TAHAR: A Transferable Attention-Based Adversarial Network for Human Activity Recognition with RFID

Human activity recognition based on radio frequency identification (RFID) has become an essential part of Internet of Things. At present, most RFID-based human activity recognition research is focused on domain-specific recognition. However, existing solutions usually match global features for domain adaptation when solving cross-domain problems, and lack consideration of untransferable features, which lead to degradation of recognition accuracy. This paper proposes a novel human activity recognition model TAHAR which adapts transferable attention and adversarial learning to eliminate the negative influence of untransferable features and domain-specific features. In TAHAR, the feature extractor extracts spatio-temporal information of phase and Received Signal Strength Indicator from the RFID signal. Then we utilize an attention module to weight features to minimize the influence of untransferable features and negative transfer. Additionally, discriminators and batch spectral penalization are used to remove domain-specific information, thereby enhancing the transferability and discriminability. Results show that TAHAR achieves an accuracy and recall of 90.17% and 85.46% respectively, achieving an outstanding performance compared with several state-of-the-art methods.