This work presents a framework for estimating body center of mass (COM) dynamics using wrist-worn inertial measurement units (IMUs). The approach combines a simplified kinematic model (KM) with human kinematic model-based neural networks (HKM-NN) to address the limitations of relying solely on IMU measurements. The KM model uses reductive assumptions to solve dynamic equations based on wrist IMU data. The HKM-NN models leverage both grey-box and black-box modeling techniques.
The research utilized a dataset of 10 healthy volunteers performing six gait activities and a sit-to-stand (SS) transitional movement. Wrist IMU measurements and ground-truth COM measurements were collected. The HKM-NN models included serial learning (ser-), simultaneous learning (sim1- and sim2-) approaches.
Evaluation of the models demonstrated that the HKM-NN models significantly improved COM acceleration estimation. The models achieved error ranges of 5.3% to 9.3% for gait activities and a best error of 3.9% for the SS movement. The sim1- and sim2- approaches exhibited greater robustness under Gaussian perturbations, while the KM model showed stronger robustness under salt-and-pepper noise.
These findings highlight the value of combining biomechanical understanding with data-driven learning for wearable sensing applications where measurement conditions are often imperfect and noisy. The hybrid approach offers improved accuracy and robustness compared to simpler models. Source: https://arxiv.org/abs/2609.12304