Comparison of Visual Simultaneous Localization and Mapping Methods for Fixed-Wing Aircraft Using SLAMBench2

Visual Simultaneous Localization and Mapping (VSLAM) algorithms have evolved rapidly in the last few years, however there has been little research evaluating current algorithm\u27s effectiveness and limitations when applied to tracking the position of a fixed-wing aerial vehicle. This research looks to evaluate current monocular VSLAM algorithms\u27 performance on aerial vehicle datasets using the SLAMBench2 benchmarking suite. The algorithms tested are MonoSLAM, PTAM, OKVIS, LSDSLAM, ORB-SLAM2, and SVO, all of which are built into the SLAMBench2 software. The algorithms\u27 performance is evaluated using simulated datasets generated in the AftrBurner Engine. The datasets were designed to test the quality of each algorithm\u27s tracking solution, as well as finding any dependence on camera field of view (FOV), aircraft altitude, bank angle, and bank rate. Through these tests, it was found that LSDSLAM, ORB-SLAM2, and SVO are good candidates for further research, with MonoSLAM, PTAM, and OKVIS failing to track any datasets. All algorithms were found to fail when the capturing camera had a horizontal FOV of less than 60 degrees, with peak performance occurring at a FOV of 75 degrees or above. LSDSLAM was found to fail when the aircraft bank angle exceeded half of the camera\u27s FOV, and SVO was found to fail below 450 meters altitude. The simulations were also tested against a comparable real world dataset, with agreeable results, although the FOV of the real world dataset was too small to be a particularly useful test. Further research is required to determine the applicability of these results to the real world, as well as fuse VSLAM algorithms with other sensors and solutions to form a more robust navigation solution