Evaluation of PoseNet for 6-DOF Underwater Pose Estimation

Autonomy in underwater intervention operations requires localization systems of high accuracy. State-of-the-art methods rely on computer vision to provide the necessary localization accuracy. However, traditional computer vision solutions rely on hand-crafted features, which often exhibit low robustness to variations in the lighting conditions. Furthermore, the most common image localization method, Perspective-n-Point (PnP), relies on specific knowledge of the distances in the scene, something which is not always available. Recent advances within deep learning, in particular, convolutional neural networks (CNNs), have resulted in promising methods for pose estimation based on imagery input. This article investigates the potential of applying a specific CNN architecture, named PoseNet, to estimate the 6-DoF pose between an underwater vehicle and a fixed object, without the need for artificial markers.