IRI Estimation across Varying Point Densities in LiDAR-Derived Point Clouds: Effects of Profile Extraction and Transformer-Based Classification

Abstract

The international roughness index (IRI) is a widely adopted metric for quantifying road surface irregularities and assessing pavement ride quality. Traditional IRI estimation methods have relied on vehicle-mounted inertial profilers that collect data along wheel paths. While effective, these methods require multiple passes, partial lane closures, or controlled traffic conditions to capture shoulders and adjacent lanes. To overcome these limitations, recent studies have explored using light detection and ranging (LiDAR) point clouds for roughness estimation. However, three challenges remain largely unaddressed: the effect of profile extraction methods on IRI consistency, the impact of nonuniform point density with distance from the LiDAR scanner, and the lack of semantic classification in IRI estimation pipelines. This study addresses these gaps by evaluating IRI estimates derived from LiDAR-acquired point clouds across different lateral segments of multilane highways. The first objective was to evaluate how the choice of longitudinal profile extraction method, namely, raw extraction, mesh-based interpolation, and piecewise cubic Hermite interpolating polynomial (PCHIP) smoothing, influences the stability of IRI estimates, under varying point densities and lateral distances from the LiDAR sensor. The second objective was to evaluate whether automatically classified point clouds, produced by a transformer-based segmentation model, can reliably support IRI estimation in place of manually annotated point cloud data. IRI was estimated using three methods. Profiler viewing and analysis software (ProVAL) software was used for profiles with uniform spacing obtained via mesh-based and PCHIP interpolation. In parallel, two analytical methods were employed across all extraction techniques: a semianalytical quarter-car simulation, and a Fourier-based method. Results showed that mesh-based extraction outperformed raw and PCHIP interpolations across all lateral positions, producing the most stable estimates, particularly when paired with ProVAL or the quarter-car model. Within 5 m of the LiDAR sensor, transformer-labeled data achieved mean absolute error (MAE) values near 0.06    mm / m , while in 5–10 m lane segments, MAEs remained below 0.7    mm / m , confirming reliable IRI estimation with automated labeling given sufficient classification precision and point density. The Fourier method also offered a computationally efficient alternative, showing reasonable agreement with the quarter-car model. These findings demonstrate the potential of integrating three-dimensional surface reconstruction and semantic segmentation for scalable and context-aware pavement condition assessment across the full roadway cross section.