Multi-person localization and track assignment in overlapping camera views

The assignment of multiple person tracks to a set of candidate person locations in overlapping camera views is potentially computationaly intractable, as observables might depend upon visibility order, and thus upon the decision which of the candidate locations represent actual persons and which do not. In this paper, we present an approximate assignment method which consists of two stages. In a hypothesis generation stage, the similarity between track and measurement is based on a subset of observables (appearance, motion) that is independent of the classification of candidate locations. This allows the computation of the K-best assignment in low polynomial time by standard graph matching methods. In a subsequent hypothesis verification stage, the known person positions associated with the K-best solutions are used to define the full set of observables, which are used to compute the maximum likelihood assignment. We demonstrate that our method outperforms the state-of-the-art on a complex outdoor dataset.