基于未标定相机双目视觉的3D重构技术研究

发布时间：2018-03-08 13:10

本文选题：双目视觉　切入点：三维重构　出处：《哈尔滨工业大学》2017年硕士论文　论文类型：学位论文

【摘要】：基于未标定相机双目视觉的3D重构技术就是利用双目视觉测量原理,使用处于不同位置的参数未知的两相机或同一相机经旋转、平移拍摄同一场景,获取同一场景的两幅图像,根据两图像的视差来建立场景的三维模型。由于成像设备具有价格低廉,应用范围广等优点,基于双目视觉的3D重构技术已经成为三维重构的重要研究方法。如何正确标定相机、提高图像匹配的精度、精确估计基础矩阵、求解精确的空间三维点坐标,是其中的研究重点与难点。本文针对上述几个问题进行了研究。首先,详细介绍了相机的标定方法。针对标准测试图像,运用基于灭点的相机标定法,利用Hough变换,就可以在无需知道相机的运动姿态和相关参数的情况下实现相机的标定;针对自拍摄图像,采用张氏平面模板标定法,运用标定模板,提高了标定精度。其次,比较SIFT算法和SURF算法的性能,在时间要求不高的情况下,从匹配点成功对数方面考虑,采用SIFT算法作为后续工作的研究基础,提出一种改进的误匹配剔除方法。在求取基础矩阵之前,剔除一些明显的误匹配点对,提高了基础矩阵的求解精度。然后,提出了一种改进的基础矩阵鲁棒性估计方法,该方法结合RANSAC算法与M-Estimator估计算法的优点,能够更加精确的估计基础矩阵,使用估计结果约束匹配点对,再次剔除误匹配点对,提高了特征点对的匹配精度。最后,采用运动恢复的方法,结合相机内参与基础矩阵,恢复相机的旋转矩阵和平移向量。针对标准测试图像,求解空间三维点坐标,根据场景深度一致性,剔除严重偏离主场景的匹配点对;为更清晰的表达场景几何结构信息,对空间三维点进行三角剖分;为恢复具有真实感的三维场景,利用OpenGL实现纹理贴图。为了验证算法的通用性,针对自拍摄图像,完成上述过程,实现了对自拍摄场景的有效恢复。对三维模型进行误差分析,结果表明,本文三维重构算法可靠。
[Abstract]:The technology of 3D reconstruction based on binocular vision of uncalibrated camera is to use the principle of binocular vision measurement. Two cameras with unknown parameters in different positions or the same camera are rotated, and then the two images of the same scene are obtained by the translation of the same camera. Based on the parallax of the two images, the 3D model of the scene is built. The technology of 3D reconstruction based on binocular vision has become an important research method of 3D reconstruction. How to calibrate the camera correctly, improve the accuracy of image matching, estimate the basic matrix accurately, and solve the exact coordinates of 3D points in space. Firstly, the calibration method of the camera is introduced in detail. For the standard test image, the camera calibration method based on the vanishing point is used, and the Hough transform is used. The camera can be calibrated without knowing the motion attitude and related parameters of the camera, and the calibration accuracy is improved by using the Zhang's plane template calibration method. Comparing the performance of SIFT algorithm and SURF algorithm, considering the logarithm of the matching point in the case of low time requirement, the SIFT algorithm is used as the basis of further research. An improved method for eliminating mismatch is proposed. Some obvious mismatch points are eliminated before the foundation matrix is obtained, which improves the accuracy of solving the fundamental matrix. Then, an improved method for estimating the robustness of the basic matrix is proposed. Combined with the advantages of RANSAC algorithm and M-Estimator estimation algorithm, the method can estimate the basic matrix more accurately, use the estimation results to constrain the matching point pairs, eliminate the mismatched point pairs again, and improve the matching accuracy of the feature point pairs. The rotation matrix and the translation vector of the camera are restored by the method of motion recovery, and the 3D coordinates of the space point are solved according to the consistency of the depth of the scene, according to the standard test image, the rotation matrix and the translation vector of the camera are recovered by the method of motion recovery. The matching points that deviate seriously from the main scene are eliminated, the spatial 3D points are triangulated to express the geometric structure information of the scene more clearly, and the realistic 3D scene is restored. Texture mapping is realized by OpenGL. In order to verify the generality of the algorithm, the process of self-shooting image is completed, and the effective restoration of self-shooting scene is realized. The error analysis of 3D model shows that, The three-dimensional reconstruction algorithm in this paper is reliable.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】