Computer Vision
Bridge Crack AI
Automated drone-based bridge crack detection with dual-stage deep learning and real-time severity assessment.
About the Project
What We Built
An end-to-end automated bridge inspection system that eliminates the safety hazards, cost, and subjectivity of traditional manual surveys. A Unity 3D simulation environment controls a virtual drone traversing a bridge structure, streaming frames to a FastAPI backend. A lightweight TFLite classifier acts as a gatekeeper, passing only crack-positive frames to a PyTorch U-Net for pixel-level segmentation. Severity is scored in real time and a comprehensive PDF inspection report is generated automatically at mission end.
The Problem We Solved
Challenge, Solution & Outcome
Traditional bridge inspections require scaffolding, lane closures, and specialist engineers working at height creating safety risks, high costs, and subjective assessments. Even drone-assisted surveys still require human review of hours of footage, creating a time and accuracy bottleneck.
A simulation-first automated inspection pipeline where a drone streams frames to a FastAPI backend running a dual-stage deep learning system. The TFLite gatekeeper filters healthy frames efficiently, while the U-Net segmentation model performs precise crack localisation only when needed, with automated severity scoring and report generation eliminating manual review entirely.
The system demonstrates fully automated crack detection with real-time severity classification, removing human subjectivity from structural assessment. The dual-stage architecture reduces unnecessary model inference, and the automated PDF output gives engineers a structured, auditable inspection record without manual documentation effort.
Technology
Tech Stack Used
Capabilities
Key Features
Unity 3D Engine simulates realistic drone physics, bridge geometry, lighting, and surface textures. C# scripts handle flight controls, camera capture, and HTTP communication with the Python backend at configurable frame intervals.
A lightweight TFLite model performs fast binary classification on every incoming frame. Frames classified as crack-free are discarded immediately, avoiding the compute cost of segmentation on healthy concrete.
Crack-positive frames are passed to a ResNet34-backbone U-Net model. The network produces a binary mask highlighting the exact pixels where cracks are detected, enabling sub-millimetre localisation.
Post-processing calculates the ratio of crack pixels to total frame pixels. Results are classified as Minor under 2%, Moderate between 2% and 5%, or Severe above 5%, each with a recommended maintenance action.
Processed frames return to the Unity dashboard as red-highlighted overlays with metadata including confidence score, severity level, and GPS-equivalent drone telemetry updating the operator view in real time.
ReportLab compiles a full mission summary at session end, including total frames analysed, crack frame count, severity distribution, drone flight path, and side-by-side original and annotated image pairs.