This thesis deals with intelligent videosurveillance, and focus on the supervision of camera networks with nonoverlapping fields of view, a classical constraint when it comes to limitate the building instrumentation. It is one of the use-case of the pedestrian re-identification problem. On that point, the thesis distinguishes itself from state of the art methods, which treat the problem from the descriptor perspective through image to image signatures comparison. Here we consider it from a bayesian filtering perspective : how to plug re-identification in a complete multi-target tracking process, in order to maintain targets identities, in spite of observation discontinuities. Thus we consider tracking and signature comparison, at the camera level, and use that module to take decisions at the network level. We describe first the classical re-identification approaches, based on the description. Then, we propose a mixed-state particle filter framework to estimate jointly the targets positions and their identities in the cameras. A second stage of processing integrates the network topology and optimise the re-identifications in the network. Considering the lack of public data in nonoverlapping camera network, we mainly demonstrate our approach on camera networks deployed at the lab. A publication of these data is in progress.
