【作者】 杨青青；

【导师】 张明；

【作者基本信息】 浙江大学 ， 信号与信息处理， 2013， 博士

【摘要】 精确快速地获取三维场景的深度信息是很多三维应用所需的关键技术。随着三维技术的发展,诸如三维表面重建、三维电视、增强现实等应用都对深度图提出了新的要求：在精度上需要更精确的稠密深度信息,在速度上需要更高的深度获取效率。三维电视技术结合了计算机三维动画,真实场景深度获取和虚实融合等技术,舒适的观感需求对高效精确的深度获取技术提出了更高的要求。作者所在的研究团队提出了创新性的“自然三维电视”概念,旨在通过舒适的观感还原真实的三维场景。同时本文作者还提出了使用“视频+深度+增强信息”作为三维内容的表达方式。自然三维电视要求获取真实的场景信息并提取场景对应的深度信息,依靠基于深度图像的绘制方法重构出多个视点图像,实现自然的立体观看感受。而基于深度图像的绘制方法严重依赖于深度图的精确性,且实时的应用场景要求深度获取算法有更高的运算速度。立体匹配技术一直以来是应用最为广泛的深度获取算法之一,由于其被动获取深度的特性和生成稠密深度的能力,立体匹配技术能够适应多种场景的应用。但是由于该技术本身是一个逆向求解的病态问题,问题求解需要在一定的假设下进行。经过多年的发展,通过立体匹配技术获取稠密视差图,进而获取精确稠密的深度图的技术已经获得了长足的发展。这些算法主要分为使用全局优化的全局算法和注重局部匹配代价优化的局部算法。本文在现有算法的基础上对立体匹配算法进行了深入研究,主要创新点和贡献如下：1.本文提出了一种基于数字图像滤波器的局部立体匹配算法框架,实现了滤波器设计和立体匹配算法的紧密结合。算法框架将局部立体匹配中的两个关键问题：匹配代价聚合和视差图后处理统一建模为在匹配代价空间或视差空间中的保持边缘的平滑滤波问题。可以通过设计高效的图像滤波算法提升局部立体匹配算法的精度和速度。2.在基于图像滤波的局部立体匹配框架内,作者提出了一种“双层局部自适应”思想指导下的自适应指导图像滤波器。在原有的指导图像滤波算法基础上,提升了代价滤波的精度,从而提高了匹配算法的整体性能。同时,作者针对算法的并行运算结构特点,在NVIDIA CUDA并行计算平台上设计了算法的并行实现。提出了高效“并行积分图像”算法,有效提升了算法在并行平台上的运行效率,达到了半实时视频帧率的处理速度。3.为进一步提升代价滤波的性能,本文基于图像形态学的理论,提出了一种基于正交测地距离权重的全图像指导滤波器。该滤波器能够使用尽可能多的支持像素点支持图像滤波,同时能够有效地保持图像边缘。将该滤波算法应用于匹配代价滤波和视差后处理滤波,都获得了明显的效果提升。作者还根据正交测地距离计算的特点,设计了高效的算法,将原本复杂的滤波操作运算降到了极低的操作数。4.针对视差后处理的关键问题,本文提出了两种后处理滤波方法。使用层级联合双边滤波器对视差图直接处理,可以有效解决遮挡区域和误匹配的视差重计算问题;使用基于正交测地距离权重的全图像指导滤波器在变换后的匹配代价卷上进行滤波,在修复遮挡区域和误匹配视差的同时,还能有效处理大范围低纹理区域的错误匹配。更多还原

【Abstract】 One of the most important technology in many3D applications is acquiring the depth of3-dimensional scenes quickly and accurately. With the development of3D technology, new requirements are raised by many new applications such as3D surface modeling,3D television and argument reality. These applications usually require more accurate dense depth map on one hand, and more efficient depth generation algorithms on the other hand.3DTV integrates many technologies such as3D computer animation, real scene depth acquisition and virtual-real fusion. The demand of flexible viewing experience requires more advanced depth generation algorithms. Based on the classic3DTV, our research team proposed a novel concept named "natural3DTV" which aims to reproduce the real3D scenes using a more comfortable way. And the author also proposed using "video+depth+enhanced information" to represent the3D scenes. Natural3DTV system captures the actual scene, generates the corresponding depth, and synthesizes multiple virtual view images using the depth image-based rendering (DIBR) technology. And the quality of virtual views generated by DIBR heavily depends on the accuracy of depth maps. Some real-time applications also ask for more efficient algorithms.Stereo matching is one of the widely used depth acquisition methods. It can generate dense depth maps and requiring no active projection instruments. But the problem itself is ill-posed and must be solved under certain assumptions or constraints. Many algorithms were proposed and the technology has been well developed. These algorithms can be mainly categorized into two classes:global methods that rely on the global optimization and local algorithms emphasize on local cost aggregation. Based on these existing methods, the author made a deep study of stereo matching algorithms. The main contributions of this doctoral thesis are as follows:1. Based on existing local stereo matching algorithms, the author successfully integrates the digital image filtering technology and proposed a novel local stereo matching framework based on image filtering. The framework uses image filtering models to carry out two main steps in local algorithms:matching cost aggregation and disparity post-processing. This allows us to improve the accuracy and efficiency of stereo matching algorithms by applying efficient image filters.2. Within the novel local stereo matching framework, the author proposed a new adaptive guided image filtering method directed by the novel concept of "two-level local adapta-tion". The new cost filtering method improved the accuracy of disparity maps compared to methods using the original guided image filter. The author also designed the work-efficient "parallel integral image" parallel algorithm and implemented the whole algorithm on the NVIDIA CUDA parallel computing platform. The speed of the whole matching algorithm is greatly improved.3. The author proposed a novel full-image guided filtering method based on orthogonal geodesic distance weight. This filtering method can utilize as many supporting pixel-s as possible and ensures strong edges being kept. By applying it to cost filtering and post-processing, both present good performance. According the the property of the or-thogonal geodesic distance, the author also designed a fast implementation to reduce the computational amount.4. To deal with the key problem of disparity post-processing, the author also proposed two post-processing filtering methods. One method uses novel hierarchical joint bilateral fil-tering to process the disparity map directly, the other method operates on the transformed cost volume by applying full-image guided filtering. Both methods can deal with occluded regions and mismatched pixels. The later method performs well even at large low-texture regions.更多还原