Three-Dimensional Reconstruction and Texture Characterization of Asphalt Pavement Surfaces Using Binocular Stereo Vision

Abstract

Accurate three-dimensional characterization of asphalt pavement texture requires both reliable surface reconstruction and quantitative interpretation of reconstructed geometry. This study evaluated a passive binocular stereo vision workflow integrating camera calibration, image preprocessing, AANet-based disparity estimation, point-cloud processing, reference-plane correction, surface reconstruction, and texture characterization. Calibration using 16 stereo checkerboard pairs yielded a mean reprojection error of 0.12 px and a baseline of 60.110 mm. On Scene Flow samples, AANet achieved an endpoint error of 2.37 px and a >1-px error rate of 9.2%, compared with 9.71 px and 28.7% for SGBM. Reference-block verification produced absolute height errors of 0.07–0.16 mm and a mean relative error of 4.57%. For nine representative pavement stereo pairs, valid-pixel ratios were 74.75%–87.66% and direct grid coverage was 77.86%–90.62%. Across six specimen series and eight loading stages, reconstructed MTD decreased by 38.9%, 37.7%, and 33.3% for AC-13, SMA-13, and OGFC-13, respectively. Mean specimen-level correlations between reconstructed MTD and measured TD reached 0.865, while fractal dimension–BPN correlations ranged from 0.454 to 0.940. These results demonstrate the potential of passive binocular vision for quantitative asphalt pavement texture characterization under controlled laboratory conditions.