Integration of multisensor data for overcrowding estimation

A system for interpretation of complex scenes is presented. The main characteristics of the system are based on virtual multisensor input, knowledge-based and multilevel architecture and a Bayesian network used to implement the inference mechanism and to manage the model network. The specific application of the system is crowd evaluation in an underground station environment to detect a dangerous situation by using optical sensors. The multilevel architecture of the system is modeled as a probabilistic network of passing-message nodes. Each node corresponds to a virtual distributed processor that is used to obtain the probabilistic value of the locally detected crowding level. The network updating mechanism is presented. By using several low level algorithms suitable features are extracted from images. The virtual sensor models are described