Odometry for autonomous navigation in GPS denied environments

Abstract

The mining industry is dangerous and puts human lives at risk daily, including hazardous environments such as: poor air quality due to gas leaks and dust, unstable structural mine adits after controlled detonations or natural disasters, and possible entrapment. Robots that can safely navigate into underground mining environments can conduct reconnaissance to inspect these hazardous environments reducing the risk to human lives. This thesis presents methods to enable autonomous navigation in underground mines, to include: 1) the system design for a flying platform, 2) computer vision techniques to extract the real-time pose of a moving robot, and 3) a Map Free LiDAR Odometry (MFLO) method. The flying platform system design focused on autonomously navigating in an underground mine. The complete system incorporates multiple sensors, an on-board embedded system, electrical connections, cabling, and an on-board power management system. Software was developed that integrates the sensors and fuses the measurements to be utilized for real-time odometry, obstacle avoidance, and control updates. A health monitor node was expanded to further ensure the safety of the aircraft. Computer vision strategies were developed to calculate the real-time pose of a moving robot with respect to a known static robot's position. The methods are: 1) ArUco Marker Identification, and 2) LED marker identification. Results are captured for both ArCuo and LED marker identification methods. Lastly, a real-time method to extract 3D ego-motion using a range flow constraint equation was developed. The method is map free, computationally light-weight, and reliable. MFLO is designed to operate in GPS-denied and light-deficient environments, making it ideal for small autonomous systems operating in underground mines. The range flow approach presented here performs up to 0.46\% position accuracy for an underground mine environment with a computation time of 20--96\,ms, depending on sensor resolution.