Detecting Stress During Real-World Driving Tasks Using

Physiological Sensors
Jennifer A. Healey, Rosalind W. Picard
Cambridge Research Laboratory
HP Laboratories Cambridge
HPL-2004-229
December 17, 2004*

driver, stress, This paper presents methods for collecting and analyzing physiological
traffic, automobile, data during real world driving tasks to determine a driver's relative stress
physiology, sensor, level. Electrocardiogram, electromyogram, skin conductance and
signal, computer, respiration were recorded continuously while drivers followed a set route
affect, EKG, skin through open roads in the greater Boston area. Data from twenty- four
conductivity drives of at least fifty minute duration were collected for analysis. In
Analysis I features from five minute intervals of data were used to
distinguish three levels of driver stress with an accuracy of over 97%
across multiple drivers and driving days. In Analysis II, continuous
physiological features were correlated with a continuous metric of
observable stressors showing that on a real-time basis metrics of skin
conductivity and heart rate were most closely correlated with driver stress
level. Such automatically calculated physiological features could be used
to help manage non-critical in-vehicle information systems and improve
the driving experience.

* Internal Accession Date Only

To be published in IEEE Transactions on Intelligent Transportation Systems
Approved for External Publication
 Copyright IEEE
1 Introduction
The increasing use of on-board electronics and in-vehicle information systems has made the
evaluation of driver task demand an area of increasing importance to both government and
industry[1] and understanding driver frustration has been listed by international research groups
as one of the key areas for improving intelligent transportation systems[2]. Protocols to measure
driver workload have been developed using eye glance and on-road metrics, but these have been
criticized as too costly and difficult to obtain [3], and uniform heuristics such as the 15-Second
Rule for Total Task Time, designed to provide an upper limit for the total time allowed for
completing a navigation system task, do not provide flexibility to account for changes in the
driver’s environment [3]. As an alternative, this study shows how physiological sensors can be
used to obtain electronic signals that can be processed automatically by an on-board computer
to give dynamic indications of a driver’s internal state under natural driving conditions. Such
metrics have been proposed for fighter pilots[4] and have been used in simulations[5], but have
not been tested on stress levels approximating a normal daily commute using sensors that do
not obstruct drivers’ perception of the road.
This experiment was designed to monitor drivers’ physiologic reactions during real-world driv-
ing situations under normal conditions. Performing an experiment in real traffic situations en-
sures that the results will be more directly applicable to use in these situations; however it im-
poses constraints on the kinds of sensors that can be used and the degree to which experimental
conditions can be controlled. Within these constraints, two types of analysis were performed
on the collected signals. Analysis I was designed to recognize three general stress levels: low,
medium, and high using five minute intervals of data from well defined segments of rest, city
and highway driving. For this analysis, features from all sensors were combined using a pattern
recognition technique and the different types of segments were recognized. Analysis II was de-
signed to give a more detailed account of how individual physiological features vary with driver
stress at each second of the drive, including those segments of the drive between the rest, city
and highway segments. For this analysis a continuous metric of observed stressors was created
by scoring video tapes from individual drives. This metric was then correlated with features
derived from each of the sensors on a continuous basis.
Historically, stress has been defined as a reaction from a calm state to an excited state for the
purpose of preserving the integrity of the organism. For an organism as highly developed and
independent of the natural environment as socialized man, most stressors are intellectual, emo-
tional and perceptual[6]. Some researchers make a distinction between “eustress” and “distress,”
where eustress is a good stress, such as joy, or a stress leading to an eventual state which is more
beneficial to the organism[7], however in this paper we will refer to stress only as distress, stress
with a negative bias, particularly distress caused by an increase in driver workload. There have
been a number of studies that link highly aroused stress states with impaired decision making
capabilities[8], decreased situational awareness[9] and degraded performance[10] which could
impair driving ability.
This paper presents a method for measuring stress using physiological signals. Physiological
signals are a useful metric for providing feedback about a driver’s state because they can be col-
lected continuously and without interfering with the driver’s task performance. This information

