Efficient Human Activity Recognition: Machine Learning at the Sensor Level

Wearables equipped with Inertial Measurement Units (IMUs) can be used for Human Activity Recognition (HAR). Typically, activity classification occurs within the microcontroller embedded in the wearable device. In this paper, we demonstrate the feasibility of performing real-time HAR at the extreme edge, hence directly within the IMUs. The evaluation has been carried out on two 6-axis IMUs developed by STMicroelectronics (LSM6DSV16BX and LSM6DSO16IS) and using data collected through an earbud. Specifically, this study analyzed two tasks. The first, for dynamic activities recognition, achieved 99.2% accuracy using a Convolutional Neural Network (CNN) implemented into the Intelligent Sensor Processing Unit (ISPU) of the LSM6DSO16IS. The second, for posture recognition, achieved 79.1% accuracy using the Machine Learning Core (MLC) of the LSM6DSV16BX. Moreover, both IMUs successfully executed the algorithms, consuming between 590 μA and 620 μA.