Ground Plane Estimation for Obstacle Avoidance During Fixed-Wing UAV Landing

The automatic crash-landing of a fixed-wing UAV is challenging due to high velocity of approaching aircraft, limited manoeuvrability, lack of ability to hover and low quality of textural features in common landing sites. Available algorithms for ground plane estimation are designed for automatic cars, robots, and other land-bound platforms, where textural features can be easily compared due to low altitude, or for rotor-wing UAV that can hover. Their usefulness is limited when quick manoeuvre is needed to avoid collision with obstacles. Due to developments in parallelised, mobile computational architectures, an approach based on dense disparity estimation becomes available assuming proper constraints on ground plane transformation phase space. We propose an algorithm utilising such constraints for ground plane estimation on the often prohibitively time-consuming task of disparity calculation as well as plane estimation itself using a pyramid-based approach and random sample consensus in order to discard pixels belonging to obstacles. We use Inertial Navigation System, commonly available in fixed-wing UAVs, as a source of such constraints and improve our estimation in subsequent frames allowing for stable flight trajectory as well as detection of obstacles protruding from the ground plane.