基于并行计算的点云配准算法研究

发布时间：2018-04-09 14:04

本文选题：OpenCL　切入点：ICP　出处：《北方工业大学》2016年硕士论文

【摘要】：随着三维扫描技术的发展成熟与普及,三维应用技术得到了较好的发展。尤其是在逆向工程、文物保护、缺陷检测、医学手术、3D打印以及游戏娱乐等领域获得了广泛的关注与研究。由于部分遮挡等问题,所以为了获得完整的点云数据,需要进行多次扫描拼接配准。而三维点云配准一直以来都是三维扫描问题中的关键技术。本文在前人的研究基础上,主要研究了点云配准技术以及GPU并行技术,研究并实现了采用近似最近点的ICP迭代最近点并行配准算法与Kinect现场取景扫描配准建模系统。本文的主要研究成果如下：针对kd-tree搜索算法回溯查询效率低且其与ICP迭代最近点点云配准算法结合不紧密等问题,本文提出了基于ICP迭代最近点点云配准算法特性的kd-tree近似最近点查找算法及其GPU并行实现,通过非递归建树,动态控制建树层次,控制回溯空间等步骤完成对应点集的获取。最后通过实验证明所提出近似最近点查找算法,有效的提高了对应点集的查找效率,进而提高了ICP迭代最近点点云配准算法的时间效率。针对ICP迭代最近点点云配准算法对大数量集数据配准效率低以及当前GPU算法的设备局限性,本文提出了将CPU下的近似最近点ICP配准算法,通过OpenCL通用GPU并行平台移植到GPU中,采用GPU对算法进行并行加速。最后通过实验证明本文所实现GPU下的ICP迭代最近点点云配准算法,具有较好的通用性与时间效率。设计并实现了Kinect现场取景扫描配准建模系统,该系统主要实现通过现场数据采集完成模型配准。通过结合Kinect深度相机、特征提取粗配准算法以及本文所实现算法,所实现的系统包括数据采集,三维重建,模型预处理,模型配准等功能模块。系统实验效果图,显示所实现系统能够处理复杂场景,并对噪声具有一定的鲁棒性。
[Abstract]:With the development and popularization of 3D scanning technology, 3D application technology has been well developed.Especially in reverse engineering, cultural relic protection, defect detection, medical surgery 3D printing and game entertainment and other fields have received extensive attention and research.Due to partial occlusion and other problems, in order to obtain complete point cloud data, it is necessary to perform multiple scanning stitching registration.And 3D point cloud registration has always been a key technology in 3D scanning.Based on the previous researches, this paper mainly studies the point cloud registration technology and GPU parallel technology, and studies and implements the ICP iterative nearest point parallel registration algorithm and the Kinect scene scanning registration modeling system using approximate nearest points.The main research results of this paper are as follows: aiming at the low efficiency of backtracking query of kd-tree search algorithm and the problem that it is not tightly combined with the nearest point cloud registration algorithm of ICP iteration,In this paper, based on the characteristics of ICP iterative nearest point cloud registration algorithm, the kd-tree approximate nearest point search algorithm and its GPU parallel implementation are proposed. The corresponding point sets are obtained by non-recursive tree building, dynamic control of tree building hierarchy and control of backtracking space.Finally, it is proved by experiments that the proposed approximate nearest point search algorithm can effectively improve the search efficiency of the corresponding point set, and then improve the time efficiency of the ICP iterative nearest point cloud registration algorithm.In view of the low efficiency of ICP iterative nearest point cloud registration algorithm for large number of data sets and the equipment limitation of current GPU algorithm, this paper proposes an approximate nearest point ICP registration algorithm based on CPU, which is transplanted to GPU through OpenCL general GPU parallel platform.GPU is used to accelerate the algorithm in parallel.Finally, it is proved by experiments that the ICP iterative nearest point cloud registration algorithm implemented in this paper has good generality and time efficiency.The Kinect scene scanning registration modeling system is designed and implemented. The system mainly realizes the model registration through field data acquisition.Through the combination of Kinect depth camera, feature extraction rough registration algorithm and the algorithm implemented in this paper, the system includes data acquisition, 3D reconstruction, model preprocessing, model registration and other functional modules.The experimental results show that the system can deal with complex scenes and is robust to noise.
【学位授予单位】：北方工业大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP338.6

【相似文献】