3D Perception for Safe-Gap-Aware Autonomous Drone Navigation

Flying a drone through narrow gaps, as in racing or complex environments, poses significant planning challenges. Trajectories must account for gap geometry (shape, size, orientation) for proper drone reorientation. Existing methods often fail to accurately estimate gap size. Moreover, the time required for gap size estimation is relatively high. These limitations prevent small and mini drones, which have limited onboard computing resources, from autonomously navigating through narrow gaps while maintaining steady velocity. To address this challenging problem, we present a model capable of performing the following tasks on the fly: a) estimate the gap shapes, b) compute their sizes, and c) build a 3D model of the obstacle. The proposed system has been evaluated both in simulation and in real-world environment containing narrow gaps. The obtained results indicate that the proposed approach is up to 74.86% more accurate than YOLOv8 in estimating gap size and achieves up to 62.49% faster execution time in real-world settings.