Nova metodologia para exploração e mapeamento autônomo de VANTs em terrenos desconhecidos usando deep learning através de estratégia orientada por fronteiras

Abstract

Unmanned Aerial Vehicles (UAVs) play a crucial role in exploring unpredictable terrains, such as disaster sites and search zones. To operate safely and efficiently in dynamic environments, these robots require precise autonomous navigation and efficient trajectory planning. This work presents an innovative strategy for the autonomous exploration of UAVs in unknown environments, focusing on boundary detection and utilization. The research proposes advanced methods based on deep learning techniques for accurate boundary detection, aiming to expand the unknown map as efficiently as possible. Additionally, a new decision-making policy is introduced, enabling the robot to select the most suitable boundary for exploration using information derived from a neural network model specialized in this task. The results were evaluated in three distinct scenarios: an indoor simulated environment, an outdoor simulated environment, and a real-world outdoor experiment. The proposed methodology reduced exploration time by approximately 25% indoors and 51% outdoors compared to a recent state-of-the-art approach. In addition to time efficiency, other metrics—such as average distance traveled and the number of decisions made—also improved across all scenarios. Another relevant aspect of this research was the demonstration of a weak correlation between the number of detected boundaries and the time required to identify them, highlighting the robustness of the proposed methodology. The observed results indicate that boundary detection time remains nearly constant, regardless of the number of boundaries present, making this approach particularly suitable for unknown and highly complex environments.