Vision-Based Receding Horizon Control for Micro Air Vehicles in Urban Environments

This paper presents a vision-based receding horizon control algorithm for enabling a micro air vehicle to fly autonomously through an unknown urban environment. A vision-based geometry estimation algorithm is used to generate a map of the environment and provide obstacle avoidance constraints on permissible MAV trajectories. The geometry estimation process is comprised of three components: a feature point tracker, a structure from motion algorithm, and an adaptive learning process. The feature point tracker is used to track points in the images collected by an onboard camera. The three-dimensional positions of these points, a representation of the scene structure, are calculated using structure from motion. The mathematical learning algorithm is used to generate an adaptive three-dimensional obstacle map from the collection of measured points in the environment. The receding horizon controller computes the control sequence that minimizes a cost functional over a finite time horizon while enforcing obstacle avoidance constraints. The control is applied over a subset of this horizon after which a new optimization is performed. Therefore, the receding horizon approach can easily incorporate new information from the updated obstacle map each time an optimization is performed. The vision-based receding horizon control strategy is demonstrated in simulated flights of a micro air vehicle through a virtual environment. These simulations assume the vehicle motion is known and use synthetic images from the virtual environment. The simulations results show that the vision-based receding horizon control algorithm was able to successfully navigate the vehicle to a goal location while avoiding buildings in the environment.