In-Sensor Movement Variability Tracking

Movement variability tracking traces the spatiotemporal movement of limbs against a pre-recorded action template. This paper introduces an ultra-low-power, low footprint, and data-efficient movement variability detection algorithm implemented on-chip on an inertial sensor die. During the training phase, the user attaches an inertial sensor to the target limb and performs limb movements under supervision. The algorithm within the sensor automatically segments the region of interest with maximal entropic density and generates approximate gravity vector templates of the action primitive. During the inference phase, weighted similarity metrics provide the latent geometric distance between temporal portions of the stored templates and live template as heatmaps. These heatmaps provide quantitative and actionable information on which parts of the movement trajectory to correct. The algorithm requires under 6 seconds of training data, under 7 kB of memory and 0.2 mA of current on-sensor, and has a temporal resolution of 0.5 seconds. The concept benefits a broad application spectrum, including sports analysis, exercise monitoring, rehabilitation, and gait tracking.