IMU-Kalibrierung und Lernverfahren mittels Transformer-Netzwerk für die Lokalisierung

Inertial localization approaches play a central role in navigation systems and are usually based on the double integration of dynamic acceleration sensor values. However, these methods face significant challenges in accurately tracking complex movements, such as those that occur in smartphones or wearables, and tend to drift over short periods of time. Furthermore, the desire to implement localization with these commercially available devices raises the question of whether individual calibration of the integrated inertial sensors (Inertial Measurement Unit, IMU) is necessary, which would be difficult to implement in practice. This work aims to overcome these challenges through a state-of-the-art supervised learning approach. A novel transformer network is used to solve both calibration and localization simultaneously. The use of learning approaches as a sensor calibration method can improve the accuracy of the prediction. The integration of knowledge-based features into the transformer model enables an improvement in the accuracy and reliability of position determination. The transformer-based attention implementation of this approach combines physical and learning-based methods into an advanced method that enables accurate and reliable localization in complex device placements. In addition, comprehensive analytical evaluations with realistic quality assessment using multiple metrics are presented.