A fuzzy classifier for visual crowding estimates

A trainable vision-based system is presented, which is able to perform reliable, real time estimates of the crowding level present on the platforms of underground stations. Taking as input standard the B/W images of the scene, a classification of the crowding level is performed in terms of five qualitative crowding classes, ranging from no people to overcrowding. Visual feature extraction and fuzzy classification methods employed are described in detail, as well as the procedure adopted to train the hyper basis functions neural classifier. Experiments and results obtained on real data are reported, with some emphasis on the possibility of empirically estimating the generalization capability of the proposed system.