Classification of Hand-Based and Non-Hand-Based Physical Activities Using Wearable Sensors

This study presents the application of machine learning techniques for physical activity recognition on data obtained from wearable sensors. For this purpose, it proposes the separate classification of hand-based (i.e., eating, writing, clapping) and non-hand-based (i.e., walking, jogging, sitting) human activities recorded by the accelerometer and gyroscope sensors of smartphones and watches. Different machine learning algorithms were compared to build the most appropriate model for the application. The experimental results showed that building two separate classification models for hand-based and non-hand-based activities achieved better accuracy (94.96%) than a single classification model (92.10%) which covered all activities.