Generalised vertical projection histograms using multi-plane homology

Vertical projection histograms are an efficient shape representation for 2D binary silhouettes and have been widely used in pedestrian localisation for video surveillance. The weakness of this method is that it is not invariant to rotation. In this Letter, a generalised vertical projection histogram is proposed to solve this problem, in which the homology transformations of the foreground silhouettes, for a set of parallel planes, are warped to, and accumulated, in the original foreground map. Then a method, similar to the vertical projection histogram, is carried out to localise the pedestrians in the foreground silhouettes. The algorithm applies an integrated approach using both image projection and geometric projection. Its value is demonstrated in a case study on pedestrian localisation with cast shadows.