True scaled 6 DoF egocentric localisation with monocular wearable systems

Abstract

In this work we present a novel approach to obtain scaled odometry and map estimates when performing monocular SLAM with wearable cameras. After proving first that the oscillation of the body during walking can be observed in the odometric estimate from a monocular SLAM algorithm, we develop a method to estimate the walking speed from the frequency of this oscillation. Having the real walking speed, a scale factor can be dynamically computed to obtain a true scaled estimate of the map and visual odometry, avoiding scale drift on long term trajectories. Although the algorithm requires the person to be walking in order to estimate the scale, the experiments, carried out in outdoor and indoor environments and with different types of cameras, show that our method is reliable and robust to challenging situations like stops, changes in pace or stairs, and provides a significant improvement with respect to the initial unscaled estimate. It also outperforms state-of-the-art solutions to correct the scale drift in monocular SLAM, giving in addition the absolute scale of the trajectory and the 3D observed scene.