Segmenting light detection and ranging (LiDAR) point cloud of building accurately is the important section in the reconstruction of three-dimensional model. In view of the complex roof structure of complex buildings and poor segmentation accuracy of the existing algorithms, the authors put forward a kind of algorithm of region growing with the basic element of triangles to segment the point cloud of the building. First of all, Delaunay triangulation network is constructed, correlation is set up among laser points, unit normal vectors of triangles are calculated, initial partition is conducted on point cloud with the character that vectors in unit vector approach of triangles on the same plane of the building are basically consistent; then, because dispersion and deviation of point cloud could produce many disheveled triangles, dissection is conducted on points that are composed of disheveled triangles; based on good robustness of random sample consensus (RANSAC) algorithm, boundaries of planes of the building combining are obtained with Alpha Shape algorithm, plane and isolated point are combined in over-segmentation. The test result shows that the point cloud segmentation on the roof of the building is ideal in integrity, accuracy and quality with the method put forward in this paper.