Simulation based learning of operator-environment interactions for dynamic crowd monitoring by means of a Bayesian Dynamical Switching Model

Human behaviour analysis has important applications in many emergency management problems as Intelligent Video Surveillance (IVS) for crowding situations. In many VS systems, supervision from a human operator is needed; for example, in overcrowding situations, the experience of a security operator is crucial in order to redirect people flow for the maintenance of an acceptable safety level. An automatic system may assist human decisions relying on real time processing of huge data amount coming from intelligent heterogeneous sensor networks. Interactive and Cognitive Environments (ICEs) can be designed with the help of Cognitive Systems (CS), in order to support or even to substitute human operators in analysing sensor data and taking appropriate decisions. Interactions between a human operator and a dynamic environment need to be modelled. To this end, Dynamical Bayesian models can be used within a CS to learn different interaction strategies. However, the availability of an interactive crowd simulators is necessary for the off-line operator training and CS learning. This paper describes an innovative and complex structure based on coupled Switched Event Dynamic Bayesian Networks (SE-DBNs) able to learn the cause-effect relationships between user actions and changes in crowd configurations. This learning phase represents an effective knowledge transfer from human operator towards automatic systems called Cognitive Node (CN). It is shown that Switched E-DBNs, within a Dynamical Switching Model (DSM) framework for interaction description, can be used to represent and predict possible operators’ actions. Results are presented, where crowd behaviour is modelled by means of Social Forces and can interact with a human operator within a visual 3D simulator. The way anomalies are detected and consequently handled is demonstrated in both the user-support and automatic modes.