基于非量测相机无结构影像的运动估计算法研究

发布时间：2018-06-27 23:20

本文选题：由运动恢复结构 + 无结构影像　；参考：《中国测绘科学研究院》2017年硕士论文

【摘要】：由运动恢复结构,即structure from motion是计算机视觉和摄影测量中的核心问题,它是仅仅通过一组具有重叠度的影像或视频来恢复相机拍摄瞬间的位置和姿态同时获得场景的稀疏三维结构信息的处理过程,此过程的准确性和有效性将直接影响后续的密集重建。虽然这个问题在2000年左右整个流程就基本确定下来,多视图几何重建也已被研究多年,但是由于重建过程中设置的参数较多流程较为复杂,很难找到全局最优解,大规模影像的重建速度依然存在瓶颈,至今国内外依然有较多研究者关注此问题。目前,国内还没有一款较为成熟的通用的基于多源影像建模的商业软件,国外商用软件依然占据主导地位,基于这些背景,有必要在此基础上开展基于任意拍摄的无结构影像的稀疏三维重建算法研究。传统的摄影测量中空中三角测量过程,相当于计算机视觉中的structure from motion需要严格的相机标定参数和规则的航带信息,其过程是先做单个模型的相对定向再进行模型间的连接,而且定向都是沿着摄影测量坐标系中的X方向进行。然而在计算机视觉领域,拍照的过程较为随意,一般也不需要影像的任何先验信息,绝大多数情况下可以从影像的EXIF信息中估计相机的参数,然后再bundle adjustment中进行参数的进一步优化,此方法对获取影像设备的要求降低,处理过程中的人工干预少自动化程度较高。所以,针对当前多数影像,如手机影像、数码相机影像、甚至互联网影像,同时也包括无人机影像和航空影像,有时存在很难获取相机检校参数或航带信息不规则的情况,采用计算机视觉中的这种方法可以很好的解决这一问题。基于此,本文主要研究内容包括:1.系统总结基于影像重建的基础理论和算法原理:介绍计算机视觉中经典的三种基于影像的重建算法并比较每种算法的优缺点和适用条件,归纳出基于影像重建的一般方法和流程,并全面地总结当前国内外最先进的几大商业软件和开源软件的功能和特点,在此基础上概括出基于影像重建的两大关键问题:一是面对大数据量的影像,如何找到这些影像的对应关联信息,即data association,二是采取何种策略来重建这些影像,不同的策略可能导致不同的结果。2.详细分析了两张影像到三张影像的重建算法:此过程对应于摄影测量中的相对定向和绝对定向,分别研究了基于8点法和5点法相对定向并比较其特点,在已知相机参数和未知相机参数的情况下采取不同的相对定向算法,并将该算法与RANSAC算法相结合,获得鲁棒解;并且引入了计算机视觉中一种新型的在已知相机参数情况下3点绝对定向算法,实验证明此算法简便有效,可利用在序列影像重建过程中。研究一种在匹配几何约束阶段通过多模型约束来选取影像对的最优几何模型,为后续的重建过程奠定基础。3.对比了三种不同的特征提取算法(SIFT、SURF、ORB)对像对运动恢复的影响:传统的基于影像重建大多仍然采用SIFT,由于其稳定性和鲁棒性,但是SIFT的速度较慢,在影像数据量较大和实时性方面有所欠缺。本文将另外两种特征提取算法SURF和ORB应用到影像重建中,并在文中详细比较了三种算法的性能和对重建的影响,给出了在保持其重建稳定性的情况下提高速度的策略。4.实现并改进了经典的渐进式序列影像重建算法:利用SIFT和SURF算法分别进行特征提取和匹配,在获得所有影像的匹配信息后,作为整体输入生成track信息,用指定和自动搜索两种方法找到初始像对,在确定初始像对之后,用剩余影像和之前像对有最多2D-3D对应点来确定下一张影像,每确定一张影像再对该影像的位姿,即旋转和平移做一次bundle adjustment,然后再根据设定的参数来确定是否做整体的bundle adjustment,结果证明此方案可行并达到预期效果。提出一种基于特征点分布的影像选取策略,与传统的基于最多2D-3D特征点数量的方法相比,可以一定程度上提高重建的精确性和鲁棒程度。
[Abstract]:The motion recovery structure, structure from motion, is the core problem in computer vision and photogrammetry. It is the process of recovering the sparse three-dimensional structure information of the scene by only a set of overlapped images or videos to restore the camera's position and posture at the same time. The accuracy and effectiveness of this process will be There is a direct impact on subsequent intensive reconstruction. Although the whole process is basically determined around 2000, multi view geometric reconstruction has been studied for many years. However, it is difficult to find the global optimal solution because of the more complex parameters set in the reconstruction process. The reconstruction speed of the large pattern image still has a bottleneck. There are still many researchers at home and abroad concerned about this problem. At present, there is no more mature commercial software based on multi source image modeling in China. Foreign commercial software still occupies the dominant position. Based on these background, it is necessary to carry out a sparse three-dimensional reconstruction algorithm based on unstructured images based on arbitrary photographing. In the traditional aerial photogrammetry, the process of aerial triangulation is equivalent to the structure from motion in computer vision, which requires strict camera calibration parameters and rules. The process is to make the relative orientation of a single model first and then to connect the model between the models, and the orientation is carried out along the X direction in the photogrammetric coordinate system. However, in the field of computer vision, the process of taking pictures is more random and generally does not need any prior information of the image. In the vast majority of cases, the parameters of the camera can be estimated from the EXIF information of the image, and then the parameters are further optimized in the bundle adjustment. This method reduces the requirements for the acquisition of the image equipment and the process of processing. Therefore, for most of the current images, such