TrackLet: Data-Driven Inertial Tracking on Your Own IMU Data

Empowered by deep learning, the data-driven smartphone inertial tracking has attracted much attention due to its high accuracy and robustness. However, training a one-size-fits-all model requires a significant amount of inertial measurement unit (IMU) data with accurate labels, which incurs high costs of annotation and computation. In this work, we propose a new solution for IMU tracking that trains a specific model for each individual on their own IMU data. Individual IMU data can be opportunistically collected and automatically annotated with coarse-grained locations provided by smartphones (lite labels). To learn from noisy lite labels, we propose a lightweight and noise-resistant framework TrackLet for personalized data-driven IMU tracking. Two effective techniques are designed to combat the noise in lite labels, namely CNAL (Chained-Noise Adaptation Layer) and CoAdapt (Cooperatively Adaptive small-loss selection and weighting). Extensive experiment results demonstrate that TrackLet achieves high accuracy yet at a low cost, outperforming its state-of-the-art counterparts.