Object Recognition by a Flying Robot Using Image Processing Algorithms

Abstract

In the past few years, researchers and scientists have paid particular attention to unmanned aerial vehicles, especially multirotors. Various research groups around the world are working on these types of drones, which are used in various fields such as rescue operations, civil operations, surveying, and sampling of infected areas. In most applications, it is necessary to determine the position of a particular object using these robots. To do this, we extract features from the images sent by the robot and match them to the known image features to identify the object we are looking for. One application of this matching is to compare two or more images and stitch them together to extract two-dimensional or three-dimensional maps. In this study, image processing algorithms including SIFT, SURF, ORB, and BRISK were implemented on three types of hardware including Raspberry Pi3, Odroid C2, and Intel NUC, and they were evaluated in terms of speed and accuracy using the Python programming language and OpenCV machine vision library. Subsequently, the results of the study were analyzed, and appropriate algorithms were proposed in accordance with the applications of the flying robot. The ORB algorithm was found to be the fastest, while the SIFT and BRISK algorithms were the slowest. The Intel NUC hardware had the least mass-to-speed efficiency, and Odroid C2 averaged the best mass-to-speed efficiency. On the other hand, the SIFT algorithm provided the most precision in finding image properties among the investigated algorithms, and the ORB algorithm provided the least precision in finding objects in the images.