This paper presents the specifications and architecture for a stereovision sensor to be used in intersection assistance. The intersection problem imposes a wide field of view, reasonable accuracy for the typical intersection length, and fast response time. The image and 3D data provided by the low level routines are used to generate two kinds of environment descriptions an unstructured description, composed of elevation maps, occupancy grids, and polylines delimiting obstacle areas and curbs, and a structured description, composed of lanes, cuboidal objects, and classified pedestrians. The descriptions can be further combined, and additional data sources can be used, in order to provide a complete and accurate description of the intersection.