Research On The Detection Algorithm Of Abnormal Crowd State

In recent years, with the scale of the urban public spaces is more and more big, the crowded degree is higher and higher, the frequency of crowd abnormal events such as Robbery, Brawl, trample increased rapidly, brought serious damage to public’s life and property security. Therefore, this paper proposes a crowd abnormal state detection algorithm for high density crowd based on graph analysis, this algorithm has a good generalization ability.This paper mainly aims at crowd evacuation, aggregation, riot three abnormal states detection, and designed the corresponding detection methods. Research content involves the moving target detection, the construction of a crowd network diagram, diagram analysis of feature point, unsupervised clustering of feature point, crowd characteristics analysis, etc. The main research work is as follows:1) For the moving target detection and information extraction problem, due to the movements of individuals in the high density are similar to many flowing particles which are gradually moving. We use the method of KLT to extract and track feature point, which has many advantages such as high calculation speed, high accuracy, suitable for more complex situations and so on.2) Due to the crowd can simulate as many individual through interaction to form a network model and can use the graph of graph theory to describe, as well as individuals in the crowd maintain local interaction among neighbors with a fixed number of neighbors on topological distance. We combined feature points extracted by KLT and feature points adjacent set extracted by KNN to build a crowd network diagram. In addition, this paper defines a standard for behavior consistency between feature points, and puts forward a description of the degree of aggregation, which measures the extent of intensive distribution of the crowd.3) The traditional clustering algorithms often need to set the cluster number beforehand, or require a large number of training samples, aiming at this problem, we put forward a kind of unsupervised clustering algorithm for crowd feature points. Consistency between feature points is calculated on the basis of the crowd network diagram by analyzing the correlation between the clustering feature points.4) A class-based analysis method of the characteristics of the crowd is proposed in this paper, the characteristics include the degree of aggregation, velocity, variance of the velocity direction and the number of clusters. The crowd characteristics are compared with characteristics of the three kinds abnormal state, to determine whether an abnormal state of the crowd happened. Experimental results show that the results of the proposed algorithm are consistent with human perception, the accuracy of the algorithm is relatively high.