An important challenge of mobile robotics is to allow efficient and safe navigation in human populated environments. This requires the robots to perceive human presence and motion, and more generally to model and exploit the human flows in the environment. In this context we address the problem of modeling human flows from the perception of mutli-robot systems. This leads us to propose a two layer grid model of human motion probabilities. The first grid is computed from the observations and the second introduces a human motion prediction that allows to accelerate the flowmodeling. We study the advantage of computing the proposed flow grid in path planning and navigation tasks inside populated indoor environment. We compare the model with two other A*-based models in simulations involving 200 moving persons and 4 collaborative robots.