The research presented in this thesis focuses on the problem of the perception of the urban environment which is complex and dynamic in the presence of noisy and incomplete exteroceptive measurements obtained from on-board sensors. The problem is formalized in terms of sensor data fusion with a spatial representation of the environment. This work has been carried out for the autonomous navigation of intelligent vehicles within the national project ANR CityVIP. After having considered various formalisms to represent uncertainty, a fusion of spatio-referenced grids managing uncertainty with belief functions is studied. This system is capable of merging multi-layers and multi-echoes lidar measurements, obtainedat different time indexes to build a dynamic local map as a discrete evidential occupancy grid. The main advantages of belief functions are, firstly, to represent explicitly ignorance, which reduces the assumptions and therefore avoid introducing wrong a priori information and, secondly, to easily use conflicting information to determine the dynamics of the scene such as movements of the cells. The formalism of evidential occupancy grids is then presented in details and two multi-layers and multi-echos lidar sensor models are proposed. The propagation of the information through geometrical transformations is formalized in a similar way of image transformation framework. Then, the implementation of the approach is described and the injection of prior geographic information is finally investigated. Most of the works presented have been implemented in real time on a vehicle and many tests in real conditions have been realized. The results of these researches were presented through five international conferences [Moras et al., 2010, Moras et al., 2011a, Moras et al., 2011b, Moras et al., 2012], [Kurdej et al., 2012] and the experimental vehicle was presented at the official demonstration project CityVIP in Paris and at the IEEE Intelligent Vehicles Symposium 2011, in Germany.
