Robust Relocalization and Its Evaluation for Online Environment Map Construction.

Area of Science:

• Computer Vision
• Robotics
• Photogrammetry

Background:

• Robust camera orientation tracking and relocalization are crucial for creating surround-view panoramas with single cameras.
• Existing methods often struggle with tracking loss and environmental changes like lighting variations.

Purpose of the Study:

• To present a robust methodology for camera orientation relocalization using virtual keyframes.
• To improve the speed and robustness of relocalization, especially during tracking failures and under challenging lighting conditions.

Main Methods:

• Utilizing virtual keyframes strategically placed in environment maps for online map construction.
• Implementing a system variant for relocalization only upon tracking failure, using efficient image-patch descriptors.
• Employing illumination normalization and saturated area removal to enhance robustness against lighting changes.

Main Results:

• The proposed system demonstrates practical usefulness through comparisons of different variants using three evaluation methods.
• The approach effectively re-estimates camera orientation by matching frames against known-orientation keyframes.
• Evaluations over indoor and outdoor sequences show improved relocalization rates against ground truth.

Conclusions:

• The virtual keyframe approach offers a robust and efficient solution for camera orientation relocalization.
• The system variants provide practical benefits for surround-view panorama acquisition.
• The proposed techniques enhance system performance in diverse and challenging environments.