WiKnow: A Human Activity Recognition Method in Office Scene with Coordinate Attention from WiFi Channel State Information

In the modern information society, people spend more and more time in the office. Therefore, understanding the various human activities of users in office scenarios is important to enable health monitoring, smart home, and human-computer interaction. Current human activity recognition (HAR) approaches for office scenarios are often characterized by privacy invasiveness, high cost, insufficient accuracy, or the need for additional wearable devices. To this end, we propose WiKnow, a non-contact radio frequency (RF) human activity recognition system based on WiFi channel state information (CSI), which is based on denoising CSI amplitude data and converting it into a 3D tensor based on the CSI data characteristics, inputting it into a designed convolutional neural network (CNN), and improving the performance of the network through the coordinated attention (CA) mechanism, to understand the corresponding human activities. We conducted experiments on our dataset and a public dataset, and the accuracy reached 98.61% and 97.92%, respectively, and the experimental results verified the effectiveness of the method in recognizing human activities.