VisLanding: Monocular 3D Perception for UAV Safe Landing via Depth-Normal Synergy

Zhuoyue Tan, Boyong He, Yuxiang Ji, Liaoni Wu

TL;DR

VisLanding is a monocular 3D perception framework that enables UAVs to identify safe landing zones (SLZ) using only a single camera. By integrating depth and surface normal information into a dual-flow optimization process, it achieves significantly better robustness and "zero-shot" generalization across different environments compared to traditional vision-based methods.

What Problem Does This Paper Address?

Autonomous landing is a critical yet high-risk task for Unmanned Aerial Vehicles (UAVs), especially in unknown or complex terrain where pre-defined markers aren't available. To land safely, a UAV must identify areas that are flat, obstacle-free, and sufficiently large.

Current solutions generally fall into two camps:

  1. Multi-sensor fusion: Using LiDAR or RGB-D cameras. These are accurate but expensive, heavy, and power-hungry.
  2. Pure vision-based methods: These use standard cameras (monocular vision) but often struggle with "scale ambiguity" (not knowing how far away or how large an object actually is) and lack geometric reliability.

The authors identify a major gap: standard semantic segmentation models (which just label pixels as "grass" or "building") tend to overfit to specific textures and fail when moved to a new environment (the "domain gap"). VisLanding aims to solve this by using 3D geometric priors—depth and surface normals—to guide the identification of safe zones.

Methods

VisLanding builds upon Metric3D V2, a state-of-the-art foundational model for monocular 3D perception. The core innovation is a dual-flow optimization framework:

  • Depth-Normal Flow: This branch inherits Metric3D’s ability to jointly optimize depth and surface normals in a "canonical camera space," which helps resolve scale ambiguity.
  • Safe Landing Zone (SLZ) Flow: The authors added a dedicated branch that treats landing zone identification as a binary semantic segmentation problem (Safe vs. Unsafe). This branch doesn't just look at RGB pixels; it performs "geometric fusion," using the updated depth and normal information from the first branch as auxiliary features.
  • Iterative Refinement: The model uses a Dense Prediction Transformer (DPT) with iterative optimization steps ($T$). At each step, the model refines its understanding of the 3D geometry and the safety mask simultaneously.
  • Area Estimation: By combining the predicted depth, surface normals, and the camera's intrinsic parameters, the system can project pixels into 3D space to calculate the actual physical area (in square meters) of a candidate landing site in real-time.

Datasets

The researchers utilized two primary datasets to prove their model's effectiveness:

  • WildUAV: A real-world dataset captured from a UAV perspective. The authors manually annotated 1,329 images for safe landing zones based on "flat and obstacle-free" criteria.
  • Semantic Drone Dataset: Used specifically for cross-domain evaluation. The model was trained on WildUAV and then tested on this dataset (which features urban settings) without any further training to see how well it generalizes to new environments.

Key Insights and Ideas

  • Geometry over Texture: Traditional models like SegFormer rely heavily on visual textures. When the "look" of the ground changes, they fail. VisLanding relies on the geometry (is it flat? is it a continuous surface?). Because physics stays the same across domains, the model is much more robust.
  • The Power of Foundation Models: By building on Metric3D V2, the authors leverage "zero-shot" capabilities. The model already "understands" 3D space from being pre-trained on massive datasets, which the authors then specialize for the UAV landing task.
  • Safety Redundancy: In real-flight tests, the system showed "conservative estimation" at long distances—meaning it underestimated the available area. This provides a built-in safety margin for autonomous decision-making.

Key Results

  • Superior Generalization: In cross-domain tests (moving from wild to urban environments), standard models like SegFormer saw their performance (mIoU) plummet by 64%. VisLanding’s performance only dropped by 33%, maintaining a much higher Mean Intersection over Union (52.44% vs. ~32-35% for competitors).
  • Qualitative Coherence: VisLanding produces clean, continuous masks of safe zones, whereas baseline models often produced "fragmented" or "noisy" outputs when faced with unfamiliar terrain.
  • Real-World Accuracy: In flight tests, the area estimation became increasingly accurate as the drone approached the target, successfully identifying landing zones of approximately 34m².

Limitations and Future Work

  • Computational Overhead: The model uses a Vision Transformer (ViT) backbone, which is computationally "heavy." Currently, it runs at less than 5 FPS on an NVIDIA Jetson Orin Nano. This is too slow for high-speed autonomous flight.
  • Dataset Scale: While the results are promising, the authors acknowledge the need for even larger, more diverse datasets to further validate performance.
  • Next Steps: The team plans to explore model distillation (shrinking the model while keeping the accuracy) and transfer learning to make the system fast enough for real-time onboard processing on low-power drone hardware.

Summary updated

Abstract

This paper presents VisLanding, a monocular 3D perception-based framework for safe UAV (Unmanned Aerial Vehicle) landing. Addressing the core challenge of autonomous UAV landing in complex and unknown environments, this study innovatively leverages the depth-normal synergy prediction capabilities of the Metric3D V2 model to construct an end-to-end safe landing zones (SLZ) estimation framework. By introducing a safe zone segmentation branch, we transform the landing zone estimation task into a binary semantic segmentation problem. The model is fine-tuned and annotated using the WildUAV dataset from a UAV perspective, while a cross-domain evaluation dataset is constructed to validate the model's robustness. Experimental results demonstrate that VisLanding significantly enhances the accuracy of safe zone identification through a depth-normal joint optimization mechanism, while retaining the zero-shot generalization advantages of Metric3D V2. The proposed method exhibits superior generalization and robustness in cross-domain testing compared to other approaches. Furthermore, it enables the estimation of landing zone area by integrating predicted depth and normal information, providing critical decision-making support for practical applications.
BibTeX
@inproceedings{tan2025vislanding,
  author = {Zhuoyue Tan and Boyong He and Yuxiang Ji and Liaoni Wu},
  title = {{VisLanding: Monocular 3D Perception for UAV Safe Landing via Depth-Normal Synergy}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2025},
  doi = {10.1109/iros60139.2025.11247278},
}

Loading related content...

Figure