Computer Vision + Applied ML
Real-Time Defect Detection for Infrastructure Inspection
Built an inspection system that classifies structural defects from live survey video, so findings arrive while the crew is still on site instead of weeks later in a report.
90%
classification accuracy
4x
faster processing
30%
lower inspection costs
25%
faster stakeholder decision-making
The problem
Infrastructure inspection produces far more footage than anyone can review. A survey generates hours of video, a specialist reviews it afterwards, and the finding that mattered gets confirmed long after the crew has packed up and left. Review is also inconsistent, because two reviewers looking at marginal footage do not always agree and there is no cheap way to check which of them was right.
What I built
Built a detection pipeline that runs against the live feed during the survey rather than against the archive afterwards. Predictions are surfaced next to the footage in an analytics view, so an operator can confirm or dismiss a finding while still on site and while re-inspecting is still cheap. The reporting that used to be assembled by hand afterwards is generated from the same record the operator worked against.
Technical approach
- Several models combined rather than relying on one classifier, because the defect classes fail in different ways and no single model was reliably best across all of them
- A larger model's behaviour distilled into a smaller one so it could keep up with a live feed instead of running offline afterwards
- Predictions surfaced next to the footage they came from, so an operator confirms or dismisses a finding rather than being asked to trust a score
- Reporting generated from the record the operator already worked against, which removed a manual transcription step and the errors that came with it
- Model output translated into the terms the client actually made decisions in, rather than reported as accuracy figures nobody could act on