Local Terrain Mapping for Obstacle Avoidance using Monocular Vision
Open Access
- Author:
- Marlow, Sean Quinn
- Graduate Program:
- Aerospace Engineering
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- None
- Committee Members:
- Jack Langelaan, Thesis Advisor/Co-Advisor
Jack Langelaan, Thesis Advisor/Co-Advisor - Keywords:
- occupancy grid
optical flow
obstacle avoidance
rotorcraft - Abstract:
- The motivation behind the research described in this thesis is to be able to navigate a small unmanned aerial vehicle (UAV) through complex environments. Missions envisioned for small UAVs now require low altitude flights among many obstacles. These obstacles can be simple shapes (telephone poles) or more complex shapes (corners and protrusions of an urban canyon). This thesis focused on the problem of estimating obstacle locations and safe navigation to a known goal location. The close proximity and complex nature of these environments requires a system of navigation and obstacle avoidance using onboard sensors. However, payload limitations of small UAVs place significant restrictions on sensor size, weight, and power consumption. This thesis describes the development and simulation of an algorithm that enables safe navigation using only a monocular camera and GPS corrected inertial navigation system (GPS/INS). The obstacle estimation problem poses many challenges. First, the camera measurements (bearings to obstacles and bearing rates of the obstacles due to motion of the vehicle) are heavily corrupted with noise. This greatly reduces the certainty of information obtained from the camera. Second, the equations governing the vehicle motion and vision measurements are highly non-linear. The combination of noisy measurements with non-linear equations leads to significant uncertainties in the estimation problem. Traditional feature-based mapping techniques (like the Kalman Filter) become intractable in this environment as too many features must be mapped to resolve the complex shaped obstacles. A local occupancy grid is instead used to store estimates of occupied space. The local occupancy grid has a limited size that can be specified depending on the sensor field of view and computational condiserations. The origin of the grid is fixed to the vehicle and in this application the orientation is fixed to an inertial frame. By keeping the grid centered on the vehicle, vehicle motion must be accounted for in the grid with a motion update step. A potential field trajectory generator provides the means for vehicle navigation and obstacle avoidance. The algorithm performance is examined through simulations in an environment modeled after the McKenna Military Operations in Urban Terrain site at Ft. Benning, GA.