💻 computer science

AMFD-YOLO: An Adaptive Multi-scale Feature Modulation Network for Small Object Detection in Airport Surface Surveillance

This paper proposes AMFD-YOLO, an enhanced YOLO11-based network incorporating a Multiscale Occlusion-aware Attention Module, a Feature Distribution Modulation activation function, and a P2 detection layer to significantly improve the accuracy of detecting small and occluded airport surface targets under complex environmental conditions.

Original authors: Tian Luan, Fan Yang, Huanxia Wei, Weijun Pan

Published 2026-11-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tian Luan, Fan Yang, Huanxia Wei, Weijun Pan

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the vast, moving landscape of an airport, safety depends on seeing everything, everywhere, all at once. From the massive jets taxiing to the tiny ground crew members guiding them, every object must be tracked to prevent collisions and ensure smooth operations. For decades, airports have relied on radar and radio signals to monitor this activity, but these tools often struggle to identify exactly what an object is or to see the small details that matter most. To fill this gap, engineers have turned to computer vision, teaching cameras to "see" and recognize aircraft, vehicles, and people just as a human would. However, teaching a computer to spot a person from a high tower, where they might appear as just a few pixels on a screen, is incredibly difficult. The background is cluttered with complex textures, the weather changes constantly, and objects often hide behind one another, causing standard vision systems to miss them entirely.

A team of researchers from China and the United Kingdom has developed a new way to help computers see these small, hidden targets more clearly. They created a system called AMFD-YOLO, which is built upon a powerful, modern detection framework known as YOLO11. The core problem they tackled is that standard systems often lose the fine details of small objects as they process an image, much like how a blurry photo makes it hard to read a small sign. To fix this, the researchers added three specific upgrades to the system's "brain." First, they introduced a mechanism that helps the computer pay extra attention to parts of an image that are blocked or hidden, allowing it to guess the full shape of an object even when only a piece is visible. Second, they adjusted how the system handles light and shadow, making it more adaptable to the harsh glare of a sunny day or the dim light of a rainy night. Finally, they added a dedicated layer to the system specifically designed to look at the sharpest, most detailed parts of the image, ensuring that tiny targets like a single person are not overlooked.

When the team tested this new system on a dataset of real airport surveillance images, the results were striking. The improved system was able to detect small objects with far greater accuracy than the previous version. In particular, the ability to find people, which is notoriously difficult for computers, improved dramatically. The new system found people correctly about 47.5 percent of the time, a massive leap from the baseline system, which missed almost everyone. It also became much better at spotting vehicles and aircraft, even when they were crowded together or partially hidden. The researchers found that by combining these three upgrades, the system could handle the chaotic, busy environment of an airport apron much better than before, reducing the number of missed targets and false alarms.

The study highlights that while the new system is a significant step forward, it is not yet a perfect solution for every possible situation. The researchers noted that their testing was done on a specific set of images that, while diverse, did not cover every extreme weather condition or every possible camera angle found in the real world. Furthermore, the added layers that make the system so good at seeing small details also require more computing power, which could make it harder to run on the smaller, portable devices often used in the field. The team plans to gather more varied data and work on making the system faster and lighter so it can be deployed more widely. For now, however, this work demonstrates a clear path toward smarter, safer airport surveillance, proving that with the right adjustments, computers can learn to see the small things that keep the world moving.

Technical Summary: AMFD-YOLO for Airport Surface Surveillance

Problem Statement
Accurate detection of airport surface targets—including aircraft, ground vehicles (tow tractors, catering trucks, refueling trucks, baggage carts), and ground staff—is critical for intelligent surveillance and automated ground operations. However, existing computer vision methods face significant challenges in this domain. The primary difficulties include:

Small Object Detection: Targets often occupy very few pixels in wide-area surveillance images, leading to severe feature loss during network down-sampling.
Complex Environments: Surveillance is hindered by adverse weather (haze, rain, snow), nighttime conditions, uneven illumination, and complex, time-varying backgrounds.
Occlusion and Scale Variation: Frequent mutual occlusion and large variations in target scale (from massive aircraft to tiny personnel) degrade detection performance.
Background Clutter: Distinguishing targets from similar textures and crowded apron/taxiway scenes remains difficult for standard models.

