多视图三维重建中图像配准和光束法平差过程的并行算法研究

发布时间：2019-04-16 15:50

【摘要】：近年来,随着计算机视觉的发展和成熟,多视图三维重建以其成本低廉,操作方便的优势倍受关注,目前已广泛应用于数字城市、医疗成像、虚拟现实等领域。图像特征提取、特征匹配和光束法平差是多视图三维重建的几个关键步骤,但其中也在一些问题。首先,特征提取阶段通常采用的SIFT算法并不能适用所有场景。其次,随着GPU在计算能力和存储器带宽上展现出巨大的优势,基于GPU加速的多视图三维重建算法成为研究的热点。然而特征匹配算法迁移至GPU平台的较少,并且由于硬件架构的不同,并行算法的可靠性以及有限的显存对光束法平差并行算法进行了极大的制约。为此本文展开了以下研究:(1)本文提出了一种Harris-Laplace特征结合SIFT描述的图像配准并行算法。特征提取阶段,使用改进的Harris-Laplace算法提取出对图像亮度、旋转和尺度具有不变性的特征点,然后SIFT描述子对特征点进行描述。在特征点匹配阶段则利用双向匹配法和极线几何约束对特征点进行粗匹配和精匹配。在分析该算法并行性的基础上,采用CPU_GPU协同处理技术,从任务的性质和传输时间两方面考虑,在CPU和GPU端进行合理的任务划分,提升了整体算法的运行效率。(2)BA问题时常采用LM将其线性化,并用PCG算法求解法方程。本文先是利用PCG算法不需要显式存储法方程系数矩阵的特性,将其分解为简单的雅各比矩阵和向量相乘的问题。然后在BA并行算法预处理过程中增加了一个过滤步骤,剔除了由数值转换导致的错误点,保证精度的情况下最大限度发挥GPU高峰值的单精度浮点运算能力。最后在深入分析雅各比转置矩阵与原矩阵联系的基础上,提出了一种光束法平差并行算法。对涉及到雅各比转置矩阵的运算重新进行并行化设计,无需存储雅各比转置矩阵即可解决光束法平差问题。从实验结果分析可知,本文提出Harris-Laplace特征结合SIFT描述的图像配准并行算法,提升了算法的整体效率,准确度也得到可靠保证。本文提出光束法平差并行算法在获得较好加速效果的同时,也大幅缩减了算法占用的显存空间。借助GPU的强大的并行计算能力,多视图三维重建实时性得到了很好的满足。
[Abstract]:In recent years, with the development and maturity of computer vision, multi-view 3D reconstruction has been widely used in many fields such as digital city, medical imaging, virtual reality and so on because of its low cost and convenient operation. Image feature extraction, feature matching and beam adjustment are the key steps in multi-view 3D reconstruction, but there are also some problems. First of all, the SIFT algorithm usually used in the feature extraction phase is not suitable for all scenes. Secondly, with the huge advantages of GPU in computing power and memory bandwidth, the multi-view 3D reconstruction algorithm based on GPU has become a hot research topic. However, the migration of feature matching algorithm to GPU platform is less, and because of the different hardware architecture, the reliability of parallel algorithm and the limited display memory greatly restrict the parallel algorithm of beam adjustment. The main contents of this paper are as follows: (1) A parallel image registration algorithm based on Harris-Laplace feature and SIFT description is proposed in this paper. In the feature extraction phase, the improved Harris-Laplace algorithm is used to extract the feature points which are invariant to the brightness, rotation and scale of the image, and then the feature points are described by the SIFT descriptor. In the stage of feature point matching, two-way matching method and polar geometric constraint are used for coarse matching and fine matching of feature points. On the basis of analyzing the parallelism of the algorithm, the CPU_GPU cooperative processing technology is used to divide the tasks reasonably on the CPU and GPU sides, considering the nature of the task and the transmission time. The efficiency of the whole algorithm is improved. (2) LM is often used to linearize the BA problem and PCG algorithm is used to solve the equation. In this paper, we first decompose the PCG algorithm into a simple Yakubi matrix and a vector multiplication problem by taking advantage of the characteristic of the coefficient matrix of the equation without the need of explicit memory method. Then a filtering step is added to the pre-processing process of the BA parallel algorithm, which eliminates the error points caused by the numerical conversion and maximizes the single-precision floating-point operation capability of the GPU with high peak value under the condition of ensuring the accuracy. Finally, on the basis of deeply analyzing the relationship between Yakubi transposable matrix and original matrix, a parallel algorithm of beam adjustment is proposed. The parallel design of the operation involving Yakubi transposable matrix can solve the adjustment problem of beam method without storing Yakubi transposable matrix. From the analysis of the experimental results, this paper proposes a parallel image registration algorithm based on Harris-Laplace feature and SIFT description, which improves the overall efficiency of the algorithm and ensures the accuracy of the algorithm. In this paper, a parallel algorithm for beam adjustment is proposed, which not only achieves better acceleration effect, but also greatly reduces the memory space occupied by the algorithm. With the powerful parallel computing ability of GPU, the real-time performance of multi-view 3D reconstruction is satisfied.
【学位授予单位】：江西理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】