Reliable and efficient perception and reasoning in dynamic and densely cluttered environments are still major challenges for driver assistance systems. Most of today's systems use target tracking algorithms based on object models. They work quite well in simple environments such as freeways, where few potential obstacles have to be considered. However, these approaches usually fail in more complex environments featuring a large variety of potential obstacles, as is usually the case in urban driving situations. In this paper, we propose a new approach for robust perception and risk assessment in highly dynamic environments. This approach is called Bayesian occupancy filtering; it basically combines a four-dimensional occupancy grid representation of the ecobstacle state space with Bayesian filtering techniques.