We aim to combine aspects from the fields of pervasive computing, pattern recognition and machine learning to achieve the following key objectives:

• Provide robust distributed architectures for synchronising and integrating the data from multiple sensors into a single, cohesive stream.
• Enable wide area tracking across multiple rooms and multiple stories.
• Assign the tracked target to nodes with the best view in a dynamic fashion.
• Develop probabilistic, scalable and modular models for pattern recognition at varying resolutions and under conditions of uncertainty in the sensor data.
• Utilise local pattern recognition to determine the global state.

We will investigate techniques to coordinate multiple cheap cameras to deal with complex spatial temporal scenarios in wide-area scenes such as for surveillance and outdoor broadcasting. The main issues we address are:

• Representation of the domain knowledge. That is, actions and interactions of a group of tracked objects (people in this case). In particular, the representation must encode the inherent spatio-temporal characteristics of the domain, deal with uncertainties of noisy observations and represent coupling of goals and sub-goals to allow for recognition of multi-object actions.

• Coordination of multiple cameras. Each camera has limited resources, and the problem is for each camera to decide the best course of action (which objects it should track), so that the overall objective of tracking all objects reliably is best satisfied.

• Recognition of complex multi-object actions. That is, to be able to interpret group behaviour where a group consists of more than one interacting object.