Road hazard detection in low visibility

A research-led guide to detecting potholes, debris, damaged road surfaces, pedestrians and stopped vehicles when fog, rain, glare or darkness makes driving less predictable.

• Computer vision
• Object detection
• Adverse weather
• Edge AI
• Road safety

Watch: related videos

These YouTube videos are embedded from the links supplied for this topic. Use the player controls to watch without leaving the page.

1 · Hazard categories

2 · Why adverse conditions break detectors

Object detectors learn statistical patterns from their training images. Fog reduces contrast; rain and spray add streaks and occlusion; nighttime footage introduces sensor noise and uneven illumination; wet roads create reflections that resemble objects. A model tested only on clear daytime images can therefore appear strong while failing under distribution shift.

• Fog and haze: distant edges fade and objects blend into the background.
• Rain and spray: streaks, droplets and windshield artefacts obscure boundaries.
• Night and glare: highlights saturate while nearby regions remain dark.
• Motion blur: vehicle movement smears small hazards across pixels.
• Domain shift: camera, region, road material and weather differ from the training set.

3 · A practical research pipeline

1. Define scope and classes. Decide whether the task covers potholes, debris, vehicles, pedestrians or all of them. Avoid merging visually or operationally different hazards into one vague label.
2. Build a representative dataset. Include clear, foggy, rainy, dusk and night scenes; multiple cameras and roads; near and distant objects; and difficult negative examples such as puddles, shadows and road markings. Document licensing and annotation rules.
3. Split by source, not adjacent frames. Put entire drives, videos, locations or collection sessions into one split. Randomly splitting neighbouring video frames can leak near-duplicates into training and inflate scores.
4. Train a baseline detector. Start with a well-documented real-time detector (for example a YOLO-family model or RT-DETR variant), then compare with a second baseline under the same data and compute budget.
5. Test enhancement as an ablation. Compare raw images against a fixed dehazing / low-light enhancement pipeline. Enhancement can improve visibility, but it can also hallucinate edges or amplify noise.
6. Evaluate per condition and class. Report precision, recall, AP at IoU thresholds, mAP, latency and performance for each weather/lighting subgroup—not just one aggregate accuracy.
7. Validate in the real environment. Test on unseen roads and cameras, measure false alarms per distance or time, and keep a human safety process. A model output is not a guarantee that a road is safe.

4 · Model design choices

5 · Evaluation: what should be reported?

• Precision: of predicted hazards, how many are correct?
• Recall: of annotated hazards, how many did the system find?
• AP / mAP: detection quality across confidence thresholds and classes; state the IoU protocol.
• Latency and FPS: measured on the target hardware, including preprocessing.
• False-alarm rate: especially important for road alerts and driver trust.
• Robustness slices: report fog, rain, night, glare, distance, occlusion and camera separately.
• Calibration: check whether confidence scores correspond to empirical correctness.

Do not report a benchmark number until the dataset split, class mapping, hardware, model version and evaluation script are recorded. No experimental score is claimed on this page.

6 · Dataset and reproducibility checklist

• Record dataset name, version, source, license and class definitions.
• Publish train/validation/test split rules and avoid near-duplicate leakage.
• Preserve original images and record any enhancement parameters.
• Report class counts and long-tail/rare-hazard performance.
• Save random seeds, code version, model weights and inference settings.
• Inspect false negatives manually, especially rare but high-consequence hazards.

7 · Limitations and safety

Camera-only systems cannot reliably infer every hazard's depth, severity or drivability from a single image. Fog enhancement does not restore information that was never captured. Public benchmark performance may not transfer to local roads, monsoon conditions, unfamiliar cameras or night scenes. Treat this as decision support and research—not an autonomous driving safety guarantee.

Research references

• Research literature: object detection in adverse weather (arXiv search)
• BDD100K dataset — diverse driving conditions
• ACDC adverse-condition driving dataset project
• COCO dataset and evaluation context
• Ultralytics detection documentation (baseline implementation reference)

← Browse all research topics