Synchronisation and matching in camera networks

Networks of cameras are composed of nodes with the capability of performing local processing that helps transferring the minimum amount of information for completing network tasks. Two important challenges for camera networks are synchronisation (temporal video alignment) and object matching (inter-camera association and object re-identification). In this seminar I will present a video alignment method based on observing the actions of a set of articulated objects that is applicable to general and unconstrained scenarios in a way that is not feasible with current state-of-the-art approaches. The method uses high-level video analysis (object actions) and does not impose constraints on the relative pose or motion of the cameras, on the structure of the time warping between the videos and on the amount of overlap among the fields of view. Next I will discuss a feature selection method that minimizes the data needed to represent the appearance of an object by learning the most appropriate features for person re-identification. For each feature, the computational cost for extraction and storage are considered together with its performance to select cost-effective good features. This selection allows us to improve the re-identification while reducing the bandwidth for communicating the features over the network. The methods will be illustrated with several examples from on-going and recently completed projects.