Answers are grounded in this paper and cite their sources.
The Physics-Guided Domain-Quality Selective Network (PG-DQSNet) combines CSI motion representations, domain supervision, and class-conditional review scores for cross-domain fall recognition. Packet-centered amplitude dynamics and sanitized relative-phase differences feed a multi-scale temporal encoder. The selector uses capture quality, calibrated probability, dropout disagreement, and prototype distance. We evaluate 2011 recordings with 429 demonstrated falls from three participants and three rooms using acquisition-group-separated fitting, calibration, and test partitions. Matched deterministic inference gives person/room-disjoint macro F2 of 0.608/0.537 against amplitude-only input at 0.526/0.373. The original frozen policy gives recall of 0.941/0.762 with specificity of 0.064/0.319. At a descriptive 10% test false-positive rate, matched PG-AP recall is 0.111/0.189. Learned selection gives room-disjoint AURC of 0.507 against confidence ordering at 0.653; person-disjoint values are 0.686/0.681. Frozen review thresholds defer 19.7%/33.8% of recordings. Class-specific counts, matched-inference controls, perturbation tests, and split ledgers document the measured recognition and review behavior.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: