A Visual Odometry Artificial Intelligence-Based Method for Trajectory Learning and Tracking Applied to Mobile Robots

Abstract

Autonomous systems have demonstrated high performance in several applications. One of the most important is localisation systems, which are necessary for the safe navigation of autonomous cars or mobile robots. However, despite significant advances in this field, there are still areas open for research and improvement. Two of the most important challenges include the precise traversal of a bounded route and emergencies arising from the breakdown or failure of one or more sensors, which can lead to malfunction or system localisation failure. To address these issues, auxiliary assistance systems are necessary, enabling localisation for a safe return to the starting point, completing the trajectory, or facilitating an emergency stop in a designated area for such situations. Motivated by the exploration of applying artificial intelligence to pose estimation in a navigation system, this article introduces a monocular visual odometry method that, through teach and repeat, learns and autonomously replicates trajectories. Our proposal can serve as either a primary localisation system or an auxiliary assistance system. In the first case, our approach is applicable in scenarios where the traversing route remains unchanged. In the second case, the goal is to achieve a safe return to the starting point or to reach the end point of the trajectory. We initially utilised a publicly available dataset to showcase the learning capability and robustness under different visibility conditions to validate our proposal. Subsequently, we compared our approach with other well-known methods to assess performance metrics. Finally, we evaluated real-time trajectory replication on a ground robot, both simulated and real, across multiple trajectories of increasing complexity.