DOVO: Mixed Visual Odometry Based on Direct Method and Orb Feature

Abstract

In this paper, in order to get real-time environment information and pose estimation of robot, a novel visual odometry method called DOVO is proposed. First, the ORB feature of image frame is computed. Then, based on the number of key point that ORB feature gets, we set a threshold K to determine the reliability of pose estimation using ORB feature. If the number of key point is smaller than the threshold K, direct method is used to keep trace of the camera and estimate the camera pose by optimizing the photometric error based on luminosity of the scene is constant. If the number of key point is larger than threshold K, then pose estimation is computed by optimizing reprojection error. We use TUM dataset to make experiments to show this method guarantees pose accuracy and real-time performance.