Airborne light detection and ranging technology is widely used in the urban 3D reconstruction in recent years,and building roof planes segmentation is a basic and crucial step in the automatic reconstruction.In this paper a coarse-to-fine method was proposed to segment building roof planes automatically.Firstly,the normal and planar equations of each point were calculated via the method of the low-rank subspace clustering framework with prior knowledge(LRSCPK).Secondly,the buildings were coarsely segmented based on the normal vector consistency and the distance from the point to the plane.Thirdly,the planar equation of the cluster with the largest number of points in the coarse segmentation was estimated,and the points satisfying the planar equation were extracted from the points without fine segmentation.Then the extracted points of the clustering with the maximum number of points were projected on the fitting planes,on which the plane was segmented finely.Lastly,the rest roof planes were extracted by repeating the coarse and fine steps above.The method was validated by using several buildings with different structures,and segmented results were compared with other three algorithms.The results showed that compared with other three algorithms,the building roof planes were accurately segmented by using the proposed method,without spurious roof planes.The method was able to accurately segment the roof of buildings with different structures and provide accurate planes for 3D reconstruction.