散乱点云特征提取和聚类精简技术研究

发布时间：2018-07-31 08:17

【摘要】：近年来,反求工程技术在汽车制造、航天航空和医学等领域得到了广泛的应用,而随着三维扫描技术的进步,通过三维扫描获取的点云数据量十分庞大,在实际的应用中存在着数据冗余的问题,因此对点云数据的精简成为了当前的热点研究问题。本文的研究内容主要分为点云特征提取、点云聚类和点云精简三个方面,具体内容如下:首先,在点云特征提取方面,针对以往散乱点云特征提取算法存在尖锐特征点提取不完整以及无法保留模型边界点的问题,提出了一种基于多判别参数混合方法的散乱点云特征提取算法。该方法对于每个k邻域计算数据点曲率、点法向与邻域点法向夹角的平均值、点到邻域重心的距离、点到邻域点的平均距离,据此四个参数定义特征阈值和特征判别参数,特征判别参数大于阈值的点即为特征点。以上四个参数中,曲率、法矢夹角和数据点到其邻域点的平均距离三个参数参与检测曲面的尖锐点,而点到邻域重心的距离则主要用于识别边界数据点,同时也能为检测曲面尖锐点提供一定的作用。实验结果表明,与已有算法相比,该算法不仅可以有效提取尖锐特征点,而且能够识别边界点。其次,在点云聚类方面,针对传统K-means聚类算法应用于点云数据时存在迭代收敛时间长、多次运行的聚类结果具有随机性以及聚类效果较差的问题,提出了一种基于自适应八叉树的点云K-means聚类算法。该方法利用自适应八叉树为K-means聚类提供与点云密度分布相关的初始化聚类中心和K值,然后迭代输出聚类结果。实验表明该方法在聚类的评价函数值和运行时间上都优于传统的K-means聚类,而且消除了多次运行时聚类结果的随机性。最后,在点云精简方面,首先利用本文提出的散乱点云特征检测方法提取点云特征点,然后对点云进行基于自适应八叉树的K-means聚类操作,最后在不包含特征点的聚类中以距离聚类重心最近的数据点代替整个聚类,其他数据点删除。为了保留模型的细节特征,在包含特征点的聚类中,选择该聚类所包含数据点中曲率差值最大的两个点作为新的初始化聚类中心再次进行聚类细分,直到聚类中数据点的最大曲率差小于阈值或者聚类中只有一个数据点为止,最终同样以距离聚类重心最近的数据点代替整个聚类。通过实验对比,精简后的数据点分布均匀没有空洞,精简误差以及用于片状点云时因为边界收缩而产生的误差较小,从而能够使精简算法适用于封闭及片状的点云数据类型。
[Abstract]:In recent years, reverse engineering technology has been widely used in the fields of automobile manufacture, aerospace and medicine, but with the progress of 3D scanning technology, the amount of point cloud data obtained by 3D scanning is very large. There is a problem of data redundancy in practical applications, so the reduction of point cloud data has become a hot research issue. The research contents of this paper are mainly divided into three aspects: point cloud feature extraction, point cloud clustering and point cloud reduction. The specific contents are as follows: first, in point cloud feature extraction, Aiming at the problem that the feature extraction of scattered point cloud is incomplete and unable to preserve the boundary points of the model, a new feature extraction algorithm for scattered point cloud based on mixed multi-discriminant parameters is proposed. This method calculates the curvature of data points for each k neighborhood, the average angle between normal point and neighborhood point, the distance from point to center of gravity, and the average distance from point to neighborhood point. According to the four parameters, the characteristic threshold and characteristic discriminant parameters are defined. The point at which the feature parameter is greater than the threshold is the feature point. Among the above four parameters, curvature, normal vector angle and the average distance from the data point to its neighborhood point participate in the detection of sharp points of the surface, while the distance from the point to the center of gravity of the neighborhood is mainly used to identify the boundary data points. At the same time, it can also provide a certain role for the detection of sharp points of surfaces. Experimental results show that the proposed algorithm can not only extract sharp feature points effectively, but also recognize boundary points. Secondly, in the aspect of point cloud clustering, the traditional K-means clustering algorithm has the problems of long iterative convergence time, randomness and poor clustering effect when it is applied to point cloud data. A point cloud K-means clustering algorithm based on adaptive octree is proposed. In this method, the adaptive octree is used to provide the initial cluster center and K value related to the point cloud density distribution for the K-means clustering, and then the clustering results are output iteratively. The experimental results show that the proposed method is superior to the traditional K-means clustering in evaluation function and running time, and the randomness of the clustering results at multiple runs is eliminated. Finally, in the aspect of point cloud reduction, the point cloud feature points are extracted by using the scattered point cloud feature detection method proposed in this paper, and then the point cloud is extracted by K-means clustering operation based on adaptive octree. Finally, in the clustering without feature points, the nearest data point is replaced by the whole cluster, and the other data points are deleted. In order to preserve the detailed features of the model, the two points with the largest curvature difference among the data points included in the cluster are selected as the new initialization clustering center to subdivide the cluster again. Until the maximum curvature difference of the data points in the clustering is less than the threshold or there is only one data point in the clustering, the data points closest to the center of gravity of the clustering are also replaced by the data points of the whole clustering. The experimental results show that there are no holes in the uniform distribution of the reduced data points, and the reduction error and the error caused by the boundary contraction when used in the flake point cloud are smaller, so that the reduced algorithm can be applied to the closed and flaky point cloud data types.
【学位授予单位】：西南科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP311.13

【相似文献】