Professional Documents
Culture Documents
Detected Pulses From The Motion of The Head To Estimate Heartbeat Rate
Detected Pulses From The Motion of The Head To Estimate Heartbeat Rate
Detected Pulses From The Motion of The Head To Estimate Heartbeat Rate
Abstract—The beating of the heart sends blood throughout with not only timely samples but also long-term trends and
the body to feed the cells and puts pressure on the blood vessel statistical analyses for a more comprehensive assessment of
walls. Therefore, monitoring this process is extremely important the treatment process. The non-invasive measurement method
in the diagnosis and treatment of the disease. In the past, people is highly applicable because it is non-invasive and non-
did this by counting pulses of artery or wearing of heart monitor obtrusive.
on the body. In both of these cases, physical contact with the
body is required to get the heart rate. Today, with the These contactless methods are largely based on computer
advancement of science and technology, we can overcome these vision techniques or algorithms using artificial intelligence.
limitations based on computer vision techniques. This paper Among them, one type can be mentioned is
proposes one such method - measuring heartbeat rate through photoplethysmography (PPG). Its basic principle is to use red
frequent pulses from the motion of blood vessel on the head. We or infrared light to shine on the body, where there are dense
extract them from a regular computer's camera in real-time. vascular areas such as the face, hands, etc. Then, the system
Our method identifies and records the oscillation of the feature will pick up variations in reflected light and analyzed for heart
points. Then, using the algorithm principal component analysis rate parameters. Overall, it meets the requirements of a non-
decays their trajectories into a group of component motions. contact heart rate measurement system. However, it requires
Thus, we can determine the best component corresponding to the use of expensive specialized light sources to improve
heartbeats based on its frequency. In order to do that, we built accuracy. At the same time, PPG is very sensitive to
an electrocardiogram and analyzed it to identify the peaks of the
movement. To overcome the above disadvantages, another
trajectory, which corresponded to the heartbeats. Based on that
data, we calculated the subject's heartbeat rate. When evaluated
method is proposed based on the hypothesis of the periodicity
on 10 subjects, our approach reports heartbeat rate roughly the of the blood circulation system. Each time blood is pumped
same as a conventional sphygmomanometer. This result shows into the arteries to the organs, a subtle change is caused. It
that the proposed method is correct, robust, and economically could be a change in skin color or the location of blood
suitable. However, in order to achieve high technical efficiency vessels. These changes are usually invisible to the naked eye,
without changing the economy, this method needs to be further but they can be viewed with techniques such as Eulerian video
improved in the future. magnification. The resulting periodic changes are then used to
measure cardiovascular indices.
Keywords— Electrocardiogram, Heartbeat rate, Head motion
detection, Trajectory, Feature point tracker, Principle component In this paper, we use microscopic head vascular
analysis, Discrete cosine transforms, Photoplethysmograph. oscillations related to the cardiac cycle as above. However, in
order to extract heart rate information at a low cost without
I. INTRODUCTION compromising accuracy, we use the camera of an ordinary
Heartbeat rate is obviously a vital sign of the human body's computer and calculate the measurement in real - time.. Our
activity and its measurement is of great importance in many algorithm detects and tracks the pulse from this movement.
applications, such as fitness assessment, training programs, From a biomechanical point of view, due to the spine as a
and especially in medical diagnosis. For example, in the pillar, the vertical axis of the head is less likely to have large
medical diagnosis of an operation, heartbeat rate, along with amplitude movements that affect the oscillations related to the
other parameters, is used as a crucial sign that helps to assess heart rate. Meanwhile, the horizontal axis tends to capture
the condition of the cardiovascular system. In this example, it most of the swaying of the dynamic balance. Therefore, the
can be used also for ensuring the safety of the process. If vertical direction is the best axis to measure the pulses needed
heartbeat rate goes beyond the normal range, it is no longer for heart rate analysis. However, besides blood circulation,
safe to continue surgery without further intervention. Not only respiration is another source of involuntary head movements.
that, but heartbeat rate also has a particularly important Thus, after monitoring the head oscillations, we proceed to
meaning in health monitoring, especially for premature filter out the high frequencies due to the respiratory activity
neonates and the elderly. Because they have delicate skin and and the subject's intentional movement. These pulses are then
are vulnerable to the traditional sensors, a non-invasive passed through several other algorithms to increase accuracy.
heartbeat rate extraction method is more suitable. In it, we used the Discrete Cosine Transform (DCT). About
Furthermore, as the population ages, continuous or at least DCT, we will detail it in the following section. After obtaining
frequent monitoring outside the clinic can provide doctors an electrocardiogram (ECG), the heart rate is calculated.
Although it has not yet reached absolute accuracy, the • Using an additional moving average filter before the
results achieved are very positive and there is a lot of potential Butterworth filter that is employed in Balakrishnan
for development in the coming time. The method that we give [2].
above in addition to the technical advantages mentioned • Replacing the FFT of Balakrishnan's system with
above, the cost of the system is also relatively cheaper than
DCT.
other types available on the market.
• Using webcam in laptop instead of high quality
The remainder of this paper is organized as follows: camera.
