Two-stream prototype network for Wi-Fi CSI-based cross-domain human activity recognition

Abstract

Human activity recognition (HAR) using Wi-Fi channel state information (CSI) faces severe challenges in cross-domain generalization and data scarcity. Existing methods either rely on complex hardware deployment or suffer from insufficient spatiotemporal feature extraction, leading to poor performance under domain shifts. To address these issues, this paper proposes a lightweight two-stream feature extractor, Recurrent Convolutional Recursive Transformer—Multi-Scale Convolution Augmented Transformer (RCRT-MCAT), for few-shot cross-domain HAR. The model decouples CSI signals into a temporal stream and a channel stream to separately mine complementary spatiotemporal information. The MCAT branch employs multi-scale convolution and adaptive attention to capture fine-grained temporal patterns. The RCRT branch adopts recurrent convolution and recursive Transformer to efficiently model spatial dependencies across antennas and subcarriers. An evaluation framework is established on two public datasets, SignFi and Wiar, covering four experimental settings: in-domain recognition, cross-environment recognition, cross-environment cross-user recognition, and cross-dataset recognition. Experimental results demonstrate that, on the most challenging 76-way cross-environment gesture recognition task, the proposed model achieves an accuracy of 77.1% under the 1-shot setting, representing a 24.2 percentage point improvement over the FewSense model. When the number of samples is increased to 5-shot, the accuracy rises sharply to 91%, which is 28.2 percentage points higher than FewSense. In the cross-dataset recognition scenario, our model reaches a fine-tuned accuracy of 77.8% on user a2, 13.5% higher than FewSense. The average unfine-tuned accuracy across all users is 62.2%.