Uncertainty-Aware Vision-Based Landing-Site Perception for Autonomous UAV Landing in Urban Environments

Abstract

Autonomous UAV landing in urban scenes requires a perception module that can identify candidate landing surfaces, reject structural and dynamic hazards, and express uncertainty before a downstream controller commits to a landing maneuver. This study reformulates UAV landing perception as a unified three-class landing-safety segmentation problem by relabeling UAVid, UDD6, and VDD into candidate landing area, structural obstacle, and critical hazard classes. The Landing3 dataset is constructed with an implementation-consistent relabeling protocol in which critical hazards override other labels and only sufficiently large connected components of source-specific candidate classes are retained as landing candidates. Two real-time segmentation models are evaluated under this unified task: PIDNet, a CNN-based multi-branch model with explicit boundary modeling, and SCTNet, a Transformer-guided model with semantic alignment for long-range context modeling. A post hoc conformal prediction (CP) module then converts softmax outputs into pixel-wise prediction sets, and the final candidate landing region is extracted only from pixels whose prediction set is the singleton candidate-landing class. Experiments compare the two models in terms of best validation mIoU, class-wise IoU, resolution-dependent accuracy–speed trade-offs, power-constrained FPS and complexity, qualitative candidate-area visualization, and CP-derived safe-area quality. A fixed-checkpoint sensitivity analysis further shows that the comparative model ranking remains stable across the tested connected-component threshold settings. SCTNet provides a lighter model and higher throughput under most tested resolutions and power limits, whereas PIDNet preserves higher safe-area recall, safe IoU, and spatial coherence after CP filtering. These results show that reliable UAV landing perception requires joint consideration of cross-dataset task definition, real-time model efficiency, and uncertainty-aware candidate-area extraction, rather than semantic segmentation accuracy alone.