A real-time vision system for crowding monitoring

In this paper, the approach used to estimate the number of people for planning purposes in DIMUS (ESPRIT project P-5345) is described. Crowd estimation is based on the image-processing and inference phases applied to the acquired data. Images come from a set of visual b/w camera oriented towards a zone to be monitored. Some significant features extracted from each acquired image are related to the number of people present in the monitored scene using the nonlinear models obtained by means of dynamic programming in an off-line training phase. The present approach, employing previously obtained estimates, improves the accuracy of estimation, with respect to an evaluation based only on present available data, and can predict crowding values without using new data, between two successive acquisitions. Results obtained after an extended test phase in a station of Genova's underground are reported.<>