Multi-Level and Multi-Scale Representation Learning for Cross-Person Activity Recognition in Wireless Networks

Human activity recognition (HAR) plays a crucial role in the development of intelligent systems, capable of automatically detecting and classifying human activities. This technology is vital across a range of applications, including smart homes, healthcare monitoring systems, and interactive gaming. Among existing HAR methods, WiFi-based methods, utilizing channel state information (CSI), have emerged as a significant advancement over vision-based or wearable device-based methods, owing to their non-intrusive nature, cost-effectiveness, and privacy preservation. However, the effectiveness of HAR methods can be hindered by the physiological differences among individuals. Specifically, a model trained on the CSI data from one person often underperforms when recognizing activities performed by another person. To overcome the challenge of cross-person activity recognition (CPAR), a multi-level and multi-scale representation learning method (M2RL) is proposed. This method combines learnable multi-level wavelet decomposition with a multi-scale residual convolutional network (MSRN) to extract more robust and discriminative features from CSI data for CPAR. Experimental results demonstrated that our proposed method achieves a recognition accuracy improvement of over 4% compared to conventional deep learning models, such as LSTM and BiLSTM. The codes can be downloaded from https://github.com/nuptzzh/CPAR.