People counting by learning their appearance in a multi-view camera environment

摘要

We present a people counting system that, based on the information gathered by multiple cameras, is able to tackle occlusions and lack of visibility that are typical in crowded and cluttered scenes. In our method, evidence of the foreground likelihood in each available view is obtained through a bio-inspired mechanism of self-organizing background subtraction, that is robust against well known foreground detection challenges and is able to detect both moving and stationary foreground objects. This information is gathered into a synergistic framework, that exploits the homography associated to each scene view and the scene ground plane, thus allowing to reconstruct people feet positions in a single "feet map" image. Finally, people counting is obtained by a k-NN classification, based on learning the count estimates from the feet maps, supported by a tracking mechanism that keeps track of people movements and of their identities along time, also enabling tolerance to occasional misdetections. Experimental results with detailed qualitative and quantitative analysis and comparisons with state-of-the-art methods are provided on publicly available benchmark datasets with different crowd densities and environmental conditions.