Small Scale Crowd Behavior Classification By Euclidean Distancevariation-Weighted Network

Crowd behavior analysis is a key research topic in the field of computer vision. The tradition method of crowd behavior analysis can be divided into two categories, i.e. micro and macro analysis. Microscopic analysis uses the trajectories and gestures to identify crowd behavior. This kind of method is very effective for these scenes which number of people is few and body movements can be obtained. However, the macroscopic properties of a crowd will be lost. Macroscopic analysis is to regard a crowd as a whole, it analyzes the behavior from external performance of the overall crowd. This kind of method is suitable for processing large scale crowd with a common motion pattern. However, it can not work well if the size of a crowd is small or the motion pattern is loose. Since small scale crowd often appear with small groups which crowd behavior has correlation. To reveal more properties of a small scale crowd the micro-based method and macro-based method should be integrated to analysis the behavior of a crowd.Network is a suitable tool to reveal the macroscopic properties of a system by responding the interaction between microcosmic individuals. Network can be expressed as a graph which can be used to describe the individuals(nodes) and the interaction between individuals(edges). Through the edges and weight of each edge, network can reveal the microcosmic influence degree between each node. The statistical properties of a network can be used to reflect the macroscopic property of the network. In this paper, we use network to represent and analyze small-scale crowd behavior. A method which use Euclidean distance variation-weighted network to recognize the crowd behavior is proposed in this paper. Firstly, the trajectories and location information of the individuals in crowd are captured by tracking every pedestrian. Furthermore, by calculating the Euclidean distance variation between two individuals the interaction between individuals is gained. Then, the Euclidean distance variation-weighted networks of five typical crowd behavior including gather, meet, together, separation and dispersion are constructed. The nodes represent individuals in the crowd and the weight of each edge represents the extent of interaction between individuals. Finally, the characteristic parameters of crowd networks including the path length and network weights are extracted and the crowd behavior is classified by using k-nearest neighbor method.Experimental result shows the proposed method can effectively express and recognize small-scale crowd behavior. The lowest classification accuracy of the proposed method can reach 86.8% in the five kinds of crowd behavior.