as mobile phone images, digital camera images, and even Internet images, including UAV images and aerial images, sometimes it is difficult to obtain camera calibration parameters or aerial information irregularities. This method can be used in computer vision. Based on this, the main contents of this paper include: 1. systematically summarize the basic theory and algorithm principle based on image reconstruction: introduce three classic image based reconstruction algorithms in computer vision and compare the advantages and disadvantages and applicable conditions of each algorithm, and generalize the general methods and processes based on image reconstruction. And summarize the functions and characteristics of the most advanced commercial and open source software at home and abroad. On this basis, we summarize the two key problems based on image reconstruction: first, how to find the corresponding correlation information of these images, that is, data association, and what kind of strategy to reconstruct the image based on the image of large amount of data. Some images, different strategies may lead to different results,.2. detailed analysis of the reconstruction algorithm of two images to three images: this process corresponds to the relative orientation and absolute orientation in photogrammetry, and studies the relative orientation of the 8 point method and the 5 point method, respectively, and compare the characteristics of the camera parameters and the unknown camera parameters. Different relative orientation algorithms are adopted, and the algorithm is combined with the RANSAC algorithm to obtain a robust solution. And a new 3 point absolute orientation algorithm in the case of the known camera parameters in the computer vision is introduced. The experiment proves that the algorithm is simple and effective. In the process of sequence image reconstruction, a kind of matching geometry is studied. In the beam phase, the optimal geometric model of image pair is selected by multiple model constraints, which lays the foundation for the subsequent reconstruction process..3. contrasts the effect of three different feature extraction algorithms (SIFT, SURF, ORB) on motion recovery. The traditional image based reconstruction is still mostly SIFT, because of its stability and robustness, but the speed of SIFT. In this paper, two other feature extraction algorithms, SURF and ORB, are applied to the image reconstruction. In this paper, the performance of the three algorithms and the effect on the reconstruction are compared in detail, and the implementation of the strategy.4. to improve the speed in the condition of maintaining the stability of the reconstruction is also given. The algorithm of the progressive sequence image reconstruction: using SIFT and SURF algorithm for feature extraction and matching respectively. After obtaining the matching information of all the images, the track information is generated as the whole input, and the initial image pairs are found by two methods of designation and automatic search. After the initial image pairs are determined, there are the most 2D-3 with the remaining images and the previous images. The D corresponding point determines the next image, and each image is determined to do a bundle adjustment on the position of the image, that is, rotation and translation, and then determines whether to do the whole bundle adjustment according to the set parameters. The result proves that the scheme is feasible and achieves the expected effect. A image selection based on the distribution of feature points is proposed. Compared with the traditional method based on the number of 2D-3D feature points, the strategy can improve the accuracy and robustness of reconstruction.
【学位授予单位】：中国测绘科学研究院
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;P23

【相似文献】