Design and implementation of a new vision-based navigation for Unmanned Aerial Vehicle.

Overview & Synopsis

The use of GPS is the most efficient way to locate a drone. However, problems with GPS signal reception is an obstacle to estimate drone positioning. This paper presents another process of geolocation, without GPS, which will allow the drone to locate itself on long routes, by vision-based absolute matching. The purpose of this paper is to introduce the development of a robust program ensuring the real-time geolocation of a drone in cruise flight on any type of topography. The aim of this operation is to develop algorithmic architectures that allow improving the geolocation performance by using image-matching method based on the selection of correct descriptors matchings. The challenge is to have generic models to predict the performance of a computing architecture for an algorithm. This study shows how to optimize this matching process to limit inaccuracies and drifts while keeping a correct position calculation time. This algorithm is based on a new approach, totally absent in the literature, using the closed geometric envelope, formed by descriptors of points of interest in images. A geometric envelope comparison method is applied between any taken image of the ground by the camera of the drone and images of a database (DB), beforehand embedded, to select correct descriptors matchings. Therefore, the number of descriptor matching should not exceed 6, with the execution time of the order of 140ms. Nevertheless, to consider images with many descriptors, such as those detected on cities, it is necessary to upgrade this original algorithm to associate a known "similarities between transformations" mathematical method which is explained afterwards. So, for the number of descriptors matching over 6, the method by similarities succeeded in eliminating the false matches and thus rectifying the drift in 98% of the cases in about 5ms. Finally, this study measures impacts of seasonal changes, fog, contrast, and blur; between image taken by the camera of the drone and the images in the DB; on the performance of geolocation.