RIOT: Recursive Inertial Odometry Transformer for Localisation from Low-Cost IMU Measurements.

Area of Science:

• Robotics and Autonomous Systems
• Sensor Fusion
• Machine Learning for Navigation

Background:

• Inertial localisation is crucial for ego-motion estimation when external sensors are unavailable.
• Low-cost inertial sensors suffer from bias and noise, causing unbounded errors and making direct position integration difficult.
• Traditional methods require prior system knowledge and are limited by predefined dynamics.

Purpose of the Study:

• To develop a data-driven deep learning solution for inertial odometry that overcomes limitations of traditional methods.
• To integrate traditional recursive state estimation principles into a deep learning framework for improved accuracy.
• To create pose-invariant deep inertial odometry frameworks capable of learning motion characteristics and sensor errors.

Main Methods:

• Proposed a novel approach combining traditional recursive state estimation with deep learning.
• Incorporated true position priors during the training process using inertial measurements and ground truth displacement data.
• Developed two end-to-end frameworks utilizing self-attention mechanisms for spatial feature and long-range dependency capture.

Main Results:

• The developed deep learning models learned motion characteristics and systemic error bias and drift.
• Evaluated against a Gated Recurrent Unit, demonstrating superior performance.
• Achieved a sequence length weighted relative trajectory error mean of ≤0.4594 m across various users, devices, and activities.

Conclusions:

• The proposed learning process effectively enhances deep inertial odometry models.
• The end-to-end frameworks provide accurate and pose-invariant ego-motion estimation.
• This data-driven, recursive approach offers a robust solution for inertial localisation challenges.