Packets Partition Network for WiFi-Based Human Activity Recognition
Pu Li, Senquan Yang, Songxi Hu, Xiuwei Yu · 2024
The utilization of WiFi signals for Human Activity Recognition (HAR) has attracted considerable interest due to its non-intrusive nature and respect for privacy. Empowered by the strong feature extraction ability of Convolutional Neural Network (CNN), many CNN-based models have been proposed to improve the recognition accuracy of HAR tasks. Recent research has integrated attention mechanisms—previously validated for their outstanding efficacy in vision-related tasks—into HAR, However, existing deep learning models take every packets of CSI data to extract discriminative feature. They neglect the continuity of CSI data, which includes not only along the packets dimension but also the channel dimension. In this paper, we propose a Packets Partition Network (PP-Net), which uniformly partitions the convolution feature along the packets direction to extract partitioned features. A custom loss function based on Cross-Entropy is used to train the model, optimizing it to minimize the distance between predictions and ground truth labels. When compared to state-of-the-art HAR methodologies, our PA-Net exhibits superior performance, achieving accuracies of 99.0% and 97.7% on the UT-HAR and NTU-HAR datasets, respectively. This demonstrates exceptional ability of proposed method to manage the continuity of CSI packets within HAR tasks.