Joint Multi-person Detection and Tracking from Overlapping Cameras

We present a system to track the positions of multiple persons in a scene from overlapping cameras. The distinguishing aspect of our method is a novel, two-step approach that jointly estimates person position and track assignment. The proposed approach keeps solving the assignment problem tractable, while taking into account how different assignments influence feature measurement. In a hypothesis generation stage, the similarity between a person at a particular position and an active track is based on a subset of cues (appearance, motion) that are guaranteed observable in the camera views. This allows for efficient computation of the K-best joint estimates for person position and track assignment under an approximation of the likelihood function. In a subsequent hypothesis verification stage, the known person positions associated with these K-best solutions are used to define a larger set of actually visible cues, which enables a re-ranking of the found assignments using the full likelihood function.

We demonstrate that our system outperforms the state-of-the-art on four challenging multi-person datasets (indoor and outdoor), involving 3-5 overlapping cameras and up to 23 persons simultaneously. Two of these datasets are novel: we make the associated images and annotations public to facilitate benchmarking.