Local Feature Extraction And Applications In Large-scale Point-Clouds

With the rapid development of laser scanning technology,related applications of 3D point clouds have become increasingly popular in various fields.Local feature extraction of 3D point cloud is a key scientific issue and plays an important role in the application of point cloud registration,semantic segmentation of point cloud,and 3D object recognition and classification.However,when dealing with the point clouds with noise,outliers,severe occlusion,and uneven distribution of point density,the local feature extraction algorithms still can not work well.This thesis proposes a theoretical framework for keypoint detec-tion and local feature description and then develops a pairwise point cloud registration method and a 3D road boundary extraction method.The contributions of this thesis can be summarized as:(1)This thesis proposes a covariance-based local feature extraction framework for keypoint detection and local feature description;(2)This thesis develops a game-theoretic matching method for pairwise registration of point clouds;(3)This thesis develops a supervoxel algorithm and a graph cuts algorithm to effectively extract 3D road boundaries.In terms of local feature extraction of point clouds,this thesis proposes a feature fusion algorithm based on covariance matrix.The basic idea of the algorithm is to fuse the ge-ometric features,color features,and intensity features contained in the point cloud data through the statistical characteristics of the covariance matrix.The experiments on three point cloud models demonstrates the effectiveness and superiority of the proposed algo-rithm.For pairwise point cloud registration,this thesis proposes a registration framework based on a non-cooperative game.The algorithm firstly uses covariance feature descriptors to perform feature matching to obtain the correspondence set.Then,by applying global geometric constraints through a non-cooperate game,a large number of false correspon-dences are iteratively excluded until they converge to be compatible with each other.The plausible true correspondences are finally isolated to calculate the transformation matrix.Experimental results demonstrate that the registration algorithm proposed in this paper is superior to the state-of-the-art registration algorithms in terms of both precision and ef-ficiency.The average registration errors achieved on three models are 0.46cm,0.32cm and 1.73cm,respectively.The computational times cost on these models are about 288s,184s,and 903s,respectively.Moreover,the test on real scenes verifies that the registra-tion algorithm proposed in this thesis can be used for the registration application of real large-scale point clouds.In terms of 3D road boundary extraction,this thesis proposes a new algorithm based on supervoxel segmentation and graph cuts.First,the algorithm improves the supervoxel seg-mentation algorithm based on the local feature extraction to segment the point clouds into a series of facets.Then the ?-shape algorithm is used to extract the boundaries between ad-jacent non-coplanar facets,and a graph cut model is built to extract road boundary points.The proposed method is tested on two datasets,and achieves an average precision over 98%,an average recall over 95%and an average F1-measure over 97%on two datasets.The effectiveness and superiority of the proposed method over the state-of-the-art methods is demonstrated.