Dense Point Cloud Reconstruction of Buildings from Sparse Binocular Stereo Vision Images

Dense 3-dimensional point cloud reconstruction from sparse images is essential for accurate building measurements in civil engineering. However, traditional methods are often time-consuming, requiring dense image capture or specific viewpoints. To address these limitations, a binocular stereo vision-based method for building point cloud reconstruction is proposed. The approach integrates an automatic disparity-range estimation algorithm with an improved semi-global block matching method to generate high-quality disparity maps. Additionally, an image-segmentation-based method is employed for background noise removal. Finally, a cross-enhancement mechanism that accounts for rigid-body motion constraints is incorporated for precise point cloud registration from sparse, low-overlap images. Compared with traditional methods, the proposed coarse-registration algorithm achieves significant performance gains, reducing average rotation and translation errors by 92.21% and 13.68%, respectively, while shortening the computation time by at least 43.55%. In the evaluated outdoor facade and indoor beam-column component, component-level dimensional measurement errors remained within 3%.