A Panoramic Localizer Based on Coarse-to-Fine Descriptors for Navigation Assistance

Area of Science:

• Computer Vision
• Robotics
• Geographic Information Systems

Background:

• Visual Place Recognition (VPR) enables precise localization by matching query images to database images.
• Existing VPR methods struggle with appearance variations (illumination, viewpoint, seasons) and limited Field of View (FoV).
• Forward-facing cameras in current VPR systems offer narrow FoV and are susceptible to viewpoint changes.

Purpose of the Study:

• To propose a novel panoramic localizer for robust Visual Place Recognition.
• To leverage omnidirectional perception using 360° panoramas for enhanced localization.
• To address the limitations of narrow FoV and viewpoint variations in existing VPR algorithms.

Main Methods:

• Developed a panoramic localizer employing a coarse-to-fine descriptor strategy.
• Utilized NetVLAD descriptors for robust coarse matching across panoramas.
• Integrated Geodesc keypoint descriptors for detailed fine-stage matching.

Main Results:

• The proposed system demonstrates high recall and strong generalization capabilities across diverse visual appearances.
• Experiments on public and real-world datasets validate the system's effectiveness.
• The panoramic localizer achieved accurate localization despite significant appearance variations.

Conclusions:

• The panoramic localizer offers a significant advancement for Visual Place Recognition.
• Leveraging 360° panoramas overcomes limitations of traditional VPR systems.
• The system is suitable for integration into mobile navigation devices for various applications.