Performance Analysis and Optimization of Eulerian Video Magnification for Vital Signs Estimation in Resource-Constrained Embedded Systems

This work presents a comprehensive computational performance and accuracy analysis of an embedded system implementing Eulerian Video Magnification (EVM) for vital signs estimation. The system was evaluated on a Raspberry Pi 4 and PC, comparing five face detection models with MediaPipe selected for optimal efficiency. Results demonstrate that the implementation enables real-time heart rate (HR) estimation with an end-to-end processing rate of 44.69 ± 0.91 FPS. The system achieved perfect face detection (100% detection rate) with minimal positional jitter (0.006 ± 0.019 pixels). HR estimation accuracy showed a mean absolute error of 12.39 ± 11.28 BPM, with 71.78% of measurements falling within 15 BPM of reference values and 46.21% within 5 BPM. The system maintained thermal stability without throttling at an average CPU frequency of 1.68 GHz, confirming the feasibility of EVMbased vital sign monitoring in resource-constrained embedded systems while providing realistic performance benchmarks.