Human-like Driving for Autonomous Vehicles using Vision-based Road Curvature Modeling

Most autonomous vehicles use GPS to determine vehicle location and heading. Using GPS for vehicle autonomous driving posts a few challenges. It does not look far ahead of the vehicle and requires frequent adjustment of vehicle heading. This results in an unstable control system and increases the chance of unstable driving behavior. Unlike this kind of passive or reactive control system, human drivers look far ahead of the road to determine the road curvature and actively adjust vehicle heading and turning angle in advance. Visual information is essential to enabling this kind of human-like driving behavior for autonomous vehicles and should be an important supplement to the GPS-based system, especially in GPSdenied environments. Road lanes are often curved, making vision-based detection of smooth and continuous curves in front of the vehicle a challenging task. Furthermore, commonly used computer vision algorithms such as edge detectors or Hough transform for line or curvature detection are not robust in changing lighting conditions. This paper presents a vision algorithm designed specifically for detecting and modeling road curvature for humanlike active steering control and heading adjustment for autonomous vehicles. The proposed algorithm has been tested for paved and unpaved road conditions and shown very good results.