An important challenge for mobile robots is to navigate efficiently in human populated environments. In this context, we examine how human presence grids can be extended to model human motions, considering only embedded sensors. The proposed flow grid computes in each cell a discrete distribution of the human motion. The model is defined to take into account the most recent observations, so as to adapt to changes. More, it is expanded with a predictive motion pattern. Then we revisit the cost function of the A* pathplanning algorithm to take into account the risk of encountering humans. We compare the standard A* with variants exploiting the human presence likelihood [1] and the proposed flow grid. Experiments in simulation show that the Flow grid A* is able to compute paths minimizing the risk of navigating against human flows, and to adapt to their variations. Experiments with a mobile robot confirms the ability of the model to map human flows and to optimize paths.