Keyline Detection with YOLO: Runway Perception for Vision-Based Aircraft Landing

Leander Fritsch, David Rügamer, Alexej KlushynarXiv 2026 · arXiv

TL;DR

This paper introduces object keyline detection, a novel task that extends object detectors like YOLO to predict a fixed, semantically ordered set of infinite lines (keylines) for each object instance. By treating lines as first-class geometric features alongside keypoints, the authors significantly improve the robustness of vision-based aircraft landing systems, ensuring reliable perception even when runway corners are truncated or occluded.

What Problem Does This Paper Address?

In safety-critical applications like autonomous aircraft landing, precise pose estimation is vital. Traditionally, this relies on detecting keypoints (like runway corners). However, keypoints are fragile: if a corner is cut off by the camera's field of view or hidden by a cloud, the geometric constraint is lost.

While general line-segment detectors exist, they typically produce an unordered "soup" of lines across the entire image. They don't know which line belongs to which object, nor do they assign semantic meaning (e.g., "this is the left edge of runway 2"). This paper bridges the gap between object detection and line extraction, providing a "line-level analog" to semantic keypoints that remains observable even when endpoints are missing.

Figure 1. A partially visible runway during final approach. The two near corners are truncated by the image boundary, preventing keypoint-based pose estimation, while four of five runway keylines remain observable, providing line constraints for pose estimation.

Methods

The authors extend the YOLO-NAS architecture to support keyline prediction through several key innovations:

  • Continuous Hesse-Normal Parameterization: Instead of predicting a single, unstable angle $\theta$, the model predicts a signed distance $\rho$ and a unit normal vector $[\cos \theta, \sin \theta]$. This representation is continuous and avoids the "wrap-around" discontinuity (where $0^\circ$ and $360^\circ$ are numerically far but geometrically identical) that plagues standard angle regression.
  • Object Keyline Similarity (OKLS): Inspired by the standard Object Keypoint Similarity (OKS) used in pose estimation, the authors propose OKLS. It measures the agreement between a predicted infinite line and ground-truth support endpoints. This serves as a unified evaluation metric, a target assignment signal during training, and a regression loss.
  • Architectural Integration: The authors compared "Split" vs. "Fused" prediction heads. In the Fused variant, keypoints and keylines share the same geometric branch, which proved to be highly efficient, adding almost no latency compared to a keypoint-only model.
Figure 4. Head architecture variants for joint keypoint and keyline prediction: (a) Split: uses a dedicated keyline branch; (b) Fused: shares a single geometric branch across tasks.

Datasets

The research primarily uses a curated proprietary dataset of real aircraft approach images.

  • Training set: 6,730 images from 230 trajectories across 95 airports.
  • Validation/Test sets: ~700 images each, from airports strictly disjoint from the training set to prevent geographic data leakage.
  • The dataset includes annotations for 4 runway corners (keypoints) and 5 semantic lines (left, top, right, bottom edges, and the centerline).

The authors also benchmarked against general-purpose line detectors like M-LSD and LINEA using an "oracle matching" protocol to provide a fair comparison on their semantic task.

Key Insights and Ideas

  • Lines are more robust than points: The "aha moment" is that while a keypoint provides zero information if occluded, a partially visible edge still provides a strong geometric constraint on orientation and perspective.
  • Semantic ordering matters: By assigning a fixed index to each line (e.g., index 0 is always the left edge), the model can learn specific visual priors for each feature, making it much more accurate than general line detectors that treat all lines as equal.
  • The "Lever-Arm" Effect: OKLS is superior to standard coordinate losses because it naturally accounts for the fact that small angular errors matter more for long, distant lines. It penalizes errors based on how much they displace the line relative to its physical support.
Figure 3. Our proposed Object Keyline Similarity (OKLS) uses the orthogonal distances p1 and p2 from ground-truth support end- points (e1, e2) to the predicted line ˆl as its residual.

Key Results

  • Superior Parameterization: The continuous Hesse-normal form achieved the highest accuracy (55.7 $AP_{KL}$) compared to point-based or polar-angle alternatives.
  • Improved Constraint Availability: For pose estimation (PnPL), having at least three features is a requirement. Using keypoints alone, this was possible in only 71.3% of test images. Adding keylines boosted this "availability" to 99.3%.
  • Efficiency: The Fused head architecture added keyline capabilities with a negligible latency increase of only 0.1 ms (on an NVIDIA Tesla T4), making it practical for real-time flight systems.
  • Outperforming Baselines: Even when giving general line detectors (like LINEA) the advantage of "oracle matching," the proposed instance-centric approach achieved significantly higher Average Precision.
Figure 5. Keyline predictions on a two-runway scene (San Francisco International Airport). The zoomed view (right) shows all five predicted keylines per runway with accurate geometric alignment, illustrating the multi-instance prediction capabilities of our method.

Limitations and Future Work

While highly effective for runways, the authors note several areas for further exploration:

  • Scope: The method assumes objects have a fixed, semantically ordered set of linear features. It may struggle with deformable objects or categories where the number of edges varies.
  • Downstream Evaluation: The study evaluates performance in the image domain ($AP$). The next logical step is integrating these keylines into a full 6-DoF pose estimation pipeline to measure the actual reduction in aircraft positioning error.
  • Geometric Constraints: Currently, the model predicts lines independently. Future work could enforce global constraints, such as ensuring parallel runway edges share a common vanishing point.

Summary updated

Abstract

Modern object detectors can predict keypoints as per-object geometric features, but keypoints are inherently vulnerable to occlusion and truncation. Object edges can provide complementary information, often remaining observable even when partially occluded or when their defining endpoints leave the field of view. We therefore introduce object keyline detection, the task of predicting a fixed, semantically ordered set of infinite lines per object instance. Unlike the outputs of general line-segment detectors, each keyline is associated with an object instance and identified by a semantic index, providing a line-level analog of keypoints. To solve this task, we extend YOLO with keyline prediction parameterized in continuous Hesse-normal form $[ρ,\cosθ,\sinθ]$, and introduce Object Keyline Similarity (OKLS), which serves as an evaluation metric, target assignment signal, and regression loss. We instantiate this formulation on runway perception in vision-based landing, a motivating example containing semantic linear features that remain partially observable throughout the approach. Experiments show that our representation achieves the highest detection accuracy among four evaluated line parameterizations, and keylines integrate directly into standard YOLO heads as additional output channels without degrading keypoint performance or adding meaningful latency.
BibTeX
@misc{fritsch2026keyline,
  author = {Leander Fritsch and David Rügamer and Alexej Klushyn},
  title = {{Keyline Detection with YOLO: Runway Perception for Vision-Based Aircraft Landing}},
  year = {2026},
  eprint = {2611.09316},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.09316},
}
Figure