CAJ | 학술논문

该文利用运动恢复结构（structure from motion，SFM）、多视图立体视觉（multi-view stereo，MVS）技术，提出了一种坡面侵蚀沟三维模型的快速重建方法。首先对普通相机拍摄的照片采用尺度不变特征变换（scale-invariant feature transform，SIFT）完成特征点的提取与描述，随机采样一致性算法（random sample and consensus，RANSAC）过滤掉最近邻匹配（nearest neighbor，NN）产生的误匹配点；然后通过SFM方法，迭代求解出相机矩阵和三维点坐标，用光束法平差（bundle adjustment，BA）进行非线性优化，确保误差的均匀分布和模型的精确；再使用基于面片的多视图立体视觉算法（patch-based multi-view stereo，PMVS），在局部光度一致性和全局可见性约束下，以SFM生成的稀疏点云为种子面片开始扩散，完成点云稠密重建。将照片快速重建方法获取的点云与地面激光扫描仪（terrestrial laser scanner, TLS）获取的点云及实测数据进行比较，结果表明，照片重建方法生成的点云稠密且能够完整展示侵蚀沟的发育形态，与 TLS点云间的平均距离为0.0034 m，照片重建与三维激光扫描方法对侵蚀量的估算相对误差为8.054%，提取的特征线匹配率达89.592%。研究结果为侵蚀沟监测提供了参考依据。
해문이용운동회복결구（structure from motion，SFM）、다시도입체시각（multi-view stereo，MVS）기술，제출료일충파면침식구삼유모형적쾌속중건방법。수선대보통상궤박섭적조편채용척도불변특정변환（scale-invariant feature transform，SIFT）완성특정점적제취여묘술，수궤채양일치성산법（random sample and consensus，RANSAC）과려도최근린필배（nearest neighbor，NN）산생적오필배점；연후통과SFM방법，질대구해출상궤구진화삼유점좌표，용광속법평차（bundle adjustment，BA）진행비선성우화，학보오차적균균분포화모형적정학；재사용기우면편적다시도입체시각산법（patch-based multi-view stereo，PMVS），재국부광도일치성화전국가견성약속하，이SFM생성적희소점운위충자면편개시확산，완성점운주밀중건。장조편쾌속중건방법획취적점운여지면격광소묘의（terrestrial laser scanner, TLS）획취적점운급실측수거진행비교，결과표명，조편중건방법생성적점운주밀차능구완정전시침식구적발육형태，여 TLS점운간적평균거리위0.0034 m，조편중건여삼유격광소묘방법대침식량적고산상대오차위8.054%，제취적특정선필배솔체89.592%。연구결과위침식구감측제공료삼고의거。
Based on Structure from Motion(SFM) and Multi-View Stereo(MVS) techniques, this paper proposed a rapid 3d reconstruction method of slope eroded gully. Firstly, feature points were extracted and described by using the Scale-Invariant Feature Transform(SIFT), and then Random Sample and Consensus(RANSAC) algorithm was applied to filter inaccurate matching points generated by Nearest Neighbor(NN) algorithm; Secondly, in the condition that there were no camera parameters and scenario-based three-dimensional information, SFM was used because it provided a solution to iterate and get camera matrix and 3d point coordinates. During the iterating process, Bundle Adjustment(BA) algorithm was used for nonlinear optimizing and to ensure symmetrical distribution of the error in order to keep precision of the reconstructed model;After that, with the constraints of local photometric consistency and global visibility, Patch-Based Multi-View Stereo(PMVS) algorithm was adopted to expand sparse point cloud generated by SFM. Thus far the dense reconstruction of point cloud had finished. In order to validate the rationality and accuracy of using this method to monitor gully erosion, indoor runoff scouring experiment was conducted in“hydrology and water resources”laboratory at Xi’an University of Technology. Photos used in the reconstruction were taken by Canon 550d SLR camera. Because modeling process relied on tracking with the oriented point on the subject to determine the final 3d model of point set, so two adjacent photos’ differential seat angle can’t be too large, in case of losing trace points. Reasonable selection of photo shooting location, trajectory and angle should be considered according to the experimental environment and conditions. This paper used the VisualSFM software to complete detecting and matching of feature points, sparse reconstructing of point cloud as well as self-calibrating of camera;used CMVS and PMVS2 tools to finish dense reconstruction, and Meshlab to achieve visualization. After the finish of procedures mentioned above, three-dimensional model of slope eroded gully was built. At the same time, Trimble TX5 Terrestrial Laser Scanner(TLS) was used to obtain a referential point cloud data of the eroded gully in the experiment. After the preprocessing, point clouds obtained by SFM-MVS technique and terrestrial laser scanner were segmented the same area of gully head and reduced the point number to ten thousand. Comparing reconstructed point cloud with point cloud obtained by terrestrial laser scanner and measured data showed that, dense point cloud generated by photo reconstruction method can completely show the developing form of the gully, especially can achieve a better result in the wall, ridge and rolling area than point cloud obtained by laser scanner. Calculation and analysis showed that average distance between scanned and reconstructed point cloud was 0.0034m. Respectively generated digital elevation models for two point clouds by Kriging interpolation method, and computing results indicated that the relative error of erosion estimating was 8.054%. Based on the slope map generated by DEM, characteristic lines were extracted, of which the matching rate was 89.592%. The influence of Pixel value on reconstructing process and result was discussed at the end of the paper. The results of the study provided a reference for monitoring gully erosion.