室外自然光照成像计算与阴影处理技术研究
文献类型：学位论文
| 作者 | 段志刚
|
| 学位类别 | 博士 |
| 答辩日期 | 2016-11-30 |
| 授予单位 | 中国科学院沈阳自动化研究所 |
| 导师 | 唐延东 ; 田建东 |
| 关键词 | 光谱辐照度计算 图像成像计算 阴影生成 阴影检测 光照不变图像 阴影特征 道路检测 导航线检测 |
| 其他题名 | Imaging Computing and Shadow Processing of Outdoor Natural Illumination |
| 学位专业 | 模式识别与智能系统 |
| 中文摘要 | 复杂多变的光照环境给计算机视觉算法及应用（如特征提取，目标分割与识别、测量）带来诸多问题，降低了其算法的鲁棒性及环境自适应性。寻求该问题的有效解决方案一直是计算机视觉及相关学科的重要研究内容。我们从基本的大气物理、物理光学原理及物理成像机理出发，研究图像中光照变化问题，基于可计算光谱辐照度提出了光照成像的计算方法。以此为基础，对计算机视觉领域中的光照处理，如阴影生成、阴影检测、道路监测等进行研究，提出了新的模型算法。论文中的主要研究内容及成果如下：针对计算机视觉及其应用，给出了基于有效SPD 计算的室外光照成像计算。室外光源的光谱辐照（SPD）随着时间和大气条件的变化而变化，这导致场景的变化和常见的自然光照现象，例如暮光，阴影和霾/雾。计算室外光源在不同时间（或天顶角）和在不同大气条件下的SPD对于计算机视觉非常重要。SPD可计算方法可以应用于计算机视觉任务，包括光谱反演计算，图像成像计算和阴影图像处理。通过太阳位置的计算，给出了阴影生成中图像阴影的方向和阴影长度的计算方法。利用阴影衰减特性, 提出了一种新的图像阴影生成算法。结合图像阴影边缘的处理，可生成更接近物理真实的图像阴影。与基于图形学的阴影生成方法和通过抠图合成的阴影生成方法相比，本算法无需进行复杂的运算和任何抠图操作，只需要基于原始像素值、阴影三色线性衰减和日光与天空光SPD 比率便可以计算出生成阴影的像素值。这一具有物理光学背景的阴影生成算法，提高了图像阴影生成的视觉效果。通过相应的实验结果，验证了该算法的有效性。针对室外光照条件下阴影的快速高效检测问题，提出了一种基于正交分解的阴影检测方法。根据图像光照的正交分解方法，得到一幅彩色光照不变图像和一幅光照变化图像。通过K-means 算法将彩色光照不变图像分类为几个区域，每个区域具有一致的反照率。根据分类结果，对光照变化图像采用EM 算法进行高斯混合建模，提取阴影区域。最后采用形态学算子对提取的阴影区域进行优化。本方法不需要复杂的特征算子学习过程，大大降低了算法的时间复杂度，而且不需要任何先验知识，可以直接应用到实时场景处理中。提出了四个具有物理特性的图像阴影特征和基于区域与机器学习的阴影检测算法。四个新的图像阴影特征来自于日光SPD 和天空光SPD 的比率计算。我们发现，尽管日光SPD 和天空光SPD 分布是完全不同的，但是在RGB 图像中，每个通道的SPD 比率大致近似于常数，并满足一定的统计规律。由此得到四种新的具有物理特性的阴影特征。在新的物理特征基础上，结合图像的纹理等特征，我们提出了一种新的基于区域和机器学习的单幅室外图像的阴影检测方法。通过实验和比较，表明了我们的方法优于目前几种经典的阴影检测方法。在智能交通及机器人自主导航等系统中，提出了一种基于光照不变图像分割和投票函数的图像道路及视觉导航线提取算法。该算法首先利用正交分解的方法获取彩色光照不变图像，并对其进行分割，然后通过构造的投票函数及道路判别准则，提取道路区域，最后通过扫描道路定位点并对其进行最小二乘拟合提取导航线。本方法不需要大量样本进行学习。通过实验验证，所提算法与现有两种算法相比，在检测精度和速度上均具有明显优势，并且算法复杂度相对较低，能够有效的解决各种室外光照环境下道路及导航线的提取问题。 |
| 英文摘要 | The complicated and changeful illumination environments bring many problems to computer vision and its applications, such as feature extraction, object segmentation, recognition and measurement. They degrade the performance of the algorithms in computer vision and the adaptiveness of these algorithms to the environments. It is one of the important research topics how to make the algorithms of image processing robust to the changeful illumination environments. We develop the research on the illumination problems in image processing based on theories of atmospheric physics, physical optics, imaging mechanism and from the view of characteristic analysis of physical imaging. Based on computable spectral irradiance, we present a computing method for illuminant imaging. On this basis, we study the illuminant processing in computer vision, such as shadow generation, shadow detection and road navigation, and propose some new corresponding models and algorithms. The main research contents and achievements include the following aspects: For computer vision and its applications, we propose a computing method for outdoor illuminant imaging based on simple and effective SPD calculating method. The Spectral Power Distributions (SPD) of outdoor light sources are vary over time and atmospheric conditions, which causes the appearance variation of a scene and common natural illumination phenomena, such as twilight, shadow, and haze/fog. Calculating the SPD of outdoor light sources at different time (or zenith angles) and under different atmospheric conditions is of important to computer vision. Computable spectral irradiance method can be applied in computer vision tasks including spectral inverse calculation, imaging computing and shadow image processing. By the calculation of the solar position, we also present a computing method of shadow direction and shadow length in an image for shadow generation. An image shadow generation method is proposed with shadow attenuation