The use of autonomous pan-tilt cameras as opposed to static cameras can dramatically enhance the range and effectiveness of surveillance systems, but effective tracking in such pan-tilt scenarios remains a challenge. Existing approaches for constructing mosaiced background models require accurate camera motion parameters, and online updates for the background model in the presence of scene activity, as well as real-time tracking of targets in the presence of partial occlusions have not been solved. In this paper we propose a model that requires no camera motion parameters, the background is learned online, and the solution is integrated with target tracking. Camera egomotion is estimated as the dominant cluster mean for a mixture of Gaussians learned over point correlations between consecutive frames. Putative target regions are detected as changes over the learned background model GMM mosaic. In scenes involving multiple agents, a particular target is identified based on pre-defined appearance priors, and this target is kept in the image center as it moves in the scene, occasionally encountering occlusions. The camera pan-tilt control is achieved using dynamic error expectation to drive proportional-integral action. Results (validated ROC curves against hand-groundtruthed data) are presented from different imaging and task conditions.