Automatic reconfiguration of video sensor networks for optimal 3D coverage

Abstract

During the last years, the research in the field of video analytics has focused more and more on video sensor networks. Although single-sensor processing is still an open research field, practical applications nowadays require video analysis systems to explicitly consider multiple sensors at once, since the use of multiple sensors can lead to better algorithms for tracking, object recognition, etc. However, given a network of video sensors, it is not always clear how the network should be configured (in terms of sensor orientations) in order to optimize the system performance. In this work we propose a method to compute a (locally) optimal network configuration maximizing the coverage of a 3D environment, given that a relevance map of the environment exists, expressing the coverage priorities for each zone. The proposed method relies on a transformation projecting the observed environment into a new space where the problem can be solved by means of standard techniques such as the Expectation-Maximization algorithm applied to Gaussian Mixture Models.