Real-Time Two-Stage Screening Framework for UAV Autonomous Landing in Unknown Environments

Abstract

This study addresses the challenge of autonomous landing of unmanned aerial vehicles (UAVs) in unknown environments by proposing a vision-based system that utilizes a single RGB-D camera. The core of our approach integrates an enhanced real-time semantic segmentation model based on the DDRNet architecture, incorporating structural re-parameterization (RDBlock), effi cient context aggregation (Fast-DAPPM), and boundary optimization (BOM). The model achieves a 97% inference speed improvement (65 FPS vs. 33 FPS for the baseline) while increasing mean intersection over union (mIoU) by 1.04 percentage points on AeroScapes. It also achieves 73.23% mIoU on the combined Drone-Landing dataset. The landing framework combines semantic-based candidate selection with depth-based geometric verifi cation in a two-stage process. A safety region defi ned by the UAV’s physical dimensions is updated in image coordinates using depth measurements, while a four-subdomain fl atness assessment evaluates local terrain conditions. During descent, potentially dynamic obstacles entering the selected landing region trigger hovering and, if occupancy persists, a climb followed by landing-region reassessment. Flight experiments demonstrate the feasibility of the system in representative static multitarget, dynamic-intrusion, and gully-terrain scenarios. Implemented on edge computing platforms, our solution provides a computationally efficient approach for autonomous UAV landing, integrating semantic risk screening, geometric verifi cation, and obstacle monitoring within a unifi ed perception and control framework.