Automated pothole detection and volume assessment using PointPSSN and smartphone LiDAR point clouds

The rapid detection and assessment of potholes are critical for ensuring road traffic safety. However, point cloud-based techniques relying on surveying vehicles or drones are often expensive and may be limited by roadside obstruction or narrow roadways. This study proposes a novel approach for assessing road potholes using point cloud data collected by smartphone LiDAR. The method integrates the Point Pothole-Specialized Segmentation Network (PointPSSN), a lightweight point cloud segmentation model designed to achieve high accuracy with low parameter complexity and rapid inference, together with scale-adjustable voxelization for assessment. The PointPSSN model incorporates a Geometric Feature Encoder module to capture the geometric attributes of potholes by extracting local geometric features. Neighbor Finder module identifies and aggregates neighboring points that provide more significant information. Experiments were conducted using a smartphone LiDAR device within a 7.28 km² area of Wuchang District, Wuhan, China, encompassing diverse road conditions. A dataset of 1040 potholes was constructed for model training and evaluation. The results demonstrate that the PointPSSN model achieves a segmentation accuracy of 97.336 %, precision of 91.322 %, recall of 79.888 %, an F1-score of 85.223 %, and an intersection-over-union (IoU) of 74.251 %. Notably, the accuracy, F1-score, and IoU surpass the performance of state-of-the-art models by 0.233 %, 1.336 %, and 2.006 %, respectively. In terms of efficiency, PointPSSN requires only one-seventh of the FLOPs and one-fifteenth of the parameters of state-of-the-art models, while achieving an 18.37 % faster inference speed. Furthermore, the average relative errors in depth and volume assessment using voxelization methods are 9.08 % and 9.04 %, respectively.