A multimodal wearable system for Human Activity Recognition during daily activities

This study proposes a wearable system consisting of inertial and electromyography (EMG) sensors for the recognition of activities of daily living in the context of neurodegenerative diseases. The value of Human Activity Recognition (HAR) rests in its ability to test motor cognitive abilities objectively and precisely at home and to keep track of activities performed, for instance when it's necessary to follow the impact of pharmacological therapy. Current technologies for HAR primarily rely on two types of systems: camera-based systems and wearable sensor systems. In this study, a wearable multimodal system based on the fusion of EMG and inertial sensors was designed and developed. Data have been collected from 21 young subjects performing 5 gestures and, after a data pre-processing, a multi-modal dataset has been created. Then, classification has been performed using two machine learning techniques, i.e. a multiclass Support Vector Machine (SVM) and a Random Forest (RF). The best combination of sensors is analyzed, considering both classifiers’ accuracy and wearability. Following the proposed approach, levels of accuracy of 79% are reached using EMG and the inertial sensor on the dominant index finger and wrist, whereas, using EMGs and only the hand's wrist sensor achieves an accuracy of 78%, showing that an excellent yet non-invasive activity recognition system can be obtained with just a few sensors. Finally, this system can be a reliable solution to be used both in clinics and home care.