Parameter-Efficient Domain Adaption for CSI Crowd-Counting via Self-Supervised Learning with Adapter Modules

Device-free crowd-counting using WiFi Channel State Information (CSI) is a key enabling technology for a new generation of privacy-preserving Internet of Things (IoT) applications. However, practical deployment is severely hampered by the domain shift problem, where models trained in one environment fail to generalise to another. In this paper, we present a two-stage framework for cross-domain CSI-based crowd counting that combines self-supervised contrastive pretraining with parameter-efficient adapter-based fine-tuning and an interpretable stateful counting logic. The learnt representations are effectively transferred between environments and tasks. In the novel WiFlow dataset presented, the approach achieves an MAE of 0.44 in a 10-shot setting, while on the public WiAR benchmark it achieves 98.8% accuracy. We further introduce a Generalisation Index (GI) to summarise relative cross-domain performance; results indicate strong transfer for event classification, while counting error remains more sensitive to domain-specific temporal effects. Furthermore, our framework sets a new state-of-the-art public WiAR benchmark with 98.8% accuracy. A systematic component-wise ablation isolates the contribution of each element: self-supervised pre-training provides the largest single performance gain, ADDA alignment bridges cross-dataset domain gaps, adapter fine-tuning matches full fine-tuning to within 1% (98.84% vs. 99.67%) while training 97.2% fewer parameters, and the stateful counter reduces final occupancy MAE by 57% over raw classification output alone. Our work provides a practical and scalable approach for developing robust sensing systems, demonstrating promising transfer capability across diverse indoor environments.