A distributed online learning tracking algorithm

In this paper we introduce a way of tracking people in an indoor environment across multiple cameras with overlapping as well as non-overlapping fields of view. To do so, we use our distribution model called SpARTA and an extended Tracking-Learning-Detection algorithm. A big advantage in comparison to other systems is that each camera node learns the tracked person and builds a database of positive and negative examples in real time. With these datasets we are able to distinguish different people across different nodes. The learned data is shared across nodes, so that they improve each other while tracking. In the main part we present an experimental validation of the system. Finally, we will show that distribution of tracking data improves tracking across multiple nodes considerably with regard to partial occlusion of the tracked object.