CASIA OpenIR  > 毕业生  > 硕士学位论文
光场三维显示与平行深度获取
马魁1,2
学位类型工学硕士
导师耿征
2018-05-24
学位授予单位中国科学院研究生院
学位授予地点北京
关键词光场 三维显示 投影阵列 平行理论
其他摘要

三维技术的迅猛发展,带领我们进入了一个崭新的三维时代。其中三维电影和虚拟现实逐渐成为我们休闲放松的日常娱乐组成。光场是物理世界中光线的集合,光场显示旨在于重现物理世界中光场分布,因其可在无需辅助设备的裸眼情况下更加真实的重建视觉三维立体空间,依靠其自身的技术优势引起了科研机构的广泛关注。本文以光场理论为基础,将全光函数简化到光场四维模型空间中,深入探究了光场显示与标定的原理,提出了创新性算法和优化系统,并实际搭建了物理样机进行验证;探究了多视点渲染技术中图像深度信息获取的技术方案。 

在光场显示方面,搭建了基于投影阵列的真三维显示系统,我们将光场重构模型引入切分拼接算法中,提出了一种新型场景切分-拼接模拟光场重构的显示算法。该方法利用光场思想分析显示屏幕对入射光线和出射光线的光路影响,对视点处形成的图像光线进行光路反跟踪,追寻到投影仪的出射光线点,从而形成对应的映射关系。其优点在于能够在投影仪个数有限的前提下大幅度的增加空间视点个数,并且在硬件参数未知的情况下通过标定校正实现较好的三维显示效果,增加系统的鲁棒性。 

在光场显示标定方面,我们提出了一种基于光路采集的自标定解决方案,实现一种复制性强、成本可控并且分辨率高的可移动三维显示装置。该方法是基于手机的高分辨率显示屏,通过对屏幕子像素点的光路采集,并考虑光学器件对光线散射的影响,将模型数学化并使用包围盒优化求解得到最优标定结果。基于此方法,设计了一套标定盒硬件系统,只需将手机放置在标定盒中,计算机启动相机捕获程序并生成标定图像传输给手机端显示,标定完成后系统会自动生成两张标定后的合成索引图用于实际三维场景显示图像合成中使用。其优势在于消除了手动校准的步骤,能够在三维显示设备的光学参数未知的情况下进行标定,实现了便捷标定与高质量的三维显示效果。 

三维显示内容可以通过相机对真实场景进行采集通过自由视点渲染技术处理得到,在场景数据处理时,主要是对场景的深度进行提取,然后进行虚拟视点的合成,为显示内容的渲染提供技术支持。因此,我们对图像深度信息的获取进行研究。目前随着深度学习技术的深入研究和应用,单目视觉深度研究取得了很大进展,但仍存在相关理论不完善和数据集规模多样性的缺乏等缺陷,因此将复杂系统建模与调控的ACP理论推广到单目深度获取这一领域,采用平行视觉的思路进行研究,提出基于平行理论的单目深度获取的基本框架和关键技术,并按照ACP三步曲进行了相关工作的展开,通过虚拟场景来模拟和表述复杂真实环境,利用计算实验对各种单目深度模型进行建模和测试,最终利用平行执行来实时调优模型,以达到对复杂场景的智能分析。 

最后,我们总结了全文工作并展望了相关工作的发展趋势。

;

The rapid advances on three-dimensional imaging technologies have led us into a new three-dimensional era. For example, three-dimensional movies and virtual reality have enhanced our daily entertainment experience. The light field is the collection of light rays in the physical world. Since the light field display aims at reproducing the light field distribution in the physical world, it is able to reconstruct more realistically three-dimensional objects for naked eye visualization without using any auxiliary device. The unique advantages of light field techniques have gained wide spread applications in numerous research fields by many research institutions. To further improve the light field theory, this thesis aims at developing a simplified light field function with a four-dimensional space model. We performed extensive study on the techniques which were applicable to light field display and calibration. We proposed an innovative light field rendering algorithm, optimized design of a 3D light field display system, and actually built a physical prototype for carrying out verification experiments. We explore the technical solution of the scene depth information acquisition in multi-viewpoint rendering system.

For light field display, we designed and built a truth three-dimensional display system based on a projection array. We introduced the light field reconstruction model into the segmentation and stitching algorithm. We also proposed a new display algorithm for scene segmentation-splicing simulation light field reconstruction. The method utilized the light field idea to analyze the effect of the display screen on the light path of the incident light. And the outgoing light performed the optical path anti-tracking on the image light which formed at the view point. By tracing the exit light point of the projector, it corresponded to the mapping relationship of the stroke. The advantage of this algorithm was that the number of spatial viewpoints could be greatly increased with fewer number of projectors. Besides, We achieved a good quality of three-dimensional display through calibration correction, even the hardware parameters were unknown. This novel design enhanced the robustness of the system.

For light field calibration, we proposed a self-calibration solution based on light field acquisition. This new method could be readily used for a mass-producible, cost-effective, high-resolution mobile 3D display devices, such as the high-resolution display for mobile phones. By acquiring light field at sub-pixel level and taking into account the influence of the optical device on the light scattering, our model is mathematically optimized by a bounding box to obtain the optimal results. To experimentally validate our technique, a calibration box hardware system was designed. During calibration we only need to place the mobile phone in the calibration box, and the computer could generate calibration images automatically and send them to the mobile phone display. Upon the completion of the calibration process, the system automatically generated two index maps. These index maps were used in the actual three-dimensional scene display image synthesis. The advantage of this algorithm and calibration technique was that the steps of manual calibration were eliminated, and the calibration could be performed without knowing the optical parameters of the three-dimensional display device. Finally, a high quality three-dimensional display visual effect was achieved.

The 3D display content can be captured by the camera arrays through the free viewpoint rendering technology. We mainly research the scene depth acquisition in the scene data processing. After the scene depth acquisition, the virtual viewpoint can be rendered. So it provides the technical support for the display content rendering. Therefore, we research the acquisition of the image depth information. We applied the deep learning technology and the parallel theory of complex system modeling to the monocular depth acquisition. We developed the parallel vision concept based on the parallel theory of the monocular depth acquisition. The parallel theory trilogy is used to carry out related work and validated via simulated virtual scene. We described the complex real environments by using computational experiments to model and testing various monocular in-depth models. By using the parallel execution, a real-time tune model achieved the intelligent analysis of complex scenes.

Finally, we summarized our work and made a few observations on the development trend of related work.

语种英语
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/20993
专题毕业生_硕士学位论文
作者单位1.中国科学院大学
2.中国科学院自动化研究所
推荐引用方式
GB/T 7714
马魁. 光场三维显示与平行深度获取[D]. 北京. 中国科学院研究生院,2018.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
光场三维显示与平行深度获取.pdf(2984KB)学位论文 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[马魁]的文章
百度学术
百度学术中相似的文章
[马魁]的文章
必应学术
必应学术中相似的文章
[马魁]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。