Computer Vision System for Landing Platform State Assessment Onboard of Unmanned Aerial Vehicle in Case of Input Visual Information Distortion

The paper describes a computer vision system for organizing a safe landing of an unmanned aerial vehicle in conditions of potential distortions of the input video information. A sequence of methods for image preprocessing was proposed. Firstly, it is necessary to conduct a contrast enhancement using histogram equalization. After that, a Gaussian filter should be applied to remove an extra noise. Neural network YOLOv3-tiny was trained to recognize the state of the landing platform - open or closed. The achieved recognition accuracy on the test sample was 0.96. The algorithm was implemented in Jetson Nano and the achieved frame processing time is equal to 0.5 seconds.