Abstract: In order to realize the cultural heritage digital protection of Mogao Grottoes in Dunhuang, a real scene 3D modeling method based on the four-camera vision system was designed. At first, aiming at the problem of low accuracy of the feature points matching in binocular stereo vision system, a four-camera stereo vision system and high-accuracy stereo matching strategy adapted to it were designed. Using the redundant design and self-checking function of the fourcamera vision system, the high-accuracy spatial point cloud data can be obtained. Then, aiming at the problem of error connection between the uncorrelated objects in the Delaunay triangulation results, the length of triangle constraint condition in the image coordinate system and the world coordinate system were increased. It can be used to prevent the error triangle. The condition can be obtained adaptively according to the point cloud data distribution characteristic. The system has been applied to the 3D reconstruction of the 172nd cave in Mogao Grottoes. The results show that the system can reconstruct the complex scenes with multiple objects and large depth changes efficiently. The 172nd cave has 41649 effective point clouds, and the 3D modeling time is about half an hour.