A Sift-based Registration Method Of Terrestrial Laser Scanning Point Clouds

Laser scanning is an emerging fast data acquiring method which could retrieve3Dinformation of target surface accurately without touching. It has become one of the basismethods for3D reconstruction of complicated targets. Because of terrestrial laser scanners’work mode and self occlusion on the silhouette of the object, in many cases the object hasto be scanned from different viewpoints to acquire its full data. In order to rebuild realisticdetailed3D models, all point clouds must be transformed to a consistent coordinate system.This procedure is called point cloud registration. Since the accuracy of point cloudsregistration directly affects the quality of final3D models, point cloud registration researchis always a hot topic in point clouds processing. But rapid modeling using point clouds ishindered by the existing method’s limitations in accuracy, efficiency and range ofapplications.Considering such a situation, after analyzing present point cloud registration methods,a SIFT-based registration of terrestrial laser scanner point clouds is proposed in this paper.First, in order to display the implicit structured information visually, a novel image methodfor storing point cloud quickly and accurately which lays the foundation of point cloudregistration is developed through2.5D characteristics of terrestrial laser scanning data.Second, the method is brought forward in this paper. Geometric invariant is determined onthe basis of analyzing3D geometrical models, fast computation and storage of normalinformation is achieved based on3D coordinates and image pyramids in different scalesare built from the convolution of a variable-scale Gaussian. Then, keypoints are foundedby detecting local maxima and minima in different scale space in order to acquire stablefeatures. Meanwhile, invariant descriptor vector is calculated according to histogramstatistics on normal using statistical classification. Matching features are extracted bycomparing Euclidean distance between different invariant descriptor vectors as thesimilarity measure. In this process, registration of point clouds is accomplished. Lastly, thefeasibility and efficiency of our proposed method is demonstrated by field derived buildingdate.