作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程

• 图形图像处理 • 上一篇    下一篇

基于光线的全局优化多视图三维重建方法

陈 坤,刘新国   

  1. (浙江大学计算机辅助设计与计算机图形学国家重点实验室,杭州 310058)
  • 收稿日期:2012-12-11 出版日期:2013-11-15 发布日期:2013-11-13
  • 作者简介:陈 坤(1988-),男,硕士,主研方向:计算机视觉;刘新国,教授、博士、博士生导师
  • 基金资助:
    国家自然科学基金资助项目(61379068);中央高校基本科研业务费专项基金资助项目(2011QNA5029)

Global Optimized Multi-view 3D Reconstruction Method Based on Rays

CHEN Kun, LIU Xin-guo   

  1. (State Key Lab of CAD&CG, Zhejiang University, Hangzhou 310058, China)
  • Received:2012-12-11 Online:2013-11-15 Published:2013-11-13

摘要: 利用光线跟踪原理,提出一种全局优化的多视图三维重建方法。根据图像轮廓得到物体的包围盒,采用体素离散物体所在的几何空间。从相机中心向图像上每个像素发射一条光线,为确定光线达到的体素,使用归一化互相关(NCC)值度量光线-体素的一致性,并估计采样空间中面片的法向信息,以提高NCC值的可信度。设计基于因子图的全局优化模型得到物体体素,针对光线因子的特殊性,设计一种高效的置信度传播算法,使重建方法的时间复杂度从指数阶降为线性阶。实验结果表明,与基于马尔可夫场的重建方法相比,该方法的鲁棒性较好,可提高重建模型的准确度和完整性。

关键词: 三维重建, 因子图, 可视壳, 多视图立体匹配, 归一化互相关

Abstract: This paper presents a multi-view 3D reconstruction method based on rays. It generates the bounding volume for the object in the images using silhouettes, and discretizes space of the object into voxels. For each pixel in the image, a ray is generated from the center of the camera. To determine the voxels for the rays, Normalized Cross-correlation(NCC) value is used to measure the consistence between the rays and the voxels, where the patch normal is estimated to improve the confidence of the corresponding NCC value. A global optimization model based on factor graph is designed to extract the voxels belonging to the reconstructed object. In addition, an efficient belief propagation algorithm is proposed based on the characteristics of ray factor, which reduces the computational complexity from an exponential one to a linear one. Experimental results show that the proposed method is more robust than the previous reconstruction method based on Markov Random Field(MRF), and achieves better reconstruction result in terms of both accuracy and completeness.

Key words: 3D reconstruction, factor graph, visual hull, multi-view stereo matching, Normalized Cross Correlation(NCC)

中图分类号: