Personal assistance, delivery services, and crowd navigation through robots fleet are complex activities that involve human-robot interaction and fleet coordination. Human location estimation is one of the key factors in assisting robots in their tasks. This paper proposes an efficient process for propagating human presence probability based on partial observation of humans by the robot fleet. This process provides real-time information about the most probable region on the map where humans can be found. We propose a new problem representation allowing us to efficiently parallelize the propagation. To deal with the learned model and the real time robot observations, we propose to include a Gaussian rotation probability process (VonMises) combined with the previous learned observation to adapt the propagation. A set of experiments has been conduced with simulated environments that include real data allowing us to evaluate the model and to compare with the standard approaches.