Wi-Fi Sensing Enabled Violent Activity Detection in a Smart Home

In recent years, Wi-Fi sensing technology has been used extensively for human activity recognition (HAR), people counting, movement detection, and a few other applications for smart homes. Wi-Fi sensing is evolving as an alternative to video-based surveillance. It does not violate privacy and can work effectively in non-line-of-sight conditions and dark and dusty environments. Violent activity among people present in a room is an undesirable activity and needs to be monitored for initiating preventive and timely action. This paper presents a model for non-invasive, privacy-preserved, and non-line-of-sight monitoring of violent activities involving two persons in a room. We use Intel 5300n network interface card to extract the received Wi-Fi signal's channel state information (CSI). CSI is fine-grained information corresponding to each sub-carrier containing the received signal's crucial amplitude and phase information. We performed numerous experiments and collected the CSI data corresponding to different activities, including violent ones. We then pre-processed the CSI data using the Hampel filter and wavelet denoising techniques. Then we used different machine learning models like LDA, SVM, KNN, Naive-Bayes, and Deep neural network architectures for classifying violent and non-violent activities. It is found that the convolutional neural network (CNN) model outperforms all other machine learning models with an accuracy of 95.59%.