Autonomous navigation among humans is, however simple it might seems, a difficult subject which draws a lot a attention in our days of increasingly autonomous systems. From a typical scene from a human environment, diverse shapes, behaviours, speeds or colours can be gathered by a lot of sensors ; and a generic mean to perceive space and dynamics is all the more needed, if not easy. We propose an incremental evolution over the well-known occupancy grid paradigm, introducing grid cell propagation over time and a limited neighbourhood, handled by probabilistic calculus. Our algorithm runs in real-time from a GPU implementation, and considers completely generically space-cells propagation, without any a priori requirements. It produces a set of belief maps of our environment, handling occupancy, but also items dynamics, relative rigidity links, and an initial object classification. Observations from free-space sensors are thus turned into information needed for autonomous navigation.