# Advanced signal-processing framework for remote photoplethysmography-based heart rate measurement: Integrating adaptive Kalman filtering with discrete wavelet transformation

## TL;DR

This paper introduces a new signal-processing framework for noncontact heart rate measurement using a camera, which improves accuracy by reducing noise and motion artifacts.

## Contribution

The novel framework combines discrete wavelet transform and adaptive Kalman filtering for enhanced heart rate estimation in rPPG.

## Key findings

- The proposed DWT-RAKF framework achieved a MAE of 0.72 and RMSE of 1.14 bpm on the PURE dataset.
- It outperformed existing methods in real-world conditions with varying skin tones and lighting.
- The algorithm showed MAEs of 0.94 and 1.11 bpm for fair-skinned and dark-brown subjects, respectively.

## Abstract

Remote photoplethysmography (rPPG) is a noncontact camera-based optical technique which analyses skin-color variations due blood volume changes beneath the skin and estimates vital signs such as heart rate (HR). Driven by the increasing demand for long-term unobtrusive vital sign monitoring systems, rPPG has grown significant interest in clinical and nonclinical domains. However, conventional rPPG methods are often limited by following challenges such as motion artifacts (MA), ambient light intensity and weak signal quality. Moreover, their overall accuracy is significantly impacted by demographic and skin-tone variations. To address these limitations, an advanced signal-processing framework integrating discrete wavelet transform (DWT) for denoising and signal-to-noise-ratio enhancement with residual-based adaptive Kalman filtering (RAKF) for frame-wise temporal consistency and MA reduction is proposed (DWT-RAKF). Further, a multi-channel fusion strategy is integrated with dual-stage band-pass filtering technique to isolate the HR signal while effectively discarding unrelated signal components. Our proposed framework is evaluated on both public and custom datasets. Regarding the PURE dataset, the proposed framework obtained a mean absolute error (MAE) and a root mean square error (RMSE) of 0.72 and 1.14 bpm respectively, outperforming several conventional state-of-the-art methods. To further evaluate its real-world performance, intra-dataset testing is implemented using custom dataset comprising subjects with varying skin-tones and under natural lighting conditions. The results revealed that the proposed algorithm obtained the lowest MAEs of 0.94 and 1.11 bpm for fair-skinned and dark-brown subjects respectively, indicating that the integration of the proposed signal filtering strategy with rPPG achieved effective real-time HR measurement.