Identifying IMU Noise Characteristics from USV Trajectories

Inertial navigation is crucial for Unmanned Surface Vehicles (USV) as it enables tracking of the vehicle's position, speed and orientation independent of any external infrastructures, such as Global Navigation Satellite System (GNSS), which are often not accessible in difficult maritime environments. Without GNSS, the vehicle usually relies on the Inertial Navigation System (INS), utilizing Inertial Measurement Unit's (IMU) data by applying dead reckoning to the accelerometer and gyroscope measurements over time. While INS can be reliable for short missions, errors can accumulate over time. Investing in more expensive sensors can achieve higher precision, but it's more cost-effective to complement INS with advanced estimation method to enhance accuracy. This paper introduces a data-driven approach that leverages deep learning to improve IMU noise modeling and enhance navigation accuracy for surface vehicles. Three deep learning architectures based on Convolutional Neural Network (CNN), Residual Network (ResNet) and Bidirectional Long-Short Term Memory (Bi-LSTM) network, were implemented and compared. The proposed approach employs gyroscope and accelerometer measurements from a lower-quality IMU as input and aims to match the angle and position estimates obtained from the ground truth trajectory, showing an end-to-end approach to position estimation directly from raw IMU data.