Scaling Activity Recognition Using Channel State Information Through Convolutional Neural Networks and Transfer Learning

Unobtrusive sensing is receiving much attention in recent years, as it is less obtrusive and more privacy-aware compared to other monitoring technologies. Human activity recognition is one of the fields in which unobtrusive sensing is heavily researched, as this is especially important in health care. In this regard, investigating WiFi signals, and more specifically 802.11n channel state information, is one of the more prominent research fields. However, there is a challenge in scaling it up. Transfer learning is rarely applied, and when applied, it is done on filtered/modified data or extracted features. This paper focuses on two aspects. First, convolutional networks are used across multiple participants, days and activities and analysis is done based on these results. Secondly, it looks into the possibility of applying transfer learning based on raw channel state information over multiple participants and activities over multiple days. Results show channel state information is accurate for single participants (F1-score of 0.90), but sensitive to different participants and fluctuating WiFi signals over days (F1-score of 0.25-0.35). Furthermore, results show both clustering and transfer learning can be applied to increase the performance to 0.80 when using minimal resources and retraining.