基于单目图像序列的场景稠密三维重建技术研究

发布时间：2018-11-10 12:06

【摘要】：基于图像序列的场景三维重建技术是计算机视觉领域中的一个重要研究课题。它用目标的二维图像序列和摄像机内外参数来恢复物体的三维信息。与传统的利用激光三维扫描或建模软件得到三维模型的方法相比,基于图像序列的三维重建方法具有重建约束限制条件少、自动化程度高、成本低、适宜对大尺度场景进行重建的优点。因此,该方法在逆向工程、机器人定位与导航、虚拟现实、增强现实、增强虚拟环境以及军事等领域具有广阔的应用前景。然而长期实践表明,基于图像序列的三维重建方法面临着诸多难题,如摄像机外参标定技术的算法复杂且准确度不高,无法对摄像机位姿提供精确的估计,不准确的特征匹配导致得不到精确的三维信息。图像特征点检测与匹配算法,面对非合作环境,难以在弱纹理或纹理重复区域检测和匹配可靠的特征点对,致使点云和曲面层次上的重构都出现较大的失真,重建目标表面出现空洞现象严重,这成为突出的技术性障碍。此外,在图像特征匹配过程中的特征点匹配对数与运算速度也很难兼顾。本文针对上述问题展开了研究,主要贡献有:(1)设计了一种基于运动法的摄像机位姿跟踪算法,采用SIFT算法进行图像特征点的检测与匹配,运用基于RANSAC框架的五点算法估计两视图下摄像机的相对位姿,采用基于RANSAC框架的透视N点法进行多视图位姿跟踪,有效提高了位姿估计精度。(2)提出了一种基于光流反馈的稠密三维重建方法,能够实现对场景准确、快速的三维立体化建模。采用TV-L1光流算法建立更加稳健的同名像点匹配关系,对得到的稠密匹配点进行采样,进而构建稀疏点云和初始网格。从运动视觉分析的角度寻求多视重构的求解方法,将重建模型反馈至重建过程,用各视图像的偏差驱动模型变形。将粗糙的、不准确的原始网格曲面经过致密的非刚性变形,调整为精确的曲面。(3)在统一设备架构下,利用图形处理器(GPU)对重建算法中的光流匹配、合成图像、网格调整等部分进行并行加速,显著提高了重构算法的实时性。本文针对三维重建中的外参求解精度、特征匹配准确度、三维重建的稠密性、重建算法的实时性等问题,提出了有效的解决方案。室内真实场景下的重建结果证明了所提算法的可行性和准确性。
[Abstract]:Scene 3D reconstruction based on image sequence is an important research topic in the field of computer vision. It uses the object's two-dimensional image sequence and the camera's inside and outside parameters to restore the object's three-dimensional information. Compared with the traditional method of obtaining 3D model by laser scanning or modeling software, 3D reconstruction method based on image sequence has fewer constraints, higher automation and lower cost. It is suitable for large scale scene reconstruction. Therefore, this method has a broad application prospect in reverse engineering, robot positioning and navigation, virtual reality, enhanced virtual environment, military and other fields. However, long-term practice shows that the 3D reconstruction method based on image sequence faces many difficulties, such as the complexity and inaccuracy of camera external parameter calibration technology, which can not provide accurate estimation of camera pose. Inaccurate feature matching results in the lack of accurate 3D information. In the face of non-cooperative environment, it is difficult to detect and match reliable feature point pairs in weak texture or texture repeat regions in image feature point detection and matching algorithm, which results in a large distortion of point cloud and surface level reconstruction. The appearance of void on the surface of the reconstructed target is serious, which becomes a prominent technical obstacle. In addition, it is difficult to match logarithm and operation speed of feature points in the process of image feature matching. The main contributions of this paper are as follows: (1) A camera pose tracking algorithm based on motion method is designed. SIFT algorithm is used to detect and match the image feature points. The five-point algorithm based on RANSAC framework is used to estimate the relative pose of the camera in two views, and the perspective N-point method based on RANSAC framework is used to track the position and pose of multi-view. The accuracy of pose estimation is improved effectively. (2) A dense 3D reconstruction method based on optical flow feedback is proposed, which can realize accurate and fast 3D 3D modeling. The TV-L1 optical flow algorithm is used to establish a more robust matching relation of the image points with the same name. The dense matching points are sampled and the sparse point clouds and initial grids are constructed. From the point of view of motion vision analysis, the solution of multi-view reconstruction is sought, and the reconstruction model is fed back to the reconstruction process, and the model deformation is driven by the deviation of each visual image. The rough and inaccurate original mesh surface is adjusted to the precise surface through dense non-rigid deformation. (3) under the unified equipment architecture, the (GPU) is used to match the optical flow in the reconstruction algorithm, and the image is synthesized. Grid adjustment and other parts of the parallel acceleration, significantly improve the real-time reconstruction algorithm. In this paper, an effective solution is proposed to solve the problems such as the precision of extrinsic parameter solution, the accuracy of feature matching, the density of 3D reconstruction and the real-time performance of reconstruction algorithm. The results of indoor real scene reconstruction prove the feasibility and accuracy of the proposed algorithm.
【学位授予单位】：北京工业大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】