Fig. 2: Tracking of arm pose from a single wrist-based IMU using parametrized arm-pose coordinate system and deep learning.
A Schematic diagram showing a parameterized model for arm pose. Positive θ indicates moving the corresponding limb upward. B Deep learning architecture diagram for tracking upper arm orientation (θu). The inputs at each timestep are forearm acceleration and orientation (i.e., yaw, pitch θf, and roll) represented as a unit quaternion. These inputs are fed through a fully connected (FC) layer followed by two bidirectional LSTM (BiLSTM) layers. The latent feature vector for the current timestep is then passed through a final FC layer to predict the upper arm orientation quaternion, normalized to the unit norm. The orientation quaternion is finally used to calculate θu. (C) Histogram of absolute errors for θu prediction for the model pretrained on the Virginia Tech Natural Motion Dataset alone (“pretrained”) and after fine-tuning on in-house training data (“fine-tuned”). D Time series of predicted and measured θu for a representative participant from a test fold. E Mean inference time for the arm-tracking model with 32- and 8-bit weight precision when run on a commercial smartphone. The dashed line indicates the cardiac cycle duration for a heart rate of 240 beats per minute (BPM). Data were represented as mean ± standard error of the mean (n = 50 for 32-bit and 8-bit).
