For any autonomous vehicle, understanding which area around the vehicle is free or drivable is a key component. It is important to find road boundaries such as ditches, curbs or guard rails. Finding small objects on the road, that might be blocking the vehicle path, is also critical. Most prototypes for autonomous vehicles feature laser scanners for this purpose. We propose a stereo vision based system as redundancy or as a cost efficient replacement for laser scanners in the application of drivable surface determination. The system generates a detailed map that indicates which area in front of the vehicle that is considered to be drivable and non-drivable.
The base for this map is a dense 3D point cloud generated from the stereo vision system. The left and right images are used to create a disparity image which is then translated into a 3D point cloud. The 3D points are used to generate a detailed elevation grid of the observed area. The gradient of this grid is then used for the judgement of drivable/non-drivable. The 3D point cloud and the drivability map is generated in real time at 22 Hz on production intent hardware.
We demonstrate the performance on a variety of use cases. We show that stereo vision is a strong candidate to replace the laser scanners used in many autonomous driving applications. At the same time our stereo vision system features many other applications, such as EuroNCAP AEB, lane tracking and traffic sign recognition.