基于伪特征点的树点云配准算法研究

发布时间：2018-04-25 00:07

本文选题：三维点云数据 + 伪特征点　；参考：《西北农林科技大学》2017年硕士论文

【摘要】：为了快速从树点云中得到完整的三维形态点云,点云配准必不可少。目前专家学者提出许多点云配准算法,但是树表面粗糙,枝干纤细且相互遮挡,三维扫描仪获得的点云不完整且存在噪点,现存的算法并不能完全适应树点云的独特特点。基于此,本文提出一种基于伪特征点的树点云配准算法,该算法分为初始配准和精配准。论文的主要创新点及其研究内容如下:(1)提出一种伪特征点提取算法。针对树结构复杂,特征点提取困难的问题,采用伪特征点提取算法提取树点云的伪特征点,该方法通过一次分簇、二次分簇、计算伪特征点等步骤完成伪特征点提取,达到使用较少的点精细显示点云特征的目的,得到较好的伪特征点集。实验结果表明,与基于几何特征的特征点提取算法相比,伪特征点提取算法更适应于树点云的特征。(2)提出一种基于伪特征点的树点云配准算法。针对获取点云数据稠密,配准较为耗时的问题,在初始配准中使用提取的伪特征点粗略的调整两片点云的位置,减少精配准的迭代次数,提高配准效率。针对树点云中的噪声点影响提取对应点对正确率的问题,本文在初始配准及其精配准中采用邻域信息分布的相似性来筛除错误的对应点对,提高对应点对的正确率。并针对初始配准和精配准所使用数据的不同特点,分别使用夹角和距离来度量邻域分布的相似性,提高对应点对的正确率,从而改善配准精度。(3)针对基于伪特征点的树点云配准算法的配准性能验证问题,使用有叶及无叶树点云验证该算法的有效性;使用非树点云验证该算法的可扩展性;并在相同实验环境与实验数据的前提下,与其它配准算法比较的方法,验证该算法的优越性。实验表明,在相同迭代次数的前提下,该算法的配准误差比ICP(Iterative Closed Point)算法的配准误差减少41.1%,比SICP(Sparse ICP)算法的配准误差减少16.8%。另外,论文还使用盆栽模型、Bunny等模型来验证算法的通用性。实验表明,该算法也能够配准非树点云,具有较强通用性。
[Abstract]:In order to get a complete three-dimensional morphological point cloud from the tree point cloud, registration of point clouds is essential. At present, experts and scholars have proposed many point cloud registration algorithms, but the tree surface is rough, the branches are thin and each other is obscured. The point cloud obtained by the 3D scanner is incomplete and has noise. The existing algorithms do not fully adapt to the unique special characteristics of the tree point cloud. Based on this, this paper proposes a tree point cloud registration algorithm based on pseudo feature points, which is divided into initial registration and fine registration. The main innovation points and their research contents are as follows: (1) a pseudo feature point extraction algorithm is proposed. The pseudo feature point extraction algorithm is used to extract tree points for the problem of complex tree structure and the difficulty of extracting feature points. The pseudo feature point of a cloud is extracted from a cluster, two clusters and a pseudo feature point to extract the pseudo feature points. A better pseudo feature point set is obtained. The experimental results show that the pseudo feature extraction algorithm is better than the geometric feature extraction algorithm based on the feature point extraction algorithm. To adapt to the feature of tree point cloud. (2) a registration algorithm for tree point cloud based on pseudo feature points is proposed. In order to obtain the dense data of the point cloud, the registration is more time-consuming. In the initial registration, the extracted pseudo feature points are used to roughly adjust the position of two point cloud, reducing the number of iterations of the precise registration and improving the registration efficiency. The noise point affects the correct rate of the extraction of the corresponding point. In this paper, we use the similarity of the neighborhood information distribution to screen out the corresponding point pairs in the initial registration and the fine registration, and improve the correct rate of the corresponding point pairs. The similarity of the cloth improves the correct rate of the corresponding point pair and improves the registration accuracy. (3) the validity of the algorithm is verified by using the leaf and leaf free tree point cloud to verify the validity of the registration performance verification problem based on the pseudo feature point based tree point cloud registration algorithm, and the scalability of the algorithm is verified by using non tree point cloud, and the experimental data are also used in the same experimental environment and experimental data. Under the premise of comparison with other registration algorithms, the superiority of the algorithm is verified. The experiment shows that the registration error of the algorithm is less than that of the ICP (Iterative Closed Point) algorithm by 41.1%, and the registration error of the SICP (Sparse ICP) algorithm is less 16.8%. than the SICP (Sparse ICP) algorithm. The paper also uses the pot model, Bunny and other models are used to verify the universality of the algorithm. Experiments show that the algorithm can also register non tree point clouds, and has strong versatility.

【学位授予单位】：西北农林科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：S126

【参考文献】