| Collective behavior Collective behavior, which reveal similar or consistent behavior patterns of individual collection in crowded senses. Collective behavior exists widely in social life and has been the research highlights since recent years. There are many reason caused by these collective behaviors, and one of important issues addressed by researchers is how collective action can be more accurate and quick description and said. Recognition of collective behavior is of great significance and role of anomaly detection and video surveillance, which are closely related to people’s lives. So, the recognition of collective behavior has attracted more and more attention of researchers. In the collective behavior recognition in video scenes, similarity measure between the complex and interactive individuals is the key problem which affecting the clustering results. As a classical clustering algorithm for segmentation of graphics, the Normalized Cut will not be effective to analyze the collective behavior and recognition when solve complex scenarios or more irregular movement problems.In this paper, we study the targeted research and present an algorithm which can handle multiple complex collective behavior recognition based on hierarchical cluster analysis with the background of this significant difficulty and problem, and the work is as the following:(1) This paper has proposed a composite similarity measure method based on Delaunay Triangulations to research the multi-angle analysis and calculation between the individuals, such as distance, orientation and topological adjacency, so as to measure similarity from multiple perspectives;(2) This paper has proposed a hierarchical clustering analysis method to identify the patterns of collective behavior. Firstly, we blend in the orientation information and adjacency topology relationship between the individuals based on the existing methods. Secondly, we cluster the movement in the video scene with improved Normalized Cut, Finally, merge the neighbor clustering subclasses according to adjacency relations between individuals so as to achieve the effect of the hierarchical clustering analysis;(3) This paper has proposed a merge adjacent sub clustering algorithm. According to classical clustering segmentation algorithm of Normalized Cut, the limitation of the number of clusters need to be set in advance, the algorithm has made use of adjacent entropy to merge adjacent cluster and eventually determine the number of collective behavior clusters automatically after judging the child after clustering adjacency. In dealing with a variety of complex scenes, the algorithm has better accuracy and stability with experimental validation in many situations of data sets.A sequence of experiments and evaluations based on the actual video data set verify the accuracy and efficiency of the method. In addition, we discuss parameters in the experiments, and compare the results of existing research methods. |
