Polynomial curb detection based on dense stereovision for driving assistance

A real-time algorithm for curb detection in traffic scenes, based on dense stereovision, is proposed. Curbs are modeled as cubic polynomial curves. 3D points from stereovision are transformed into a Digital Elevation Map (DEM), in order to have a compact representation of the 3D space. Curb points are detected as the cells of the DEM that present a specific height variation. Only curb points that are temporally persistent and non-occluded are considered. Relevant cubic polynomials are computed from the set of curb points by a RANdom SAmple Consensus (RANSAC) approach. For each relevant polynomial, the curb patch is extracted by analyzing the DEM along the polynomial curve. Finally, the vertical location and height of each curb are computed based on the local elevation data.