Behavior Recognition For Small Scale Crowd Based On Causal Complex Network Analysis

The crowd behavior analysis is one of the most important subjects of the vision-basedhuman activity recognition, and it has received more attention from international anddomestic academics in recent years. The application areas of vision-based human activityrecognition mainly are intelligent monitoring and intelligent traffic, thus behavior analysisand understanding is one of the hotspots of the research. Presently, many domestic andinternational scholars carried on a great deal of research to single person actionrecognition and massive crowd behavior recognition, while researches on small scalecrowd behavior are relatively fewer. But as we all know, the small scale crowd behavioroccur much more often in real scenarios. Therefore, the researches on small scale crowdbehavior have practical significance and wide application prospect. This paper is mainlyfocuses on how to recognize the small scale crowd behavior effectively.The small scale crowd behavior have both the micro-level behavior feature and themacro-level behavior feature, from this point, this paper puts forward a behaviorrecognition algorithm which involve the micro-level features and the macro-level features,that is the behavior recognition algorithm for small scale crowd based on casual complexnetwork analysis. The complex network can characterize the relationship of each target ata macro level, so we need to find a way to express the relationship of each target.Therefore, the researches start with the microscopic feature-trajectory, and the GrangerCausality Test are employed to measure the interaction between the targets, and theconsequence will be the basis for construct complex network. In the process of networking,each target will be the node of complex network, and the interaction between the targetwill determine whether to connect the corresponding nodes with edges. After that, thestatistical properties of complex networks are calculated and will be the behaviorcharacteristics, the statistical properties are the Average Path Length, the Betweeness andthe Clustering Coefficient. In the experiments，in order to verify the correctness of ourmethod, two kinds of classification algorithm is employed to classify the crowd behavior,they are k-Nearest Neighbor classification algorithm and Support Vector Machineclassification algorithm. Six kinds of representative crowd behavior are selected to test the algorithm in the experiment, they are Gather, Chat, Split, Linger, Meet and Walk-Together,the number of participants is4~8in the video of we captured.Finally, through statistical analyze the results of the experiment, it shows that thealgorithm can express and recognize the crowd behavior effectively, compared with thesingle network features, the features of network which combine the pair-causality andgroup-causality is much better on the recognition rate. In the experiment, the mostbehaviors’ recognition rate is more than80%, and some individual behaviors’ recognitionrate is stays above90%.