Research on the intelligent recognition method for pavement pothole diseases based on 3D point cloud technology

Potholes, as one of the most destructive pavement defects, severely compromise driving safety and comfort. Traditional manual detection methods are inefficient and fail to meet modern road maintenance requirements. While two-dimensional image-based pothole detection improves efficiency, it cannot capture depth information and suffers from poor spatial coverage. This study focuses on analyzing road surface potholes using 3D point cloud data. High-resolution 3D point cloud data is rapidly acquired via 3D laser scanning technology, generating comprehensive geometric information. A specialised dataset was constructed by integrating public data and expanding sample size through data augmentation techniques. Addressing the limitations of traditional algorithms in identifying potholes within complex natural environments, this study proposes a receptive field optimisation method based on deep learning for point cloud segmentation of pavement potholes. Experimental results show that the PointNet++ model optimised with receptive field achieves an Intersection over Union (IoU) of 79.02%, Precision of 87.47%, Recall of 89.11% and F1-score of 88.28%. This study indicates that the model can accurately segment pothole regions even under limited interference scenarios in this research. The proposed method provides a technically feasible and highly accurate solution for intelligent pavement defect identification.