【作者】 刘家朋；

【导师】 方涛；

【作者基本信息】 上海交通大学 ， 模式识别与智能系统， 2018， 博士

【摘要】 三维重建在虚拟现实、增强现实、三维打印、计算机辅助设计、图形学、动画、计算机视觉、医学图像重建、数字多媒体、工业逆向设计、文物博物馆及数字城市等领域有广泛的应用,一直以来是计算机视觉领域研究的热点。由于三维重建的基本信息是相机采集的图像数据,在实际应用中,由于光照环境复杂或物体表面反光等原因,造成采集的图像整体光照不均或者部分信息丢失,进而影响物体或场景三维重建的效果与质量,光照不均匀是三维重建最直接和最具挑战性问题之一。为此,本文对光照不均匀条件下的高精度相机标定,高动态物体三维重建以及三维模型无缝纹理映射等关键技术进行了研究,主要工作与创新如下:(1)为了提高光照不均匀条件下的相机标定精度,系统分析了不均匀光照对角点提取数量、角点提取精度、标定模板识别以及相机参数解算优化等各个方面的影响,并提出了一种基于棋盘格的光照不均匀条件下的高精度相机标定。首先,为了提高不均匀光照下的角点提取数量,利用颜色恒常性原理,对反射图像进行Harris角点提取,使在光照不均匀的条件下,不经过图像预处理,也能准确地对角点进行稳定的提取。其次,为了进一步提高角点提取精度,将棋盘格角点先验知识与二次曲面拟合相结合,克服光照不均匀导致的角点缺失或者位置不准,使提取的角点具有亚像素精度。为了不均匀的光照条件下亦能稳健的对标定模板进行识别,设计了一种含有中心定位圆的标定模板,并将中心圆定位与Delaunay三角剖分相结合来进行稳定地标定模板识别。最后,在标定参数解算阶段,由于不均匀的光照导致角点空间分布与镜头径向畸变不一致,对传统的相机参数优化流程进行了改进,以降低标定过程中的畸变误差。与多种方法对比测试表明,在各种光照条件下,提出的算法具有较好的稳定度与标定精度,可以在实际复杂光照现场提供高精度的相机标定。(2)通过分析不正确曝光对相位恢复产生的影响,提出一个基于高动态范围成像技术的相位恢复与三维重建算法。区别于传统的相移方法利用图像的灰度值来进行物体的相位恢复与三维重建,研究利用图像的照度值E进行相位恢复与三维重建。首先,采集多种曝光的条纹图像序列,然后,采用高动态范围重建技术恢复对应于每一个相移步骤的每一个像素位置的E值。最后,根据最小平方逼近原理由恢复的所有相移步骤的E值图像计算相位值。图像照度值E值图像具有曝光不变性,提出的算法为反光或者表面颜色复杂的物体三维重建提供了一个新的思路。而且,当由多曝光图像恢复E值来重建图像时,相机的动态范围被放大,从而抑制了随机噪声影响,重建出的三维数据也具有更好的质量。与多种基于高动态范围成像技术的三维重建算法对比实验表明,针对各种类型的物体,基于照度值E值可以实现稳定和准确的相位恢复与三维重建。(3)在对重建出的三维模型数据进行纹理映射时,从不同角度及距离拍摄的纹理数据,常常具有较大的光照与颜色差异。为此,提出了一种基于多分辨率分解与融合的无缝纹理映射算法,该算法在保持一定清晰度的前提下消除了不同角度拍摄的纹理数据的颜色光照缝隙,即使对于构网误差较大的区域也能得到较为满意的结果。与传统的纹理映射方法相比,提出的算法考虑到了纹理像素的所有可见图像集,并通过多参数加权弱化了对纹理像素贡献小的图像,从而在消除纹理缝隙的同时保证纹理图像清晰度。除多参数加权外,在生成最终纹理像素时,还基于多分辨率分解与融合,进一步消除纹理边界光照不一致的影响以及由于多参数加权导致的纹理细节模糊,使得最终合成的纹理图实现无缝的纹理映射及细节保存。与最优参考图像法等多种方法对比测试表明,即使存在较大的光照差异,本章提出的算法在保证颜色一致性的同时,仍能实现良好的细节保存。在光照不均匀的条件下,本文提出的相关算法取得了较好的效果。然而,由于三维重建技术本身的复杂性以及现实中物体类型的多样性,本文提出的算法针对非常发光的物体,尤其是镜面反光物体重建,仍有一定局限性。更多还原

