In this paper, an automatic building extraction process based on MVS point clouds is proposed to automatically extract building point clouds from urban MVS dense point clouds of complex scenes by projection, morphological expansion and contour extraction techniques. Aiming at the deficiency of Poisson surface reconstruction, this paper proposes a surface model optimization method based on RANSAC fast fitting. The method generates the optimized surface model through the filter denoising process and chunked RANSAC fast fitting. Finally, a workflow for the 3D reconstruction of urban buildings based on the MVS point cloud is proposed. In the analysis for the urban 3D modeling technique, the average error of the model after reconstruction is only 0.731%, and the measurement errors in the three-dimensional directions of length, width, and height are less than 5 cm. and the time consumed before and after the optimized method in this paper is reduced by an average of 3.09 s. Therefore, this study provides a simple and efficient method for the automatic extraction and 3D reconstruction of urban buildings.