【作者】 郭龙源；

【导师】 杨静宇；

【作者基本信息】 南京理工大学 ， 模式识别与智能系统， 2009， 博士

【摘要】 计算机视觉主要研究如何利用计算机实现人的视觉功能,即利用二维投影图像实现对客观世界三维场景的感知、识别和理解。立体匹配是计算机视觉和非接触测量研究中最基本的关键问题之一,该技术通过像点的视差来获取深度或距离信息,可以为三维重建、机器人导航、自主车导航等提供有用的信息。在实际应用中,因为变形,扭曲,遮掩等情况的影响,立体匹配是较难彻底解决的一个病态问题。在计算机视觉技术中,双目视觉更接近于人的双眼视觉原理,并且在实际应用中更容易实现。本课题对双目视觉的立体匹配相关理论及一些方法进行了研究,并给出了阶段性的成果。基于区域相关匹配的方法是传统的匹配方法,但计算量较大,影响了其在实践中的广泛应用。为了减少算法的计算量,提高算法速度,根据视差梯度和搜索范围的关系,本文提出了一种基于视差梯度的可变搜索范围区域匹配方法。核心思想是每个点的匹配搜索范围可以由该点的视差梯度和它前一点的视差来确定,而不必每个点匹配时都要在最大估计视差范围内搜索。这样可以去除冗余搜索,减少匹配时间,提高匹配精度。图像变换是许多图像处理分析中的一种有效手段,根据Rank变换的原理,并引入了Census和色差梯度的约束条件,本文提出了基于Rank变换的彩色图像匹配方法。结果表明采用Rank变换特征作为匹配基元比直接采用像素灰度值作为匹配基元得到的视差图更精确,并在一定程度上抑制了噪声对匹配结果的影响;另外汉明码距离约束条件和色差梯度的引入减少了误匹配的发生,从而进一步提高了算法的正确匹配率。红外图像高噪声,低分辨率的特点,使得对它们进行视差匹配时,基于灰度的区域匹配方法很难得到较好的效果。本文提出了一种基于相位一致性变换的匹配方法。经相位一致性变换后的图像,噪声得到了一定的抑制,图像的特征更加明显,基于变换后图像的匹配效果更好。障碍物检测是自主车(ALV)、机器人导航以及汽车辅助驾驶研究中重要的研究课题。基于立体视觉的障碍物检测是目前障碍物检测中最常用的方法。本文采用了一种基于彩色图像的障碍物检测的方法,利用彩色图像HSV空间对图像进行分割,并通过立体匹配得到障碍物位置。该方法对于检测道路,车辆,房屋等人工目标和抑制自然背景有较好的效果。更多还原

【Abstract】 Computer vision mainly studies how to realize human visual function with computers. The main idea is to perceive, identify and understand the three-dimensional scene from two-dimensional projective images. Stereo matching is one of the key issues in computer vision and non-contact measurement. The information of depth or distance can be obtained from disparity of the points. It is useful in three-dimensional reconstruction, robot navigation, vehicle navigation, and so on. Stereo matching is an ill-posed problem with the influence of distortions and occulsions. In computer vision, binocular vision is similar to the mechanism of human binocular vision, and easy to achieve in practical applications. The relevant theories and approaches of binocular stereo matching have been studied in this thesis, and some progressive achievements have been made.Area correlation is traditional approach in stereo correspondence. But its wide application is affected by the heavy computational burden. To reduce computation and enhance algorithm speed, the paper presents a variable search region area-matching approach based on disparity gradient, according to relationship between disparity gradient and search region. The core of this method is that the search region of each point in one image can be ascertained by itself disparity gradient and its former point’s disparity. Not all points need matching in the largest search region under the present algorithm. So, the algorithm can decrease redundant search, shorten the matching time and improve accuracy of matching.Image transformation is widely and effectively used in image processing. According to the principle of rank transformation and census constraint and color difference gradient constraint, the paper presents a color image matching algorithm based on rank transformation. The experiment results show that the disparity of rank transformation image is more precise than that of intensity image. At the same time, the matching result is more robust by noise influence to a certain extent. In addition, census constraint and color difference gradient constraint further reduce outlier and further enhance right matching ratio.Infrared images have higher noise and lower resolution than visible images. These features makes area matching base on intensity more difficult to obtain a good disparity image. After analyzing the phase congruency transformed image, the paper presents a novel area matching method. By this way, noise in images is restrained and the images features are more distinct. The result of correspondence based on transformation image is better.Obstacle detection is an important research subject in ALV, robot navigation and driver assistant system. The vision-based approach is the most commonly method in obstacle detection. The paper present an obstacle detection algorithm based on color image. The image is segmented under HSV of color image first. And the obstacle locations are obtained with stereo correspondence. The approach is good for detection of path, vehicle and building and can restrain the nature background.更多还原