Rasik Labs
← Back to projects

Infrastructure Intelligence

Road and infrastructure damage detection

Turn dashcam footage and drone imagery into a prioritised repair list, automatically detecting cracks, potholes, and surface damage in real time.

Problem

Road authorities inspect thousands of kilometres manually: slow, expensive, and inconsistent. A single inspector misses hairline cracks that become potholes within a season. By the time damage is reported, repair costs have multiplied. The challenge is a system that processes dashcam or drone imagery in real time, classifies damage by type and severity, and produces a prioritised repair list, turning reactive maintenance into proactive asset management.

System demo

Inference running on SKU-110K public dataset · YOLOv8m pretrained checkpoint · Annotations rendered with OpenCV

Architecture

Step 1

Dashcam / drone feed

Step 2

Frame extraction + preprocessing

Step 3

YOLOv8 damage detection

Step 4

Damage classification

Step 5

Severity scoring + GPS tag

Step 6

Priority repair list

Step 7

Dashboard / GIS export

MetricsPrototype benchmarks on RDD2022 validation set — to be replaced with production results

mAP@50

73.4%

Precision

76.8%

Recall

71.2%

F1 score

73.9%

Inference (GPU)

38ms

Damage classes

4

Tech stack

Python 3.11YOLOv8mOpenCVAlbumentationsRDD2022 dataset47,420 images across 6 countriesFastAPIStreamlitadaptable to your existing GIS or asset management platformDocker

Edge deployment — works where connectivity doesn't

Road inspection vehicles and drones operate in areas with no reliable internet. We deploy optimised model variants directly onto onboard hardware, enabling real-time detection and GPS coordinate logging with no cloud dependency. Results sync automatically when connectivity is restored.

NVIDIA Jetson OrinRaspberry Pi 5Qualcomm AI Hub (QCS series)...and other devices per your requirements

Production considerations

  • Input source flexibilitydashcam video, drone footage, or static survey images; frame extraction rate configurable per use case
  • Weather and lighting variationaugmentation pipeline covers rain, shadow, and low-light conditions across 6 countries in training data
  • GPS taggingdetections tagged with coordinates from vehicle or drone telemetry for direct GIS integration
  • Severity scoringdamage instances scored by area, class, and density to produce a prioritised repair queue, not just a raw detection list
  • GIS / asset management integrationoutput feeds existing road asset platforms via REST API or GeoJSON export
  • Retraining workflownew road surface types or country-specific damage patterns added with a labelled sample set; built into every delivery

Explore this project

Live demo and source code links will be added as they become available.

View demo — Coming soonView repository — Coming soon