition mechanism, and put forward a complete crowdabnormal event detection method. The experiments show that our method has achieved good results in detecting the common abnormal human behavior such as thefight, panic and trample.3. As we know, the traditional optical flow algorithm just takes a simplegrayscale consistency hypothesis when establishing the corresponding relationsbetween adjacent frames of pixels. While the light changes, even if no movement,the optical flow still exists. In another hand, in the area in which the famine ofgrayscale changes appears, target motion is often undetectable. In view of thatproblem, we proposes a feature extraction algorithm based on SIFT. In establishingthe corresponding relationship between adjacent frames of pixels, we use SIFTdescriptor instead of the simple grey value to obtain motion and directioninformation accurately and describe this information by the weighted histogram.Then we achieve the recognition of the abnormal human behavior by train thehistograms with Hidden Markov Model（HMM）. The experiment results show thatour method can effectively recognize the abnormal behavior in video scene.4. Due to characteristic of high behavioral features dimension，large dataquantity and unstable characteristics endemic structure in the crowd behaviordetection, we propose a new the abnormal human behavior method based on Locallylinear embedding（LLE）and sparse representation. We improve the instability of thelocal manifold structure by adding the LLE regularization. The local linearembedding regularization can effectively preserve the local manifold structure of testsample, and improve sample discriminant ability. Experimental results show that thelocal linear embedding algorithm of sparse representation can effectively improvesample discriminant ability and get good experiment effect in the crowd behaviordetection.