Public Infrastructure
Crowdsensed road surface monitoring at state scale
A road network of roughly 589,000 km was monitored by manual photo-and-GPS reports: weeks to action, inconsistent data, and no way to separate genuine defects from ordinary vibration. We built a crowdsensing pipeline that reads smartphone accelerometer, GPS and camera data as vehicles drive, classifies the road surface it is travelling on, and distinguishes potholes from speed bumps and driver handling — the false-positive problem that makes accelerometer-based detection unreliable in the field. Smartphone imagery serves as the source of truth, so the models are retrained against evidence rather than assumption.
Modelling approach follows published research on deep learning for road surface monitoring — Varona, Monteserin & Teyseyre, Personal and Ubiquitous Computing (2020).
- Focus
- Deep learning (CNN, LSTM, reservoir computing), crowdsensing, computer vision
- Engagement
- Proof of concept and MVP — a PWA, live at machinelearnspothole.site
- Outcome
- 93% accuracy on pothole and surface classification, with an imagery-backed retraining loop
A phone in a moving vehicle senses the road through the suspension. Each trace is plotted against road position, so every spike sits directly beneath the defect that produced it.
The tyre bridges what is narrower than its contact patch, so a street gutter registers differently from a pothole of the same depth — and a speed bump produces a clean signature that looks a great deal like a defect. Separating those three is the classification problem.
Detecting a pothole is the easy half. Not calling a speed bump a pothole is the half that decides whether a road authority can act on the data — so the evaluation is reported per class, against five alternative models, rather than as one headline number.
Accuracy by model
Stability event classification. The CNN is compared against two other deep models and three non-neural baselines on identical data.
Where the errors go
Overall accuracy hides which mistakes a model makes. The diagonal is correct; everything off it is a misclassification, and the shade is that error's share of the true class.
The classes that actually matter
Per-class F-measure on the stability-event task. Every model finds potholes. Only one holds up on speed bumps and street gutters — the two events that generate false pothole reports.
Varona, Monteserin & Teyseyre — “A deep learning approach to automatic road surface monitoring and pothole detection”, Personal and Ubiquitous Computing 24:519–534 (2020). Figures reproduced from the published results.