Long-term visual localization in dynamic benthic environments: the SEALOC dataset, footprint-based ground truth, and...

Area of Science:

• Robotics and Autonomous Systems
• Marine Science and Oceanography
• Computer Vision and Image Processing

Background:

• Long-term visual localization is crucial for cost-effective optical benthic monitoring using autonomous underwater vehicles (AUVs).
• Existing research is limited by a lack of curated datasets and precise georeferencing for benchmarking.
• Accurate ground-truthing requires precise geometric information, including 3D seafloor image footprints.

Purpose of the Study:

• To address the understudied nature of long-term visual localization in benthic environments.
• To introduce SEALOC, a novel curated dataset for benchmarking visual localization in the benthos.
• To present a new method for ground-truthing visual localization results in near-nadir underwater imagery.

Main Methods:

• Developed SEALOC, a dataset featuring georeferenced AUV imagery from five benthic sites revisited over six years.
• Included raw and color-corrected stereo imagery, camera calibrations, and sub-decimeter registered camera poses.
• Implemented a novel ground-truthing method estimating 3D seafloor footprints to link overlapping camera views.

Main Results:

• Benchmarked eight state-of-the-art visual place recognition (VPR) methods on the SEALOC dataset.
• Observed significantly lower Recall@K performance compared to existing terrestrial and underwater benchmarks.
• Demonstrated that traditional distance-threshold ground-truthing can overestimate VPR performance in rugged terrains.

Conclusions:

• The SEALOC dataset and novel ground-truthing method advance long-term visual localization research in dynamic benthic environments.
• Findings indicate a need for more robust VPR methods tailored to the complexities of underwater settings.
• The study provides a foundation for future research in autonomous underwater navigation and mapping.