Crawling 2021 !!link!! — Fu10 The Galician Night

Paper Title: FU10 The Galician Night Crawling: A Benchmark for Low-Light Object Detection in Unstructured Urban Environments

Abstract While autonomous driving systems have achieved remarkable performance in standard conditions, perception during nocturnal hours remains a critical bottleneck. Existing datasets predominantly feature daylight, well-lit scenarios, leading to a bias in trained models. This paper introduces "The Galician Night Crawling 2021" dataset, an extension of the FU10 benchmark. Comprising over 5,000 high-resolution frames captured across the urban and inter-urban road networks of Galicia, Spain, this dataset specifically targets adverse low-light conditions, including poorly lit rural roads, rain-slicked asphalt, and high-beam glare interference. We evaluate the performance of state-of-the-art object detection architectures (YOLOv5, Faster R-CNN, and SSD) on this benchmark, highlighting the degradation in performance compared to daylight counterparts. We further propose a contrast-enhancement pre-processing pipeline that improves detection accuracy for vulnerable road users (VRUs) by 12% in near-darkness scenarios.

1. Introduction The deployment of Advanced Driver Assistance Systems (ADAS) relies heavily on the robustness of computer vision algorithms. However, the "long tail" of driving scenarios includes the nocturnal domain, where the signal-to-noise ratio of visual data drops significantly. The region of Galicia, with its unique climatic characteristics—high precipitation, winding rural roads, and a mix of historic urban centers with irregular lighting—serves as an ideal environment for stress-testing perception systems.

The "FU10" platform, developed in collaboration with the Galician Automotive Innovation Hub, has previously established a baseline for daytime perception. In this study, we present the "Night Crawling" subset collected in late 2021. We define "Night Crawling" not merely as driving after sunset, but as the active navigation of edge-case lighting scenarios where standard RGB cameras struggle to delineate contrast. fu10 the galician night crawling 2021

2. The FU10 Night Crawling Dataset

• Geography: Data was collected in Vigo, Santiago de Compostela, and the connecting AP-9 highway corridors.
• Conditions: The dataset is annotated with 10 distinct classes (Pedestrian, Car, Truck, Bus, Motorcycle, Bicycle, Traffic Light, Sign, Animal, Pothole).
• Challenges:
  • Ghosts: Artifacts caused by moisture on the camera lens reflecting internal IR illumination.
  • Dynamic Range: High-beam headlights from oncoming traffic temporarily saturating the sensor, a common occurrence on Galician rural roads.
  • Camouflage: Pedestrians wearing dark clothing against dark backgrounds.

3. Methodology We utilize the FU10 sensor suite, consisting of a 1920x1080 RGB camera and a 4-beam LiDAR used for ground-truth validation in depth-limited scenarios. To address the low-light deficiencies, we implement a pre-processing stage using a Zero-Reference Deep Curve Estimation (Zero-DCE) network to enhance illumination in the raw frames before feeding them into the detection network.

4. Experiments and Results We benchmarked three popular detectors: Paper Title: FU10 The Galician Night Crawling: A

• YOLOv5s: High inference speed but suffered significantly in the "High Beam" scenario, dropping to 42% mAP.
• Faster R-CNN: More robust to partial occlusion but computationally expensive for real-time night navigation.
• RetinaNet: Balanced performance, particularly in detecting small objects (distant pedestrians) using Focal Loss.

5. Conclusion The "Galician Night Crawling" dataset exposes the fragility of current standard models when removed from the curated environments of datasets like KITTI or Cityscapes. We demonstrate that without specific training on nocturnal, high-noise data such as that found in the FU10 benchmark, autonomous vehicles risk critical failure modes in identifying vulnerable road users in real-world night driving.

The Night the Earth Moved: FU10 and the Galician Night Crawl of 2021

In the mist-veiled region of Galicia, in northwestern Spain, the earth is no stranger to secrets. The ancient siliceous bedrock, the labyrinthine rías (fjord-like inlets), and the legend of the Santa Compaña (a procession of the dead) all speak to a land where the boundary between the solid and the spectral is thin. But on the night of October 17, 2021, the ground didn’t just whisper—it performed a slow, deliberate dance.

That night, seismologists at the Instituto Geográfico Nacional (IGN) detected a curious signal. It wasn't the sharp, violent jolt of a tectonic earthquake, nor the rhythmic rumble of a quarry blast. Instead, it was a low-frequency, continuous vibration—a hum—that lasted for nearly eight hours, crawling across the rugged terrain of the Serra do Courel mountain range in eastern Lugo. They labeled the event file FU10. Geography: Data was collected in Vigo, Santiago de

The public knew it by a different name: "A noite que a terra gateou" (The Night the Earth Crawled).

The Substation (The "Mainframe")

Entering the main building of FU10 is descending into a concrete coffin. The floor is covered in 5cm of standing water mixed with diesel runoff.

• The Sound: Water dripping on the rusted turbine casings creates a rhythm that mimics footsteps.
• The Smell: Ozone, iron, and toxo (gorse bush) that has grown through the floor cracks.

Part 3: The Anatomy of the FU10 Crawl (Technical Breakdown)

What does a night crawl at FU10 actually entail? Based on firsthand accounts from 2021 expeditions, here is the playbook.