基于摄像机位姿估计的数学模型, 提出了一种检测摄像机位移前后目标图像特征点的方法, 通过求解摄像机发生位移前后的相对位姿矩阵来解决应用视觉图像获得点云的初始配准问题。首先, 介绍了摄像机位姿估计模型, 包括本质矩阵、旋转矩阵以及平移矩阵; 然后, 介绍了SURF算子的特征点检测、描述和匹配的方法, 在此基础上面向双目视觉和单目结构光系统, 分别提出了摄像机位移前后目标图像SURF特征点匹配和深度估计模型; 最后, 分别进行双目视觉和单目结构光系统点云的获取、位移前后目标图像特征点检测匹配和深度估计实验, 应用摄像机位姿估计模型求解旋转矩阵和位移矩阵, 并对位移矩阵进行统计分析剔除粗差。实验中采用基于点云空间特征点和基于图像的方法进行对比, 点云对应特征点均方误差缩小至12.46 mm。实验结果验证了方法的可行性, 表明本文的点云初始配准方法能较好地获得点云精确配准初值。

Based on the mathematical model of camera pose estimation, a feature point detection method of camera images before and after movement was proposed to obtain relative posture matrix of camera before and after movement, which can solve initial registration problem of point clouds derived from machine vision. Firstly, the estimation model of camera pose was introduced, including essential matrix, rotation matrix and translation matrix. Secondly, the detection, description and matching of feature points for SURF operator were introduced. On this basis, SURF feature points matching of camera images before and after movement and depth estimation model were respectively proposed for binocular vision and monocular structured light system. Finally, the acquisition of point clouds derived from binocular vision and monocular structured, feature points detection and matching of camera images before and after movement as well as camera depth estimation were realized experimentally. The mathematical model of camera pose was estimated to solve the rotation matrix and the translation matrix, and the residual analysis was carried out on the translation matrix for eliminating gross errors. In the experiment, the method of initial registration of point cloud based on feature point and based on images as contrast, the results show that the mean square error of the corresponding feature points are reduced to 12.46 mm. The result verifies the feasibility of the method, and indicates that the point registration method can obtain good initial values for accurate point cloud registration.