PULSE
Prototype-Guided Learning of Temporal CSI Features for Few-Shot Activity Recognition
We present PULSE, a lightweight pipeline for activity recognition from Wi-Fi CSI. From real-world captures we form three physics-guided inter-frame descriptors—real-part correlation, complex Euclidean change, and frame energy—over short windows; a compact 1D-CNN learns 128-D embeddings.
To adapt across environments, we use training-free few-shot prototypes in cosine space built from only K labeled target windows. Two plug-in normalizers—Support-Stat Adaptive Normalization and Prototype Fusion—further improve robustness, and an optional conformal rejection provides calibrated “unknown” decisions. On our real-world CSI data, PULSE achieves 98.9% in-domain accuracy; with K=5 it attains 99.7% known-class accuracy without fine-tuning. PULSE is compute-efficient, requires no per-environment retraining, and offers a clear LOEO protocol for reproducible few-shot RF sensing.
