Fig. 1. Flowchart of the proposed method

Fig. 2. Point cloud displayed with intensity as eigenvalue

Fig. 3. Neighborhood point selection method. (a) KNN; (b) spherical neighbor

Fig. 4. Point cloud displayed with omnivariance as eigenvalue

Fig. 5. Point cloud displayed with planarity as eigenvalue

Fig. 6. Point cloud displayed with verticality as eigenvalue

Fig. 7. Improved RandLA-Net. (a) Network structure; (b) MFF; (c) LFA

Fig. 8. Toronto-3D semantic information

Fig. 9. Local road marking point cloud

Fig. 10. WHU-MLS semantic information

Fig. 11. WHU-MLS Local road marking point cloud

Fig. 12. Toronto-3D semantic segmentation results. (a) RandLA-Net; (b) proposed method

Fig. 13. WHU-MLS semantic segmentation results. (a) RandLA-Net; (b) proposed method

Fig. 14. Segmentation results at different sampling rates. (a) 0.06 sampling rate; (b) 0.02 sampling rate

Fig. 15. Toronto-3D road marking extraction results. (a) Truth value; (b) method of reference [8]; (c) method of reference [9]; (d) proposed method

Fig. 16. WHU-MLS road marking extraction results. (a) Truth value; (b) method of reference [8]; (c) method of reference [9];(d) proposed method