An efficient method for extraction of planar features from point clouds was proposed based on the concepts of self-adaptive octree voxel generation and voxel-based region growing. The proposed method involved the selection of correlated thresholds through statistics of voxel information. A voxel-based region growing approach was employed for planar feature extraction, instead of a point-based one. A point cloud was voxelized in initial voxel width and the geometrical features for each voxel were calculated, including the normal vector, eigenvalue, and three dimensionality features. The terminal constraints for octree subdivision were thereby determined through statistics and a list of octree voxels with inhomogeneous sizes was obtained after subdivision. Furthermore, planar facets were extracted through voxel-based region growing at different levels associated with the corresponding statistical threshold constraints. Evaluation experiments were performed by analyzing four different types of point clouds. The obtained experimental results show that the precision and recall rates can reach 95%, which indicates that the proposed method is insensitive to data quality and can be adaptive to various laser-scanned point cloud data. The proposed method can therefore achieve fine planar feature extraction results with high operating efficiency.