Improving activity recognition with context information

Activity recognition driven by sensor data has been heavily focused on in recent years, especially as wearable sensors become more common and popular personal equipments. As known, recognizing individual activities is one of the typical machine learning applications, which includes several basic phases, such as data collection, feature extraction, model training and performance evaluation. Originally, after data being collected and preprocessed, features are manually extracted for model training. As this may lead to a time-consuming and boring job, feature learning approaches were investigated and received a great success, especially when Deep Neural Networks, a powerful model for feature representation, were employed into this task. That is, classifier could be established with raw data. Those features, no matter manually selected or learned from raw data, depend on separate frames that were split from previous collected long data. Since human activity represented by sensor data is actually a time series signal, context information plays an important role, how to take advantage of knowledge implied in the inter-frames becomes significant. Unlike previous attempt that only enlarges the length of each frame, in this research, a new method that models context information with neighboring frames is investigated for activity recognition tasks. Based on the Daily and Sports Activities, a publicly available dataset, experiments are performed with three typical classifiers for activity recognition. The results show that the proposed method could significantly improve the recognition performance and the more the context information are considered, the higher the recognition accuracy will be.