It has been shown that multi-people tracking could be successfullly formulated as a Linear Program to process the output of multiple fixed and synchronized cameras with overlapping fields of view. In this paper, we extend this approach to the more challenging single-camera case and show that it yields excellent performance, even when the camera moves. We validate our approach on a number of basketball matches and argue that using a properly retrained people detector is key to producing the probabilities of presence that are used as input to the Linear Program.