Much research have in resent years been directed at sensing the presence and state of people and many possible applications exist. One application whose importance is continuously increasing with the average age of the western societies increases is assistive living environments for elderly.

In this project a person tracking system for multi-camera environments is developed with this application in mind. Tracking is based on two different cues or modalities: Foreground and feature points. These compliment each other well since foreground is present whenever a person is moving, while feature points initialised on a person can be tracked indefinetly if he is stationary.

Foreground is first found in 2D for each camera by estimating a background model using Gaussian Mixture Models. Then foreground from all cameras are combined yielding one single voxel-based 3D foreground. Feature points are initialised for each camera on persons and tracked individually in 2D using the KLT-tracker. Finally, layered sampling is applied to fuse information from foreground estimation and feature points into one particle filter for each present person.

The system is implemented to run on one computer using video recordings, and a part of the system is implemented to run live distributed between multiple computers. Tests on the CLEAR 2007 data set prove that the combination of the two modalities provide better performance than a tracking system based solely on one of them.