A Survey on Hand-based Behavioral Activity Recognition

Wrist-worn devices such as smart-watches can be a perfect way to collect personalized user data, since these devices are embedded with various sensors such as the accelerometer and gyroscope which can sense user's hand movement, as well as help in identifying the activity performed by the user. Additionally, smart-watches are light and unobtrusive unlike the wristbands used in past studies which require the user to be connected by several wires and bulky sensors. In this study, we examined the different neural network adopted for the recognition of hand-based behavioral activity to recognize six complex hand motion activities such as smoking, drinking, eating, typing giving a talk and writing with a pen in a non-controlled environment, using accelerometer and gyroscope sensor signals acquired from smart-watch. Unlike other machine learning methods which use hand crafted features, CNN can learn features automatically from the data as well as capture local dependency.