单目视觉SLAM车载摄像机快速位姿估计及景物重构

针对单目视觉同时定位与地图构建(simultaneous localization and mapping, SLAM)应用,提出一种快速车载摄像机位姿估计及景物结构3D重构算法。在无具体标定物的情况下,利用移动机器人二自由度运动导致摄像机采集视图中对应极点的特殊性质,标定出真实摄像机与虚拟摄像机间的相对姿态,并利用主动视觉方法进一步标定出移动机器人坐标系与虚拟摄像机坐标系间的相对位移;构造无穷单应变换,将真实视图中通过SIFT算法得到的假设欧氏匹配点集转换为对应虚拟摄像机的虚拟假设欧氏匹配点集,并利用基于RANSAC的归一化三点算法快速地对本质矩阵进行估计和分解;利用已观测到的路标三维信息,递归地剔除本质矩阵分解出的平移量的尺度不确定性,并利用线性三角形法重构出景物结构。实验结果表明:该算法在保证运算精度的同时,具有更快的运算速度。

Fast pose estimation for on-board camera and scene reconstruction in monocular vision SLAM

According to the simulaneous localization and mapping(SLAM), a fast pose estimation of on-board camera, together with 3D reconstruction of scene structure algorithm was proposed. The special properties of the Euclidean epipoles corresponding to the mobile robots 2-DOF movement were utilized to calibrate the relative pose information between real camera coordinate and virtual camera coordinate without the utilization of specific calibration object, and the active vision method was utilized to further calibrate the relative position information between mobile robot coordinate and virtual camera coordinate; An constructed infinite homography was adopted to turn the hypothesis Euclidean point correspondences obtained by SIFT algorithm into the virtual hypothesis Euclidean point correspondences, and the RANSAC based normalized 3-point algorithm was implemented to estimate and decompose the essential matrix; The previous 3D information of the observed landmarks were adopted to eliminate the scale uncertainty of the translation vector acquired by essential matrix decomposition, and the scene structure was reconstructed by linear triangulation method. Experimental results show that the proposed algorithm has the advantages of high precision, as well as the low computational complexity.