Comparative Analysis of Upsampling Algorithms for Pavement Point Clouds: Toward 3D Pavement Distress Inspection

The accurate representation of pavement condition is crucial for road agencies to make appropriate maintenance decisions. With the great development of three-dimensional (3D) measurement technologies and devices in the past decades, acquiring 3D information of pavement has become much more accessible in the practice of road maintenance. However, the insufficient density of point clouds collected by economical devices still limits their utilization in pavement condition evaluation. Fortunately, many point cloud upsampling methods have been successfully developed to convert sparse and nonuniform points to dense representations, while their performance in the pavement area still needs validation. This study aims to examine existing upsampling algorithms for pavement point clouds based on public datasets and evaluate their effectiveness through visualization of point clouds collected by us. Specifically, a dataset comprising 10,026 samples of point clouds selected from public recourse was developed for training and evaluation. Three distinct point cloud upsampling methods, Poisson reconstruction, cubic spline interpolation, and deep learning-based methods, were compared on the dataset. Moreover, the point clouds of pavement collected by ourselves were upsampled to evaluate their performance. The results indicated that the deep learning method demonstrated superior performance in the P2F metric, reducing it by 5.9% compared to the traditional cubic spline method. The deep learning methods enhance the information capture capability of low-density point cloud devices, which will facilitate the application of low-cost equipment for road surface detection, thereby contributing to the development of efficient road surface detection and management systems.