SIA OpenIR  > 光电信息技术研究室
基于特征匹配的多视图三维重建
其他题名Multi View 3D Reconstruction Based on Feature Matching
孙会超1,2
导师惠斌
分类号TP391.41
关键词特征提取与匹配 增量式三维重建 种子点 准稠密扩散
索取号TP391.41/S96/2018
页数57页
学位专业控制工程
学位名称硕士
2018-05-17
学位授予单位中国科学院沈阳自动化研究所
学位授予地点沈阳
作者部门光电信息技术研究室
摘要基于特征匹配的多视图三维重建属于被动式三维重建,该技术从同一场景不同视角的图片恢复目标的三维结构。目前,基于特征匹配的多视图三维重建主要包括图像点特征提取与匹配、多视图稀疏重建、稠密化重建等部分。本文对基于特征点的多视图三维重建涉及到的关键技术进行研究,主要的研究如下:(1)研究了图像特征点提取与匹配算法,提出了一种改进的混合尺度空间Harris角点检测和SIFT特征描述的图像匹配算法。该算法在高斯尺度空间分层检测Harris角点,构建出角点的多尺度响应金字塔,在多尺度空间检测出角点特征;然后采用SIFT特征描述的方式为角点构建特征描述符;最后采用近邻之比的双向匹配策略进行图像特征匹配。实验证明该算法用于多视图图像匹配,具有尺度、光照、模糊不变性,在保持匹配精度的同时,能大大减少匹配时间。(2)本文采用基于两视图重建的增量式三维重建方法。首先对两视图进行重建,输入图像的特征匹配对,采用RANSAC算法估计出基本矩阵并去除外点;然后通过读取图片的EXIF信息实现相机的自标定并得到本质矩阵;之后由本质矩阵分解得到摄像机投影矩阵,并通过线性三角形法计算出特征点的空间坐标,得到稀疏点云;最后累加迭代不同视角的点云并进行优化,得到多视图稀疏三维点云。(3)为了从少量图像恢复场景可视化三维点云,提出了一种改进的准稠密扩散算法。准稠密扩散主要包括种子点选取、生长扩散和匹配过滤三部分。该算法采用尺度不变Harris角点作为准稠密扩散的种子点;通过视差梯度约束生长种子点;最后通过极约束和置信度限制进行匹配过滤;实验证明改进的准稠密扩散算法用于准稠密三维重建,可以有效的实现少量视图获取空间立体目标的可视化三维结构,并提高重建的精度。
其他摘要Multi view 3D reconstruction based on feature matching is a passive method of 3D reconstruction, which reconstructs 3D structure from different views of the same scene. At present, multi view 3D reconstruction based on feature matching mainly includes image point feature extraction and matching, multi view sparse reconstruction, dense reconstruction and so on. In this paper, the key technologies of multi view 3D reconstruction based on point feature are studied. The main research is as follows: (1) The algorithm of image point feature extraction and matching is studied, a mixed multi scale space Harris corner detection and SIFT feature description algorithm is proposed. This algorithm detects Harris corners in the Gauss scale space, and detects the feature of the corner points in the multi-scale space. Then construct the feature descriptors for the corner points by SIFT feature description. Finally the bidirectional matching strategy of is applied to match the image features. The experiments show that the algorithm has the invariance of scale, illumination and fuzzy on image matching, it can reduce the matching time greatly while maintaining the matching precision. (2) The multi view sparse 3D reconstruction method based on image feature is studied. We applied incremental 3D reconstruction method to recover 3D point cloud. First, we reconstructe 3D point cloud from the matches of two views, the fundmental matrix is estimated by RANSAC algorithm and the outliers is removed. Then the camera’s focal length is read by the EXIF information of the picture. Then the essential matrix is estimated and the camera projection matrix is decomposed by the essential matrix. And the linear triangulation is used to calculate the 3D points. Finally, point clouds with different viewpoints are iterated and optimized, and multiple view sparse 3D point clouds are obtained. (3) The 3D point cloud visualization is studied, a quasi dense propagation algorithm based on scale invariant Harris corner is proposed. Quasi dense propagation method includes three parts: seed selection, seed propagation and matches filtering. The algorithm uses the scale invariant Harris corner as the seed point, and obtain more matching through growth transmission and. And filtering the matches by epipolar constraint. Experiments show that the algorithm can effectively achieve the visual 3D point cloud of the target.
语种中文
产权排序1
文献类型学位论文
条目标识符http://ir.sia.cn/handle/173321/21826
专题光电信息技术研究室
作者单位1.中国科学院沈阳自动化研究所
2.中国科学院大学
推荐引用方式
GB/T 7714
孙会超. 基于特征匹配的多视图三维重建[D]. 沈阳. 中国科学院沈阳自动化研究所,2018.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
基于特征匹配的多视图三维重建.pdf(2986KB)学位论文 开放获取CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[孙会超]的文章
百度学术
百度学术中相似的文章
[孙会超]的文章
必应学术
必应学术中相似的文章
[孙会超]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。