WiFi sensing based macro and micro human activity recognition using machine learning framework

Abstract

Human activity recognition (HAR) is a rapidly advancing field with applications spanning newlinehealthcare, smart environments, and human-computer interaction. Traditionally, HAR newlinesystems have relied on wearable devices or cameras, which can be intrusive and privacyinvasive. newlineHowever, the emergence ofWi-Fi sensing as a non-intrusive, device-free alternative newlinefor recognizing human activities is a significant advancement. By leveraging the ubiquitous newlinenature of Wi-Fi and the fine-grained information contained in Channel State Information newline(CSI), Wi-Fi sensing enables the detection and classification of human activities in various newlineenvironments, offering a new level of convenience and privacy. newlineThis research explores the application of Wi-Fi sensing technology for human macro newlineand micro activity recognition in real-time environments. By analyzing Channel State newlineInformation (CSI) from Wi-Fi signals, we have developed a model that demonstrates a high newlinelevel of accuracy in detecting and classifying macro activities such as walking, sitting, or newlinefalling. Our proposed multi-stage deep learning model, which uses Squeezenet, Alex Net, newlineand Google Net, significantly reduces false negatives, particularly for critical activities such newlineas falls, while maintaining a balance with false positives, thereby ensuring the system s newlinereliability. newlineExhaustive experiments are conducted to collect the CSI values for friendly and violent newlinebehavior between two participants. It was observed that the CSI amplitude of violent activity newlineis high when compared with non-violent activities. We proposed a CNN model to distinguish newlinebetween violent and non-violent behaviors with an accuracy of 95.59%. Additionally, we developed a two-phase deep neural network that identifies the individual newlineand then classifies the activity they are performing, enhancing the accuracy of combined newlineparticipant and activity recognition from 81.4% to 83.9%. A two-phase multi-stage deep newlineneural network model is proposed to minimize false negatives. This multi-stage model newlineuses transfer learning-based