Self-supervised learning for using overhead imagery as maps in outdoor range sensor localization.

Area of Science:

• Robotics
• Computer Vision
• Geospatial Analysis

Background:

• Traditional outdoor vehicle localization relies on pre-built sensor-specific maps.
• Acquiring and maintaining these maps can be costly and inconvenient.
• Publicly available overhead imagery offers a potentially ubiquitous and affordable alternative map source.

Purpose of the Study:

• To develop a robust vehicle localization method using overhead imagery as a map.
• To address the challenge of differing sensor modalities between overhead images and ground-based sensors (radar, lidar).
• To enable self-supervised learning for localization without requiring metrically accurate ground truth data.

Main Methods:

• A learned metric localization approach is proposed.
• The method leverages self-supervised learning for efficient training.
• It specifically addresses cross-modality data fusion, including millimeter-wave radar.

Main Results:

• The method demonstrates robustness and versatility across various sensor configurations.
• Localization errors are comparable to prior supervised methods.
• Effective cross-modality localization is achieved without pixel-wise aligned ground truth for supervision.

Conclusions:

• Overhead imagery can serve as a viable and cost-effective map for vehicle localization.
• The proposed learned metric approach successfully bridges the modality gap between overhead and ground sensors.
• Millimeter-wave radar proves particularly valuable for this localization task due to its resilience to environmental conditions.