Subgraph Learning for Topological Geolocalization with Graph Neural Networks.

Area of Science:

• Robotics
• Computer Science
• Artificial Intelligence

Background:

• Spatial cognition, including self-localization and navigation, requires efficient learning methods.
• Mimicking human-like spatial abilities is a significant challenge in artificial intelligence and robotics.

Purpose of the Study:

• To propose a novel approach for topological geolocalization on a map.
• To utilize motion trajectory and graph neural networks for enhanced localization accuracy.

Main Methods:

• Encoding motion trajectories as path subgraphs, where nodes represent turning direction and edges represent relative distance.
• Training a graph neural network on these subgraphs for multi-class classification to determine location.
• Testing the method on small, medium, and large map datasets using simulated and real-world trajectories.

Main Results:

• Achieved 93.61% accuracy on small datasets, 95.33% on medium datasets, and 87.50% on large datasets with simulated trajectories.
• Demonstrated comparable accuracy on actual trajectories generated by visual-inertial odometry.
• The approach leverages graph neural network capabilities for effective spatial learning.

Conclusions:

• The proposed method offers an efficient learning approach for topological geolocalization.
• It requires only a 2D graph map and an affordable sensor for relative motion trajectory data.
• This technique enhances spatial cognition capabilities by effectively modeling motion trajectories.