Abstract: We propose a novel method based on the adaptive outlier removal and a decoupling algorithm to address the problems associated with inlier selection in visual odometry. The Oriented FAST and Rotated BRIEF (ORB) operator is used to detect the feature points. The re-projection error normalized by optical flow is used as the inlier selection criterion, which is independent of the depth and position of the feature points. Ego-motion estimation is divided into two steps. First, the five-point RANSAC method is modified and combined with spherical linear interpolation (SLERP) to estimate the rotation parameters. Then, the motion-only bundle adjustment is combined with a normalized reprojection error to remove outliers and obtain the optimized translation parameters. The iteration stops when it reaches the maximum or the set of inliers remains unchanged. Experiments are conducted on the public data set KITTI, and the results show that the method has a significant improvement in terms of outlier removal compared with the conventional method, which uses a re-projection error with a fixed threshold and simultaneously estimates six degrees of freedom. Furthermore, the overall accuracy is improved compared to other similar works.