This paper proposes an improved greedy projection triangulation algorithm to address issues such as lengthy processing time, rough surface quality, and hole formation. Firstly, the voxel filter is improved, the optimal voxel grid is calculated with the number of point clouds as the threshold, and the adjacent points of the center of gravity are used instead of voxels to achieve down-sampling. Subsequently, statistical and Gaussian filtering are applied to smooth the simplified point cloud. The octree is adopted to replace the k-d tree for neighborhood point cloud search, and a moving least squares method optimized for octree is utilized for normal estimation, which reduces backtracking time and mitigates normal ambiguity. Finally, based on Delaunay′s spatial region growth algorithm, the plane is used Triangulation, obtaining a triangular mesh surface according to the topological relationship of points in the plane. The experimental results show that, compared with the traditional greedy projection triangulation and Poisson algorithm, the proposed method has better reconstruction surface smoothness and reduces the number of holes while maintaining the algorithm′s speed and local details.