Vision-based geometry estimation and receding horizon path planning for UAVs operating in urban environments

This paper presents a receding horizon control strategy to enable small unmanned air vehicles to fly autonomously through complex urban environments. An adaptive, multiresolution-based learning algorithm is employed to estimate the 3D geometry of the environment. This learning algorithm generates an adaptive approximation based on measurements of the 3D positions of static points in the environment, obtained via feature point tracking and structure from motion. A receding horizon path planning strategy is used to select a series of locally-optimal path points in front of the vehicle. These path points are selected in such a manner that the vehicle approaches an overall target point while avoiding static obstacles that may lie in its path. These obstacles are estimated via the adaptive learning algorithm, which provides constraints for the path planner. A standard waypoint controller is then used to fly the vehicle through the selected path points. This vision-based control strategy is demonstrated in simulations of a small UAV flying through a virtual urban environment