Coarse-to-Fine vision-based localization by indexing scale-invariant features.

Area of Science:

• Computer Vision
• Robotics
• Geographic Information Systems

Background:

• Accurate camera localization is crucial for autonomous systems and augmented reality.
• Existing methods often face trade-offs between speed and precision.
• Novel approaches are needed to improve global localization efficiency and reliability.

Purpose of the Study:

• To present a novel coarse-to-fine global localization approach.
• To leverage object recognition and text retrieval techniques for camera localization.
• To achieve fast and reliable camera pose estimation in diverse environments.

Main Methods:

• Utilizing Harris-Laplace interest points as natural landmarks.
• Employing scale-invariant feature descriptors for landmark characterization.
• Indexing landmarks into a location vector space model (LVSM) and a location database.
• Implementing a two-stage localization process: coarse (LVSM) and fine (database voting).
• Verifying localization using epipolar geometry and recovering camera position via essential matrix decomposition.

Main Results:

• The integrated coarse-to-fine approach enables fast and reliable global localization.
• The system demonstrated efficiency and reliability in both indoor and outdoor tests.
• The method effectively combines the speed of coarse localization with the accuracy of fine localization.

Conclusions:

• The proposed coarse-to-fine global localization method offers an efficient and reliable solution.
• The integration of different localization strategies enhances performance.
• This approach shows significant potential for applications requiring accurate camera pose estimation.