A method for large-scale scene 3D reconstruction based on crowdsourced imagery

Abstract

Aiming at the problem of high data demand for three-dimensional (3D) reconstruction based on a single data source, a large-scale scene 3D reconstruction method using crowdsourced images is proposed. Firstly, dense point clouds are extracted from satellite images and aerial images respectively. For satellite images, dense point clouds are extracted via pixel-wise matching on the basis of DEM-assisted geo-positioning accuracy improvement method. For aerial images, the positioning parameters of each image are obtained by the structure from motion (SfM) model, and then dense point clouds are extracted by using the multi-view dense point cloud generation method which integrates multiple strategies. Subsequently, multi-source point clouds are calibrated and merged by using a hierarchical registration strategy. After that, large-scale scene surface reconstruction, model refinement and automatic texture mapping are implemented, and Internet images are employed to repair holes in 3D models and textures. Experimental results demonstrate that this method can make full use of the advantages of multi-source data including satellite images, aerial images and Internet images. Through information complementarity, it overcomes the limitations imposed by stringent requirements for single-source data. The reconstruction yields high fineness and completeness, enabling complete and accurate 3D modeling of large-scale scenes.