A Robust Rotation-Equivariant Feature Extraction Framework for Ground Texture-Based Visual Localization

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Ground texture-based localization offers robust pose estimation but struggles with large rotations and sparse textures.
• Traditional feature extraction methods lack reliability in challenging ground texture scenarios.

Purpose of the Study:

• To develop a learning-based feature extraction framework, GT-REKD, to overcome limitations in ground texture-based localization.
• To achieve rotation-invariant keypoint and descriptor extraction for enhanced localization accuracy.

Main Methods:

• Implemented a learning-based framework (GT-REKD) utilizing group-equivariant convolutions over the cyclic rotation group.
• Incorporated directional attention and orientation-encoding heads for precise feature extraction.
• Generated dense keypoints and descriptors invariant to 0-360° in-plane rotations.

Main Results:

• GT-REKD achieved 96.14% matching in pure rotation tests and 94.08% in incremental localization.
• Demonstrated low relocalization errors of 5.55° and 4.41 pixels.
• Consistently outperformed baseline methods under extreme rotations and sparse textures.

Conclusions:

• GT-REKD provides a robust solution for visual localization challenges posed by rotations and sparse textures.
• The framework shows significant applicability to visual localization and simultaneous localization and mapping (SLAM) tasks.