Contrastive Learning for Image Registration in Visual Teach and Repeat Navigation.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Visual Teach and Repeat (VT&R) navigation is a versatile robotic method.
• Existing VT&R systems struggle with environmental changes like day-night and seasonal variations.
• Accurate path following requires robust image representations insensitive to illumination and environmental shifts.

Purpose of the Study:

• To develop a method for calculating horizontal displacement between images for VT&R.
• To enhance the robustness of VT&R systems against environmental and illumination changes.
• To improve the accuracy of robot path traversal in dynamic conditions.

Main Methods:

• Utilized a fully convolutional neural network (FCNN) for dense image representation.
• The FCNN model was trained to be robust to environmental and illumination variations.
• Investigated the generation of synthetic training data to further improve model robustness.

Main Results:

• Achieved state-of-the-art performance on datasets with seasonal and day/night variations.
• Demonstrated that the model can generate effective additional training data.
• Successfully validated the method in a real-world mobile robot experiment.

Conclusions:

• The proposed FCNN-based approach significantly improves VT&R robustness.
• The method offers a practical solution for reliable robot navigation in changing environments.
• The self-improvement capability through synthetic data generation shows promise for future research.