Dashcam Lane Retroreflectivity Estimator

I need a computer-vision pipeline that takes ordinary dash-cam footage and automatically returns numeric retroreflectivity values for every visible lane marking. No handheld retroreflectometer data should be required at inference time; the entire estimation must rely solely on the image stream.

Core requirements
• Input: single-frame or video images captured from standard dash-cams.
• Operating envelope: robust under daytime, nighttime, rain, fog, and other weather variations with minimal performance drop.
• Output: a clean numerical table (CSV or JSON) reporting per-frame or per-segment retroreflectivity metrics—ideally the common RL value in mcd·m⁻²·lux⁻¹—plus time stamp and pixel/roadway coordinates so results can be mapped back to the footage.

Technical outline
A deep-learning approach (e.g., PyTorch or TensorFlow) combined with classical image-processing in OpenCV is acceptable, provided the final model is light enough for near-real-time execution on a laptop or edge GPU. Feel free to draw on publicly available roadway datasets, synthetic augmentation, or domain-adaptation tricks to cover night and adverse-weather scenarios.

Deliverables
1. Fully commented source code and trained model files.
2. Inference script that accepts MP4 or image sequences and writes the numeric report.
3. Brief documentation describing data preprocessing, model architecture, and a reproducible training procedure.
4. Short demo video or notebook showing the system handling both clear-day and night footage.

Acceptance
I will run the script on a fresh set of dash-cam clips; results must be consistent (same order of magnitude) with spot measurements we’ve collected on a small validation stretch of road.