The full text of this item is not available at this time because the student has placed this item under an embargo for a period of time. The Libraries are not authorized to provide a copy of this work during the embargo period, even for Texas A&M users with NetID.
Generalization in Wi-Fi-Based Single Occupant Activity Recognition through Activity-Specific Feature Representation
Abstract
Occupant activity recognition (OAR) has been essential for various human-centric applications, such as elderly healthcare, building energy management, and home security. With the rapid deployment of Wi-Fi networks, exploiting commercial Wi-Fi devices has great potential to support device-free OAR systems. These systems use Channel State Information (CSI), representing how human activity-based environmental changes affect the Wi-Fi signals propagating through physical space. However, the current state-of-the solutions have some practical limitations. First, actual housing environments generate a significant difference in the performance of CSI-based OAR due to various environmental factors, such as structural building materials, surrounding Wi-Fi interference, housing layout, and population density. This research proposes, Wi-Sensing, an approach to exploit spatial-temporal features extracted from multiple receivers deployed throughout an indoor space to recognize an individual’s daily activities accurately. This approach uses a Short-Time Fourier Transform (STFT) to convert time-series CSI data into image data. The converted image data from each receiver are then integrated as large image data, which preserves the spatiotemporal information of all the receiver data. Furthermore, Wi-Sensing exploits a Convolutional Neural Network (CNN) as a feature extractor for large image data and Long Short-Term Memory (LSTM) to classify target activities in daily living (ADLs). As a result, Wi-Sensing recognizes an individual’s ADLs performed at various locations at an average of 96% accuracy and provides consistent performance in different housing environments.
The second limitation of existing CSI-based OAR solutions is performance degradation over time in a housing environment. This is because the nature of Wi-Fi signals has temporal variation, and physical environmental changes affect Wi-Fi propagation over time. Therefore, the distribution of CSI data collected in an actual housing environment continuously changes, and the pre-trained model cannot recognize ADLs for long-term periods. Therefore, this research proposes CSI-based Long-term Occupant Activity Recognition (CLOAR), an approach to extract temporal-invariant and activity-specific features in a semi-supervised meta-learning manner. This approach leverages labeled source data and unlabeled target data with pseudo labels. In addition, it synthesizes numerous query datasets using mix-up-based data augmentation, generalizing a CSI-based OAR model during training. As a result, CLOAR provides an average of 91.09% activity recognition accuracy for target data with different statistical characteristics from source data.
Lastly, previous CSI-based OAR solutions focused on recognizing key ADLs (e.g., falling, sleeping, and eating), but an individual’s daily routine comprises many activities that can gradually vary across multiple days or times. Although assessing an individual’s ADL routine variability can serve as an essential indicator for various human-centric applications, it is challenging to recognize all the activity that occurred in a day. Therefore, this research proposes an approach to assess an individual’s ADL routine variability from directly processing CSI data without activity recognition. This approach uses an autoencoder algorithm that reduces the dimension of input CSI data and reconstructs the original CSI data. First, this research trains an autoencoder-based model to learn a normal ADL routine pattern. Then, the trained model provides significant reconstruction errors for abnormal ADL routines. Finally, this research correlates the size of reconstruction errors and the level of ADL routine variability.
Although Wi-Fi signals are vulnerable to environmental changes, the information about an individual’s bodily movements remains in CSI data. This research extracts activity-specific features from CSI data in various actual housing environments, demonstrating the feasibility of using CSI data for OAR in practical applications.
Subject
channel state information (CSI)occupant activity recognition
routine variability assessment
smart home healthcare
Citation
Lee, Hoonyong (2022). Generalization in Wi-Fi-Based Single Occupant Activity Recognition through Activity-Specific Feature Representation. Doctoral dissertation, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /197993.