Fast and robust b-spline terrain estimation for off-road navigation with stereo vision

Estimating the ground surface in order to distinguish between obstacles and free space is a requirement for most autonomous navigation tasks. Assuming a flat world is often sufficient in urban scenarios. In rural areas and for off-road navigation, however, this assumption doesn´t hold and estimating a dedicated model of the terrain is necessary. We propose to use an advanced 2.5D occupancy grid to accumulate surface heights over time. The ground is analytically described by a rational b-spline surface. In a fast 2-pass approach the ground is first modeled very coarsely to obtain the overall shape and to distinguish surface and obstacle cells. In the second step, a higher resolution model is estimated using only the surface cells. To further improve the robustness of the estimated ground surface, information from previous time steps is reused. Finally, navigation relevant data such as free space, obstacles and terrain slope can be computed based on the estimated ground surface.