Research On Point Cloud Data Processing Technology In 3D Reconstruction

With the point cloud data model more and more widely being used in three-dimensional solid modeling, 3D reconstruction have been a hot research in many application fields, such as reverse engineering, industrial inspection, preservation of cultural relics, medical image processing and so on. The original point cloud data is often faced many problems caused by device shaking, human factor, surroundings factor, etc. These problems can lead the three-dimensional point cloud data polluted with some noise. Furthermore, the huge amount of data will also affect the efficiency of the display,storage and processing in model reconstruction. The study of three-dimensional point cloud data processing has important significance.Point cloud data processing involve point cloud data denoising, point cloud data simplification, point cloud data registration, feature recognition, regional split, geometry estimation and model reconstruction. The main works of this paper focus on three points as follows:Firstly, a point cloud denoising algorithm based on improved K-means clusterin g is proposed to extract the desired target point cloud from noisy scattered point cl oud data. This algorithm can automatically identify and remove the local outlier poi nts while preserving the original useful cloud point data as possible.The experiment al results show that it does not need to traverse the global data during iterating, w hich can reduce the number of iterations and the time consumption, and improve th e denoising efficiency.Secondly, a point cloud simplification algorithm, combined with point cloud curvature and the normal vector, is proposed to simplify the densely point cloud data. This algorithm can build spatial topological relationship of scattered point cloud data by the improved K-means clustering algorithm. Then, a feature factor, calculated from the weighted point cloud curvature and the normal vector, is used to determined which points are feature points. Finally, the feature points are conserved by setting a threshold of proper feature factor to simplify the point cloud data in clusters. The experimental results show that the algorithm improves the simplify efficiency, also can preserve the detail characteristics of point cloud data and geometric shapes.Finally, in the process of point cloud data registration, the speed of the existing Iterative Closest Point(ICP) algorithm searching the corresponding points is slow and the efficiency of point cloud data registration is low. A new point cloud data registration algorithm based on boundary detection is proposed, which combined with the point cloud center of mass distance feature. On the premise of implementing the coarse registration and obtaining good tracking, the boundary detection in terms of centroid distance is adopted to extract point cloud edges from the point cloud data in different angles,On this basis, the algorithm uses the K-D tree to search proximate points between the two point clouds boundary, and obtains the transformation matrix by the unit quaternary method. Accordingly, it can rapidly realize point cloud data accurate registration. Compared with the classical ICP algorithm and the existing improved ICP algorithm, the experimental results show that this algorithm based on precision effectively increases the registration efficiency, which is more suitable for large amount of point cloud data registration.