Towards Learning-Based Relative Underwater Pose Estimation Using a Single Light

Underwater localization is a challenging but an essential task for Autonomous Underwater Vehicles (AUV) operations. In this paper, we purpose utilizing learning-based methods to perform relative location estimation base on an active light source. This approach avoids the large latency associated with acoustic-based localization and the reduced detection range inherent in other vision-based techniques that rely on passive features or fiducial markers. Two styles of neural networks, a feedforward neural network (FNN) and a convolution neural network (CNN), with variation on the configurations, are implemented. Additionally, image processing pipeline for light features and feature concatenation of camera exposure time for CNN are discussed. Finally, the neural networks are evaluated on their capability of estimating relative location using a light in an indoor tank environment.