基于连续深度融合的多视图三维重建研究

发布时间：2018-04-21 02:09

本文选题：三维重建 + 立体匹配　；参考：《浙江大学》2013年博士论文

【摘要】：随着影视、动漫与游戏行业的蓬勃发展,其对高真实感三维场景重建的需求越来越多。而在文物数字化等领域,对于三维模型重建要求更高,从三维重建的逼真度要求上升到了对三维形体准确度及表面色彩保真度的要求。基于多视图立体匹配的三维重建是实现上述需求的一种重要方法,可直接计算得到包含准确色彩纹理的三维模型结果。准确性、鲁棒性以及计算效率是评价基于多视图立体匹配三维重建的重要标准。图像的畸变、随机噪声、重复纹理以及物体间的遮挡等因素影响了多视图立体匹配算法的鲁棒性和重建结果的准确性。本文主要从提升算法鲁棒性和重建结果的准确性两个方面来深入研究面向复杂场景的三维重建方法：一方面,研究高质量的深度图计算以及融合算法,通过对影响深度计算准确性的一些因素进行建模,提高计算结果的准确性。另一方面,研究基于连续优化的深度计算方法,利用连续优化计算鲁棒性高的特点,来提高三维重建算法的鲁棒性。具体地,本文研究图像的径向畸变矫正、基于非凸连续优化的深度计算、基于凸连续优化的深度计算以及基于连续深度图融合的多视图立体匹配。主要工作与创新包括： ●提出了一种基于矩阵QR分解的图像径向畸变矫正算法,解决了现有三维重建管线中畸变参数计算不够鲁棒的问题,提升了多视图三维重建算法的鲁棒性和重建结果的准确性。通过将畸变参数计算转化成矩阵分解问题,简化了参数的计算过程。 ●提出了一种基于对称连续优化的深度图计算方法,使能量泛函的解更趋于全局最优解,有效的提高了深度图的质量。通过将立体匹配问题转化成连续马尔科夫随机域的形式,建立了基于对称连续优化的深度计算模型。在模型的数据项中,引入颜色一致性约束和梯度一致性约束,提高了算法的准确性。设计了基于多层图像金字塔的迭代计算框架,有效地提高了计算出的深度图的质量。在匹配泛函模型的设计中,还引入了左右一致性约束,进一步提升了深度计算结果的准确性。 ●提出了一种基于凸优化的深度图计算方法,有效地提高了深度计算过程的鲁棒性和计算结果的准确性。针对物体间的相互遮挡等原因导致深度并不是严格连续的问题,提出了分段连续假设条件下的深度图计算方法将深度计算问题转化成自由不连续泛函模型来实现深度的计算,同时在泛函模型中引入了图像分割的先验知识,有效地抑制深度图在图像低频区域的噪声。通过利用将泛函模型松懈成凸泛函的方法,确保了深度图的计算过程不依赖初始值,提升了算法的鲁棒性,提高了深度图的质量。 ●提出了一种基于连续深度图融合的三维重建方法,提高了重建模型的准确性。通过利用左右一致性信息来控制深度图不同区域的更新速度,提高了深度图的质量。设计了一种利用近邻图像信息和深度信息进行深度图优化的机制,进一步提高了深度图的质量。综合利用左右一致性约束信息、点的法向量信息以及相机的视角信息有效解决了深度融合过程中的去噪问题。
[Abstract]:With the vigorous development of animation and game industry, more and more demand for 3D scene reconstruction of high realism is needed. In the fields of digitalization of cultural relics, the reconstruction of 3D model is more demanding. The requirement of three-dimensional reconstruction from the fidelity of 3D reconstruction to the accuracy of three-dimensional shape and the surface color fidelity. 3D reconstruction of body matching is an important method to realize the above requirements. It can directly calculate the results of 3D model containing accurate color texture. Accuracy, robustness and computing efficiency are the important criteria for evaluating 3D reconstruction based on multi view stereo matching. Blocking factors affect the robustness of the multi view stereo matching algorithm and the accuracy of the reconstruction results.
In this paper, the methods of 3D reconstruction for complex scenes are studied in two aspects: the robustness of the lifting algorithm and the accuracy of the reconstruction results. On the one hand, the high quality depth map calculation and the fusion algorithm are studied. By modeling some factors that affect the accuracy of the depth calculation, the accuracy of the calculation results is improved. In order to improve the robustness of the 3D reconstruction algorithm, the depth calculation method based on continuous optimization is studied.
Specifically, this paper studies the correction of radial distortion of images, depth calculation based on non convex continuous optimization, depth calculation based on convex continuous optimization and multi view stereo matching based on continuous depth map fusion. The main work and innovation include:
An image radial distortion correction algorithm based on matrix QR decomposition is proposed to solve the problem that the distortion parameter calculation in the existing 3D reconstruction pipeline is not robust enough to improve the robustness of the multi view 3D reconstruction algorithm and the accuracy of the reconstruction results. The calculation process.
A depth map calculation method based on symmetric continuous optimization is proposed, which makes the energy functional solution more global optimal solution and improves the quality of the depth map effectively. By transforming the stereo matching problem into the form of the continuous Markov random domain, a depth calculation model based on symmetrical continuous optimization is established. With the introduction of color consistency constraint and gradient conformance constraint, the accuracy of the algorithm is improved. An iterative calculation framework based on the multi-layer image Pyramid is designed to effectively improve the quality of the calculated depth map. In the design of the matched functional model, the left and right constraints are introduced, which further improves the depth calculation results. Accuracy.
A method of depth map calculation based on convex optimization is proposed, which effectively improves the robustness of the depth calculation process and the accuracy of the calculation results. In view of the problem that the depth is not strictly continuous for the reasons of mutual occlusion among objects, a depth figure calculation method is proposed under the condition of piecewise continuous hypothesis. It is converted into a free discontinuous functional model to realize the calculation of depth. At the same time, the prior knowledge of image segmentation is introduced in the functional model, and the noise of the depth map is effectively suppressed in the low frequency region of the image. By using the method of reducing the functional model into a convex functional, the calculation process of the depth map does not depend on the initial value, and the algorithm is improved. The robustness is improved and the quality of the depth map is improved.
A three-dimensional reconstruction method based on continuous depth map fusion is proposed to improve the accuracy of the reconstruction model. By using the information of the left and right consistency to control the update speed of the different areas of the depth map and improve the quality of the depth map, a mechanism of depth map optimization is designed by using the information of adjacent images and depth information. The quality of the depth map is improved step by step. Using the information of the left and right consistency constraints, the point information of the normal vector and the camera's angle of view information can effectively solve the denoising problem in the process of depth fusion.

【学位授予单位】：浙江大学
【学位级别】：博士
【学位授予年份】：2013
【分类号】：TP391.41

【参考文献】