快速鲁棒的大场景三维重建

CORC > 自动化研究所 > 中国科学院自动化研究所 > 毕业生 > 博士学位论文

题名	快速鲁棒的大场景三维重建
作者	崔海楠
学位类别	工学博士
答辩日期	2016-05
授予单位	中国科学院大学
授予地点	北京
导师	胡占义
关键词	从运动恢复结构先验信息融合三维建模广义摄像机模型匹配点轨迹选取
中文摘要	随着图像传感器技术的快速发展，大量高分辨率的图像对基于图像的快速大场景三维重建算法提出了巨大的挑战。本文主要针对大场景三维重建中速度和精度的两个问题进行系统研究，提出多种快速鲁棒的全局式大场景三维重建算法，主要工作包括以下几方面： 1. 针对有先验信息的图像重建，提出了一种融合有噪先验信息的大场景三维重建方法。首先利用图像拍摄时的先验信息，比如GPS信息、IMU信息以及指南针信息等，初始化摄像机的拍摄位姿。然后，根据初始的摄像机旋转矩阵，迭代地判别外极几何图中边的潜在内点，并利用非线性最小二乘方法优化摄像机旋转矩阵。在此基础上，根据初始的摄像机位置，迭代地发现匹配点轨迹中的潜在内点，通过反复捆绑调整优化摄像机的位置和空间三维点云。大量不同类型图像重建实验表明，我们的方法与很多优秀的重建方法在重建精度和场景结构上相当，但是在大场景重建速度和鲁棒性上我们表现更好。 2. 针对多摄像机拍摄系统，提出一种全局式融合广义摄像机模型的大场景三维重建方法。在多摄像机系统拍摄图像的过程中，多个摄像机内部的相对旋转矩阵和相对平移关系是保持不变的。我们将此先验信息融入到摄像机位姿参数优化中，通过引入广义摄像机模型，使得最终捆绑调整得到的摄像机位姿更加符合真实拍摄的物理模型。通过对街景车和无人机倾斜射影两种多摄像机拍摄系统进行真实图像测试，该方法不仅提高了三维重建系统的速度，而且重建得到的场景结构和精度也得到一定的提升。 3. 为了提高大场景三维重建中捆绑调整的速度，提出一种基于外极几何图覆盖来进行匹配点轨迹选取的方法。首先，我们根据匹配点轨迹的长度、尺度和平均重投影误差三项准则，对匹配点轨迹的可信度进行排序。然后，通过覆盖外极几何图的生成树以保证摄像机之间的连接关系。最后，随着多次外极几何图生成树的构建，得到一组匹配点轨迹的子集用以进行后续的迭代三角化和捆绑调整。大量图像重建实验表明，重建系统在融合该方法后，不仅提升了三维重建系统的速度，而且极大地降低了大场景三维重建时所需要的内存消耗，使得重建系统可拓展性更强。 4. 针对没有先验GPS信息或者仅有部分GPS信息的图像重建问题，提出一种全局式的摄像机位姿求取方法。首先，利用外极几何图上相对旋转与绝对旋转矩阵的关系，求取摄像机的绝对旋转矩阵。然后，基于RANSAC技术，利用2点法重新计算外极几何边上的相对平移方向，并根据优化后的相对平移方向和匹配点轨迹计算外极几何边模长的比值。最后，利用该比值和相对平移方向，通过凸的L1问题优化求解摄像机的绝对位置。通过在网络抓取图像和无人机图像数据上的测试，该方法与很多优秀的重建方法在重建精度和场景结构上相当，但是在有部分GPS信息时，它可以内在地融合这些信息进行求解。
英文摘要	With the development of camera sensor technology, a large number of highresolution images bring a big challenge to the large-scale 3D reconstruction. This thesis is focused on both the reconstruction accuracy and computation efficiency. To this end, we propose a series of fast and accurate large-scale 3D reconstruction algorithms. Our main contributions include: 1. For the images with auxiliary imaging information, an efficient, accurate and scalable reconstruction method is proposed by fusing these noisy prior information. Firstly，initial camera poses are obtained by the auxiliary imaging information, including GPS、IMU and compass angles. Then, based on the initial camera rotations, potential epipolar geometry edge inliers are iteratively distinguished and used for optimizing global camera rotations by non-linear least square algorithm. Given the global camera rotations and initial camera locations, potential tracks inliers are iteratively found and used for optimizingThe camera poses and scene structure by bundle adjustment. Extensive experimental results show that our method performs similarly or better than many of the state-of-art reconstruction approaches, in terms of reconstruction accuracy and completeness, but is more efficient and scalable for large-scale image datasets. 2. For the multi-camera system, a large-scale scene reconstruction method is proposed by fusing generalized camera model in a global manner. Since the relative rotations and translations in the multi-camera system are fixed during the image capturing process, we fuse this prior model information into the optimization algorithm to inherently make the final camera poses satisfy the real image capturing model. According to the experimental reconstruction results on both street view system and oblique airborne photogrammetry system, our method not only accelerates the speed of reconstruction, but improves the scene structure and reconstruction accuracy in some degree. 3. To improve the computation efficiency of bundle adjustment, we propose an efficient tracks selection method by covering the epipolar geometry graph. Firstly, according to three selection rules: Redundancy, Accuracy and Connectivity, the tracks priority are ranked by their lengths, scales and costs. Then, the tracks selection problem is solved by covering the spanning tree of the epipolar geometry graph. Finally, with the coverage of multiple spanning trees, a compact set of tracks is obtained for the final bundle adjustment. Extensive experiments show that the new reconstruction system with our tracks selection module performs similarly or better than many of the state-of-art approaches in terms of reconstruction completeness, but is more efficient and scalable for large-scale image datasets. 4. For the images without or with a part of GPS information, we propose a robust global camera poses estimation method. Firstly，by minimizing the differences between relative rotations and absolute rotations on the epipolar geometry edges, the whole absolute camera rotations are obtained. Then, with RANSAC technique, the relative translations are re-estimated by a 2-point algorithm. With refined translations and tracks, the scale-ratio between two translations is computed by constructing adjacent triangles. Finally, with the scale-ratio estimations and refined translations, the camera centers estimation problem is tackled by solving convex L1 optimization problems. According to the experimental reconstruction results on both Internet images and unmanned aerial vehicle images, our method performs similarly or better than many of the state-of-art reconstruction approaches, but it could inherently fuse the auxiliary GPS information.
语种	中文
内容类型	学位论文
源URL	[http://ir.ia.ac.cn/handle/173211/11719]
专题	毕业生_博士学位论文
作者单位	中国科学院自动化所
推荐引用方式 GB/T 7714	崔海楠. 快速鲁棒的大场景三维重建[D]. 北京. 中国科学院大学. 2016.