Multi-domain Feature Extraction for Human Activity Recognition Using Wearable Sensors

The extraction and use of features from the raw sensor data plays an extremely crucial role in determining the recognition performance of an activity recognizer. Existing studies aim to train an accurate prediction model by extracting different features, however, few of them systematically investigate the power of features from different domains when they are used separately or jointly. To this end, we conduct a comparative study on multi-domain feature extraction for human activity recognition. Specifically, we first extract features from the time-, frequency-, and wavelet-domains, and then use different combinations of the three domain features to build activity recognizers. Finally, comparative experiments are performed on two activity recognition datasets and four classification models are used to avoid selection bias. Results indicate the superiority of using time-domain or frequency-domain features over wavelet features in terms of prediction performance and also show that the simultaneous use of multi-domain features generally generalizes better across datasets and classifiers, indicating that they, to a certain extent, contain complementary feature information.