An Improved ORB-SLAM2 With Refined Depth Estimation

Abstract

We present a novel VSLAM system based on ORB-SLAM2. Firstly, the uncertainties of the position estimate of the 3D feature points by triangulation are computed through the propagation of covariance matrix. Then, the value of the triangulated depth and the variance of depth are updated by an information refining method from multiple triangulations. In addition, the map points with large uncertainties are excluded. And, the uncertainties of map points are used to select keyframe. As a result, more accurate estimation of the map points and the robot poses at the keyframes can be obtained. We test the proposed algorithm on the TUM open source datasets and show that the proposed algorithm is superior to the ORB-SLAM2. A real-time experiment is executed using the Pioneer 3-AT robot in the indoor environment.