【摘要】 针对视觉里程计中的特征点漂移与累积误差问题,提出一种基于双向重投影的双目视觉里程计方法。为降低特征点匹配的高错误率及改善匹配效率,提出一种改进的随机采样一致(RANSAC)算法。该算法依据特征点寿命长短、环形匹配、优质匹配等方法选取优质特征点以降低特征点样本容量,从而改善RANSAC算法的效率;为抑制视觉里程计累积误差、提升位姿估计精度,提出一种双向重投影的位姿估计算法,该方法通过解算下一帧图像中的最小化重投影误差得到相机位姿以更新空间点集,更新后的空间点重新投影至当前帧图像以提升位姿解算精度。最后利用KITTI数据集进行了仿真。结果表明该算法的150 m内定位误差在X方向小于1 m,在Z方向小于0.9 m;相对于ORB-SLAM2算法,定位精度提升了约20%,系统的稳定性更好,有效削弱了累积误差,验证了该算法的正确性与可行性。更多还原

【Abstract】 Aiming at the problem of feature points drift and error accumulation in visual odometer, a binocular vision odometer method based on bidirectional reprojection is proposed. In order to reduce the high error rate of feature point matching and improve the matching efficiency, an improved RANSAC(Random Sample Consensus) algorithm is proposed. The algorithm selects high-quality feature points according to the feature-point longevity, the ring matching, and the high-quality matching to reduce the sample size of feature points, so as to improve the efficiency of the RANSAC algorithm. In order to suppress the accumulated error of visual odometer and improve the position and posture estimation accuracy, a bidirectional projection algorithm for pose estimation is proposed. By resolving the minimized reprojection error in the next frame, the position and posture of the camera are obtained to update the space point set, and the updated space points are reprojected to the current frame image to improve the accuracy of the position and posture. Finally, simulations are carried out by using the KITTI data. The result shows that the position error of the algorithm(in the distance of 150 m) is less than 1 m in the X direction, and is less than 0.4 m in the Z direction. Compared with the ORB-SLAM2 algorithm, the position accuracy is improved by about 20%, the stability of the system is better, and the cumulative error is effectively weakened, which verify the correctness and feasibility of the algorithm.更多还原