⚡ electrical engineering

A Task-Driven Evaluation of UAV Detection and Tracking under Synthetic Fog

This paper presents a task-driven evaluation framework demonstrating that while synthetic fog severely degrades UAV detection and tracking, fog-inclusive training offers the most robust performance gains, whereas test-time image restoration yields significant benefits primarily when detectors are trained solely on clean data.

Original authors: Amir Pouladi, Vesal Ahsani, Haijun Li, Homayoun Najjaran, Afzal Suleman

Published 2026-07-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Amir Pouladi, Vesal Ahsani, Haijun Li, Homayoun Najjaran, Afzal Suleman

Original paper licensed under CC BY 4.0 (http://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

Imagine you are trying to spot a tiny, distant drone flying high in the sky. On a clear day, it's like looking for a speck of dust on a clean window; you can see it easily. But now, imagine thick fog rolls in. The window becomes a blurry, white sheet. The drone doesn't disappear, but it loses its edges and blends into the background, making it nearly impossible for a computer to find it.

This paper is like a lab experiment where researchers built a "fog machine" for computer vision to test how well different tools can find these drones when the weather turns bad. They didn't just look at how clear the picture looked; they asked a more practical question: "Does making the picture clearer actually help the computer find the drone?"

Here is how they did it and what they found, explained through simple analogies:

1. The Setup: Building a Fake Fog Machine

Real fog is hard to study because you can't control it perfectly, and it's dangerous to fly drones in it for testing. So, the researchers built a digital fog machine.

The Recipe: They took clear photos of drones and used a "depth map" (a digital guess of how far away everything is) to apply a mathematical formula. This formula adds fog that gets thicker the further away the object is, just like real fog.
The Result: They created a library of images ranging from "lightly misty" to "blindingly thick fog."

2. The Three Tools They Tested

The researchers tested three different strategies to help the computer see the drone:

Strategy A: The "Clean Room" Student. They trained a computer detector only on clear, sunny photos. Then, they threw foggy photos at it to see how it panicked.
Strategy B: The "All-Weather" Student. They trained the computer on a mix of clear photos and foggy photos. This student learned what fog looks like from day one.
Strategy C: The "De-hazing Glasses." They took the foggy photos and ran them through a special software (called a "dehazer") to try to scrub the fog off the image before the computer tried to find the drone.

3. The Big Discovery: Seeing vs. Finding

The most important finding is a bit counter-intuitive. Usually, we think: "If I clean up the picture, the computer will do a better job."

The paper says: Not necessarily.

The "Clean Room" Student: When this student faced fog, they failed miserably. They missed the drone constantly. However, when they put on the "De-hazing Glasses" (Strategy C), they got much better. The glasses helped them see the drone again.
The "All-Weather" Student: This student was already tough. They didn't need the glasses as much. In fact, sometimes the glasses made things slightly worse or didn't help at all. Why? Because the "De-hazing" software sometimes smudged the edges of the drone while trying to remove the fog, confusing the detector that was already used to foggy shapes.

The Analogy: Imagine you are trying to find a friend in a crowd.

If you are blindfolded (trained only on clear days), putting on glasses (dehazing) helps you see them.
But if you are already used to searching in a crowded, noisy room (trained on fog), putting on those same glasses might just distort your view and make it harder to recognize your friend's face.

4. The Tracking Problem: Losing the Trail

The researchers also tested "tracking"—keeping an eye on the drone as it moves.

The Chain Reaction: Tracking is like a relay race. The "Detector" is the runner who spots the baton (the drone). The "Tracker" is the person who keeps running with it.
The Result: If the fog is so thick that the Detector misses the drone (drops the baton), the Tracker has nothing to hold onto. No matter how good the Tracker is, they can't follow a ghost.
The Lesson: The biggest factor in successful tracking wasn't the tracker's skill; it was whether the detector was trained to handle the fog in the first place.

5. The Bottom Line

The paper concludes with a few key takeaways for anyone building systems to find drones in bad weather:

Training is King: The best way to handle fog is to train your computer on foggy images from the start. It's like teaching a driver to drive in the rain rather than just teaching them on a sunny day and hoping they can figure it out later.
Cleaning isn't a Magic Wand: Just because an image looks clearer to a human eye doesn't mean the computer will find the object better. Sometimes, the process of cleaning the image creates new "artifacts" (glitches) that confuse the computer.
Missed Detections are the Enemy: Fog doesn't usually trick the computer into seeing things that aren't there (false alarms); it makes the computer miss things that are actually there. This is the main reason tracking fails.

In short, the paper argues that instead of just trying to "fix" the picture after the fact, we should teach our computers to be comfortable in the fog from the very beginning.

Technical Summary: A Task-Driven Evaluation of UAV Detection and Tracking under Synthetic Fog

Problem Statement

Reliable detection and tracking of small Unmanned Aerial Vehicles (UAVs) in open airspace are critical for situational awareness and sense-and-avoid systems. However, fog severely degrades performance in sky-dominant, long-range imagery by reducing target-background contrast and suppressing high-frequency details essential for recognizing small, distant objects. While single-image dehazing has advanced significantly, existing benchmarks often fail to demonstrate that improvements in image-level restoration metrics (e.g., PSNR, SSIM) translate to proportional gains in downstream perception tasks. Furthermore, there is a lack of systematic, task-level studies connecting depth-aware fog synthesis, image restoration, and UAV detection/tracking within a controlled protocol, particularly for scenarios where fog effects are strongly depth-driven.

