Advertisement

Automated Heart Rate Measurement Using Wavelet Analysis of Face Video Sequences

  • Amruta V. MoreEmail author
  • Asmita Wakankar
  • Jayanand P. Gawande
Conference paper
Part of the Lecture Notes in Networks and Systems book series (LNNS, volume 33)

Abstract

To overcome the drawbacks of the conventional heart rate measurement method, a new approach is developed to measure cardiac pulse automatically using video imaging technique and wavelet analysis. In this paper, the color video images of the human face are used for detection of cardiac pulses. The specific region of interest (ROI) in face image is detected to obtain red, green, and blue intensity signals. Next normalized red, green, and blue intensity signals are decomposed using discrete wavelet transform (DWT) to obtain approximate and detail coefficients. Then, the specific frequency band from decomposed signal is obtained with the help of bandpass filter using Hamming window function. The cardiac pulse is measured with the help of pulse frequency in power density spectrum of filtered signal. The cardiac pulse measured with help of this system is compared with heart rate measured from reference ECG signal of the same object. This technique improves the accuracy from 73.14 to 89.86% if forehead of the subject is considered instead of face.

Keywords

Cardiac pulse Discrete wavelet transform (DWT) Face detection Power spectrum density (PSD) 

1 Introduction

The continuous assessment of cardiovascular function in human being is necessary for the diagnosis of chronic diseases. The measurement of heart rate is one of the simplest methods to identify the risk factor in cardiovascular diseases. The electrocardiogram (ECG) is the conventional technique to measure the cardiac pulse. In ECG, the electrodes are attached to the arms, legs, and chest of the patient with the help of adhesive gel which is not comfortable and can cause skin irritation. So, it has become a need to measure the cardiovascular pulse of patient so that the system will provide suitable comfort level to patient and system will give more accurate heart rate. So new system is proposed which is a noncontact type automated system and it accurately measures the heart rate of patient [1, 2].

Poh et al. [3] have developed a new methodology that overcomes problems of motion artifacts by applying independent component analysis on recorded video signals and a finger blood volume pulse (BVP) sensor is used to correlate the cardiac pulse signals from the face video. Lewandowska et al. [4] explained algorithm which measures pulse rate directly from face image captured from webcam by using principal component analysis. After some enhancement, the proposed technique helps in monitoring person at home. Gault and Farag [5] explained the noninvasive and measurement methods for cardiac pulse measurement from human face using thermal IR camera. Heart rate is calculated from 512 video frames with the help of wavelet analysis, vascular mapping, and blood perfusion modeling.

In this paper, we present automated and noncontact type method which measures cardiac pulse from video images based on wavelet analysis. Face video is recorded to capture subsequent changes in the amount of reflected light for cardiovascular pulse measurement. When ambient light is incident on face, some amount of light is absorbed and some amount of light is reflected. Subsequent changes in amount of reflected light related to every cardiac cycle can be used for cardiovascular pulse measurement. The rest of the paper is organized as follows: Sect. 2 describes automatic cardiovascular pulse detection algorithm is explained in detail. The flow diagram represents the step of the algorithm. Experimental setup, which is followed during experimentation and experimental results showing results of face detection and different RGB traces, is given in Sect. 3. Finally, Sect. 4 presents the conclusion.

2 Wavelet-Based Cardiac Pulse Detection

For every cardiac cycle, blood volume changes in the facial blood vessels. Cardiovascular pulses are measured with the help of reflected light when ambient light is incident on face [3]. Based on this principle, new system is developed for the cardiovascular pulse measurement which is noncontact type automated system. Figure 1 shows the flow diagram of cardiovascular pulse measurement algorithm. The system consists of basic webcam (Built-in camera on HP 15 ab-032tx) to record the video for analysis. All videos are recorded at 30 frames per second (fps) with pixel resolution of 640 × 480. Video is captured in color format for 60 s and saved in WMV format.
Fig. 1

Flow diagram of cardiovascular pulse measurement algorithm

Voila and Jones algorithm is used for face detection which detects faces within the video frames and highlights the face region for each video frame. For each detected face, the algorithm gives x and y coordinates of the highlighted region along with its height and width which defines box and considers it as a region of interest (ROI) for the subsequent calculation.

