Monocular Visual Odometry for Fixed-Wing Small Unmanned Aircraft Systems

The popularity of small unmanned aircraft systems (SUAS) has exploded in recent years and seen increasing use in both commercial and military sectors. A key interest area for the military is to develop autonomous capabilities for these systems, of which navigation is a fundamental problem. Current navigation solutions suffer from a heavy reliance on a Global Positioning System (GPS). This dependency presents a significant limitation for military applications since many operations are conducted in environments where GPS signals are degraded or actively denied. Therefore, alternative navigation solutions without GPS must be developed and visual methods are one of the most promising approaches. A current visual navigation limitation is that much of the research has focused on developing and applying these algorithms on ground-based vehicles, small hand-held devices or multi-rotor SUAS. However, the Air Force has a need for fixed-wing SUAS to conduct extended operations. This research evaluates current state-of-the-art, open-source monocular visual odometry (VO) algorithms applied on fixed-wing SUAS flying at high altitudes under fast translation and rotation speeds. The algorithms tested are Semi-Direct VO (SVO), Direct Sparse Odometry (DSO), and ORB-SLAM2 (with loop closures disabled). Each algorithm is evaluated on a fixed-wing SUAS in simulation and real-world flight tests over Camp Atterbury, Indiana. Through these tests, ORB-SLAM2 is found to be the most robust and flexible algorithm under a variety of test conditions. However, all algorithms experience great difficulty maintaining localization in the collected real-world datasets, showing the limitations of using visual methods as the sole solution. Further study and development is required to fuse VO products with additional measurements to form a complete autonomous navigation solution