【作者】 韩坤芳；

【导师】 沈会良；

【作者基本信息】 浙江大学 ， 物理电子学， 2011， 硕士

【摘要】 物体表面的三维信息可以通过双目立体视觉来重建,包括摄像机标定,极线校正,对应点匹配,视差计算,亚像素精细化,模型重建等过程。对应点匹配是双目立体视觉的关键。本文采用时空立体匹配,通过人为主动地向空间物体投射一系列光条模式,使得物体表面纹理信息丰富,从而降低匹配难度,提高准确率。利用各种约束条件如对称性匹配约束,视差约束,顺序性约束等可进一步提高匹配效率。时空立体匹配方法较易得到整数级视差,为精细化视差,需要对初步得到的视差进一步亚像素优化。本文还对匹配过程中光条模式的取舍,匹配窗口大小等因素对重建模型的影响进行详细的分析与结果比较。实验结果显示,恰当的光条数目与合适的窗口大小对物体形状重建起着重要的作用。通过双目立体视觉重建得到的真实物体三维模型一般具有较大的噪声,并可能存在孔洞。针对这一问题,本文提出在规则三维点集下,通过非局部(non-local, NL)方法来滤除三维模型的噪声并填补孔洞。为提高算法效率,对含孔区域和非孔区域的点集,分别做基本NL滤波和基于主成分分析(principal component analysis, PCA)的NL滤波,然后进一步通过NL方法对孔洞做填补。实验结果表明,对于多种真实物体,本文所提出的方法均能有效地重建三维模型,且与其它方法相比具有更高的精度。更多还原

【Abstract】 3-D surface can be reconstructed based on binocular stereo vision. It always includes camera calibration, image rectification, correspondance pair matching, disparity calculation, sub-pixel refinement, model reconstruction, etc. Correspondance pair matching is the most important step in reconstruction. Space-time stereo matching can effectively improve matching accuracy and reduce matching difficulty, by projecting a series of light patterns for the purpose of enriching surface texture information. At the same time, symmetry restriction, disparity restriction, order restriction and others can also improve matching efficiency. Space-time stereo matching can reach integer disparity easily. Sub-pixel refinement is needed for higher accuracy. This paper analyses and compares the amount of patterns, matching windows’size and other factors’influence in 3D reconstruction. The experimental results show that, right amount of pattern and matching windows’size play an important role in 3-D surface reconstruction.The 3D models reconstructed from binocular stereo always exhibit noticeable noise, and even contain holes. To deal with this issue, this thesis presents a non-local (NL) method to denoise and hole-fill the original model on regular vertex set. To improve computational efficiency, the proposed method uses the basic NL method and principal component analysis (PCA) based NL method to remove model noise in hole and non-hole regions, respectively. The filled holes are then processed by NL denoising to improve the consistency with the surrounding surfaces. The experimental results show that, when evaluated on different real objects with various sizes, the proposed method can reconstruct 3D models effectively and performs much better than previous methods.更多还原