(The) development of the scale-aware monocular depth estimation aided monocular visual SLAM system for real-time robot navigation

Abstract

The autonomous navigation systems in mobile robots are currently one of the hottest research topics as the unmanned mobile robot market size is forecasted to skyrocket in the near future. Many sensors, for instance LiDAR, RADAR, IMU and cameras, become the core components in such systems. Because of being rich in information, relatively cheap, lightweight, compact, ubiquitous, and having low power consumption, monocular vision-based navigation, namely visual SLAM and visual odometry, has attracted a lot of attention from robotic researchers. However, monocular vision-based navigation systems impose many challenges: scale-ambiguity, tracking robustness, and map initialization delay. In this work, we show that, by incorporating dense depth maps predicted by monocular depth estimation network to visual SLAM, the proposed RGB-Deep D SLAM framework could mitigate issues of monocular SLAM. The results, evaluated in the KITTI dataset, indicate that the proposed framework improves tracking robustness by about 12 \%. Moreover, when the depth estimation network is scale-aware, scale-ambiguity problem of the odometry is also alleviated, as evidenced in the reduction of the normalized of median value of root mean square error (RMSE) of absolute trajectory error (ATE) of about least 84 \%. Thanks to the presence of depth information, the RGB-Deep D SLAM could instantly initialize the map which solves the delay issue in the monocular visual SLAM. These results suggest that RGB-Deep D SLAM could be deployed as a monocular vision-based real-time navigation system. In addition, we verify the deployment of the proposed system by integrating it to the full navigation stack consisting of SLAM, path-planner, and controller. The autonomous navigation is successfully demonstrated in two scenarios of the unknown environment using the DJI Tello drone.