The RGB channels are obtained from the region of interest. Spatial average of all pixels present in the ROI of RGB channels is obtained to get raw RGB traces which are denoted by X1(t), X2(t), and X3(t).

The normalization is the transformation of Xi (t) to \( X_{i}^{\prime } \left( t \right) \) to obtain normalized RGB signal having zero mean and unit variance.
$$ X^{\prime } \left( t \right) = \frac{1}{{\sigma_{i} }}\left[ {X_{i} \left( t \right) - \mu_{i} } \right] $$
(1)
for each i = 1,2,3 where σi and µi are the standard deviation and mean of Xi(t), respectively. The normalized RGB signals are analyzed using two-channel wavelet filter bank consist of low pass filter (H0) and high pass filter (H1) of analysis filter bank. The analysis scaling and wavelet function are given by the two-channel filter bank [6, 7] as follows:
$$ \emptyset \left( t \right) = \frac{2}{{H_{0} \left( \omega \right)|_{\omega = 0} }}\mathop \sum \limits_{n} h_{0} \left( n \right) \emptyset \left( {2t - n} \right) $$
(2)
$$ \varphi \left( t \right) = \frac{2}{{H_{1} \left( \omega \right)|_{\omega = 0} }}\mathop \sum \limits_{n} h_{1} \left( n \right) \emptyset \left( {2t - n} \right) $$
(3)
where h0(n) and h1(n) are the analysis low pass filter (LPF) and high pass filter (HPF) coefficients, respectively [6, 7].

Bandpass filter is used to pass certain ranges of frequencies and rejects frequencies outside that ranges. In this system, the performance limits is [0.75, 4] Hz which corresponds to [45, 240] bpm to provide a wide range of heart rate. Hence, BPF with hamming window function is used to provide all frequencies present in the range of 0.75–4 Hz [8]. The power density spectrum (PDS) of decomposed signal is used to obtain all contained peaks around that frequency to calculate normalized pulse frequency (fp). Hence, periodogram is calculated to obtain normalized pulse frequency (fp) [8]. While capturing video, ECG signal of the same person is also obtained from the PowerLab software. PDS of obtained ECG signal is calculated to obtain reference pulse frequency for heart rate calculation from reference ECG signal. This reference heart rate is used to calculate proposed system’s accuracy.

3 Experimental Results

The experimentation is performed inside a room with changing amount of sunlight as the only illumination source. Ten volunteers are participated in the experimentation as per ethical standards. Subject is seated in a chair in front of a laptop, approximately 0.5 m away from the webcam. Video of one minute is recorded for all subjects. From each frame of recorded face video, center 60% width and full height of box is considered to obtain region of interest (ROI). The RGB channels are obtained from the region of interest. Spatial average of all pixels present in the ROI of RGB channels are obtained to get raw RGB Trace. These raw RGB traces are normalized by following Normalized Equation, i.e., Eq. (1). Normalized RGB signal is decomposed using the Daubechies (db1) wavelets into different levels by following traditional convolution method of 1D DWT decomposition. Approximate coefficient of second level decomposed signals is as shown in Fig. 2, which are considered for bandpass filtering. Figure 3 shows filtered decomposed signal.
Fig. 2

Second level decomposition of normalized RGB trace of forehead

Fig. 3

Filtered RGB signal of forehead

Normalized pulse frequency is obtained from the power density spectrum as shown in Fig. 4. Heart rate is obtained from reference ECG Signal as well as video signal. Both answers are compared to calculate the absolute error of this system. Heart rates of different subjects by considering face as well as forehead are calculated and compared to reference heart rate to calculate accuracy as shown in Table 1.
Fig. 4

PSD of filtered RGB signal of forehead

Table 1

Heart rate calculation from subject’s face and forehead

 

Heart rate (bpm)

MAE (%)

