Synthetic-to-Real Pipeline for Safe Landing Zone Detection

2026-06-08

AI Summary

This work addresses the challenge of enabling safe autonomous landing for unmanned aerial vehicles in non-cooperative, unstructured environments where real annotated data are scarce. To circumvent the need for ground-truth labels, the authors propose an end-to-end perception framework trained exclusively on synthetic imagery generated via a procedural rendering engine combined with domain randomization to produce semantically annotated urban scenes. The model leverages the OneFormer Transformer architecture for semantic segmentation and integrates Euclidean distance transforms with dynamic reasoning logic to accurately identify the largest safe landing zone. Evaluated on the UAVid dataset, the approach demonstrates strong segmentation performance and successfully locates collision-free landing sites in previously unseen real-world drone footage, confirming its generalization capability and feasibility for edge deployment.

Abstract

As Uncrewed Aerial Vehicles (UAVs) transition toward higher levels of autonomy, the ability to perform unassisted recovery in non-cooperative, unstructured environments becomes critical. Achieving safe autonomous landing requires high-fidelity semantic resolution to distinguish navigable terrain from hazardous obstacles, yet development is often hindered by the scarcity of annotated aerial datasets. This work proposes a comprehensive perception and data generation pipeline designed to bridge the sim-to-real gap for autonomous landing tasks. We introduce a procedural synthetic data engine that generates photorealistic urban environments with automated semantic annotations through domain randomization. A Transformer-based OneFormer architecture is fine-tuned exclusively on this synthetic data, leveraging multi-head self-attention mechanisms for global context resolution. To ensure operational safety, a deterministic landing module utilizes a Euclidean Distance Transform (EDT) and dynamic inference logic to identify the largest inscribed safe landing zones while maintaining strict clearance buffers around obstacles. Quantitative benchmarking against the UAVid dataset demonstrates robust semantic segmentation performance, while qualitative validation on real-world UAV footage confirms the system's ability to identify collision-free landing sites in unseen environments. Our results highlight the potential of high-fidelity procedural simulation to eliminate the need for manual annotation while providing robust, edge-deployable situational awareness for autonomous UAV recovery.

Problem

• Safe Landing Zone Detection
• Synthetic-to-Real
• Autonomous UAV
• Semantic Segmentation
• Unstructured Environments

Innovation

• synthetic-to-real
• procedural generation
• OneFormer
• Euclidean Distance Transform
• domain randomization