Developing Compact Models Using Regression Confidence Forge Knowledge Distillation for IMU-Based Indoor Positioning System

This letter focuses on developing practical and resource-efficient solutions for indoor positioning systems using inertial measurement unit sensors (IMU) by introducing a compact and efficient model. The model, derived from the robust neural inertial navigation (RoNIN) architecture, features a lightweight model that is achieved by reducing the number of filters. A specific knowledge distillation (KD) method, regression confidence forge (ReCoF) KD, is proposed and employed to address potential performance implications, enhancing the efficacy of the streamlined model. The smallest proposed model exhibits an 86% size reduction from RoNIN Resnet, leading to an 18.8% acceleration in inference time and 56% more power efficiency on the edge. Notably, the proposed model maintains high performance, as evidenced by its absolute trajectory error (ATE) and relative trajectory error (RTE).