Comparisons and comments of other related works are given To build our system, in addition, we also refer to a few
in the next section. Section 3 explains the used hardware related pieces of research such as rapid object detection using
structure of the proposed system. Section 4 presents some of
a boosted cascade of simple features by Viola [3], pyramidal
the main algorithms of our method. Experimental results are
reported in Section 5. Finally, the conclusion of this paper is implementation of the Lucas Kanade feature tracker by
presented in Section 6. Bouguet [4],…
893
2021 International Conference on Decision Aid Sciences and Application (DASA)
So, from the detected face, our system localizes regions of with the clearest main frequency. The average pulse rate is
interest and from there, the feature points are extracted and determined using this frequency. For more detailed analysis
tracked by an algorithm like Lucas Kanade of Bouguet [4]. and beat duration calculation, we perform peak detection in
The most stable parts of the face, most resistant to changes in the time domain. More details about this PCA algorithm will
facial expressions, are the forehead and cheek areas around the be presented in the next section.
nose. To keep these regions, we first kept 60% of the width
and 80% of the height (obtained experimentally in proportion Of the components found by PCA, the most periodic
to the human face) of the region detected by Viola's face component belonged to heart rate. To find this most periodic
detector. Then in this refined area, we remove the eye area by one, our proposed system applies DCT to all components
removing all pixels that are within 25% to 45% of the height obtained by the PCA. As mentioned above, instead of using
(obtained experimentally in proportion to the human face) of the FFT algorithm to calculate the percentage of the signal's
the refined area (Figure 2). total spectral power thanks to the frequency of the peak power
and its first harmonic, we improve by using the DCT
algorithm. After performing this substitution, our system
gives more stable and accurate results than Balakrishnan's
system.
The processed signal can be viewed as an
electrocardiogram, from which we can calculate heart rate.
IV. ALGORITHMS
A. Trajectory Generation and Smoothing
As mentioned in the previous section, we have selected
the appropriate area of the face that is related to the heart rate.
Fig. 2. The green box is returned by Viola’s face detector and the red boxes
are those that are of interest to the proposed system.
Our aim in this step is to extract the oscillation pulses in that
region which are recorded from the feature points as input
After that, our system applies a moving average and a data for our system. To do this, we use the algorithm of Lucas
bandpass filter to the vertical component of the trajectory of Kanade of Bouguet [4]. It helps us to capture the necessary
each feature point to remove irrelevant frequencies and data in each frame. However, not all frequencies of the
involuntary head motion. We found that the head movements trajectories are required or useful for pulse detection. The
associated with cardiac activity were very small and mixed resting heart rate of a normal adult is between 45 and 120
with a variety of other involuntary head movements. Based on beats/min in other words, between 0.75 and 2 Hz. Based on
the structure of the human body, the head-neck and trunk
that, we conducted a test of the data in different frequency
system can be thought of as a sequence of stacked inverted
pendulums. This construction allows for unrestricted head ranges. Finally, we have found that the frequency range from
movement in most axes. There are several sources of 0.75 to 2 Hz gives a much more accurate result. However,
involuntary head movement that complicate the isolation of with the small frequency range, the data we have is not much
movements attributable to the pulsatile activity. The first is the and affects the periodicity of the heart rate. Therefore, after
pendulum oscillatory motion that keeps the head in dynamic considering harmonics higher than 2 Hz, we decided to filter
equilibrium. Thus, the vertical direction is the best axis to the signal from 0.75 to 5 Hz because in this range the result
measure the movement of the upright caused by the pulsing is more stable with relative accuracy. This is done using an
head which is suitable for the purposes of our research because 8th order Butterworth filter.
the horizontal axis tends to capture most of the dynamic
equilibrium swaying. The second source of involuntary head B. Principle component analysis (PCA)
movement is respiratory bobbing. That's why we filter out The primary source signal of interest is the head
low-frequency motion. The actual acceleration of the movement caused by the cardiovascular pulse. The
involuntary vertical head movement measured about 0.098 characteristic point trajectories are combinations of this as
m/s2. Meanwhile, the typical duration of the left ventricular
well as other movements caused by sources such as
ejection time of a heart cycle is approximately 1/3 seconds.
Using these numbers, we can calculate a rough estimate of the respiration, vestibular activity, and changes in facial
displacement of the head to be about: expression. While we only need information related to the
heart rate, we decompose this superimposed motion into sub-
signals to isolate pulse. For the above, we treat each facial
point in each frame as a separate data point and use PCA to
Although this calculation ignores the complex structure of select for altered positions that are cardiovascular-specific
the head system, it provides an indication of how small the components. Next, we choose a dimension to project the
motion is. position time series to obtain the pulse signal. Analytically,
In the next step, the filtered trajectories are fed to PCA to given N features, we represent the N-dimensional position of
find the strongest independent components. The PCA method the head at frame t as mt = [y1(t), y2(t), · · ·, yN(t)]. The mean
allows us to isolate the motion corresponding to the pulse and and the covariance matrix of those positions are:
project it onto a 1D signal allowing us to extract individual
beat boundaries from the peaks of the trajectory. To select the
correct source for analysis and calculate the duration of each 1
� = |T| × ∑Ti=1 mi
m (2)
beat, we check the frequency spectrum and choose the source
894
2021 International Conference on Decision Aid Sciences and Application (DASA)
1
Σm = × ∑Ti=1(mt − m
� )T+1 (3) • Si with the largest Qi is the heartbeat rate signal, and
T
the heartbeat rate can be obtained as:
PCA algorithm finds the principal axes of the
transformation of the position as the eigenvalues of the FFT(IDCT (min{Khi})) × 60 bpm (7)
covariance matrix: V. EXPERIMENTAL RESULTS
895
2021 International Conference on Decision Aid Sciences and Application (DASA)
896