ResT-IMU: A Two-Stage ResNet-Transformer Framework for Inertial Measurement Unit Localization

Area of Science:

• Robotics and Artificial Intelligence
• Sensor Fusion and Navigation

Background:

• Accurate indoor positioning is crucial but challenging in complex environments.
• Existing methods struggle with precision and robustness.

Purpose of the Study:

• To propose ResT-IMU, a two-stage indoor positioning method integrating ResNet and Transformer architectures.
• To enhance trajectory prediction accuracy and robustness using IMU data.

Main Methods:

• Kalman filtering and windowing applied to IMU data.
• ResNet for motion feature extraction and velocity prediction.
• Transformer's self-attention for temporal feature analysis and direction refinement.

Main Results:

• Achieved low velocity prediction errors (0.0182 m/s on iIMU-TD, 0.014 m/s on RoNIN).
• Demonstrated significant reductions in Absolute Trajectory Error (ATE) and Relative Trajectory Error (RTE) compared to ResNet, IMUNet, and ResMixer models on benchmark datasets.

Conclusions:

• ResT-IMU offers superior accuracy and robustness for indoor trajectory prediction.
• The integration of ResNet and Transformer architectures effectively captures motion and temporal dynamics.
• Validated performance on iIMU-TD and RoNIN datasets.