Evaluation of Several Feature Detectors/Extractors on Underwater Images towards vSLAM

Area of Science:

• Robotics
• Computer Vision
• Oceanography

Background:

• Modern visual SLAM (vSLAM) algorithms rely on computer vision for mapping and trajectory generation from camera data.
• Existing research extensively evaluates feature detectors for air and ground vSLAM but lacks comprehensive underwater studies.

Purpose of the Study:

• To characterize underwater images affected by light and particles.
• To evaluate common interest point detectors and descriptors for underwater vSLAM.
• To identify challenges and limitations of feature detectors in subsea environments.

Main Methods:

• Image characterization of underwater scenes with varying light and particle densities.
• Performance evaluation of interest point detectors and descriptors based on feature matching, descriptor precision, and processing time.
• Comparative analysis of feature detector suitability for underwater vSLAM applications.

Main Results:

• Underwater images present significant challenges due to light scattering and suspended particles.
• Performance metrics (feature matches, precision, speed) differ substantially among tested detectors.
• Specific feature detectors show promise for robust underwater vSLAM.

Conclusions:

• Feature detectors are essential for underwater vSLAM, despite environmental complexities.
• The choice of feature detector significantly impacts vSLAM performance in subsea scenarios.
• Further research is needed to optimize feature detection for challenging underwater conditions.