SUBVO Dataset: Analyzing Feature Extraction for Underwater Monocular Visual Odometry

This paper addresses the challenges of underwater Visual Odometry (VO) by introducing a novel dataset collected with a monocular camera in a controlled pool environment. The dataset is complemented by preprocessing techniques that mitigate underwater imaging issues such as color distortion and blue-green color cast. A comparative analysis of feature detectors identifies AKAZE as the most effective for underwater imagery. Additionally, a genetic algorithm (GA) optimizes the RANSAC inlier threshold, enhancing essential matrix estimation. Experimental results highlight a significant reduction in root mean square error (RMSE), with AKAZE achieving a 0.07-meter error compared to traditional methods. This study provides advancements in underwater navigation and mapping, promising improvements for autonomous underwater vehicles (AUVs) and marine research.