Autonomous Navigation Algorithm of Monocular UAV Based on Depth Estimation and Robust VIO

The autonomous navigation of micro-monocular UAVs has emerged as a research focus, leveraging their compact size, lightweight design, and energy efficiency. However, the inherent scale ambiguity of monocular systems hinders the direct acquisition of dense, true-scale depth observations, posing collision risks. This paper introduces an innovative monocular UAV navigation system that integrates Visual Inertial Odometry (VIO) with a depth estimation network. The system leverages VIO for precise UAV positioning and employs the sparse, true-scale point cloud from VIO to calibrate the scale of monocular depth estimations, yielding a depth map with consistent scale continuity for downstream planning tasks. Furthermore, the depth estimations are utilized to refine VIO initialization, enhancing the algorithm's initialization speed and robustness. Experimental results demonstrate a 27.7% increase in VIO initialization velocity and a modest improvement in positioning accuracy. Additionally, our approach achieves a significant reduction in obstacle collision rates (34.9% decrease) and an enhanced flight mission completion rate (21.4% increase), outperforming current state-of-the-art end-to-end autonomous UAV navigation algorithms.