【作者】 李鑫；

【导师】 黄华； 宋学勇；

【作者基本信息】 四川大学 ， 控制工程（专业学位）， 2022， 硕士

【摘要】 三维重建作为计算机视觉的一项重要研究内容,在应急救灾、文物重建、高精地图、自动驾驶、游戏等行业有着广泛的应用。而对目标或场景进行三维重建时,数据通常无法只通过同一时间、视角的单一传感器获得,需要结合不同时间、视角以及不同传感器采集,并用多种算法完成重建。本文主要研究多视图影像三维重建算法、点云配准算法,并将其用于地质灾害现场等数据的影像三维重建及激光点云配准融合,以得到更完整的三维目标信息。本文主要研究工作如下:(1)、基于运动恢复结构的三维重建关键算法,包括相机模型、特征提取、特征匹配、多视图几何、三角化、PNP、SFM稀疏重建等,对上述理论基础进行数学推导及代数描述;对重建过程进行深入研究,对重建的各个阶段进行实验和分析,阐述实际应用过程中的优化方法,并使用公开数据集以及无人机采集的校园建筑、文家沟灾害现场影像进行稀疏以及密集重建工作,最终获得影像三维重建的稠密点云。(2)、针对多源点云配准存在噪声、部分重叠、不同模型的配准参数难确定等问题,提出一种基于贡献因子的改进Tr ICP算法。首先,提出结合改进体素降采样及随机采样的方法对点云进行降采样。然后,利用改进算法的贡献因子来保留对配准贡献度更大的点对,使用奇异值分解法(SVD)求解变换矩阵,同时计算距离曲线上的点经过原点的斜率来自动计算重叠度,实现点云的全自动配准。使用斯坦福大学的Bunny点云以及“茂县624”滑坡现场点云等数据对改进算法、Tr ICP、3D-NDT、4PCS等多个算法进行对比实验。结果表明:相对于Tr ICP,改进算法在Bunny点云以及滑坡体点云上,配准速度分别提升50%和67%,且精度更高;相对于3D-NDT与4PCS,改进算法同样有更高的精度和速度,并且对参数不敏感,并在添加大量噪声情况下仍能正确配准,这表明该算法能对含大量噪声、部分重叠、非同源的激光与影像重建点云进行可靠高效的自动配准。(3)、对结合RANSAC以及FPFH特征的点云配准算法进行了研究并探索其在配准中的优化,实验结果表明该算法在不同数据上均得到良好的初始位姿;使用FPFH+RANSAC作为初始配准,本文改进算法作为精配准算法,在WHUTLS公开数据集、校内多站激光点云数据、灾害现场非同源的激光与影像重建点云上进行实验,结果表明,在WHU-TLS上的最大轴旋转误差为0.20599°,最大平移误差0.17 m,成功配准最大176.108°的校内点云以及非同源、大量噪声的灾害现场点云。说明该方法对不同数据、大角度及噪声鲁棒,能有效地对大场景下多站激光雷达扫描点云以及影像重建点云进行配准融合,得到更完整的目标点云,为三维重建以及灾害现场的三维数据融合分析提供了基础。更多还原

【Abstract】 As an important field in computer vision,3D reconstruction is widely used in emergency relief,cultural relic reconstruction,high-precision map,autonomous driving,games,etc.When trying to reconstruct an object or a scene,the complete info of the scene can not be obtained simply from a single sensor with a same space and time,it’s necessary to combine multi-sensors to obtain data with different space and time,and use different algorithms to register and fuse the whole 3D scene.This paper mainly studies the multi-view 3D reconstruction algorithm and point cloud registration algorithm,and applies them to the registration and fusion between 3D reconstruction of geological disasters and laser point cloud to obtain more complete3 D scene information.The main research work of this paper is as follows:(1)Key algorithms of 3D reconstruction based on Structure from motion,including camera model,feature extraction,feature matching,multiple view geometry,triangulation,PNP,SFM sparse reconstruction,etc.And the theoretical basis above were mathematically deduced and algebraic described;Key steps of reconstruction were deeply studied,experiments and analyses were performed on different stages of reconstruction,the optimization methods in the process of practical application were described,and the sparse and dense reconstruction work is carried out by using the public data set,campus architecture and the disaster scene images of Wenjiagou collected by UAV,to obtain the dense point cloud from image three-dimensional reconstruction.(2)In order to solve the problems(noise,partial overlap,registration parameters determining for different models,etc.)in multi-source point cloud registration,an improved Tr ICP algorithm based on the contribution factor was proposed.First,the improved voxel down-sampling and random down-sampling methods were adopted to resample the point cloud.The contribution factor was proposed to investigate the point pairs that contributed more to the registration.The transformation matrix was solved by using singular value decomposition.At the same time,slopes between points on the distance curve and the original point were used to calculate the overlap automatically.Therefore,the automatic registration of point cloud was realized.Comparative experiments among several registration algorithms were conducted based on the Stanford University Bunny point cloud and the ’Maoxian 624’ landslide point cloud.Compared with Tr ICP,speeds of the improved algorithm on Bunny and landslide increase by 50 % and 67 % respectively and accuracies are improved;Compared with 3D-NDT and 4PCS,it also has higher accuracy and faster speed and is insensitive to parameters.In addition,it performs well even with a lot of noise.As a result,the improved algorithm can align the LIDAR point cloud and point cloud from image reconstruction effectively and automatically,which contain lots of noise,partial overlap,non-homology.(3)The point cloud registration algorithm combined with RANSAC and FPFH features is studied and its optimization in registration is explored.The experimental results show that the algorithm can obtain good initial pose on different data;Using FPFH + RANSAC as the initial registration and the improved algorithm as the fine registration algorithm,experiments are carried out on the WHU-TLS public dataset,the laser point clouds data in the school,and the non-homologous laser and image reconstruction point cloud from the disaster site.The results show that the maximum axis rotation error on WHU-TLS is 0.20599 ° and the maximum translation error is0.17 m.The point cloud from campus with 176.108 ° rotation angle and the nonhomologous point cloud from disaster site with a large amount of noise were successfully registered.It shows that this method is robust to different data,large angles and noise,and can effectively register and fuse the multi station lidar scanning point cloud and image reconstruction point cloud in a large scene,so as to obtain a more complete target point cloud,which provides a basis for 3D reconstruction and3 D data fusion analysis of disaster scene.更多还原