A Vision Based Forced Landing Site Selection System for an Autonomous UAV

Abstract

This paper presents a system overview of the UAV forced
landing site selection system and the results to date. The
forced landing problem is a new field of research for UAVs
and this paper will show the machine vision approach taken
to address this problem. The results are based on aerial
imagery collected from a series of flight trials in a Cessna
172.
The aim of this research is to locate candidate landing sites
for UAV forced landings, from aerial imagery. Output image
frames highlight the algorithm’s selected safe landing
locations. The algorithms for the problem use image
processing techniques and neural networks for the
classification problem.
The system is capable of locating areas that are large enough
to land in and that are free of obstacles 92.3% ± 2% (95%
confidence) of the time. These areas identified are then
further classified as to their surface type to a classification
accuracy of 90% ± 3% (98% confidence).
It should be noted that although the system is being designed
primarily for the forced landing problem for UAVs, the
research can also be applied to forced landings or glider
applications for piloted aircraft.