【Abstract】 3D Reconstruction has a wide application in virtual reality,augmented reality,computer graphics,CAD,animation,computer vision,medical image reconstruction,digital media,reverse engineering,digital museum and digital city etc,it has been a hot research topic in computer vision for many years.The image captured from camera is the basic data for 3D reconstruction in real application.Because of the uneven envi-ronmental illumination or surface reflectance of objects,the captured images are often bad.Uneven illumination is one of the key and fundamental problems in 3D recon-struction.Thus this dissertation mainly focus on the study of some key technologies for 3D reconstruction under uneven illumination,for example,the influence of uneven illumination on camera calibration,3D reconstruction of reflectance object,seamless texture mapping of 3D model etc..The research work and main innovations in this dissertation are as follows:(1)To improve the precision of calibration under uneven illumination,by analyz-ing the influence of uneven illumination on corner detection,sub-pixel optimization,calibration board identification and camera parameters optimization etc.systematical-ly,an high precision camera calibration method based on checkerboard pattern under uneven illumination is proposed.First,to improve the precision of corner detection,a Harris corner detection method based on color constancy is proposed,the improved algorithm is very effective even for high dynamic range image.For the optimization of sub-pixel detection,an algorithm based on the heuristic knowledge of chessboard and the quadric surface fitting is proposed.For better calibrated board identification,a new form of chessboard with a center circle is designed,by combining the location of the center circle and mesh Delaunay,the calibrated board can be identified effec-tively under uneven illumination.Finally,to overcome the inconsistency of radial lens distortion,an optimized algorithm is proposed to reduce the radial lens distortion for uneven illumination.Compared with other methods,experiments showed that the pro-posed method achieved stable and accurate calibration under complicated illumination conditions.(2)By analyzing the influence of incorrect exposure on phase recovery and the phase error introduced by saturation,an algorithm is proposed for phase recovery and 3D reconstruction based on high dynamic range imaging technologies.Unlike existing methods that use grey values to calculate phase angle,the proposed method uses E val-ues,which are the recovered irradiance from multiple exposure images,to calculate the phase angle for specular surfaces.First,a sequence of fringe images with multiple ex-posures is captured.Then,the E value maps can be rebuilt by recovering them for every pixel at each shifting step.Finally,the phase angle for each pixel is recovered from the E value.By recovering the E value from multi-exposure images,the camera’s dynamic range is enlarged and the effect of random noise is suppressed.Thus,the accuracy of phase recovery is improved.Compared with traditional phase shifting methods based on HDR imaging,the proposed method takes into account not only grey values under one exposure,but also grey values under other exposures.Experimental results show that the proposed method improves the phase recovery accuracy and achieve good re-sults for the 3D reconstruction of dark and specular surfaces.For different objects,experiments showed that the proposed method recovered the phase angle accurately based on irradiance maps,and achieved good 3D reconstruction effects compared with other HDR methods.(3)For the texture mapping of reconstructed 3D models,the texture images are captured from different views and distances,thus the images are often suffered from color reflection and bad illumination.To achieve seamless texture mapping,an algo-rithm based on multi-resolution decomposition and fusion is proposed for 3D recon-struction.Firstly,a series of image sets are produced by analyzing the visibility of triangular facets,and the image sets are clustered and segmented into a number of op-timal reference texture patches.Secondly,the generated texture patches are sequenced to create a rough texture map,then a weighting process is adopted to reduce the color discrepancies between adjacent patches.Finally,multi-resolution decomposition and fusion technique are used to generate the transition section and eliminate the bound-ary effect.Experiments show that the proposed algorithm is effective and practical in obtaining high quality 3D texture mapping for 3D reconstruction.Compared with traditional methods,it maintains the texture clarity while eliminates the color seam-s,besides,it also supports 3D texture mapping for big data application.Compared with optimal selection methods etc.,experiments showed that,even large color differ-ence exists,the proposed method achieved seamless texture mapping and image details maintenanceIn brief,the innovative work in this dissertation achieves good effects for 3D re-construction under uneven illumination.Due to the complexity of 3D reconstruction technologies and the diversity of real objects,the proposed methods are still limited for highly reflectance surface especially for mirror-like surfaces.更多还原