word原稿】基于球面全景图像的目标识别方法研究.docVIP

基于球面全景图像的目标识别方法研究 佟国峰，谷久宏 （东北大学信息科学与工程学院, 沈阳 110819） 5 摘要：本文研究球面全景图像中的物体识别问题。由于球面全景图像与平面图像的不同，所 以传统的物体识别方法不再适用。该问题具体描述为，当用户给定一张普通的平面图像时， 如何从球面全景图像数据集中找出对应的全景图像。本文利用图像特征搜索来得到相似度最 高的全景图像。根据球面全景图像的性质，推导出球面全景图像的本质矩阵，与单应矩阵， 10 利用 SIFT 算法，将图像数据重新表达为特征向量，并以 K-D 树结构进行存储。通过近邻搜 索，以球面图像的对极几何关系为约束模型利用 RANSAC 算法剔除误匹配点。根据最后的 匹配结果，找到对应的球面全景图像。接着利用球面图像之间的单应关系，在全景图像中检 测出平面图像对应的位置，并通过实验验证了本文提出的基于球面全景图像的目标识别方法 的正确性。 15 关键词：球面全景图像；本质矩阵；单应矩阵；特征提取；对极几何 中图分类号：TP391.4 Research on object recognition base on spherical panoramic image 20 TONG Guofeng, GU Jiuhong (College of Information Science and Engineering, Northeastern University, Shenyang 110819) Abstract: In this paper, we focus on the problem of object recognition in spherical panoramic image. Since the difference between spherical panoramic image and plane image, the traditional methods cou- ld not be used in this problem directly. An instance of the object is selected by a 25 user in an image. The aim is to return the occurrences of the selected object from the panorama sets quickly. The feature det- ected in the given image used to find the related panoramic images. We derive the essential matrix and homograph matrix in spherical panoramic image base on the property of panoramic image. We represe- nt the feature provided by SIFT in image as vectors of feature. A common method is to organize the de- scriptors of single or multiple images in a 30 K-D-tree structure, which allows efficient nearest-neighbor search for a given query feature. We would reject the wrong corresponding points with RANSAC to ca- lculate the homograph matrix, and then the corresponding position in spherical panoramic image will be detected. The experiments verify the validity of the method proposed in this paper. Key words: spherical panoramic image；essential matrix；homograph matrix；feature detection； 35 epipolar geometry 0 引言 为了构建一个虚拟的环境，需要获得环境的整个三维结构。如果使用普通的相机，就有 视角狭小的缺陷，这样就需要多张图像来构建整个三维环境的一部分，并且需要融合这些三 40 维数据。全景图像是一种覆盖大范围场景的宽视角图像，可以解决普通相机视角的问题