Deep Learning Approaches Assessment for Underwater Scene Understanding and Egomotion Estimation

摘要

This paper address the use of deep learning approaches for visual based navigation in confined underwater environments. State-of-the-art algorithms have shown the tremendous potential deep learning architectures can have for visual navigation implementations, though they are still mostly outperformed by classical feature-based techniques. In this work, we apply current state-of-the-art deep learning methods for visual-based robot navigation to the more challenging underwater environment, providing both an underwater visual dataset acquired in real operational mission scenarios and an assessment of state-of-the-art algorithms on the underwater context. We extend current work by proposing a novel pose optimization architecture for the purpose of correcting visual odometry estimate drift using a Visual-Inertial fusion network, consisted of a neural network architecture anchored on an Inertial supervision learning scheme. Our Visual-Inertial Fusion Network was shown to improve results an average of 50% for trajectory estimates, also producing more visually consistent trajectory estimates for both our underwater application scenarios.