01 / AI · Analysis
A YOLO extension reads runway lines for vision-based landing
The design backs up keypoints, which are prone to occlusion, with straight lines, but the abstract reports no real flight-test results.

An extension of the YOLO object detector that predicts a set of straight lines per object appeared on arXiv on 9 Nov 2026, with a comparison of four line representations. The paper, which names the task object keyline detection, said it applied the method to runway perception in vision-based aircraft landing.
The starting point is a weakness in keypoints, the coordinates of specific points on an object. The researchers wrote that keypoints are inherently vulnerable to being hidden or cut off at the edge of the frame. An object's edges, they explained, often remain visible even when partly hidden or when their endpoints leave the field of view.
The researchers gave runway perception for vision-based landing as the motivating case. A runway has straight features that stay partly observable throughout the approach. The paper's answer is to read those lines object by object.
Straight lines tied to objects and numbered in order
The paper defines keylines as a set of infinite lines predicted in a fixed number and order for each object. Unlike the output of general line-segment detectors, each keyline is tied to a specific object and identified by a semantic index. The result carries the keypoint idea over to lines.
The researchers extended YOLO to predict lines in continuous Hesse normal form, [ρ, cosθ, sinθ]. Hesse normal form describes a line by its distance from the origin and the direction perpendicular to the line.
The paper also proposed Object Keyline Similarity (OKLS). The value served as an evaluation metric, a target assignment signal and a regression loss that reduces position error.
The highest accuracy among four line representations
In results published by the researchers, this representation achieved "the highest detection accuracy" among four line representations. The abstract does not give the names or accuracy values of the other representations.
According to the researchers' published results, keylines integrated directly into standard YOLO heads, the final prediction stage of the detector, as output channels. They wrote that this neither degraded keypoint performance nor added meaningful latency.
The abstract contains no results from real aircraft flight tests. The paper's proposal is to describe objects with lines as well as points, and the application it names is runway perception.