Studies On Key Technology Of Data Processing And Feature Recognition In Point Clouds

With the advancement of 3D scanning technology, point clouds are more widely used in engineering and lives now. At the same time, those applications put forward higher demanding to the processing and reconstruction of point clouds. Point clouds with no topological information contain primary design intents and feature information, in turn processing data precisely and recognizing feature exactly is the fundamental demands,however it is a challenge to the application of point clouds in-depth. This dissertation does some researches on feature re cognition and its relative technologies. The main works are listed as follows:(1)The size of point data increases rapidly due to the advancement of scanning precision.To improve the performance of k nearest neighbors search for large scale point clouds, we present an extracting algorithm to fulfill kNN search. Based on inner product of vectors instead of distances calculation and comparison, the determining rule for extraction neighbors is constructed by using the overlap property between neighborhoods. The proposed algorithm extracts the k nearest neighbors by using the rule and improves the speed of kNN search,especial for large scale point clouds.(2) Point clouds present non-uniformity and anisotropy clearly. To create the geodesic path on point clouds,a novel approach for computing geodesic path based on non-uniform grid and forward track is proposed. The principal marching direction is selected,and the principal marching region is calculated and split into cells non-uniformly. Upon the cells, the numerical computation method of unilateral compact difference is constructed and integrated into fast marching method to evaluate the numerical value of cells. Then forward track method, using positive orientation condition and the properties of geodesics, is employed to determine the reasonable cells to form a geodesic path and the principal normal vectors of points on path are obtained at the same time. The proposed approach improves the precision of geodesic path and avoids the failure of back track method which will across the boundary of grid in the process of back track.(3)Normal is often estimated by fitting micro tangent plane at the neighbors of the current point. To estimate the normal of point clouds, a new normal estimation method based on the output of geodesic path is proposed. The current point is start point, and the progressive neighbors of the current point within its neighborhood are selected to be end point of geodesic path fast. In turn two geodesic paths are formed and principal normal vectors are obtained along the paths. The normal planes are archived by fitting the principal normal vectors, and the intersection line of two normal planes is used to approximate the normal vector of the current point. The presented method provides a solution to estimate normal vectors for point clouds,especial for sharp feature whose normal vectors estimation require good neighborhoods which are different to choose.(4)Feature extraction requires the estimation of geometric properties such as curvature and the decision of variation trend. To detect the feature point and extract feature curve, we proposed a novel feature extraction method based on discrete Laplace operator of point clouds.In view of feature points are collinear with a potential feature line locally, construct a local spherical frame for the neighborhood centered at the given point, calculating theirs coordinates,grouping the points that having approximate coordinate into the same line segment and sort these segments according their coordinate in ascending order. Discrete Laplace operator is applied to detect the potential feature points and feature curves,meanwhile the link sequence of feature points, the link region information of feature curves for the convenience of the reconstruction of feature curves are recorded in the extraction stage. The proposed approach improves the reliability of feature extraction near the sharp feature,and avoids the blindness of comparison geometric properties randomly used in the traditional way which results in the reconstruction difficulty of feature lines.(5)Recognizing the geometric shape and extracting the geometric parameters from the unstructured point clouds is the core of surface recognition of point clouds. To recognize the surface underlying point clouds, we discuss the surfaces sampled from five primitives in two stages, geometric shape recognition and geometric parameters extraction. In the first stage,Gauss mapping is employed to classify surfaces into two groups that is plane-cylinder-cone and sphere-torus, the former is distinguished from each other using the eigen saliency, and the later is done by the discrete Laplace-Beltrami operator. In the second stage,fitting method is used to extract the geometric parameters of known shapes from point regions respectively. The proposed method archives the recognition of primitives, and simplifies the distinction between the sphere and torus especially.