DC-DOES: A Dual-Camera Deep Learning Approach for Robust Orientation Estimation in Maritime Environments

Abstract

Attitude and Heading Reference Systems (AHRS) have achieved significant accuracy and reliability, making them suitable for various applications. This is possible through the integration of high-rate measurements, though they remain prone to errors, particularly sensor drift over time. As a potential solution, AHRS can be combined with complementary devices, such as camera-based systems, which have attracted attention for their cost-effectiveness and simplicity. This study introduces the Double Camera - Deep Orientation (roll and pitch) Estimation at Sea (DC-DOES), a Deep Learning model developed to enhance roll and pitch estimations obtained from conventional AHRS at sea. In comparison to previous versions, DC-DOES operates in a novel configuration utilizing a double-camera system. This system is based on a Jetson Nano embedded platform, integrating a low-cost AHRS and two synchronized cameras, resulting in a fully customizable acquisition and processing setup. DC-DOES is trained and validated on shore to assess its effectiveness and robustness in controlled conditions and will be further deployed on board for real-time applications at sea. It is trained on the Double Camera - ROll and PItch at Sea (DC-ROPIS) dataset, which was specifically collected for this purpose. Both the code and the dataset have been made publicly available to encourage further use and improvement. The results are promising, achieving a Mean Absolute Error (MAE) of approximately 1°, highlighting the potential of this cost-effective, reliable solution for orientation estimation tasks. Additionally, tests in low-light scenarios demonstrated its robustness under challenging conditions, making DC-DOES a suitable solution for maritime navigation and beyond.

Keywords

• AHRS
• dataset acquisition
• double-camera
• deep learning
• embedded systems
• orientation estimation