Research On The Key Technologies Of Three-dimensional Scattered Point Cloud Processing

As one of the most important data representation in computer graphics,the 3D point cloud data has been widely used in reverse engineering,metrology,CAD/C AM,robotics,virtual reality,3D images,cultural relic protection,indoor scene reconstruction,and many other fields.With the rapid development of 3D scanning technology,the large-scale and complex unorganized point cloud data emerges,and brings new challenge to the traditional point cloud data processing technologies,restricts its further application and development,Therefore,in order to improve the processing effect and efficiency,in this paper,the key processing technologies of the 3D scattered point cloud data are researched.The main contents are as follows.Firstly,the ICP algorithm is sensitive to the initial value and has low efficiency,in order to solve these problems,a new registration algorithm is proposed.This method makes full use of the effective information provided by RGB-D data,using feature extraction algorithm to detect the key points,determines the descriptor vector of the corresponding key points,and finds and optimizes the matching key points using the interative closest point algorithm and the principle of maximum vector inner-product.Calculates the key point curvature,and deletes the error matched points by the rule of curvature consistency,to obtain the key points with higher registration efficiency.Calculates the transformation matrix and register all the points.Experimental results verify the Effectiveness and efficiency of the algorithm.Secondly,in order to improve the detection ability and adaptivity,get feasible detection result,and solve the problem of unsatisfactory detection when the point cloud data has larger density distribution changing,a n outliers detection algorithm which is based on dynamic standard deviation threshold using k-neighborhood density constraints is proposed.This method takes full consideration of difference density,and introduces the density characteristics into the calculation of the determining threshold,using different constraints to adjust the dynamic standard deviation for outer regions and inlier regions.Experimental results verify that this method can get a better detection from point cloud with big change density distribution,and provide a feasible data for the point cloud segmentation.Thirdly,in order to improve the segmentation sensitivity to the local detail for 3D point cloud processing,especially when the surface has more and complex characteristics,a new segmentation algorithm which is based on curvature information is proposed.The curvature information was introduced into point distance calculation,meanwhile,a selection method of initial clustering centers based on cube voxel grid is proposed to avoid the dependency on initial centers and improve efficiency,and then segment the point cloud according to the K-means clustering theory.By adjusting the parameter,this me thod can be applied to the point cloud with different surfaces changing,the initial centers selection method ensures the uniqueness and effectiveness of the segmentation result,reduces segmentation overheads,and improves efficiency.Finally,a reverse reconstruction algorithm is proposed to ensure preserving the features in point cloud surface reconstruc tion.This method based on the poisson reconstruction theory,by introducing the gauss filter smoothing method to compensate the detail smoothing drawback when using the 3 mean filter method,meanwhile,based on “divide and rule” theory,a depth parameters selection is modeled to determine the reconstruction depth according to the characteristics of different point cloud regions.This method retains the detail characteristics when getting smooth surface,reduces reconstruction overheads,and improves the surface reconstruction effectiveness and efficiency.