Sample-Efficient Cross-Domain WiFi Indoor Crowd Counting via Few-shot Learning

WiFi-based indoor crowd counting (ICC) has raised increasing interest in wireless sensing applications due to its low hardware cost and privacy preservation in sensing data collection. However, most existing WiFi-based ICC models that are fined-tuned in one domain suffer from dramatic performance degradation when applied to dissimilar domains, e.g., change of device and background deployments. In this case, a large amount of data samples need to be collected to retrain the ICC models, which is costly and indeed infeasible in many practical applications. In this paper, we propose a sample-efficient cross-domain
$W$
iFi ICC method based on few-shot learning (FSL) techniques, which achieves robust ICC performance under different application setups. Specifically, we first obtain a convolutional neural network (CNN) based CSI amplitude feature extractor using sufficient data samples in the source domain. When performing ICC task in a new target domain, the proposed feature extractor can efficiently generate a low-dimensional feature map of the input CSI measurements. By doing so, we show that a lightweight logistic regression (LR) classifier suffices to produce high ICC accuracy even under a very limited target domain dataset. Experimental results show that the proposed method could achieve 78.85% and 94.18% accuracy for 0–8 people cross-domain ICC tasks under 1-shot and 5-shot learning cases, respectively.