Using image processing and YOLO to build a navigation system for a drone with an industrial application

Abstract

In this work we developed a semi autonomous drone navigation system for a cardboard box industry, to assist in counting the stock of cardboard reels. The developed methodology has four main steps, being the QR Code decoding, optical marker detection, navigation system and drone movement. For the QR Code decoding step, the pyzbar library was used. In the optical marker detection step, the YOLOv4 Tiny library was used, which uses machine learning techniques to detect objects in real time. YOLOv4 Tiny was trained using a custom dataset with images of optical markers and labels in a closed simulation environment, achieving a hit rate of 92.10%. The navigation system step is fed by the response of the neural network, in which each marker has a function associated with it. The last step depends on the navigation system, since it sends which command the drone must follow and the movement sends this command to the drone.