C-HAR: Compressive Measurement-Based Human Activity Recognition

In this paper we present C-HAR, a low-cost,low-complexity, compressive measurement-based human activity recognitionframework for embedded devices, capable of simultaneously filtering andacquiring the sensor readings of a smartwatch's accelerometer and gyroscope atsub-Nyquist rates. A software simulation of the full system obtains an accuracyof 92.0% and 88.0% for the accelerometer-based and gyroscope-based systemsrespectively, obtained at a 5 Hz sample rate. A microcontroller implementationof the system's back-end obtains an accuracy of 90.3% and 87.1% for theaccelerometer-based and gyroscope-based systems respectively with a runtimeover twice as fast as that of a comparable baseline system.