Density Estimation And Abnormal Behavior Detection Of Small Group

In the recent years,public security situation has aggravated because of the complexity of modern social environment,which also becomes one of research hotspots worldwide.For the safety of the public properties,it is of necessity to establish a smart monitoring system specializing in the detection and early warning of crowd movement in public places.In this paper,research on crowd density estimation and abnormal behavior detection has been made and the main contents are listed below.1.An improved method of foreground extraction based on Mixture Gauss Modeling is proposed.This is accomplished through binarization and gray level transformation of original image frame,median filter for smoothing image and final foreground extraction with Mixture Gauss Modeling.Experiments show that,compared with the non-improved method,this improved one can reduce the time consumption by two third,which is a successful reflection of its superiority.2.Presenting the crowd density estimation method based on the combination of pixel statistics and texture features.Summing up the number of pixel in the image of both foreground and edge part then comparing the sum with a threshold of 38000 set from previous experiments.If the sum is smaller,a fitted function is obtained using a linear regression method based on both foreground and edge pixels;then by obtaining the number of people and pixels in the image of low crowd density through artificial marking.If the sum is larger,estimating the crowd density level on the basis of texture analysis and SVM classification,which consists of first forming a 8 dimension eigenvector by extracting the 4 texture features on both 0 and 90 directional entropy using gray level co-occurrence matrix,then using this eigenvector as input into a SVM composed of 3 two class classifiers.3.A method of fighting detection based on optical flow is proposed.The L-K optical flow algorithm was used to obtain the main direction of the normalized direction histogram,and the entropy of the normalized histogram was calculated.If both met the preset requirements,the fight behavior can be assured.Finally,according to the taken video,the experiment proves the effectiveness of the proposed algorithm.When using the pixel density and texture features combined with the group density estimation method,the correct rate of output is about 80%.The detection of the simulated fighting video shows that the recognition rate of the fight is high,and the difference between the two peers and the two fights is different.The correct rate is above 75%.Correct group density estimation and abnormal behavior detection have far-reaching significance for safeguarding social stability and maintaining public safety.