Signal Processing - November 2016 - 31

vx m and used to discard strong but short baseline shift in vx m,
ensemble averaging, where the data length of each epoch
is 64 seconds. The data is divided using 50% overlap to
QI Std is the standard deviation ratio of vx m, in and vx m, and
reduce the loss of stability.
QI Std,m indicates how much the R-peak of vx m is separated
from
the noise around the R-peak. The quality index satura■ Apply the Hanning window to reduce linear trends in
tion QI sat is a binary-valued index used to detect samples,
the data.
in vx m, that belong to the upper or lower saturation region of
■ Calculate the power of LF components that is a parameter
of combined sympathetic and parasympathetic activities,
the ECG amplifier. The quality indices multiplication QI m is
the power of HF components that is a parameter of paracalculated by multiplying QI Amplitude, QI Std, and QI sat, and is
sympathetic activity, and the ratio LF/HF as a parameter of
used to remove outliers and to generate a training data before
sympathetic activity for each epoch.
applying to the principal component analysis (PCA). The
PCA and Hotelling's T-squared value [16] are used to calcu■ Calculate the natural logarithms of LF, HF, and LF/HF to
late the probabilistic distance from the QRS complex to the
approximate the distributions of LF, HF, and LF/HF to the
center of the training data, and to calculate
normal distribution.
the confidence measure (CM). A vector
Toyota attached a chest ECG detection
It is essential to design
with a small deviation to the training data
patch directly to the driver's chest [8], [19]
an analog signal
will result in a small Hotelling's T-squared
to compare the ECG obtained from the
processing circuit
value and, hence, is very likely to be a true
patch with that measured from the DSM
with a consideration of
QRS complex [16].
system. The mutual information technique
Since Ford's noncontact-based ECG
(MIT) [8], which measures the similarity
eMC to protect the DSM
acquisition
technique measures the ECGs
between two signals, is used to compare the
system from noise and
from multiple pairs of electrodes, the peaks
two ECGs. The value of MIT higher than
interference from
from different channels need to be aligned
0.047 indicates that two signals are strongly
the eCus.
before evaluating each QRS complex.
correlated, and in Toyota's study [19], the
Referring to the ANSI/AAMI EC 57 norm
MIT value of the two ECGs is much higher
[32], all R-peaks from the multiple [three in Figure 7(d)] chanthan 0.047; MIT values for HR, ln(LF), ln(HF), and ln(LF/HF)
nels within a time interval of ! 75 ms can be assumed one sinare 0.225, 0.223, 0.209, 0.184, respectively, with 95% confidence interval.
gle R-peak from the same heartbeat, and a missing R-peak in a
When the measured ECG includes too many R-peaks
channel is removed for the later processing. This leads to tripwith motion artifacts, it is impossible to continuously detect
lets of CMs based on Hotelling's T-squared values. The numthe R-peaks successfully. Therefore, in practice, PPG is
ber of CMs whose values are smaller than a specific threshold
acquired to detect the lost RR intervals in the ECG so that
determines if the R-peak is from a true QRS complex [16].
the missing RR intervals can be replaced with the PPG
In [7], Ford found the relationship between the driver's
peak intervals.
HR and driver workload from the real driving tests.
The detection rules subfunction analyzes the status of the
In the study, two men and two women (average age
driver using the ln(LF), ln(HF), and ln(LF/HF) calculated in
of 29.5 ! 2.08 years) are selected as the subjects, and the
the signal analysis subfunction and the HR obtained from RR
authors collected the ECGs while driving between Aachen
intervals [8], [19]. As explained in Figure 6(d), Toyota defines
and Brussels. The analysis of the driver's status is performed
the detection rule of abnormal physiological status based on
in both time and frequency domains. This study proves that
the HR, and demonstrates that the driver is drowsy if the LF/
the driver's HR and driver workload have a strong relationHF or HF increases steadily for more than 30 minutes.
ship, as shown in Figure 7(e), while the percentage of the
Ford introduces another DSM algorithm utilizing a signal
power in the LF segments, LFAR 6%@ , is not very related to
processing technique to remove outliers [16] and the HRV
the driver workload. In particular, the HRs of all subjects are
technique [7] to detect the driver workload. QRS complex is
about 75~86 beats/minute in the city, about 72~83 beats/minextracted from ECG by open source ECG analysis (OSEA)
ute on highway I, and about 68~84 beats/minute on highway
[32] as described in Figure 7(a), and the outliers of the extractII. The traffic congestion in the city results in a higher HR,
ed QRS complex are removed as explained in Figure 7(b)-
which indicates that the driver workload in the city is greater
(d). In the algorithm, a signal enhancement function using the
than that on the highway.
quality indices QI is applied [16]. The basic quality indices
used in the algorithm are quality index amplitude (QI Amplitude ),
EMC issues in DSM systems using
quality index standard deviation (QI Std ), and quality index
physiological signals
saturation (QI sat ), as defined in Figure 7(b), that are calcuAs the smart vehicles offer more driver-friendly functions,
lated using vx m and vx m, in , where vx m and vx m, in are the vecthe number of electrical control units (ECUs) is growing. As
a result, the number of problems caused by the electromagtors of the signal measurements within ! 200 milliseconds
netic waves from the ECUs is continuously increasing. Since
and ! 50 milliseconds around the prompt R-peak of the
the physiological signals, such as ECG, that is typically around
QRS complex, respectively. The quality index amplitude
QI Amplitude is the peak-to-peak amplitude ratio of vx m, in and
100nV~3mV are very sensitive to electromagnetic waves, it
IEEE SIgnal ProcESSIng MagazInE

November 2016

Table of Contents for the Digital Edition of Signal Processing - November 2016