【作者】 池凌鸿；

【导师】 郭立；

【作者基本信息】 中国科学技术大学 ， 电路与系统， 2011， 博士

【摘要】 本文针对计算机立体视觉中的立体匹配中存在的问题,对双目立体校正、基于马尔可夫随机场的置信传播算法的优化及其高性能实现和立体匹配用于静态场景目标提取进行了研究。本文主要工作和创新点如下:1)针对传统立体校正需要对摄像机进行标定以获取摄像机参数的问题,提出了一种无需摄像机参数的简便的立体校正方法。该方法先对待校正图像对进行特征点匹配,得到两幅图像间的空间坐标相对关系,然后利用极线几何约束将两幅图像的对应点校正到同一水平线上,完成对立体图像对的校正。实验表明,这种校正方法能有效的校正未知摄像机参数的图像对。2)针对全局立体匹配中基于马尔可夫随机场的置信传播算法中计算时间随消息迭代次数线性增长的问题,提出了一种基于自适应机制的分层置信传播方法。该方法利用分层置信传播算法中大部分消息快速收敛的性质,引入消息收敛的条件判断,在迭代上限相同情况下,减少了算法的迭代次数,缩减了整体迭代的时间。实验表明,与传统HBP相比,该方法有效的缩减了计算时间,而且计算时间对整体迭代上限不敏感。3)针对分层置信传播算法难以快速实现的问题,提出了一种利用GPU对其进行高性能的并行计算的方法。该方法结合CUDA编程的特点,分析了分层置信传播算法的特点,对其进行像素级的并行化处理,使计算的吞吐率有效增加。实验表明,在相同的匹配水平下,该方法具有比较高的加速比。4)针对目前现有的立体视频目标提取方法主要依赖于运动场,提出了一种基于精确视差的静态目标分割方法。该方法首先通过基于图像分割与平面拟合的自适应全局立体匹配算法得到得到包含场景深度信息的精确视差图。然后由视差图的特性,先对其进行前后景分离,然后对前景进行遍历,得到其类间方差最大的灰度分割方法,完成对目标的提取。实验表明,该方法分割结果准确,目标提取效果好。更多还原

【Abstract】 To solve those problems on stereo vision in computer vision, we have studied on stereo rectification, the optimization and parallelization of belief propagation based on MRF and the static object segmentation based on disparity map.The main work and innovation of this dissertation are as follows:1) A method of stereo rectification with uncalibrated cameras is proposed.In this method,we at first detect and match the interest points to achieve the spatial relationship between the two images, and then finish the rectification by finding the projections in which the epipolar lines run parallel with the x-axis according to the epipolar geometry. Experimental results show that this method can rectify the stereo pairs accurately and meet the requirement of stereo matching.2) A self-adaptive algorithm with convergence detection to reduce the computational complexity of HBP is proposed. Convergence detection is introduced to stop the iterations of messages which are already converged to optimal values. Thus the overall computational time is reduced. Experimental results show the self-adaptive algorithm reduces computational time effectively, and the computational time is insensitive with iteration up bound. The convergence detection methodology can also be applied to other HBP related applications.3) An efficient CUDA-based graphic processing unit is introduced into implementation of the belief propagation algorithm. After analysis for the belief propagation algorithm, we achieve the parallelization on pixel-level by CUDA.Experimental results show that this method can be used to speed up stereo image processing without much loss of accuracy.4) An object segmentation algorithm based on accurate disparity map is presented. The accurate disparity map is available by a stereo match algorithm including initial matching cost estimation, mismatched pixels checking, plane estimation and self-adaptive hierarchical belief-propagation. Then, self-adaptive threshold segmentation is performed on the result that is achieved from the first step. Experimental results show that the proposed algorithm is an effective object extraction method suitable for stereoscopic static scenes and image sequences with unitary global motion.更多还原