IMULet : A Cloudlet for Inertial Tracking

Inertial measurement units (IMUs) afford the problem of localisation unique advantages owing to their independence of costly deployment and calibration efforts. However, IMU models have traditionally suffered from excessive drifts that have limited their appeal and utility. Newer machine learning (ML) approaches can better model and compensate for such inherent drift at the expense of (i) increased computational penalty and (ii) fragility w.r.t. changes in the signal profile that these ML models have been trained on. In this paper we propose an edge cloud-based inertial tracking architecture that overcomes the above limitations. Our IMU tracking cloudlet is comprised of: (i) an on-device component that compresses inertial signals for wireless transmission, (ii) a cloud-side ML model that tracks the temporal dynamics of inertial signals, and (iii) a cloud-side deep latent space tracking in order to seamlessly manage model adaptation-i.e. to mitigate the fragility of ML over-specialisation. Early evaluation demonstrates the feasibility of our approach and exposes items of future research.