Answers are grounded in this paper and cite their sources.
Fall detection on wearable edge devices requires high sensitivity within tight latency, memory and energy budgets. We propose PhyCL-Net, a time-domain network without an explicit spectral feature branch. It combines a physics-guided dynamic kernel (PDK), which adapts temporal receptive fields to local signal dynamics, with jerk-aware fall-aware attention (FAA), which emphasises transient segments. A hierarchical dual-view contrastive regulariser is used only during training, and its auxiliary heads are removed for deployment. On SisFall with 12-fold leave-one-subject-out (LOSO) validation, PhyCL-Net achieves 98.21 ± 0.10% accuracy and 98.16 ± 0.10% Macro-F1 across five seeds, with 1.049M parameters and 0.15 GFLOPs. Compared with a matched spectral baseline, its median single-thread CPU latency is 31.6% lower (125.99 ms). On the empirical ROC curve, the false-positive rate (FPR) at a 95% true-positive rate (TPR) is 0.72%, but the TPR at 1% FPR is 2.73 percentage points lower than that of the spectral baseline.
No takes yet. Share an insight, caveat, or question.
Zuo et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: