A Laser Scanner Point Cloud Registration Method Using Difference of Normals (DoN) Based Segmentation
J-H. Lyu, Z-W. Wang, Y. Zhou, J. Meng and S-F. Wang
The coarse registration is required before fine registration to obtain a better initial pose; however, it is difficult for coarse registration to obtain a better initial pose when the overlapping coefficient of source and target point cloud is low. To overcome this we propose a registration method using difference of normals (DoN) based segmentation algorithm and sample consensus initial alignment algorithm. First, the DoN based segmentation algorithm is employed to segment areas where obvious normal differences from source and target point cloud. Afterwards, Gaussian models for these subsets are established to find a pair of point cloud subsets with the most similar distribution. Then the sample consensus initial alignment algorithm (SAC-IA) is employed to register the matched subsets to obtain transformations between them, applied to source point cloud to find a better initial pose. Eventually, the iterative closest point (ICP) algorithm is involved to complete the registration. The experiments show the root mean square error (RMSE) of the registered point cloud using proposed method has been improved by 0.435 m compared with that of previous algorithm when the overlapping coefficient is 5%. The analysis shows the segmented region has clearer normal features; consequently, the SAC-IA algorithm to extract a more effective feature histogram to obtain a more accurate transformation.
Keywords: Laser scanner, point cloud registration, iterative closest point (ICP) algorithm, difference of normals (DoN), segmentation algorithm, sample consensus initial alignment algorithm (SAC-IA)