A seasonally invariant deep transform for visual terrain-relative navigation

Area of Science:

• Robotics
• Computer Vision
• Planetary Science

Background:

• Visual terrain-relative navigation (VTRN) enables autonomous robotic localization using image registration.
• VTRN is highly accurate but vulnerable to seasonal changes like lighting and vegetation.
• Current methods struggle with seasonal variations or lack interpretable uncertainty.

Purpose of the Study:

• To develop a robust VTRN method that overcomes seasonal image variations.
• To integrate deep learning with classical registration for improved navigation.
• To provide accurate and reliable localization for robotic vehicles.

Main Methods:

• A novel image transform architecture using targeted deep learning.
• Conversion of seasonal imagery to a stable, invariant domain.
• Preservation of geometric structure and uncertainty estimates.

Main Results:

• Superior VTRN performance under extreme seasonal changes.
• Elimination of gross mismatches in challenging navigation tasks.
• Demonstrated ease of training and high generalizability of the proposed method.

Conclusions:

• Classical registration methods are effective for robotic visual navigation when stabilized.
• The proposed architecture consistently anticipates reliable imagery.
• This approach enhances the robustness and reliability of VTRN for diverse robotic applications.