A Human Model for Detecting People in Video from Low Level Features

A method for detecting people in video streams is presented for the context of a front view monocular camera. This paper describes the proposed human model, which combines skin color and foreground probability maps by defining the spatial relationships that exist between them. The detection is performed with a Monte Carlo simulation for the defined Bayesian framework, in order to estimate the model parameters. The detected people are then associated with signatures, that are compared between consecutive frames in order to achieve tracking. Promising results are obtained for the detection and matching of multiple people, as presented for a transportation vehicle application.