A high-accuracy low-precision machine learning system for health monitoring

Abstract

Falling is serious and can have dangerous consequences, especially for older people. Consequences can range from minor injuries such as scratches or bruises to more serious harm such as head trauma. Coupled with an ageing population, there is, therefore, a need for fall detection systems. Such systems utilise sensors such as accelerometers or depth cameras to collect data, and a threshold-based algorithm or machine learning model determines whether a fall will occur, is occurring or has occurred. In this project, a fall detection system will be developed using an accelerometer and gyroscope. Fall detection will then be performed using a machine learning model deployed on an ultra-low-power, artificial intelligence microcontroller. This eliminates the need for bulky and expensive computational hardware or cellular connection to cloud platforms for computation. A machine learning method is preferred over threshold based algorithms due to higher accuracy and robustness.