Device-Free Human Activity Recognition in Through-the-Wall Scenarios Using Single-Link Wi-Fi Channel Measurements

Device-free sensing using WiFi channel state information (CSI) is a non-intrusive method of human activity recognition in an urban built environment. In this paper, we propose a human activity recognition model for through-the-wall scenarios based on a novel feature set and machine learning models. The proposed feature set leverages time and frequency domain characteristics of single-link WiFi channel measurements, collected from a low cost IoT device. Diverse non line-of-sight conditions are analyzed such as open and closed doors, occupants in multiple rooms and transmitter-receiver spacing of up to 9m, to constitute 7 datasets. The performance of various machine learning models on the proposed feature set was compared and high prediction accuracies between 97% to 99% was obtained on Linear Discriminant Analysis (LDA) and tree-based ensemble learning models. Joint activity recognition and location classification was also examined and by combining the proposed time and frequency domain features, a mean-cross validation accuracy of 99.6% was obtained on the LDA classifier.