Automated Pavement Pothole Detection and Multi-Dimensional Indicators Extraction Using Deep Learning and 3D Reconstruction
- 6 March 2026
- journal article
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 27 (5) , 5299-5316
Abstract
Accurate detection and quantification of pavement potholes are essential for ensuring driving safety, preventing vehicle damage, and supporting intelligent pavement condition assessment. Traditional methods, which rely on image processing techniques, often suffer from low accuracy and fail to consider the structural and three-dimensional (3D) characteristics of potholes. This study proposes an ensemble solution that combines deep learning with an emerging 3D triangulation system for automated pothole detection, segmentation, and multi-dimensional indicator extraction. The YOLOX-PTSNet two-step model is employed for pothole detection and precise segmentation. The Pyramid Transformer Module (PTM) is designed to achieve a global receptive field by replacing pooling operations with pixel block fusion and adaptive dilation rate window division, thereby minimizing information loss. The Pyramid Transformer Segmentation Network (PTSNet), entirely based on Transformer mechanisms, overcomes the limitation of CNNs, whose receptive fields are constrained by convolutional kernel size, and demonstrates strong global feature extraction capabilities and robustness. The method achieved an F1 score of 91.74% and an IoU of 0.847 on the test set, showing superior performance compared to the models tested in this study. Additionally, a calculation model was developed for extracting multi-dimensional pothole indicators using pixel-level segmentation and millimeter-level 3D elevation data through 3D reconstruction. Field validation on roads in Shanghai’s Jiading District demonstrated high accuracy and consistency, with average accuracies for maximum depth, area, and volume exceeding 92%, and R2 values of 0.919, 0.986, and 0.991, respectively. This study provides an integrated sensing-and-quantification framework that enables reliable, high-resolution 3D pothole characterization to support intelligent pavement condition assessment and maintenance planning.Keywords
Funding Information
- National Natural Science Foundation of China (52372334)
- National Natural Science Foundation of China (62206201)
- Science and Technology Project of Henan Provincial Department of Transportation (2023-1-1)
- Key Research and Development Program of Sichuan Province (202303AA080016)
This publication has 42 references indexed in Scilit:
- Cost-effective LiDAR for pothole detection and quantification using a low-point-density approachAutomation in Construction, 2025
- Swin-Unet: Unet-Like Pure Transformer for Medical Image Segmentation2023
- Advanced pavement distress recognition and 3D reconstruction by using GA-DenseNet and binocular stereo visionMeasurement, 2022
- Pavement Cracking Detection and Classification Based on 3D Image Using Multiscale Clustering ModelJournal of Computing in Civil Engineering, 2020
- Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation2018
- Pothole Detection and Classification Using 3D Technology and Watershed MethodJournal of Computing in Civil Engineering, 2018
- Pothole detection on asphalt pavements from 2D-colour pothole images using fuzzy c -means clustering and morphological reconstructionAutomation in Construction, 2017
- Pothole detection using spatio‐temporal saliencyIET Intelligent Transport Systems, 2016
- U-Net: Convolutional Networks for Biomedical Image Segmentation2015
- Pothole detection in asphalt pavement imagesAdvanced Engineering Informatics, 2011