Fully Scaled Monocular Direct Sparse Odometry with A Distance Constraint

Abstract

Existing methods for solving the inherent scale-ambiguity of monocular visual odometry adopt two main techniques: using an inertial measurement unit and multiple cameras. The distance between poses, which can be obtained from GPS navigation, wheel encoders, and some types of odometry, can also restore the scale of motion. In this work, we propose a fully scaled monocular direct sparse odometry, using the distance between image frames as a constraint to maintain a trajectory and create a map at a suitable scale based on direct sparse odometry. First, the modulus of the translation vector of the current frame, related to the reference keyframe, is corrected in the tracking unit. Additionally, a distance error term is proposed and adapted into the optimization as a further constraint between keyframes. We evaluate our method on the challenging European Robotics Challenge (EuRoC) dataset; our method shows advantages in accuracy compared with the state-of-the-art OKVIS method.