People Counting In Complex Scenarios

This paper addresses vision-based people counting under the assumption of an oblique stationary camera setup without any usage of tracking. Our approach is based on motion detection and blob analysis. The estimation of the number of people in the scene is done by accumulating the estimation values of each blob. The blob specific estimation values are computed using four input values: the minimum and maximum number of people which may be present in the blob, the blob area and a predicted number number of people which is computed from the temporal change at the blob location. We have included shadow elimination using colour invariants which are obtained from the measurement of coloured object reflectance assuming a Gaussian colour model. The method was evaluated in single camera scenarios for various sequences taken in halls of shopping centres and railway stations.