Nordic Vehicle Dataset (NVD): Performance of vehicle detectors using newly captured NVD from UAV in different snowy weather conditions

Hamam Mokayed, Amirhossein Nayebiastaneh, Kanjar De, Stergios Sozos, Olle Hagner, Björn BackeWorkshops @ CVPR 2023 · DOI · arXiv

Introduction

Researchers have introduced the Nordic Vehicle Dataset (NVD), a first-of-its-kind collection of Unmanned Aerial Vehicle (UAV) imagery specifically captured in the extreme snowy conditions of Northern Scandinavia. By testing state-of-the-art detectors like YOLOv8 on this data, the study reveals a significant performance gap: standard models struggle to "see" vehicles when they are partially buried or obscured by real-world snow, highlighting a critical need for weather-specific training data in autonomous monitoring and search-and-rescue.

The Challenge of Arctic Detection

While vehicle detection from drones has advanced significantly, most existing datasets are "fair-weather" collections consisting of images taken under clear skies or in temperate climates. In the Arctic, drones face "Polar nights" (where the sun never rises), heavy snowfall, and temperatures far below freezing.

The core challenge is that snow is not merely a visual filter; it physically obscures the visual features—such as wheels, windows, and sharp edges—that computer vision models rely on to identify a vehicle. Existing models trained on "clean" data or even artificially generated snow often fail when faced with a car that appears as a white mound against a white background. This research addresses the lack of high-quality, real-world UAV data for these extreme conditions.

Methodology

The study employed a two-pronged approach involving specialized data collection and benchmark testing.

Hardware and Data Acquisition

The researchers utilized a "Freya" flying wing UAV equipped with an Exmor RS CMOS sensor. This platform captured 1080p and 4K footage at altitudes ranging between 120m and 250m.

The 'Freya' flying wing UAV used for capturing the Nordic Vehicle Dataset.

Flight Height Estimation

A notable technical contribution is the method for estimating flight altitude when telemetry data is unavailable. The authors used perspective geometry and the diagonal length of vehicle bounding boxes to mathematically derive the drone's height.

A vehicle in snowy conditions with a bounding box diagonal used for flight height estimation.

Data Augmentation and Benchmarking

To improve model robustness, the Albumentations library was used to simulate various weather effects (fog, rain, snow) on the training set, alongside YOLO’s built-in online augmentation. The researchers evaluated three prominent object detection models: YOLOv5s, YOLOv8s, and Faster R-CNN.

The Nordic Vehicle Dataset (NVD)

The NVD is the primary contribution of this work, containing 22 videos (8,450 frames) and 26,313 annotated vehicles. The data covers a spectrum of winter conditions, including overcast skies with snowfall, low light, high-contrast sunlight on fresh snow, and patchy snow cover. Unlike many datasets that only annotate cars on paved streets, NVD includes vehicles in generic locations like parking lots and fields, which is vital for search-and-rescue missions.

Key Insights and Results

The results demonstrate that detecting vehicles in the Nordic winter is a significant challenge for modern AI:

  • Real vs. Synthetic Snow: Models trained on artificially generated snow are often ineffective because real snow hides structural features in ways that simple digital overlays cannot replicate.
  • Precision-Recall Tradeoff: In snowy conditions, models tend to have much higher precision than recall. This indicates that while the model's positive detections are usually accurate, it misses a large percentage of vehicles that are heavily covered in snow.
  • Benchmark Performance: The baseline YOLOv8s achieved a precision of 65.8% but a very low recall of 22.4%. Adding data augmentation (YOLOv8s_Au) boosted precision to 77.1% and recall to 34.6%*. While improved, the model still misses nearly two-thirds of the vehicles.
  • Comparison: Faster R-CNN performed moderately with an Average Precision (AP) of 24.4%. In contrast, these models typically achieve APs in the 70-80% range on "easier" datasets like DLR 3k or Stanford Drone, proving that NVD is a much more difficult benchmark.

Conclusion and Future Work

While augmentation provides some benefits, it is not a complete solution for the visibility issues caused by heavy snow. Future research directions include integrating "desnowing" or image restoration pre-processing steps before the detection phase, expanding the dataset to include harsher lighting conditions like Polar night, and optimizing models for real-time performance on edge hardware.

Summary updated

BibTeX
@inproceedings{mokayed2023nordic,
  author = {Hamam Mokayed and Amirhossein Nayebiastaneh and Kanjar De and Stergios Sozos and Olle Hagner and Björn Backe},
  title = {{Nordic Vehicle Dataset (NVD): Performance of vehicle detectors using newly captured NVD from UAV in different snowy weather conditions}},
  booktitle = {Workshops @ CVPR},
  year = {2023},
  doi = {10.1109/cvprw59228.2023.00560},
}
Figure