Detecting Runways in Aerial Images

Andres Huertas, William Cole, Ramakant NevatiaAAAI 1987

Abstract

Automated airport mapping from aerial imagery is a complex task due to the inherent noise in low-level image segmentation and the presence of occluding objects. This paper presents a knowledge-based system that identifies runways by combining geometric feature extraction with semantic verification against FAA standards. By forming anti-parallel line pairs (apars) and utilizing orientation histograms to focus computational resources, the system effectively distinguishes runways from other linear features like roads. Evaluation on Logan and JFK International Airports demonstrates the system's robustness in handling fragmented data and surface variations.

Introduction

In the late 1980s, automated photointerpretation emerged as a critical field for infrastructure mapping. However, the transition from raw pixels to meaningful geographic regions remained a significant challenge. Runways, while geometrically simple as long rectangles, are often obscured by aircraft, fragmented by maintenance patches, or confused with taxiways and highways. This research from the University of Southern California (USC) addresses these challenges by moving beyond pure shape detection toward a "hypothesize and test" system grounded in domain-specific knowledge.

Methodology

The system operates through a multi-stage pipeline designed to handle the messiness of real-world aerial photography.

Linear Feature Extraction and Apar Formation

The process begins with the USC "LINEAR" system, which extracts line segments based on intensity edges. From these segments, the system identifies "apars"—pairs of parallel lines with opposite contrast, representing the boundaries of a runway against its surroundings.

Search Space Reduction

To manage the thousands of potential line pairs, the system employs a "focus of attention" mechanism. It generates length-weighted histograms of orientations and widths to identify dominant patterns likely to represent runways. This step allows the system to discard up to 94% of irrelevant data early in the process.

Histogram analysis of line orientations and widths used to prioritize runway candidates.

Joining, Filtering, and Semantic Verification

Fragmented apars are joined based on collinearity and "gap texture" analysis, which checks if the area between segments matches the expected runway orientation. The final verification step—the system's core innovation—checks runway hypotheses against FAA-regulated markings, such as centerlines, threshold marks, and touchdown zones. A hypothesis is only confirmed if these markings appear in the correct geometric relationship.

Results

The system was tested on two distinct datasets: Logan International Airport (Boston) and John F. Kennedy International Airport (New York).

At Logan Airport, which provided high-contrast imagery, the system successfully reduced nearly 9,500 initial apars to approximately 500 candidates. It then accurately identified the primary runways and their associated markings.

Final detection results for Logan International Airport, showing identified runways and markings.

At JFK Airport, the system faced more significant challenges, including repair patches and varying surface materials. Despite this "surface noise," the system successfully located runway ends by identifying threshold markings, proving that semantic knowledge can compensate for poor image quality.

Conclusion

By incorporating FAA specifications into the computer vision pipeline, the USC researchers transformed a difficult segmentation problem into a manageable knowledge-based reasoning task. While future work could benefit from 3D data and integrated detection of surrounding buildings, this system established a robust framework for automated infrastructure mapping in complex cultural environments.

Summary updated

BibTeX
@inproceedings{huertas1987detecting,
  author = {Andres Huertas and William Cole and Ramakant Nevatia},
  title = {{Detecting Runways in Aerial Images}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1987},
}
Figure