采用空间一致生长策略的多视图三维重建  

Multi-view Stereo Reconstruction via Homogeneous Spatial Expansion

作　　者：李亚兰[1,2] 李志扬 Li Yalan;Li Zhiyang(College of Physical Science and Technology, Central China Normal University, Wuhan 430079;School of Electronic Information and Electrical Engineering, Xiangnan University, Chenzhou 423000)

机构地区：[1]华中师范大学物理与科学技术学院,武汉430079 [2]湘南学院电子信息与电气工程学院,郴州423000

出　　处：《计算机辅助设计与图形学学报》2018年第1期124-137,共14页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(11605071);湖南省高校科技创新团队资助项目(湘教通2014207);中央高校基本科研业务费专项资金项目(CCNU16A05029)

摘　　要：针对复杂环境因素影响下场景三维重建结果难以同时获得高重建精度和高重建完整度的问题,提出一种基于空间一致生长的多视图三维重建算法.该算法对基于特征点生长的三维重建算法框架进行拓展,新增了有条件的初值矫正环节,同时对已有的生长和滤波环节进行更替和改进.首先使用SFM从输入图像中提取稀疏种子点;然后直接在固定的世界坐标系中从所有种子点出发向其邻近三维空间扩展,得到生长点的初始位置和方向,并通过优化确定生长点的最终位置和方向,在优化前和优化中根据生长点的当前位置和方向不断选取和更换最佳主、副图,提高优化质量;再利用邻域已重建点云有条件地矫正生长点的位置和方向,并以此作为初值再次优化,避免优化收敛到局部极值,提高生长点的重建精度;最后设计光滑、深度和方向三方面的自适应一致性滤波,在减少误删的同时及时删除误差点,防止误差蔓延.实验结果表明,文中算法的重建精度和完整度均明显超过当前流行的PMVS算法;与高精度DAISY算法相比,在完整度大大提高的前提下,该算法的精度与其基本持平.A multi-view stereo reconstruction algorithm based on homogeneous spatial expansion was proposed to overcome the difficulty of reconstructing3D models with both high accuracy and high completeness for scenes in complex situations.It expanded the frame work of the feature expansion algorithm with a new conditional initial value rectification step.Meanwhile the existing steps were either replaced or improved.Firstly,sparse seed points were extracted by SFM method.Then they were expanded simultaneously to their nearby3D space directly in a fixed global world coordinates system to gain initial positions and normals for growing points.The final positions and normals of the growing points were determined via optimization,during which the best reference and auxiliary views were selected and replaced according to the current status of the growing points.Next,to avoid converging into local minima and improve the reconstruction accuracy,the positions and normals of grown points were conditionally rectified taking advantage of their neighbor points.Once rectified,re-optimization was carried out with new initial values.Finally,to suppress the spreading of errors the adaptive consistent filters for smoothness,depth and direction were designed to remove outliers and keep accurate points as far as possible.The experiments showed that both the reconstruction accuracy and completeness of the proposed algorithm are obviously higher than that of the popular PMVS algorithm.With much improved completeness,its accuracy is still on a par with the high accuracy of DAISY algorithm.

关 键 词：多视图 三维重建 特征点生长 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]