How the Danish Road Directorate benchmarked RoadAI against LCMS

AI and computer vision tools have flooded the road condition assessment market in recent years, promising effortless inspections and instant insights. In practice, many of them fall short on basic repeatability, accuracy, and robustness.

Getting AI road inspection right is harder than the hype suggests, which is why RoadAI undergoes continuous testing, validation, and development in partnership with road owners. The Danish Road Directorate (DRD) study below is a good example of that approach.

Why DRD looked beyond laser scanning

Before choosing RoadAI, DRD took a hard look at the state of their road condition data and clarified what they needed from a new solution. Their exploration of AI-based condition assessment started from a very concrete challenge: they were trying to quantify the costs and benefits of their maintenance work, especially patches. In theory, the data existed; in practice, the documentation of where and how maintenance had been carried out was lagging and unorganized. Smartphone-based AI systems soon got on their radar.

DRD decided to pilot smartphone-based AI road inspections to replace traditional methods, enabling much more frequent updates, faster repairs, closer tracking of how damage develops, and better predictions of when maintenance is needed.

Inside the tender: what DRD asked AI vendors to prove

Rather than moving straight to procurement, DRD first organized a market dialogue in March 2025, where six companies presented their solutions. The questions were pointed and practical: how does the system record pavement defects, and how might it handle other assets; how can data be exported and stored in DRD's own systems; what implementation and support actions are offered; and what role could such a system play in a future asset management environment covering multiple asset types. This dialogue framed the requirements for a formal tender, published in May 2025.

The tender focused on what mattered most for DRD's immediate needs. The system had to provide information on cracks, alligator cracking, potholes and patches; other assets were deliberately left out.

Price was one factor in the selection, but DRD's award criteria also examined quality in depth, awarding the contract to the tender offering the best balance between price and analysis quality. The underlying AI models were evaluated for architecture, training data, adaptability, and transparency. Road damage detection was assessed in terms of the system's ability to accurately detect and classify damage, including sensitivity, precision, and robustness under varying conditions. Reporting was judged on clarity, completeness, customization options, and integration with existing workflows. Data visualization was assessed for how well it supported decision-making. On this basis, DRD selected Xweather RoadAI.

Acceptance testing against LCMS: Repeatability, reproducibility, and classification results

Selection was followed by structured acceptance testing, using DRD's established Laser Crack Measurement System (LCMS) based TOTALDAMAGEINDEX as a reference. The index is calculated from LCMS recordings on the automatic road analyzer (ARAN) vehicle, where defects are mapped in 1-meter grids and then averaged over 10-meter lengths. Several test sections were defined. One 500-meter municipal road section was used for repeatability and reproducibility testing, with a TOTALDAMAGEINDEX value of 1.704. Five 1,000-meter motorway and highway sections, with Total Damage Index values ranging from 0.0495 to 0.417, were reserved for classification tests.

On the municipal road, DRD examined how consistently RoadAI reported the condition across multiple passes of the same section. Five runs from five different phones were analyzed. In the tender, DRD defined an acceptance margin of 20 percent around the mean, which for these measurements corresponds to about 2.5 standard deviations. Only five points across the series lay outside this margin, corresponding to roughly 1 percent error. On this basis, DRD approved the system's repeatability. 

For reproducibility, the focus shifted to how RoadAI's index compared with the LCMS-based reference along the same section. Data were normalized, the three measured runs were averaged, and a 100-meter moving average was used to reduce high-frequency noise. In general, the curves tracked each other well; at around 300 meters, where RoadAI's values diverged, images showed that LCMS had detected joint cracking between lanes, which the lane detection model in RoadAI had filtered out to some extent. This explained the local difference seen in the reproducibility analysis.

The classification tests asked a simpler but crucial question: could RoadAI rank road sections in the same way as the reference data? Five 1,000-meter motorway and highway sections were used, each with a known Total Damage Index and a classification based on the LCMS reference — Best, Good, Worst, Fair, and Poor. When the same sections were assessed using RoadAI data, the resulting classifications matched the reference in every case. The section classified as Best in the reference remained Best with RoadAI, Good matched Good, Worst matched Worst, and Fair and Poor were likewise consistent.

From validation to a 19-vehicle rollout

The testing led to several clear conclusions for DRD. They saw that it was necessary to think beyond traditional defect-detection equipment, and that AI, machine learning, and computer vision could provide strong repeatability and reproducibility when applied through condition indices. Data collected with smartphones proved useful for network-level defect classification, even though more work is still needed to size individual defects. They also recognized that defining acceptance criteria in tenders is challenging but can be managed effectively using indices.

Looking ahead, DRD plans to use RoadAI in 19 inspection vehicles in 2026, taking advantage of the fact that these vehicles regularly operate on the same roads. RoadAI data will support a dynamic maintenance and repair plan and feed into new probabilistic deterioration models, while DRD continues to employ line-scan lasers for correlation with traditional methods, collect additional data with surveying vehicles, and share the same datasets across different departments and international collaborations.