基于多视域广角相机视频图像拼接技术研究

发布时间：2018-04-27 15:39

本文选题：视频拼接 + 畸变校正　；参考：《中国科学院研究生院(长春光学精密机械与物理研究所)》2016年博士论文

【摘要】：目前视频拼接技术是将具有重叠区域的多幅视频图像进行拼接,进而可以形成宽视场视频图像。与单目视频图像相比较,其在展现大尺寸视频画面的同时,也可以提供更加丰富的信息。随着目前多媒体移动设备不断地普及化以及廉价化,视频拼接技术显然已经成为图像处理领域研究的热点。视频拼接相对于图像拼接而言,存在着视频实时性以及同步性等难点问题。本文采用三路视频流进行相关实验,并且采用基于改进的SIFT算法完成对本文视频拼接技术的研究。本文主要的研究内容:(1)本文首先从图像拼接的相关理论进行简单的介绍,主要包括摄像机的成像模型,图像坐标变换之间的关系,以及摄像机的运动模型,图像配准的含义,并对常用的配准技术进行了简单的介绍,为后续内容作铺垫。(2)为了权衡摄像头个数以及视场之间的关系,本文选用了广角相机对其进行成像。由于广角相机所获得的图像边缘会有严重的畸变,所以首先必须对畸变进行处理,本文在选用基于直线特征的标定方法基础上,并且提出了一种弯曲测度指标函数,靠近图像中心不同距离的曲线给予不同的权重因子,使得畸变校正更加准确。(3)首先对SIFT特征提取算法进行详细的分析,并且针对视频拼接的实时性要求对其进行改进。首先对图像的感兴趣区域进行限制。由于SIFT算法提取的特征点多而且过于密集,不利于后续特征点的匹配与投影变换矩阵的计算,所以有必要对其进行改进。第一步扩大极值点的搜素范围,第二步是对提取的特征点进行距离约束。仿真实验结果表明,改进后的SIFT算法比原算法速度更快,特征点分布也更加的均匀,同时也保持该算法原有的对光照以及旋转的鲁棒性。(4)本文选用了基于次近邻与最近邻距离比值的特征点匹配算法对特征点进行粗匹配,并且选用BBF搜索策略搜索最近邻距离与次近邻距离特征点,采用RANSAC算法进行特征点的提纯,然后使用提纯后的特征点进行投影变换矩阵的估计。为了消除配准图像产生的拼接缝,本文首先对现有的图像融合算法进行分析与实验对比,总结得到,加权平均融合算法能够很好地保留图像的细节信息,获得更高的融合图像质量,在此选用此种方法对图像进行融合。(5)由于摄像头的位置是固定的,在成像过程中摄像机的参数也是固定不变的,所以只需要对首帧图像进行特征提取计算出投影变换矩阵,其余视频帧都利用此投影变换矩阵进行操作,并且采用了OpenMP并行编程来优化整个系统,提高程序运算效率。为了保持视频图像的同步采集,本文采用Directshow技术对视频图像进行采集,同时设计了一种对每一路视频在服务器中开发一个缓冲区进行视频流的存储机制。
[Abstract]:At present, the technology of video mosaic is to concatenate multiple video images with overlapped regions, which can form wide field of view video images. Compared with monocular video images, it can provide more information while displaying large scale video images. With the popularization and low cost of multimedia mobile devices, video mosaic technology has obviously become a hot spot in the field of image processing. Compared with image stitching, video stitching has many difficulties, such as real-time and synchronism. In this paper, three channels of video streams are used to carry out related experiments, and the improved SIFT algorithm is used to complete the research of the video mosaic technology in this paper. The main research content of this paper is: (1) this paper firstly introduces the theory of image mosaic, including the imaging model of camera, the relationship between image coordinate transformation, the motion model of camera and the meaning of image registration. In order to weigh the relationship between the number of cameras and the field of view, a wide-angle camera is used to image the registration technology. Because the edge of the image obtained by the wide-angle camera will have serious distortion, the distortion must be processed first. In this paper, based on the calibration method based on linear features, a bending measure index function is proposed. The curve near the center of the image gives different weight factors, which makes the distortion correction more accurate. 3) firstly, the SIFT feature extraction algorithm is analyzed in detail, and it is improved to meet the real-time requirements of video stitching. First of all, the region of interest of the image is restricted. Because the feature points extracted by SIFT algorithm are many and too dense, it is unfavorable to the matching of feature points and the calculation of projection transformation matrix, so it is necessary to improve it. The first step is to expand the search range of extremum points, and the second step is to restrict the distance of the extracted feature points. The simulation results show that the improved SIFT algorithm is faster than the original algorithm, and the distribution of the feature points is more uniform. At the same time, it also maintains the robustness of the algorithm to illumination and rotation.) in this paper, the feature point matching algorithm based on the ratio of the distance between the next nearest neighbor and the nearest neighbor is used to match the feature points rough. BBF search strategy is used to search the nearest neighbor distance and the next nearest neighbor distance feature point, RANSAC algorithm is used to purify the feature point, and then the projection transformation matrix is estimated by using the purified feature point. In order to eliminate the stitching produced by the registration image, this paper firstly analyzes and compares the existing image fusion algorithms, and concludes that the weighted average fusion algorithm can preserve the details of the image very well. To achieve higher fusion image quality, this method is used to fuse the image. (5) because the position of the camera is fixed, the camera parameters are also fixed during the imaging process. So we only need to extract the features of the first frame image to calculate the projection transform matrix, and the other video frames are operated by the projection transformation matrix, and the OpenMP parallel programming is used to optimize the whole system and improve the efficiency of the program. In order to maintain the synchronous acquisition of video images, this paper adopts Directshow technology to capture video images, and designs a storage mechanism of video streams in a buffer for each video stream in the server.
【学位授予单位】：中国科学院研究生院(长春光学精密机械与物理研究所)
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】