Objective Heart rate serves as a pivotal metric for assessing cardiovascular and cerebrovascular health. Routinely monitoring heart rate is crucial in averting cardiovascular diseases and treating chronic ailments. Image photoplethysmography (IPPG) technology enables non-invasive measurement of physiological parameters by detecting subtle changes in light intensity caused by variations in blood volume on the surface of the skin. Recently, IPPG has witnessed extensive applications in measuring diverse vital signs. However, prevailing investigations on heart rate detection using near-infrared video predominantly occur in controlled environments due to the constrained absorption of blood constituents in the near-infrared spectrum. Studies about moving objects are scant and suffer from high errors and low generalizability. This study proposes a heart rate detection method combining orthogonal projection and singular value decomposition, termed the OP-SVD method. The OP-SVD method adeptly extracts high signal-to-noise ratio IPPG signals from single-channel near-infrared facial videos, surmounting the influence of motion artifacts and achieving accurate heart rate detection. This research holds certain significance for clinical applications requiring continuous monitoring during nighttime in settings like intensive care units (ICUs) and non-contact human health monitoring during the night. Methods The proposed OP-SVD method comprises the following steps: First, the near-infrared videos underwent processing to detect and track facial landmarks. Each frame of the facial imagery is divided into N=44 regions of interest (ROI) by leveraging the positions of these facial landmarks and blood perfusion status, effectively preserving the pulse wave information. Second, spatial averaging is applied to mitigate camera quantization noise from pixels within each ROI; the resulting time-series data are concatenated to generate the raw IPPG signal matrix. The orthogonal projection algorithm is employed to eliminate motion artifacts from the raw IPPG signal matrix caused by head movement, yielding a denoised IPPG signal matrix. A fifth-order Butterworth band-pass filter is applied to each time series in the IPPG signal matrix to adapt to a broad range of normal heart rates. Additionally, a detrending filter is used to remove low-frequency drift noise from the time series. Next, the filtered IPPG signal matrix undergoes singular value decomposition for further denoising, and reconstruction is performed based on signal quality indices, yielding a clean IPPG signal matrix. Finally, the reconstructed signals are processed using a 10 s sliding window with a step size of 0.2 s to perform a fast Fourier transform (FFT). Zero-padding is applied to the windowed data to prevent spectral leakage, generating the corresponding power spectra. Heart rate estimation is obtained by identifying the frequency with the highest peak in the power spectrum. Results and Discussions The proposed OP-SVD method can achieve accurate heart rate measurements under motion conditions. Illustrated by a 30 s near-infrared video featuring a moving subject, heart rate estimation values obtained through the OP-SVD method correlate consistently with reference values (Fig. 5). Additionally, we conduct a correlation analysis between the heart rate values obtained without denoising and those estimated using the OP-SVD algorithm (Fig. 6). The heart rate estimation values from the OPSVD algorithm show a stronger concordance with the true values. It underscores that the OP-SVD method significantly improves the consistency between heart rate estimation value and true heart rate result, achieving robust heart rate measurement under motion conditions. Finally, the OP-SVD method is compared with traditional single-channel signal processing methods based on IPPG technology [single-channel filtering (SCF), empirical mode decomposition (EEMD), and single-channel independent component analysis (SCICA)]. Empirical findings substantiate that the OP-SVD-based heart rate detection method outperforms its counterparts, with a mean absolute error (MAE) of 3. 14 bit/min (Table 1). Conclusions In response to the suboptimal accuracy of video-based heart rate detection under infrared light during nighttime conditions and the prevailing challenge of achieving high-precision heart rate measurement amidst subject head movement, this study proposes the OP-SVD heart rate detection method. The proposed method combines orthogonal decomposition projection with singular value decomposition to mitigate noise stemming from head movement, facilitating high-precision and robust heart rate detection under nighttime near-infrared scenarios using video. Comparative analysis with existing classical algorithms underscores the superior performance of the proposed method, with a mean absolute error of 3.14 bit/min.
