WiFi Signal-based Human Activity Recognition using Bidirectional Mamba Models

WiFi signals, due to their non-contact, low-cost and privacy-preserving nature, have become an important data source for Human Activity Recognition (HAR). In this paper, we propose a state space model-based deep learning method for HAR using channel state information obtained from multi-antenna WiFi signals. The model is designed by combining depthwise separable convolution and a bidirectional Mamba model, aiming to extract the spatial and temporal features jointly along long sequential WiFi data frame. Experiments show that our proposed model achieves significant improvements in accuracy on three widely-used open datasets of NTU-Fi HAR, NTU-Fi Human-ID and UT-HAR, when compared to reference SOTA algorithms including BiLSTM, ProbSparse Transformer and ViT. Experimental results also show that our proposed model significantly reduces parameter size compared to reference models with comparable latent states while ensuring near-perfect test accuracy.