【作者】 张强；

【导师】 刘玉珍； 林森；

【作者基本信息】 辽宁工程技术大学 ， 通信与信息系统， 2022， 硕士

【摘要】 随着高精度测量传感器的快速发展,用三维点云描绘世界的方式开始被广泛使用。点云配准作为三维点云处理过程中的重要一环,是完成三维重建、虚拟现实、文物修复、智能交通等实际工程的关键技术。ICP算法是应用最广泛的点云配准算法,为解决该算法本身的局限性,并提高点云配准的精度和效率,本文分别从提取特征点、计算描述符、获取匹配点对和精确配准4个方面进行研究,提出了两种点云配准算法:为解决噪声干扰、数据丢失情况下ICP算法鲁棒性差,配准精度和配准效率低的问题,提出改进的基于快速点特征直方图的ICP点云配准算法。首先,融合内部形态描述子和法向矢量角变化来提取点云特征;其次,使用指数函数改进欧氏距离,作为FPFH算法的权重系数,用其进行特征点描述;然后使用双重约束和单位四元数算法完成初始配准;最后,给ICP算法构建双向k维树,并提出用距离计算每个点对的权重,作为ICP迭代误差函数的加权公式。实验数据集用斯坦福模型和两组实际点云,实验结果表明,该算法解决了ICP算法在噪声干扰、数据丢失环境下的局限性。相较于其他算法,实物点云的配准精度至少提高11%。为解决配准部分重叠物体点云时,ICP算法配准精度和配准效率低的问题,提出基于邻域点信息描述与匹配的点云配准算法。首先,在三个半径比例下根据点的曲率变化、测量角度和特征值性质提取特征点;其次,计算改进的法向量夹角、点密度和曲率值,获取多尺度矩阵描述符;然后,为描述符建立k维树获取匹配关系,并提出几何特征约束和刚性距离约束组合,剔除错误点对,实现粗配准;最后,通过k维树改进ICP算法完成精确配准。研究设计了实际物体点云配准和斯坦福模型模拟真实物体配准两组实验,实验结果表明,相较于其他算法,实际部分重叠点云配准中该算法的配准精度、效率至少提高29%、40%;斯坦福模拟实验中,该算法的配准精度、效率至少提高11%、12%。该论文有图55幅,表9个,参考文献69篇。更多还原

【Abstract】 With the rapid development of high-precision measurement sensors,3D point clouds are widely used to depict the world.As an important part of 3D point cloud processing,point cloud registration is a key technology to complete practical projects such as 3D reconstruction,virtual reality,cultural relic restoration,and intelligent transportation.The ICP algorithm is the most widely used point cloud registration algorithm.In order to solve the limitations of the algorithm itself and improve the accuracy and efficiency of point cloud registration,this paper focuses on extracting feature points,calculating descriptors,obtaining matching point pairs and accurate four aspects of registration are studied,and two point cloud registration algorithms are proposed:In order to solve the problems of poor robustness,low registration accuracy and efficiency of ICP algorithm in the case of noise interference and data loss,an improved ICP point cloud registration algorithm based on FPFH is proposed.First,extract point cloud features by fusing internal morphological descriptors and normal vector angle changes;secondly,use exponential function to improve Euclidean distance as the weight coefficient of FPFH algorithm,and use it to describe feature points;then use double constraints and unit four the arity algorithm completes the initial registration;finally,a bidirectional k-dimensional tree is constructed for the ICP algorithm,and the weight of each point pair is calculated by the distance as the weighting formula of the ICP iterative error function.The experimental data set uses the Stanford model and two sets of actual point clouds.The experimental results show that the algorithm solves the limitations of the ICP algorithm in the environment of noise interference and data loss.Compared with other algorithms,the registration accuracy of physical point clouds is improved by at least 11%In order to solve the problem of low registration accuracy and registration efficiency of the ICP algorithm when registering partially overlapping object point clouds,a point cloud registration algorithm based on the description and matching of neighborhood point information is proposed.First,the feature points are extracted according to the curvature change,measurement angle and eigenvalue properties of the points under three radius ratios;secondly,the improved normal vector angle,point density and curvature value are calculated to obtain the multi-scale matrix descriptor;then,for the descriptor establishes a k-dimensional tree to obtain the matching relationship,and proposes a combination of geometric feature constraints and rigid distance constraints to eliminate wrong point pairs to achieve rough registration;finally,the k-dimensional tree is used to improve the ICP algorithm to complete accurate registration.The research designed two sets of experiments of real object point cloud registration and Stanford model simulated real object registration.The experimental results show that,compared with other algorithms,the registration accuracy and efficiency of the algorithm in the actual partial overlapping point cloud registration are improved by at least 29% and 40%;in the Stanford simulation experiment,the registration accuracy and efficiency of the algorithm are improved by at least 11% and 12%.There are 55 figures,9 tables and 69 references in this paper.更多还原