Organization of a Safe Unmanned Aerial Vehicle Landing Using Computer Vision in Conditions of Potential Input Visual Information Distortion

The paper presents a computer vision system for safe landing of an unmanned aerial vehicle in the situation of possible input visual information distortions. An algorithm for image preprocessing was developed, consisting of increasing contrast by equalizing the intensity histogram, followed by applying a Gaussian filter to remove extra noise. The landing platform has two states: open and closed. For their recognition, the YOLOv3-tiny neural network was modified. The accuracy of recognition for the test set was 0.97. The algorithm was integrated into the Jetson Nano. The achieved frame processing time was 0.05 seconds, which allows real-time processing of the video stream.