Temporal-Frequency Attention-Based Human Activity Recognition Using Commercial WiFi Devices

Human activity recognition has been growing for decades in a variety of technological disciplines. However, in the existing WiFi-based human activity recognition systems, there are the following problems: Firstly, in the processing of channel state information (CSI) data, mainly for the removal of noise in the superimposed signal, there is no effective removal of useless multipath signals; Secondly, the data segmentation algorithm based on the empirical threshold requires manual adjustment of the threshold in different environments, resulting in poor robustness and unstable segmentation; Thirdly, simple learning classification is applied without specific design for CSI data structure and sufficiently abstracting information features. In this paper, a device-free human activity recognition system with a temporal-frequency attention mechanism is proposed, which can be deployed on commercial WiFi devices to identify human's daily activities. Firstly, the multipath signal affected by the channel change is extracted by using the difference of the propagation delay of different multipath, thereby eliminating the delay and invalid multipath signals that have undergone multiple reflections and refractions. Secondly, a neural network model based on attention mechanism is proposed, which assigns different weights to different characteristics and sequences by imitating the human brain to dedicate more attention to important information. Then, the long short-term memory (LSTM) model is used to learn the correlation features of different dimensions to realize human activity recognition. Finally, the system performance is evaluated in different environments, and the experimental results show that our syetem holds a better performance in both line-of-sight (LOS) and non-line-of-sight (NLOS) than the existing human activity recognition systems.