Realization of wearable sensors-based human activity recognition with an augmented feature group

Feature extraction is a critical stage in human activity recognition. The information carried in features directly affects the classification performance. This paper explores a new group of features for activity recognition, which have not been broadly applied in previous works in this field. The newly introduced features are related to the attitude of the on-body devices, being extracted from both time-domain and frequency-domain. Based on the collected data, we implemented certain standard data mining techniques, e.g., the Minimum-Redundancy-Maximum-Relevance (mRMR) algorithm for feature selection, and Support Vector Machine (SVM) for classification, to evaluate the performance of the hypothesis. The comparison studies suggest the augmented features perform better than the commonly used features, giving a higher recognition accuracy of 93.46%. Exploring new features without adding more sensors, while improving the accuracy significantly, enables an efficient extraction of features from limited availability of sensors.