Research On Collective Behavior Analysis Based On Neighborhood Characteristics

Group is an unifed whole part formed by interdependent and interacted individ-uals based on the common objective. Cluster, being the sum of individuals, can becomposed of multiple groups. Understanding the collective behaviors and their forma-tion mechanism is one of the fundamental problems both in social and natural science.Multiple disciplines have done a lot of research on this topic and many models andmethods have been proposed. Compared to the traditional methods, computer visionand simulation provides us with not only powerful observation tools, but also publicfacility optimization, video surveillance, anomaly detection techniques for many keyengineering applications. Meanwhile, it can save cost of human resources and meetthe needs of intelligence in future.In recent years, agent-based model has been widely used. Based on prior rules,agent-based model treats individuals in crowd as “smart particles” which can automat-ically sense the environment and make decisions. It’s in compliance with group behav-ior mechanism and can be well extended. The model has two main methods: macro-scopic collective-dynamics and microscopic interactive-dynamics approach. The for-mer focuses on global properties, such as quantifying the collectiveness of cluster,group behavior classifcation；the latter typifed social force model, which simulatethe force exerted on individuals by introduction of social psychological research. How-ever, inthepreviousworks, the twomethods havenotbeenintegrated tojointlyanalyzecollective behaviors from both macroscopic and microscopic levels.In the feld of computer vision, it is still an open problem to determine whether anindividualmovessmoothlyornotwithinhisneighborhood. Inthispaper,anorderlinessmeasurement has been proposed. The proposed measurement is on the basis of agent- based model and captures the neighborhood characteristics of collective motion. It iscalculatedbySocialForceCorrelationPropagationalgorithmefcientlyandefectivelywith the analysis of social force feld’s direction distribution. The method is diferentfrom previous works for which only the magnitude of social force feld is taken intoconsideration.By combining the two approaches in agent-based model, this paper conductedboth macro and micro-scale analysis of crowd behavior. From the macroscopic view,this paper clusters groups with diferent motion mode by dynamic clustering algorithmbased on velocity correlation. From the microscopic view, three motion modes(freewalking, marching, circling)of cluster are conducted in Self-Propelled Particles simu-lation and the joint analysis is taken.Cluster videos are mainly shot in large public places. The video quality is oftenpoor caused by factors like variations of light, rains, fog, camera quality and shootingangle. So far, no large-scale standard video database for collective behavior analysishas been established. Therefore, Self-Propelled Particles simulation is widely usedin this paper to ignore the interference factors and validate the proposed algorithm.Then, the verifed algorithm is applied to the analysis of real scene videos collected byChinese University of Hong Kong for academic research. Experimental results showthat:(1) the proposed orderliness measurement can well quantify the degree of motionsmoothness an individual following his group;(2) the proposed dynamic clusteringalgorithm can efectively detect groups with diferent motion mode;(3) by combiningthe two approaches in agent-based model, macro and micro-scale analysis of crowdbehavior is jointly conducted.