Automatic runway detection for unmanned aerial vehicles (UAV)

Abstract

Unmanned Aerial Vehicles (UAVs) are among the latest technologies seeing major research and investment to optimize and automate its processes. Although landing is one of the most complex processes to simulate, due to several variables which affect landing performance, UAVs’ limitations, and possible financial expenses, attempts at solving this problem are numerous. In this project, a system which uses computer vision and image processing to automatically detect the runway is presented. By using and adapting CORF with push-pull inhibition algorithm [1], runway edges are detected and used to determine the position of the runway threshold midpoint. A computer vision-based approach was chosen over a GPS-only system to counter the limitations presented by GPS, mainly its lack of accuracy when calculating distances and elevation. The system uses a dataset of 631 aerial images portraying the possible conditions of a runway, such as light and weather conditions and features of the runway, to maximize effectiveness and accuracy of the runway detection process. The images are obtained both from real-life landings (107 images) and from a flight simulator (324 images), to ensure that most possible scenarios are covered. A training set made up of 150 random images from the simulation dataset is used to determine the number of detected contours to consider and which set of y-coordinates should be used to select the edges which best describe the runway. Subsequently, these parameters are used to determine the distance error in pixels between the detected and the actual runway threshold midpoints for each image in the dataset. The system obtains an average distance of error of approximately 101 pixels, where day and night scenarios obtained average distance errors of approximately 31 and 23 pixels respectively, while rain and live images prove to be more challenging with average distance errors of 125 and 258 pixels respectively. Additionally, factors such as the slant distance of the UAV from the runway, image noise, visibility, and light conditions are shown to have a significant effect on the runway detection algorithm.