Predictive and adaptive maps for long-term visual navigation in changing environments
This work addresses the challenge of degraded localization reliability in long-term visual navigation caused by persistent environmental appearance changes that render feature maps obsolete. To mitigate this issue, the authors propose a time-aware dynamic feature map management mechanism that models the periodic variations in scene appearance to predict which features are likely to be visible at specific spatiotemporal conditions. Leveraging these predictions, the system adaptively selects, prunes, and updates map features in real time. Evaluated over a three-month teach-and-repeat navigation experiment, the proposed approach significantly outperforms conventional static mapping strategies, demonstrating substantial improvements in robotic localization accuracy within temporally varying environments.