Lightweight 3D Pavement Reconstruction via Computer Vision: Optimizing Feature Matching for Low-Cost Highway Inspection

As the highway networks continue to expand, the workload for pavement inspection tasks is rapidly increasing. However, existing pavement inspection methods struggle to reconcile the contradiction between obtaining rich pavement information and the associated high costs and low efficiency. This paper presents a low-cost and easy-to-deploy vision-based workflow for pavement 3D reconstruction under normal driving conditions. Pavement point clouds were obtained through feature extraction and matching, motion estimation, and 3D reconstruction. A point cloud processing workflow combining plane fitting and perspective correction to improve geometric consistency was proposed. Given the monotonous, repetitive, and low-texture characteristics of pavement scenes, this paper systematically evaluated the performance of multiple feature extraction and feature matching algorithms, including classic and deep learning-based methods. By comparing with manual measurement results and considering both reconstruction accuracy and quality, the optimal strategy was determined. The variation in model accuracy under different speeds was investigated through experiments. The results indicate that proposed workflow can maintain stable reconstruction quality without interrupting normal traffic, and the depth estimation accuracy is within engineering-level tolerance under our measurement protocol. This study enables the low-cost acquisition of 3D pavement information, which is crucial for evaluating pavement technical conditions. It provides a new pavement inspection and monitoring solution, particularly offering significant application value for the extensive network of low-grade pavements.