基于双目视觉的大场景三维重建研究与实现

发布时间：2018-01-05 15:23

本文关键词：基于双目视觉的大场景三维重建研究与实现　出处：《吉林大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着人们的认知水平和物质文化追求的提高,越来越多的3D技术相关的产品开始在人群中普及,为了获得更好的产品体验,虚拟现实三维重建技术开始重新回到人们的视野之中,也越发受到人们的重视。本文研究并实现了一个完整的大场景三维重建系统,包括立体标定,立体校正,立体匹配和场景表面三维点云重建等过程。在立体匹配阶段,针对原NLCA算法考虑因素过于单一,无法适应大场景图形规模大,多尺度,场景物体丰富,深度变化频繁等特点,本文对原算法进行了进一步的改进。首先,我们对初始代价函数进行优化,使用ADCensus代价计算方法融合像素点的梯度信息,强化像素点的边缘特征。另外,在代价聚合阶段,在构建最小生成树时,本文定义了一种同时包含了边缘和颜色权值分量的新的权值函数为相邻点对之间的边赋值。另外根据图像边缘成因,我们在使用sobel算子对图像进行边缘检测的同时,使用SLIC超像素分割块对边缘权值分量进行二次约束,极大地减少了由于除深度变化外的其他原因产生的边缘像素点在最小生成树构建过程中产生的影响。此外,在视差求精阶段,本文提出了一种基于超像素块的视差优化算法,利用同一超像素块中的像素点具有较高的相似度的特点,对视差数据进行修复,有效提高稳定点的比重,在此基础上,结合原算法中使用的基于最小生成树代价聚合的优化方案,有效提高了优化质量。在算法的最后阶段,为了提高三维重建的完整性,本文加入了一系列视差优化过程,包括前文所述的基于超像素块的视差优化过程,并对该算法无法优化的区域进行再次填充并使用改进了的中值滤波算法对不稳定像素点区域进行优化,有效减少了视差图像上的不连续块和空洞,有效解决了视差图像与参考图像所反映的场景的实际信息不一致的问题,提高三维重建过程的准确性和完整性。实验数据表明,本文较原算法而言具有明显的优势,不仅在低分率图像上取得较好的效果,对于本文所使用的高分辨率大场景图像素材同样适用,证明了本研究构建的大场景三维重建系统可靠性、准确性以及稳定性。
[Abstract]:With the improvement of people's cognitive level and material and cultural pursuit, more and more 3D technology-related products began to be popularized in the crowd, in order to obtain a better product experience. Virtual reality 3D reconstruction technology has begun to return to people's vision, and has been paid more and more attention. This paper studies and implements a complete 3D reconstruction system of large scene, including stereo calibration and stereo correction. Stereo matching and 3D point cloud reconstruction on the scene surface. In the stereo matching stage, the original NLCA algorithm is too single to adapt to large scale, multi-scale and rich scene objects. Because of the frequent change of depth, the original algorithm is further improved. First, we optimize the initial cost function and use the ADCensus cost calculation method to fuse the gradient information of pixels. Enhance the edge feature of the pixel. In addition, in the cost aggregation phase, when building the minimum spanning tree. In this paper, we define a new weight function which includes both edge and color weight components to assign the edges between adjacent points. We use the sobel operator to detect the edge of the image, and we use the SLIC hyperpixel segmentation block to perform the quadratic constraint on the edge weight component. The effect of edge pixels due to other reasons other than depth changes in the construction of the minimum spanning tree is greatly reduced.; in addition, in the parallax refinement stage. In this paper, a parallax optimization algorithm based on super-pixel block is proposed. The parallax data is repaired by using the feature of high similarity of pixels in the same super-pixel block, and the proportion of stable points is improved effectively. On this basis, the optimization scheme based on minimum spanning tree cost aggregation used in the original algorithm is used to improve the optimization quality effectively. In the last stage of the algorithm, in order to improve the integrity of 3D reconstruction. In this paper, a series of parallax optimization processes are added, including the parallax optimization process based on super-pixel blocks mentioned above. The region which can not be optimized by this algorithm is filled again and the unstable pixel area is optimized by using the improved median filter algorithm, which can effectively reduce the discontinuous blocks and holes in parallax images. It effectively solves the problem that the actual information of the scene reflected by the parallax image and the reference image is inconsistent, and improves the accuracy and integrity of the 3D reconstruction process. Compared with the original algorithm, this paper has obvious advantages, not only in the low score image achieved better results, for the use of high-resolution large scene image material is also applicable. The reliability, accuracy and stability of the system are proved.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】