Methodology

The authors propose a unified, task-driven evaluation pipeline comprising three stages: training, detection inference, and tracking inference.

1. Synthetic Fog Generation

To address the scarcity of annotated foggy UAV data, the study employs a depth-aware atmospheric scattering model (ASM) to generate synthetic fog from real clear-weather images.

Depth Estimation: Monocular depth is estimated using MiDaS to provide a dense relative depth representation.
Normalization: Raw depth maps are normalized using dataset-specific percentile bounds to ensure consistent scaling and then inverted so that larger values correspond to farther points.
Fog Synthesis: The foggy image $I(x)$ is synthesized using the equation $I(x) = J(x)t(x) + A(1 - t(x))$ , where $J(x)$ is the clean radiance, $A$ is global atmospheric light, and $t(x) = \exp(-\beta d(x))$ is the transmission map. The parameter $\beta$ controls fog severity, varied across five levels ($0.4, 0.8, 1.6, 2.4, 3.6$) to simulate light to severe degradation.

2. Image Restoration Evaluation

Representative restoration methods from three families were evaluated on the MMAUD dataset to select the best model for downstream tasks:

Classical: Dark Channel Prior (DCP) and Color Attenuation Prior (CAP).
CNN-based: FFA-Net.
Transformer-based: DehazeFormer.
DehazeFormer was selected based on superior quantitative (PSNR, SSIM) and qualitative performance in sky-dominant imagery.

3. Detection and Training Regimes

Experiments utilized the YOLO11 family (n, s, m) detectors trained on the CfAR dataset. Two primary training strategies were compared:

Clean-Only Training: Detectors trained exclusively on clear images.
Fog-Inclusive Training: Detectors trained on a mix of clean and fog-degraded images at varying ratios (30%, 50%, 70%, 100% fog).
Detectors were evaluated on clean, foggy, and restored test sets.

4. Tracking-by-Detection

Tracking performance was assessed using ByteTrack and BoT-SORT on the DUT Anti-UAV dataset. The study evaluated tracking on clean, foggy, and restored video sequences, measuring performance via Multiple Object Tracking Accuracy (MOTA) and Identity F1 Score (IDF1).

Key Contributions

Task-Driven Framework: A unified pipeline linking depth-aware fog synthesis, image restoration, object detection, and tracking, prioritizing downstream perception performance over isolated image restoration quality.
Depth-Aware Fog Synthesis: A pipeline for generating controlled fog-degraded variants of real UAV imagery using monocular depth estimation and the ASM, with generated datasets made publicly available.
Comparative Analysis: A rigorous experimental study analyzing how weather degradation and restoration affect detection robustness and temporal tracking stability, specifically comparing test-time restoration against fog-inclusive training.

Results

Impact of Fog: Fog substantially degrades both detection and tracking, primarily by increasing missed detections (False Negatives) rather than false positives. Recall drops sharply as fog severity increases, leading to significant declines in mAP and tracking metrics (MOTA, IDF1).
Restoration vs. Training:
- Clean-Trained Detectors: Test-time restoration provides consistent improvements, partially recovering performance lost to fog. However, restored performance rarely reaches the clean baseline, indicating that restoration reduces but does not eliminate the domain gap.
- Fog-Inclusive Training: Training with foggy data provides the most consistent improvement in robustness. As the proportion of fog in training increases, the performance gap between foggy and restored inputs narrows significantly. In some cases, fog-inclusive training renders restoration unnecessary or even slightly less effective, suggesting that learning directly from degraded data is superior to relying on post-hoc restoration.
Tracking Performance: Tracking robustness is governed primarily by the underlying detector's robustness. When detectors fail due to missed detections, trackers cannot recover the target. Fog-inclusive detector training transfers directly to stronger tracking performance, while the choice between ByteTrack and BoT-SORT has a secondary, sequence-dependent effect.
Stability: Fog-inclusive training significantly reduces the coefficient of variation (CV) of mAP across fog severities, indicating more stable performance under changing visibility conditions compared to clean-only training.

Significance and Claims

The paper claims that improvements in image-level restoration quality do not necessarily translate into proportional gains in downstream perception. The study demonstrates that while restoration is beneficial for detectors trained only on clean data, fog-inclusive training is the more reliable route to achieving robustness in adverse weather.

The authors conclude that for practical deployment, a system should prioritize training detectors on degraded data to reduce reliance on restoration. Restoration may still serve as a useful test-time preprocessing step for clean-trained models or in human-in-the-loop scenarios where visual clarity aids operator awareness, but it should not be viewed as a universal solution. The findings emphasize that robustness should be evaluated jointly across detection and tracking metrics rather than in isolation. The study acknowledges its limitations to controlled synthetic fog evaluation and identifies future work in validating on real adverse-weather data and exploring multimodal perception.