A method for generating tree skeletons by thinning the voxels of point cloud data has been proposed based on point cloud data acquired by a terrestrial three-dimensional laser scanner. First, the voxel space is constructed based on the point cloud data of a tree, and the voxel coordinates of point clouds are calculated simultaneously. Second, the noise points in the voxels are filtered according to the statistical information of point cloud data in each voxel. Third, the voxels without noise are thinned using thinning templates. The skeleton nodes are then fitted considering the thinned voxels. Finally, based on the connectivity of natural trees, a depth-first search algorithm is employed to connect nodes and generate tree skeletons. The proposed method is tested with a ginkgo tree and an Amygdalus triloba f. multiplex tree. The two trees are scanned by a terrestrial three-dimensional laser scanner. The effects of different parameters on the tree skeleton are analyzed by using the tree point clouds with different scanning accuracy. In comparison with the GSA method, the proposed method can reduce time consumption for the tree skeleton generation of the ginkgo tree and the Amygdalus triloba f. multiplex tree to 1/30 and 1/67, respectively. Experimental results show that the skeletons of the two trees generated by the proposed method are consistent with the original point clouds,and the proposed method is feasible and efficient.