Multi-sensor human tracking with the Bayesian Occupancy Filter

The utilisation of a network of heterogeneous sensors to monitor human activity in a large space is essential due to the important field of view to be covered and the possible cluttered environment. The interpretation of this high number of data requires fast and powerful fusion algorithms in order to make easier the next human or computer work. In this paper the utilisation of a probabilistic occupancy map is proposed to fuse data coming from infrared and visible cameras. By estimating the occupancy and the velocity of each spatial cell representing the environment and thanks to a background subtraction algorithm, it is shown that human can be efficiently tracked. The architecture presented provides necessary information about pedestrians to perform, in the very near future, a human behaviour recognition step.