Instrumentation & Measurement Magazine 23-4 - 59

Fig. 3. Acquisition platform used to generate the dataset: (a) block diagram; (b) side view photo.

participant. Subjects stayed at rest in normal conditions (they
held normal conversation with the team acquisition members) during 2 minutes. The dataset consists of 25 RGB videos
(five attempts per subject) and the HR references, and it may
be downloaded (see [13]). In this work, we used the wrist as
a body part to measure the HR, but other zones can be used
that include a high density of vessels, for example, the face
or the chest. The only condition is that the skin zones have to
be visible. For this work, we did not consider any unexpected
movements in the subject like tremors. In Fig. 3, we show the
acquisition platform used to generate the dataset, including a
block diagram and a side view photo. The camera was located
at a fixed distance of approximately 60 cm. We synchronized
the pulse monitoring device and the camera using their timestamps, previously set and validated.

Results and Discussion
In this section, we show the HR estimation results using the
proposed method and compared to the magnification method
proposed by Wu et al. [9]. In the magnification process, the
applied broad-band temporal filter has cutting frequencies
from 0.833-1.583 Hz (50-95 bpm) that allows us to cover a
sufficiently wide range. In all the experiments, we used a magnification factor of 50. On the other hand, the input DFT vector
size was 1024 samples.
Fig. 4 shows the HR estimation and the references for the
five subjects using the testing set. The dashed line corresponds
to the reference, the blue line to the Gaussian approach [9] and
the red line to the proposed Hermite approach.
Table 1 summarizes the performance of the NN using
both the Gaussian and the proposed Hermite magnification
approaches. It reports the root-mean-square error (RMSE) between the reference and the HR estimation per subject. It can
be shown that the Hermite-based magnification strategy performed better than the Gaussian magnification.
Observing the results, it is evident that our approach is
competitive with respect to the original Eulerian magnification method. The HR estimation curves track the reference
June 2020

appropriately. In consequence, the RMSE is bounded between
1.3 and 3.5 bpm.
Subsequently, to compare the performance of the Hermite
magnification approach against the Gaussian method, a robustness noise test was performed by adding Gaussian noise
with different power spectral values (σ = 5, 10, 15, 20, 25, 30) to
each original video for the two methods. This experiment was
carried out to simulate noisy environments. In addition, the
NN was trained using the original data and it was tested using
the noisy data, proving the robustness of the NN model. Fig. 5
shows the HR estimation and the references for subject 3.
Table 2 summarizes the performance of the NN using both
magnification approaches under noisy conditions. It reports
the RMSE between the reference and the HR estimation per
noise level. We can observe that RMSE increases accordingly
to the noise spectral power. The robustness of NN under noise,
as well known, depends on the input features. Hence, the extracted HT features might explain the improvement of our
approach.
Our results, on average (RMSE = 1.86 bpm), are coherent
with other results in the literature. Compared to other works
using the RMSE metric, the work of Poh et al. [6] obtained an
average RMSE of 2.29 bpm, the work of Monkaresi et al. [7] obtained an average RMSE of 3.64 bpm, and the work of Moya
et al. [8] obtained an average RMSE of 2.82 bpm. It should be
noted that our machine learning model (NN) is simpler and
had better results than the more complex convolutional neural
network proposed in [8].
One of the most critical steps in the magnification process
is the choice of the cutting temporal frequencies. In our approach, we avoid complementary steps to select the optimal
cutting frequencies (depending on the HR of the subject) by
fixing them in a wide range, leaving this task to the NN for detecting the adequate HR pulse. On the other hand, the Hermite
reconstruction is performed using the synthesis Hermite filters that reproduce the image information in a better way, in
contrast to the Gaussian approach that uses the nearest neighbor interpolation.

IEEE Instrumentation & Measurement Magazine 59

Instrumentation & Measurement Magazine 23-4

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 23-4