EdgeCaps compressed a 33.7M-parameter teacher model into a 1.59M-parameter student model, achieving 95.5% accuracy and reducing inference latency to <0.3s on a Raspberry Pi 3.
EdgeCaps enables real-time, energy-efficient, and thermally stable RF-based vital sign monitoring directly on embedded devices by compressing a high-capacity Capsule Network.
Radio frequency (RF) sensing enables unobtrusive, non-contact monitoring of vital signs and has been increasingly enhanced by deep neural networks (DNNs), which excel at extracting discriminative features from complex RF signal representations. These advances have significantly improved the accuracy and robustness of RF-based health monitoring; nevertheless, state-of-the-art DNN models remain computationally intensive and unsuitable for deployment on low-cost, resource-constrained edge devices. As a result, most existing RF sensing systems offload data to high-performance platforms, leaving real-time embedded deployment underexplored. We propose EdgeCaps, a framework that compresses a high-capacity Capsule Network (CapsNet) into a compact student model optimized for edge deployment through structured pruning and knowledge distillation. EdgeCaps transfers knowledge from a 33.7 M-parameter teacher to a 1.59M-parameter student, achieving 95.5% accuracy–within two percentage points of the teacher, while reducing parameters and computational complexity by 95%. We validate the system through continuous respiration monitoring on a Raspberry Pi 3, where the student model achieves mean inference latency below 0.3 s, RAM usage of 18%, and core temperature under 44°C, compared with the teacher’s 0.4–0.5 s latency, 22% RAM usage, and 51°C. These results demonstrate that EdgeCaps enables real-time, energy-efficient, and thermally stable RF-based vital sign monitoring directly on embedded devices. By combining CapsNet’s representational power with model compression and knowledge distillation, this work establishes a novel pathway toward scalable, low-cost, and intelligent health monitoring systems at the edge.
Ishabakaki et al. (Thu,) conducted a other in Vital sign monitoring. EdgeCaps (compact student model) vs. High-capacity Capsule Network (teacher model) was evaluated on Accuracy and computational efficiency (inference latency, RAM usage, core temperature). EdgeCaps compressed a 33.7M-parameter teacher model into a 1.59M-parameter student model, achieving 95.5% accuracy and reducing inference latency to <0.3s on a Raspberry Pi 3.