【作者】 刘艳丽；

【导师】 彭群生； 秦学英；

【作者基本信息】 浙江大学 ， 应用数学， 2009， 博士

【摘要】 增强现实是近二十年来信息领域中迅速兴起的一种技术。增强现实技术将计算机生成的景物实时地叠加到一个真实场景的画面中,通过将真实信息和虚拟信息互相融合,创造出一个和谐统一的环境。由于具有对真实环境进行增强显示的特性,增强现实技术在文化遗产保护、医疗手术规划、军事训练等领域都具有独特的优势。但是,迄今为止,增强现实技术中仍有大量的问题尚未解决。其中,实时重构室外场景的光照条件用于虚拟物体的绘制,实现虚拟物体与真实场景的光照一致性对室外场景虚实融合的真实感具有十分重要的作用。室外光照估计也是计算机视觉的研究课题之一,变化的室外光照往往导致阴影检测、物体识别、视频跟踪等算法的不稳定。实时地求解出室外场景视频中每一帧的光照条件,然后将光照进行归一化处理去除变化光照的影响可极大地提高许多算法的性能。如何在场景三维模型未知的条件下实时地估计在线视频的光照条件对计算机视觉和计算机图形学都具有十分重要的意义。本文围绕固定视点下室外场景视频序列的实时光照估计展开研究,针对现有方法的不足提出了一系列新的解决方法。本文工作的主要贡献和创新如下:·首次给出了室外场景图像的统计参数与场景光照参数之间的解析表达式。在假定太阳光为平行光,天空光为泛光的条件下,从基本的光照明模型出发,推导出了场景图像的方差和均值与场景的太阳光和天空光入射光强之间互为表达的解析表达式,揭示了图像的统计属性与场景光照之间的关系,为研究场景的光照条件提供了新的思路和途径。·基于图像的统计参数与光照参数之间的解析表达式,提出了一个估计室外场景光照的框架。该框架在离线阶段通过学习建立解析表达式,在在线光照估计阶段,根据建立好的解析表达式,利用图像的统计参数实时地计算出场景的太阳光和天空光入射光强。为了对动态场景取得稳定的光照估计结果,进一步提出了一种基于光照的空间连贯性和时间连贯性对已求的光照参数进行光顺的方法,获得了稳定的求解结果。·将天空光模型推广到更为一般的面光源模型,提出了一个基于图像分解的室外场景光照估计框架。首先将一幅室外场景的图像表示成太阳光基图像和天空光基图像的线性组合,其中组合系数是所要求解的太阳光和天空光入射光强,太阳光和天空光基图像则分别表示了场景的几何、材质的乘积。在离线阶段通过学习求解出采样太阳位置下的太阳光和天空光基图像。在在线阶段通过对采样太阳位置处场景的太阳光基图像进行更新,实时求解视频中每一帧的太阳光和天空光入射光强。·提出了一种无需离线学习的室外场景光照估计方法。首先使用上述图像分解模型,将图像表示成太阳光和天空光基图像的线性组合。然后,通过对太阳光基图像的特点进行分析,将太阳光和天空光入射光强的求解归结为一个可实时求解的能量最小化问题。该方法对场景的材质无要求,允许场景中存在具有复杂材质属性的物体。除了增强现实,该方法可广泛应用于一般的在线视频处理,如阴影检测、基于图像的重光照、颜色恒常性以及光照归一化等。本文所提出的三种估计室外光照的方法均不需要场景的三维几何信息,因而避免了对大规模室外场景进行重建时面临的种种技术挑战,使光照估计算法更为实用。更多还原

【Abstract】 Augmented reality grows rapidly during the past two decades. By overlaying the computer-generated objects onto the real scenes, augmented reality can create a harmonious environment. Since augmented reality can enhance the display of the real world, it has been widely used in protection of digital cultural heritage, medical surgery planning, military training and so on. However, many open problems in augmented reality still remain, of which estimating the illumination of outdoor scenes in real time and then shading the virtual objects to maintain illumination consistency plays an important role in achieving high realism of an augmented reality system.Also, outdoor illumination estimation is one of research topics in computer vision. Varying illumination usually severely degrades the performance of algorithms proposed in object recognition, video segmentation, video tracking, etc. If illumination is estimated in real time, then it can be normalized, and hence the performance of those algorithms can be greatly improved. Therefore, the real-time illumination estimation of outdoor scenes is of great importance for both computer graphics and computer vision.The dissertation focuses on real-time illumination estimation of outdoor videos captured under a fixed view. The contributions of the dissertation are as follows:·An analytical expression relating image statistics and scene illumination is first derived. Under the assumption that the sunlight is directional and the skylight is ambient, the thesis derives an analytical expression between the mean and deviation of an image and the sunlight and skylight of a scene. The expression provides a new way to understand the outdoor illumination.·Based on the analytical model, a framework for real-time estimation of outdoor illumination is developed. In this approach, the correlation between the illumination parameter and the image statistics is first constructed from a set of images at off-line stage. At online stage, given a new input image captured from the same view, the illumination parameter is derived from the image statistic properties via the pre-computed correlation. In order to handle occasional motions occurred in outdoor scenes, an algorithm exploiting spatial and temporal illumination coherency is proposed to smooth the estimation results.·The skylight is further extended to a uniformly distributed area light source. Then the thesis proposes a linear model to represent an outdoor image as a linear combination of the sun basis image and sky basis image. While the sunlight basis image and skylight basis image encode the effect of the scene geometry, surface reflectance, the coefficients are the sunlight and skylight which we wish to recover. Based on the linear model, a framework for estimating outdoor illumination is developed. The framework obtains sun basis image and sky basis image under some key sun positions at off-line stage. Then at online stage, by updating the sun basis image of key sun positions, the illumination parameters of every frame are obtained in real time.·A novel approach for estimating outdoor illumination without off-line learning is proposed. The model representing an outdoor image as a linear combination of sun basis image and sky basis image is first adopted. Exploiting the characteristics of the sun basis image, the computation of sunlight and skylight of every frames is finally reduced to a minimization problem which can be solved in real time. This method involves no reflectance assumption and permits objects with complex reflectance such as specular, or anisotropic surfaces. Compared with previous work, this method requires no preprocess stage and is suitable for both video post-processing and online processing such as augmented reality, shadow detection, relighting, color constancy and illumination normalization.None of the above approaches requests information of scene geometry, thus getting rid of the difficulties in reconstructing the 3D geometry of large-scale outdoor scenes, making the outdoor illumination estimation more practical.更多还原