Monocular Navigation System for Corridor Environments Based on Relative Camera Pose Estimation: An Approach Without SLAM

This paper addresses monocular navigation in monotonous corridor environments, where visual simultaneous localization and mapping are difficult. We developed a simple navigation system based on a relative camera pose estimation model using convolutional neural networks. All data collection and experiments were conducted by running the real robot. Our system is intended to be used in applications such as security robots, and the validation results show that the relative pose estimation model is sufficiently accurate, with an RMSE of less than 1 [m] in position and less than 6 [deg] in orientation. The effectiveness of the proposed system was also confirmed in a real-world experiment, in which fully autonomous navigation was performed for up to 60 minutes. In addition, the system has advantages in that it does not require complex sensor calibration and mapping, and the relative pose estimation model has good generalization performance and is data efficient. Our research is important for the realization of autonomous robot navigation using only monocular cameras, without relying on detailed metric maps.