Visual Detection of Moving Ground Obstacles for Full-Size Autonomous Aircraft Landing

Full-size autonomous rotorcraft flying multi-hop missions must be able to react quickly to unsafe landing conditions, such as moving ground vehicle activity in the landing zone (LZ). In this paper, an algorithm is presented which detects moving ground vehicles from distances of 120 m to 500 m (394 ft to 1640 ft), using monocular visual-spectrum video and pose information. The approach requires few assumptions on the object being detected (contiguous appearance and discernible contrast against the ground), allowing for detection of different sizes and shapes of moving objects. The performance of the algorithm is evaluated over a dataset of 76 real approaches, analyzing the true and false positive rates per-approach and per-frame. The algorithm is found to correctly detect hazardous landing conditions in 41/41 approaches to moving-vehicle-occupied LZs, and falsely declare unsafe landing conditions in 2/35 approaches to clear LZs.