Optimized Autonomous Drone Navigation Using Double Deep Q-Learning for Enhanced Real-Time 3D Image Capture

Abstract

The proposed system assists in the automatic creation of three-dimensional (3D) meshes for all types of objects, buildings, or scenarios, using drones with monocular RGB cameras. All these targets are large and located outdoors, which makes the use of drones for their capture possible. There are photogrammetry tools on the market for the creation of 2D and 3D models using drones, but this process is not fully automated, in contrast to the system proposed in this work, and it is performed manually with a previously defined flight plan and after manual processing of the captured images. The proposed system works as follows: after the region to be modeled is indicated, it starts the image capture process. This process takes place automatically, with the device always deciding the optimal route and the framing to be followed to capture all the angles and details. To achieve this, it is trained using the artificial intelligence technique of Double Deep Q-Learning Networks (reinforcement learning) to obtain a complete 3D mesh of the target.