Imaging Heart Rate Detection Method Based on Clustering and Adaptive Filtering
Objective In recent years,since heart rate is one of the most important indicators of cardiovascular health,non-contact heart rate measurement methods are highly attractive and popular in daily life.Non-contact imaging photoplethysmography(IPPG)has caught much attention from biomedical researchers due to its non-invasive properties without the need for high-performance hardware devices.However,during non-contact imaging where subjects are less constrained,IPPG measurement results are susceptible to interference from rigid and non-rigid movements such as head turning,smiling,speaking and eyebrow raising,and unstable lighting.For improving the IPPG technique,we propose a region of interest(ROI)selection method with a concave lens deformation algorithm and skin color pixel clustering,and an adaptive normalized least mean square(NLMS)filtering algorithm for blood volume pulse(BVP).The proposed method improves accurate ROI extraction in less constrained conditions and the performance of filtering out non-physiological signal intensity fluctuations in ROI.Meanwhile,it has advantages in accuracy and stability under motion scenes and environments with large illumination variations,holding potential significance for non-contact heart rate monitoring in telemedicine,indoor fitness,psychological testing,and unmanned vehicles.Methods We obtain the subjects'heart rates by processing the facial video images.First,the facial skin color region is distorted and expanded by adopting the concave lens deformation algorithm to increase the percentage of the skin pixel region.Next,the K-means++clustering algorithm selects skin pixels again and builds RGB channels to estimate BVP signals.Subsequently,the chrominance-based color space projection decomposition(CHROM)algorithm is applied to pre-denoise the above-mentioned BVP signal.Finally,the proposed adaptive NLMS algorithm is employed to filter out the interference of background light,and then measure heart rate by spectrum analysis.In subsequent experiments,ablation experiments are conducted on the UBFC-rPPG dataset to verify that the improved ROI dynamic extraction method can enhance the accuracy of heart rate detection.In comparison experiments,on the same dataset,the results prove that the proposed method possesses stronger robustness to color signal fluctuations caused by the subject's head movements and facial expressions.Additionally,the results of the lighting fluctuation experiment where the light intensity of the double-arm lamp is continuously adjusted to simulate the changing light scene demonstrate the feasibility and effectiveness of the proposed method.Results and Discussions In ablation experiments,the mean absolute error(MAE)of the improved ROI extraction method with a concave lens deformation algorithm and clustering algorithm amounts to 4.29 beats per minute(min-1),and the standard deviation(SD)is 2.59 min-1,with the mean absolute percentage error(MAPE)of 4.19％and the Pearson correlation coefficient r of 0.66.Our improved ROI selection method achieves the optimum in all the above-mentioned indexes.Integrated with the concave lens deformation algorithm and clustering algorithm,the proposed improved ROI dynamic extraction method can improve the accuracy of heart rate detection in less-constrained conditions(Table 1).The MAE of the proposed method is 0.92 min-1,MAPE is 1.57％,SD is 2.43 min-1,and r is 0.65 for the comparison experiments in motion scenarios,which is better than other unsupervised methods.Compared to the supervised learning methods,our method has advantages with low MAE and SD without the necessity for pre-learning and training(Table 2).Additionally,our proposed method has smaller confidence intervals,which means that the study is more robust to color signal fluctuations induced by head movements and facial expressions of the subjects(Fig.5).In the experiments with drastic lighting changes,the proposed method still possesses smaller MAE,MAPE,and SD than others.The proposed adaptive NLMS method has been proven to be significantly feasible and effective in scenarios with varying lighting conditions(Table 3).By conducting Bland-Altman analysis,the bias of our proposed method is minimal with 95％confidence interval in the-7.8 min-1 to 7.8 min-1(Fig.6).Obviously,it indicates that our method is more robust in removing non-physiological signal fluctuations caused by illumination fluctuations.Conclusions To deal with the interference caused by normal physiological motion and ambient light in the IPPG technique,we propose an ROI dynamic extraction method integrated with the concave lens deformation algorithm,K-means++clustering algorithm,and an adaptive NLMS algorithm on the BVP signals to improve the heart rate measuring stability and accuracy of this technique.Firstly,the concave lens deformation algorithm is adopted to compress facial features in each image frame,which in turn increases the pixel area of the facial skin ROI.Secondly,the K-means++clustering method is employed to resieve the facial skin regions,build the ROI rich in physiological signals,and generate BVP signals with high signal-to-noise ratios.Thirdly,the CHROM algorithm is utilized to filter out the lighting interference caused by normal physiological motion,such as head movements and facial expressions,and further obtain first-filtered BVP signals.Fourthly,the adaptive NLMS algorithm based on the mean value of the first-filtered BVP signal is introduced for adaptively filtering out the non-physiological signals caused by illumination changes from this BVP signal.Finally,to verify the feasibility and effectiveness of our method,we carry out the ablation experiments and comparison experiments between different algorithms on the UBFC-rPPG dataset and our dataset respectively.The results demonstrate that our proposed method outperforms several popular methods in the IPPG technique and solves the difficulty of accurate heart rate measurement under scenarios with large disturbances.
biotechnologyimaging photoplethysmographyconcave lens deformationK-means++clusteringnormalized least mean square algorithm
NETL
NSTL
万方数据
