Effects of Housing Environments on Exploiting Channel State Information for Human Activity Recognition Hoonyong Lee 1 , Aryan Sharma 2 , Changbum Ryan Ahn 3 , Nakjung Choi 4 , and Toseung Kim 5 1) Doctoral Student, Department of Civil Engineering, Texas A&M University, Texas, USA. Email: throughdi@tamu.edu 2) Master’s Student, Department of Computer Science & Engineering, Texas A&M University, Texas, USA. Email: aryans@tamu.edu 3) Ph.D., Assoc. Prof., Department of Construction Science, College of Architecture, Texas A&M University, Texas, USA. Email: ryanahn@tamu.edu 4) Ph.D., Member of Technical Staff., Nokia Bell Labs, New Jersey, USA. Email: nakjung.choi@nokia-bell-labs.com 5) Master’s Student, Department of Architecture and Architectural Engineering, Seoul National University, Seoul, Korea. Email: toeskim92@gmail.com Abstract: Monitoring human activities in daily living (ADL) using fine-grained Wi-Fi signal signatures has the great potential to support an intelligent eldercare system in a home environment. Such an approach analyzes Channel State Information (CSI) extracted from a sequence of Wi-Fi packets and identifies its temporal variations as a fingerprint, caused by human activities. However, the performance of the Wi-Fi-based ADL recognition can be greatly affected by the surrounding physical environment due to the way the radio wave propagates as it penetrates through an obstacle or is reflected. In particular, building materials are expected to significantly affect its performance. In this context, this paper examines the effect of the physical housing environment on the performance of Wi-Fi-based ADL recognition. ADL recognition systems were implemented at two different housing environments: a wood-frame apartment and a reinforced concrete-frame apartment, which represent typical housing environments in United States and Korea, respectively. The experimental results indicate that structural building materials, combined with other environmental factors (i.e., housing density), create a significant difference in the accuracy rate of Wi-Fi-based ADL recognition and provide insight into how such systems should be configured for homes. Keywords: Activity Recognition, Channel State Information (CSI), Machine Learning. 1. INTRODUCTION The demand for recognizing human activities in daily living (ADL) is increasing due to applications such as elder health care, smart building, and home security (Tan et al., 2018; Yousefi et al., 2017). Traditional activity recognition technologies are based on cameras and wearable sensors. However, camera-based approaches have privacy issues in addition to the line-of-sight requirement, and approaches using wearable sensors (e.g., inertial measurement units) require users to wear the sensors in order to detect activities (Tan et al., 2018; Yousefi et al., 2017; Wang et al., 2017). In recent years, using fine-grained Wi-Fi signal signatures has been the focus of human activity recognition (Yousefi et al., 2017). As Wi-Fi signals propagate throughout an indoor space, human activities may affect their propagation paths and change the received signals. Wi-Fi signal-based approaches do not require a human to be located in the line of sight or to wear sensors. In addition, such approaches can be operated with existing Wi-Fi networks in most homes without additional devices (Ali et al., 2015). In spite of these advantages, their implementation still has numerous practical challenges due to the sensitivity of fine-grained Wi-Fi signal signatures. Many factors in housing environments could potentially affect the performance of a Wi-Fi-based ADL recognition system. For example, building materials could significantly impact recognition performance since Wi-Fi signals have different propagation loss depending on building materials (Adib & Katabi, 2013; Tesserault et al., 2007). Considering the diversity of housing environments across different countries, it is critical to understand which aspects of the housing environment could potentially affect the performance of Wi-Fi-based ADL recognition. This paper examines the effects of the housing environment, in particular structural material type, on the performance of Wi-Fi-based ADL recognition. Two housing environments with different structural materials were selected: one is a wood-frame low-rise apartment in the United States and the other is a reinforced concrete-frame high-rise apartment in Korea. They represent typical housing environments in the United States and Korea. Locomotive activities are performed in these two different housing environments because this type of activity is found to be the most reliably