CIT-HAR: A High Accuracy and Lightweight Human Activity Recognition System Using CSI Heatmaps and a Hybrid Transformer Network

Device-free human activity recognition based on channel state information (CSI) has gained significant attention in the Internet of Things (IoT) domain. However, raw CSI data often contains redundant information, and existing methods struggle to efficiently extract meaningful features from diverse human activity data, leading to challenges in both recognition accuracy and computational complexity. To address these challenges, this article proposes a novel system, CIT-HAR. Firstly, a preprocessing mechanism that combines subcarrier selection and dynamic signal purification is introduced. This mechanism reduces redundant information within the raw data and enhances the discriminability of features within the selected subcarriers. The preprocessed features are then transformed into heatmaps, which serve as representative samples of the activities. Secondly, a new attention-fusion convolutional neural network (CNN) architecture, FiNet, is designed. FiNet employs dilated convolutions and efficient channel attention to enhance local features, while leveraging a hybrid CNN-Transformer structure to extract both local and global features. An attention mechanism is then applied to effectively fuse these features. Experimental results demonstrate that CIT-HAR significantly outperforms the state-of-the-art HAR methods, achieving higher recognition accuracy and shorter training time.