A Heart Rate Measurement Model Based on Koopman Predictors With FMCW Radar

This study introduces a new deep sequence model based on Koopman theory, named HRKNet, designed for heart rate (HR) measurement using frequency modulated continuous wave (FMCW) radar. The HRKNet model comprises multiple Koopman blocks, which include the Fourier filter, time-invariant Koopman predictor (KP), and time-variant KP. The Fourier filter employs frequency-domain statistical techniques to differentiate between time-invariant and time-variant components. The time-invariant KP processes time-invariant dynamics through the global learning operator, whereas the time-variant KP manages time-variant dynamics using the local computation operator within the backtracking window and produces a fit input. The model then outputs the aggregated predictions from all blocks. Validation of the model is carried out using data from various scenarios. The mean absolute error (MAE) for HR measurement is 2.68 bpm under stationary conditions and 3.18 bpm, when there is body movement and external interference. In multitarget HR measurement on a publicly available dataset, the model achieves an MAE of 2.69 bpm, outperforming other models in terms of accuracy. With the parameter configuration providing optimal performance, the model’s prediction time is 18.45 ms per time window, utilizing 24.84 MB of memory. In conclusion, the HRKNet model demonstrates notable robustness, accuracy, and efficiency across different monitoring conditions, offering a promising method for deploying HR estimation algorithms at the edge using FMCW radar. Code is available at this repository: https://github.com/licongsheng/HRKNet