A Visual Surveillance System for Incident Detection

A conceptual module capable of scene interpretation for incident detection has been implemented using a relational network approach and a concept of dynamic grouping to ease the complexity involved in computing multiple object behaviours. This system has been tested on several roundabout scenarios and has given good results using simulated perfect data. Using more realistic data necessitated the handling of multiple hypotheses on the object classes. This was easily handled using the method developed and even though this increased the complexity of the system, on-line response was maintained.