Subject

Face image

ROI forehead

Reference ECG

Face

Forehead

Subject 1

63

91

87

27.59

4.59

Subject 2

63

70

84

25.00

16.66

Subject 3

120

77

93

29.03

17.20

Subject 4

84

84

102

17.64

17.64

Subject 5

70

84

84

16.67

00.00

Subject 6

98

70

82

16.32

14.63

Subject 7

54

70

72

25.00

2.77

Subject 8

77

77

92

16.31

16.30

Subject 9

105

63

60

75.00

5.00

Subject 10

90

70

75

20.00

6.67

Mean absolute percentage error

26.86

10.14

Hence from Table 1, we can observe that the error is 26.86% which can be reduced if we consider the forehead of person instead of face. Hence after face detection, center 60% width and full height of box are considered to obtain the new region of interest (ROI). Only forehead is obtained from calculated region of interest.

4 Conclusion

After face video recording, face and forehead are detected from each frame of video. After separation of ROI of detected face into RGB channels, raw RGB trace is obtained which shows spatial average over of all pixels in the ROI and they are normalized. Wavelet decomposition is done up to second level to obtain approximate coefficient. Bandpass filter using hamming window helps to pass frequencies present in operational range and reject all out of the range frequencies. Power density spectrum gives pulse frequency which is useful in the calculation of heart rate. As compared to heart rate obtained from the reference signal, this system improves accuracy from 73.14 to 89.86% if forehead is considered instead of face to calculate heart rate. In order to improve the accuracy of this system, the various orthogonal and biorthogonal wavelets can be tested to estimate the heart rate from face video.

Notes

Acknowledgements

All authors are thankful to all the participants and the Principal Dr. M. B. Khambete of MKSSS Cummins College of Engineering for Women Karvenagar, Pune for the support.

Compliance with Ethical Standard

All procedures performed in this study involving human participants were in accordance with ethical standards as laid down in the 1964 Declaration of Helsinki and its later amendments or comparable ethical standards. Informed consent was obtained from all individual participants involved in the study.

References

  1. 1.
    Cromwell L, Weibell FJ, Pfeiffer EA (2017) Biomedical instrumentation and measurements, 2nd edn. PHI Learning Private Limited, DelhiGoogle Scholar
  2. 2.
    Carr JJ, Brown JM (2015) Introduction to biomedical equipment technology, 4th edn. Pearson PublicationGoogle Scholar
  3. 3.
    Poh MZ, McDuff DJ, Picard RW (2010) Noncontact, automated cardiac pulse measurements using video imaging and blind source separation. Opt Expr 18:10762–10774CrossRefGoogle Scholar
  4. 4.
    Lewandowska M, Ruminski J, Kocejko T (2011) Measuring pulse rate with a webcam—a non-contact method for evaluating cardiac activity. In: Proceedings of the federated conference on computer science and information systems, pp 405–410Google Scholar
  5. 5.
    Gault TR, Farag AA (2013) Fully automatic method to extract the heart rate from thermal video. In: Computer vision and pattern recognition workshops (CVPRW), IEEE conference, pp 2160–7516, Sept 2013Google Scholar
  6. 6.
    Gawande JP, Rahulkar AD, Holambe RS (2016) A new approach to design triplet halfband filter banks based on balanced-uncertainty optimization. Digit Signal Process Elsevier 56:123–133MathSciNetCrossRefGoogle Scholar
  7. 7.
    Gawande JP, Rahulkar AD, Holambe RS (2015) Design of new class of regular biorthogonal wavelet filter banks using generalized and hybrid lifting structures. Signal Image Video Process Springer 9:265–273CrossRefGoogle Scholar
  8. 8.
    Proakis JG, Manolakis DG (2012) Digital signal processing: principles, algorithms and applications, 4th edn. Pearsons Publication, 2012Google Scholar

Copyright information

© Springer Nature Singapore Pte Ltd. 2019

Authors and Affiliations

  • Amruta V. More
    • 1
    Email author
  • Asmita Wakankar
    • 1
  • Jayanand P. Gawande
    • 1
  1. 1.Instrumentation and Control DepartmentMKSSS’s Cummins College of Engineering for WomenPuneIndia

Personalised recommendations