ProbSparse Attention with Stacked Group Convolution for Wireless Signal-Based Human Activity Recognition

With the advancement of Internet of Things, WiFi signal-based Human Activity Recognition (HAR) has demonstrated great potential in various domains. Existing WiFi-based HAR systems pursue high recognition accuracy, but often struggle in achieving model lightweightness. To address this issue, we jointly consider the spatial-temporal correlations of WiFi channel state information. Our proposed HAR method employs an encoder-only Transformer with ProbSparse attention to extract crucial global features from time-series sensor data. Furthermore, it utilizes a stacked group convolutional structure to further encode local temporal and spatial features, respectively, thereby realizing effective extraction of key spatial and temporal characteristics, as well as fusion of global and local features. Experimental results demonstrate that the proposed model achieves an exceptional mean average precision exceeding 99.9% for action recognition across two public datasets: NTU-Fi HAR and NTU-Fi Human-ID, outperforming several state-of-the-art models such as ViT, TCN and BiLSTM. Meanwhile, utilizing ProbSparse attention, our model exhibits a significant improvement in training complexity compared to several state-of-the-art models such as ResNet50, vanilla Transformer, ViT and TCN.