characteristics. By means of shadow edge processing, this method can generate image shadow which is close to the real physics. Compared with shadow generation methods in Graphics or based on matting and compositing shadow generation, this proposed method does not need complex calculations or any matting operation, and only calculate the pixel values to be generated in shadow based on their original pixel values, tricolor linear attenuation of shadows and the calculation of ratios between daylight and skylight SPDs. This shadow generation method with background in physical optics enhances the visual effects of shadow generation. A series of experimental results validate this method. For detecting the shadow in outdoor illumination conditions rapidly and efficiently, a shadow detection approach based on pixel-wise orthogonal decomposition is proposed. By the orthogonal decomposition, a color illumination invariant image and an illumination variation image are obtained. By K-means algorithm, the color illumination invariant image is classified into some regions, each of which has the same spectral albedo. According to the classification results, a Gaussian mixture model with expectation maximization algorithm is proposed for modeling the illumination variation image, and then the shadow areas are extracted. The extracted shadow areas are optimized with morphological operator. The proposed method does not need complex learning process of feature operators and reduces the time complexity of computation. It also does not require any prior knowledge and can be directly applied to the real-time scene processing. Four shadow features with physical properties and a shadow detection algorithms based on region and machine learning are proposed. The four new shadow features are derived from the ratio calculation of the daylight SPD and the skylight SPD. We found that, although the daylight SPD and the skylight SPD distributions are completely different, the SPD ratio of each channel in the RGB image is approximately constant and satisfies certain statistical laws. Four new shaded features with physical properties are obtained. Based on the new physical features and texture features of image, we propose a new shadow detection method for single outdoor image based on region and machine learning. Through experiments and comparisons, it is shown that our method is superior to several classical shadow detection methods. In intelligent transport system and robot autonomous navigation system, a road and visual navigation line extraction algorithm is proposed based on illumination invariant image segmentation and vote functions. First, a color illumination invariant image is acquired using orthogonal decomposition method. Then, with the segmentation of the image, the road region is extracted through the structure of the vote function and road criterion. Finally, the navigation line is detected by scanning path location and the least squares fitting. The proposed method does not need a large number of samples to learn. Experiments validate that the proposed algorithm has obvious advantages compared with the two existing algorithms in detection accuracy and speed, and its complexity is relatively low. The proposed algorithm is effective in the road and navigation line extraction under variable outdoor illumination environment. |
| 语种 | 中文 |
| 产权排序 | 1 |
| 页码 | 101页 |
| 源URL | [http://ir.sia.cn/handle/173321/19459]  |
| 专题 | 沈阳自动化研究所_机器人学研究室 |
| 推荐引用方式 GB/T 7714 | 段志刚. 室外自然光照成像计算与阴影处理技术研究[D]. 中国科学院沈阳自动化研究所. 2016. |
入库方式： OAI收割
来源：沈阳自动化研究所
浏览0
下载0
收藏0
其他版本
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。