视频图像自动拼接的研究与实现

【作者】 王超；

【导师】 张凯；

【作者基本信息】 北京理工大学 ， 计算机系统结构， 2008， 硕士

【摘要】 全景图像拼接技术已成为虚拟现实和计算机视觉的重要场景表示方法。利用该技术对传统的视频监控系统进行升级,使其具有了全方位、立体性和实时性的优点。论文研究的系统在Linux环境下,使用视角重叠的六个摄像头和快速视频采集卡实时采集视频流,然后对具有重叠区的相邻图像分别提取特征并进行特征匹配,最终融合生成YUV格式的全景视频流供后续实时编码和播放。本论文在汲取传统方法优点的基础上,对全景视频图像的实时自动拼接技术在以下几方面进行了深入的研究。视频图像的拼接过程可分为图像配准和融合处理两个部分。在图像配准方面,主要采用基于特征的SIFT算法进行图像特征提取和匹配,并使用稳健的随机抽样一致性优化算法对匹配特征进行过滤、提纯。在图像融合方面,为了消除曝光差异和减少“鬼影”影响,采用了改进的加权平均方法和多分辨率融合方法。由于系统对视频图像拼接的高实时性要求,本文还对特征点的提取算法、匹配算法和图像融合算法进行了优化,最终实现了全景视频图像的实时自动拼接,拼接后的视频图像能够真实反映场景特征。论文还将改进后的算法与传统方法进行了比较和评估,证明了改进算法的可行性。更多还原

【Abstract】 Full-view panoramic image mosaics has become a important way to study virtual reality, computer vision. I have upgraded the traditional video surveillance systems using the technology to make it to be better. The system using six cameras with overlapping perspectives and video capture card collecting video stream on Linux operating system. Then features from the overlapping areas of adjacent video images were extracted and matched. Finally, a YUV format panoramic video to be encoded and played was generated . This paper presents a novel approach to creating full-view panoramic mosaics based on learning the traditional theory and technology. I have conducted in depth study in the following aspects. Full-view panoramic image mosaics can be divided into image registration and image blending. In the image registration, I use SIFT algorithm to extracting and matching features from continuous two images automatically and accurately. And then I use RANSAC algorithm to optimize the result of matching. In the image blending, in order to reduce the exposure differences and eliminate ghosting effects, I use the improved weighted average method and multi-resolution blending method. As high real-time requirements of the video image mosaic system, the feature extraction algorithms, matching algorithms and image blending algorithm has been optimized. Eventually I effectively achieve a panoramic video stream which can truly describe the scene of the world. The paper also compare the improved algorithm with the traditional methods to prove the feasibility of improving the algorithm.更多还原