Methodology
The authors propose AMFD-YOLO, an improved detection network based on the YOLO11 framework. The architecture integrates three specific enhancements to address the aforementioned challenges:

Multi-scale Occlusion-aware Attention Module (MOAM):
- Embedded within the detection head, this module is designed to enhance feature representation for occluded targets.
- It utilizes multiple scale-specific transformation branches with depthwise separable convolutions and residual connections to refine features.
- A lightweight channel attention mechanism employs an exponential calibration strategy to generate adaptive weights, strengthening salient foreground features while suppressing background interference.
C3k2 FDM Module (Feature Distribution Modulation):
- This module replaces the standard normalization and activation layers within the C3k2 backbone structure.
- It introduces an explicit modulation function: $FDM(x) = \gamma \odot \tanh(\alpha x) + \beta$ , where $\alpha$ , $\gamma$ , and $\beta$ are learnable parameters.
- This design allows for adaptive feature scaling and dynamic non-linear transformation, improving the model's sensitivity to distribution shifts caused by complex lighting and weather conditions.
P2 Detection Layer:
- To counteract the loss of fine-grained information in deep layers, a fourth detection head (P2) is added to the standard three-scale pyramid (P3, P4, P5).
- The P2 layer operates on high-resolution feature maps from the shallow network, preserving spatial details and contour information essential for detecting extremely small objects like ground staff.

Key Contributions

Novel Architecture: The integration of MOAM and C3k2 FDM into the YOLO11 framework specifically tailored for the unique constraints of airport surface surveillance.
Enhanced Small-Target Perception: The addition of the P2 detection layer significantly improves the detection of tiny targets that are typically missed by standard down-sampling approaches.
Robustness to Occlusion and Environment: The proposed modules demonstrate improved capability in handling mutual occlusion and varying environmental conditions (e.g., rain, night) compared to baseline models.

Experimental Results
Experiments were conducted on the ASS (Airport Surface Surveillance) dataset (specifically the ASS1 subset), which contains 2,000 images with 8,371 annotated bounding boxes across three classes: airplanes, personnel, and vehicles. The dataset includes diverse conditions such as sunny, rainy, and late-night scenarios.

Performance Gains: Compared to the baseline YOLO11, AMFD-YOLO achieved the following improvements:
- Precision: +7.5% (reaching 84.9%)
- Recall: +23.2% (reaching 76.0%)
- mAP@0.5: +23.6% (reaching 79.9%)
- mAP@0.5-0.95: +12.5% (reaching 56.0%)
Class-Specific Improvements: The most dramatic improvement was observed in the personnel category, where recall jumped from 0.48% to 47.5% and mAP@0.5 increased from 1.11% to 55.0%. This addresses a critical failure point in baseline models where small human targets were almost entirely undetected.
Comparative Analysis: AMFD-YOLO outperformed other YOLO series variants (including YOLOv5, v6, v8, v9t, v10n, v11, v12, and v26) across all metrics, particularly in high IoU thresholds and small-target detection.

Significance and Claims
The paper claims that AMFD-YOLO provides an effective technical solution for high-precision, real-time airport surface surveillance. By synergistically addressing feature loss, occlusion, and environmental variability, the method demonstrates superior reliability in visually crowded and adverse conditions.

The authors acknowledge limitations, noting that the current dataset has fixed acquisition angles and lacks sufficient extreme weather samples. Furthermore, while the P2 layer improves accuracy, it increases computational load, requiring further verification for edge device deployment. Future work is planned to focus on model lightweighting, multi-object tracking integration, and expanding the dataset to cover more diverse operational scenarios.

More like this