The paper proposes a novel visual navigation framework to address the challenges encountered by unmanned aerial vehicles (UAVs) when operating in indoor environments, including insufficient image feature extraction and poor exploration. The framework utilizes depth images, flight state data, and relative orientation information to guide the UAV towards the intended destination in an end-to-end manner. In addition, we have designed a Convolutional Block Attention Module (CBAM) with a non-sparse reward mechanism to enhance the efficiency of visual feature extraction, thereby improving the UAV’s exploratory capabilities. We have deployed the algorithm in an AirSim-based simulation environment, characterized by ever-changing obstacles, to validate its effectiveness. The results demonstrate that our algorithm enables the UAV to avoid obstacles and reach the destination effectively in unfamiliar indoor environments. This solution has significant implications for the effective deployment of UAVs in indoor settings, particularly in applications such as search and rescue, inspection, and surveillance.
Vision-Based Indoor UAV Navigation with Deep Reinforcement Learning
Lect. Notes Electrical Eng.
International Conference on Autonomous Unmanned Systems ; 2024 ; Shenyang, China September 19, 2024 - September 21, 2024
Proceedings of 4th 2024 International Conference on Autonomous Unmanned Systems (4th ICAUS 2024) ; Kapitel : 11 ; 101-112
22.03.2025
12 pages
Aufsatz/Kapitel (Buch)
Elektronische Ressource
Englisch
Indoor Point-to-Point Navigation with Deep Reinforcement Learning and Ultra-Wideband
BASE | 2021
|Vision-Based Deep Reinforcement Learning of UAV Autonomous Navigation Using Privileged Information
ArXiv | 2024
|ArXiv | 2022
|Automatic navigation using deep reinforcement learning
Europäisches Patentamt | 2023
|