A self-adaptive approach for autonomous UAV navigation via computer vision

Abstract

In autonomous Unmanned Aerial Vehicles (UAVs), the vehicle should be able to manage itself without the control of a human. In these cases, it is crucial to have a safe and accurate method for estimating the position of the vehicle. Although GPS is commonly employed in this task, it is susceptible to failures by different means, such as when a GPS signal is blocked by the environment or by malicious attacks. Aiming to fill this gap, new alternative methodologies are arising such as the ones based on computer vision. This work aims to contribute to the process of autonomous navigation of UAVs using computer vision. Thus, it is presented a self-adaptive approach for position estimation able to change its own configuration for increasing its performance. Results show that an Artificial Neural Network (ANN) presented the best performance as an edge detector for pictures with buildings or roads and the Canny extractor was better at smooth surfaces. Moreover, our self-adaptive approach as a whole shows gain up to 15 % if compared with non-adaptive methodologies.