Cross-Domain Multi-Person Human Activity Recognition via Near-Field Wi-Fi Sensing

TL;DR

This paper addresses multi-person HAR using near-field Wi‑Fi sensing with commercial devices, where cross-domain adaptation is hindered by incomplete activity categories. It introduces WiAnchor, a training framework with time-based CSI embedding, a pre-training stage that enlarges inter-class margins, a fine-tuning stage that uses anchor matching to filter subject-specific interference, and an inference strategy that combines logits with anchor similarity. A new NFS‑Fi dataset (~65k samples) collected under native traffic across diverse environments demonstrates that WiAnchor delivers over 90% cross-domain accuracy, including significant gains (∼56.8%) for categories without FT data. The approach advances practical Wi‑Fi HAR by enabling robust generalization to unseen subjects and absent categories, with implications for ISAC and security/privacy considerations in wireless sensing.

Abstract

Wi-Fi-based human activity recognition (HAR) provides substantial convenience and has emerged as a thriving research field, yet the coarse spatial resolution inherent to Wi-Fi significantly hinders its ability to distinguish multiple subjects. By exploiting the near-field domination effect, establishing a dedicated sensing link for each subject through their personal Wi-Fi device offers a promising solution for multi-person HAR under native traffic. However, due to the subject-specific characteristics and irregular patterns of near-field signals, HAR neural network models require fine-tuning (FT) for cross-domain adaptation, which becomes particularly challenging with certain categories unavailable. In this paper, we propose WiAnchor, a novel training framework for efficient cross-domain adaptation in the presence of incomplete activity categories. This framework processes Wi-Fi signals embedded with irregular time information in three steps: during pre-training, we enlarge inter-class feature margins to enhance the separability of activities; in the FT stage, we innovate an anchor matching mechanism for cross-domain adaptation, filtering subject-specific interference informed by incomplete activity categories, rather than attempting to extract complete features from them; finally, the recognition of input samples is further improved based on their feature-level similarity with anchors. We construct a comprehensive dataset to thoroughly evaluate WiAnchor, achieving over 90% cross-domain accuracy with absent activity categories.

Figures (18)

• Figure 1: Constructing dedicated links via smart devices holds promise for multi-person HAR, but the subject-specific characteristics necessitate model fine-tuning for cross-domain adaptation, which is hindered by the absence of certain activity categories.
• Figure 2: Experiments on near-field sensing. The results indicate that the subject in the near-field of its corresponding UE has a dominant influence on the CSI.
• Figure 3: Source domain accuracy vs. target domain accuracy.
• Figure 4: FT for domain adaptation. The (a) limited number and (b) absence of samples from specific categories significantly degrade accuracy.
• Figure 5: Visualization with t-SNE. The absence of category-specific data negatively affects feature extraction for all categories.