Research On Technology Of Samll Group Analysis In Video Surveillance

Recently,with rising living conditions,more and more people pay more attention on security issue gradually.Intelligent video surveillance plays a very important role in the security field.Especially,the crowd analysis has also been received considerable attention in video surveillance.Currently,researchers in the public surveillance field have found that most crowds consist of small groups rather than isolated individuals.Modern crowd theories also agree that collective behavior is the result of the underlying interaction among small groups of individual.So the small group analysis of crowd gradually become a subject of general interest.We collected and consulted the related theory of crowd analysis and relevant information,and then implemented further research for the key technology that involved in the process of small group analysis for crowd,the main research content of this paper is as follows:(1)In the aspect of the detection of small group in crowd,We proposed a novel bottom-up hierarchical clustering approach for detection of small groups of pedestrians in crowds.Instead of computing pairwise similarity between pedestrian trajectories,followed by clustering of similar pedestrian trajectories into groups,we cluster pedestrians into a group by considering only start(source)and stop(sink)location of their trajectories.And then in terms of the evaluation standard,we evaluate the result of clustering by borrowing from the evaluation algorithm of natural language text coreference resolution.We presents the proposed approach and its evaluation using different public data sets.Our algorithm shows a better performance relatively.(2)In the aspect of the leader identification of small group in crowd,we proposed a structural learning framework for that purpose,and then extracted features model based on the time-lag analysis and pedestrian position analysis for pairwise pedestrian trajectories in small group.And then we construct the hypotactic graph of small group take the features as weight value,we used the Page Rank algorithm to calculate the Page Rank score for each node in the graph.Finally,it combined the structural SVM to train by a large number of samples.Experimental results shows that our approach played a good performance relatively for the leader identification in small group.(3)In the aspect of the small group activity recognition in crowd,in order to resolve the complexity and ambiguity problems caused by a large number of pedestrian objects,we proposed a oriented graph of small group to detect the interacted group in crowd scene so as to be robust against noisy information.Firstly,we define a interpersonal interaction area based on proxemics for every human objects;and then construct interaction area of small group by the ratio of overlapping area to total area covered by interacting human objects;Finally,we construct a oriented graph of small graph by means of Granger Causality test in interaction area of small group.Eventually,we extract three kinds of features,gather and repulsion features,Granger Causality features and additional features,and then we obtained the classifiers by trained with SVM model and show a better performance relatively on the public BEHAVE dataset.