Deep activity recognition in smart buildings with commercial Wi-Fi devices. (5th October 2020)

• Record Type: Journal Article
• Authors: Zhou, Qizhen; Xing, Jianchun; Zhang, Yuhan; Yang, Qiliang
• Is Part Of: International journal of simulation and process modelling. Volume 15:Number 4(2020)
• Page Start: 369
• Page End: 378
• Publication Date: 2020-10-05
• Subjects: channel state information -- CSI -- wireless sensing -- deep learning -- principle component analysis -- PCA -- smart building; Management -- Computer simulation -- Periodicals; Mathematical models -- Periodicals; Operations research -- Periodicals; Simulation methods -- Periodicals; 003.05
• Languages: English
• ISSNs: 1740-2123

Abstract

Activity recognition acts as a key enabler of smart building applications, such as behaviour analysis, health diagnosis and user authentication. However, existing methods either require burdensome equipment, or light and line-of-sight (LOS) working conditions. To address this challenge, we propose DeepAR, a device-free human activity recognition system with prevailing Wi-Fi signals, which circumvents the use of dedicated devices. DeepAR mainly exploits two key techniques to recognise human daily activities. Firstly, a novel principle component extraction method is presented to capture the motion-induced distortions and discard the irrelevant interferences. Secondly, deep feature maps are constructed with time and frequency domain characteristics, and a deep convolutional neural network (CNN) model is further applied to classify the activity labels. DeepAR is implemented with commercial Wi-Fi devices, and the performance is evaluated through extensive experiments. Experiment results show that DeepAR can achieve an average accuracy of 98.6% in a meeting room and 96.4% in a student office.