Leveraging cross-view geo-localization with ensemble learning and temporal awareness.

Area of Science:

• Computer Vision
• Robotics
• Geospatial Analysis

Background:

• Global Navigation Satellite System (GNSS) unreliability poses challenges for autonomous vehicle localization.
• Visual self-localization using ground-to-aerial image matching faces viewpoint, environmental, and orientation data limitations.

Purpose of the Study:

• To develop a holistic approach for robust autonomous vehicle self-localization.
• To improve visual localization accuracy by addressing limitations of existing methods.

Main Methods:

• Proposed an ensemble model aggregating predictions from multiple state-of-the-art visual localization models.
• Introduced an efficient 'naive history' meta-block for temporal awareness in localization.
• Generated a new dataset derived from BDD100K for temporal awareness experiments.

Main Results:

• Achieved 97.74% Recall@1 on CVUSA and 91.43% on CVACT datasets, surpassing current state-of-the-art.
• Demonstrated that temporal awareness, via the naive history block, significantly boosts localization accuracy.
• Temporal awareness algorithm achieved 100% Recall@1 by utilizing short-term trip history.

Conclusions:

• An ensemble approach effectively combines complementary visual localization models.
• Temporal awareness is crucial for enhancing the robustness of visual localization systems.
• The proposed methods offer a significant advancement in autonomous vehicle self-localization accuracy and reliability.