A road inspection camera can miss a defect because the image is unsuitable, the device overheats or the result never reaches the server. Those failures matter to a maintenance team even when a detector performs well on a curated test set. A peer-reviewed study by Moon-Sup Lee and Seung-Yeon Han at the Korea Institute of Civil Engineering and Building Technology examines that gap through several years of vehicle-based pothole monitoring. This Civil Insight review considers what the comparisons demonstrate and how the evidence could inform an Australian trial.
A paired test on the same vehicle
In the 2024 field experiment, the researchers fitted a baseline mount and an improved mount to each of three vehicle types. The baseline sat near the rear-view mirror, while the alternative used a passenger-side windscreen position and four-fan cooling. Each vehicle made three passes along a 2 km out-and-back route in Paju. Visual inspection identified eight unfilled potholes, creating 24 encounters per vehicle. The equipment was assessed simultaneously, so both mounts experienced the same traffic, light and road conditions.
The measured outcome was the proportion of known pothole encounters correctly detected and received at the server. A detection that never arrived counted as a miss. This makes the result useful for assessing a working inspection system, but it combines image recognition, equipment reliability and data transmission in one measure.
What changed in the field
For unfilled potholes, the improved arrangement achieved detection yields of 41.7–58.3% across the vehicles. The baseline achieved 0–4.2%. The two changes were tested together, so the experiment cannot assign a separate benefit to camera position or cooling. The improvement also leaves a substantial proportion of encounters undetected or unreceived, which a maintenance team would need to consider when deciding how to use the output.
A separate part of the study evaluated a software pipeline that first proposes possible potholes and then checks the proposed regions with a classifier. In a controlled reproduction using 2,550 held-out frames, adding that verification stage reduced false detections from 2,745 to 1,996. Precision rose from 0.709 to 0.755, while recall fell from 0.663 to 0.608. These figures describe the held-out comparison and use a different measure from the route experiment’s system-level yield.
The strength of the comparison and its limits
The simultaneous mounting test is a strength because it reduces several obvious sources of variation between runs. Reporting reception failure as part of the measured outcome also exposes a failure that would disappear from a test restricted to images already stored on a computer. The small number of physical potholes is a constraint. Repeated passes provide more encounters, yet they revisit the same defects on the same short route.
The software comparison gives a clearer account of the trade-off involved in rejecting candidates. It removed false detections and some true detections. That is relevant when a crew uses alerts to decide where to inspect. The paper also reports stronger reductions from operational data, but those records lack a fully specified labelled evaluation-set size. The held-out comparison therefore provides the more traceable numerical basis for judging the filter’s effect.
What a local road agency could test
For an Australian maintenance trial, begin with the decision the system will support. A tool used to nominate locations for follow-up inspection can be assessed against the number of useful inspections it produces and the time spent checking false alerts. A tool expected to provide network coverage also needs a way to measure the defects it misses. Those two purposes place different demands on sampling and validation.
Record the capture conditions alongside results. Mount position, windscreen reflections, vehicle type, device temperature and reception interruptions can help explain why one run produces fewer usable records. A local reference survey, with known defects and repeat visits, would allow the team to distinguish an uneventful road section from a failed capture run.
The study gives a practical reason to test the complete route from the windscreen to the inspection queue. Its findings support that evaluation approach. The reported detection percentages remain tied to the study’s route, equipment and data, so a local trial needs its own evidence before the output becomes part of routine maintenance decisions.
Original paper. Lee, M.-S., & Han, S.-Y. (2026). Vehicle-Mounted Vision Sensing for Large-Scale Road Surface Monitoring: A Multi-Year Field Study of Acquisition Geometry and a Heterogeneous Detection–Classification Pipeline for Pothole Mapping. Sensors, 26(17), 5573. Version of record published 2 September 2026. DOI 10.3390/s26175573
Featured photograph. Baseline and improved camera mounts installed together in the study vehicle. Photo: Moon-Sup Lee and Seung-Yeon Han (CC BY 4.0).