1
could then be used automatically by adaptive systems in various ways to help the driver better
cope with stress. Some examples of this might include automatic management of non-critical
in-vehicle information systems such as radios, cell phones and on-board navigation aids[2].
During high stress situations cell phone calls could be diverted to voice mail and navigation
systems be programmed to present the driver with only the most critical information to help
reduce driver workload. In addition, the music selection agent agent might lower the volume,
or offer a greater selection of relaxing tunes to help the driver cope with their feelings of stress.
Conversely, in low stress situations, the car might recognize that more driver distractions could
be tolerated and provide the driver with more entertainment options.
The recognition algorithm presented in Analysis I could be run in real time by having the on-
board computer keep a continuously updated record of the data from the last five minutes of
the drive in memory and performing the analysis continuously on this window of data. Al-
though none of the physiological signals monitored here react quickly enough to contribute to
automatic vehicle control, this kind of continuous monitoring, with a one to three minute lag in
driver state assessment, is fast enough to initiate customized changes to the driver’s in-vehicle
environment to help mitigate emotional distress. For example in high stress situations, some
users might prefer visual navigation prompts to turn off or dim, since these types of warnings
have been found to have a negative impact on situational awareness[9]. Alternatively, if intelli-
gent collision avoidance were safely available in low velocity traffic jams, driving could become
completely automated in such situations and a frustrated driver could relax by watching a movie
or by working on their laptop.
A real time implementation would have been difficult to test on this driving route because the
stress levels for the driving conditions outside of the rest, city and highway segments was not
well defined by the design. To better assess the stress conditions of the entire drive, Analysis
II looked at sixteen drives individually and created a continuous record of observable stressors
from video tapes of the entire drive. This analysis also calculated continuous variables for each
of the sensors and compared them to a continuous metric stress indicators scored throughout
the entire drive. These variables were evaluated to determine which features provided the best
single continuous indicator of driver stress. In new concept cars, such as the Toyota Pod car,
continuous signals that correlate highly with stress level could be used to control the expressive
changes in the cars lights and color[11], perhaps alerting others to the extra load on that driver.
Furthermore, using aggregate continuous records of driver stress over a common commuting
path, city planners could help quantify the emotional toll of traffic “trouble spots” which could
help prioritize road improvements.

2 Driving Protocol
The driving protocol consisted of a set path through over 20 miles of open roads in the greater
Boston area and a set of instructions for drivers to follow. Although stressful events could not
be specifically controlled on the open road, the route was planned to take the driver through
situations where different levels of stress were likely to occur, specifically, the drive included
periods of rest, highway and city driving that were assumed to produce low, medium and high
levels of stress. These assumptions were validated by two methods: a driver questionnaire and a

2
score derived from observable events and actions coded from video tape taken during the drives.
The route was designed to reflect a typical daily commute so that the recorded stress reactions
would all be within the range of normal daily stress.
To participate in the experiment, drivers were required to have a valid driver’s license and to
consent to having video and the physiological signals recorded during the drive. Before begin-
ning, drivers were shown a map of the driving route and given instructions designed to keep
the drives consistent, for example, instructions were given to obey speed limits and not to listen
to the radio. During the drive, an observer accompanied the driver in the car to answer any
of the driver’s questions, to monitor physiological signal integrity and to mark driving events
in the video record. The observer sat in the rear seat diagonally in back of the driver to avoid
interfering with the drivers’ natural behavior.
All drives were conducted in mid-morning or mid-afternoon when there was only light traffic
on the highway. Two fifteen-minute rest periods occurred at the beginning and end of the drive.
During these periods the driver sat in the garage with eyes closed and with the car in idle. The
rest periods were used to gather baseline measurements and to create a low stress situation.
After the first rest period, drivers exited the garage through a narrow, winding ramp and drove
through side streets until they reached a busy main street in the city. This main street was
included to provide a high stress situation where the drivers encountered stop and go traffic
and had to contend with unexpected hazards such as cyclists and jaywalking pedestrians. The
route then led drivers away from the city, over a bridge and onto a highway. Between a toll
at the on-ramp and a toll preceding the specified off-ramp, drivers experienced uninterrupted
highway driving. This driving was included to create a medium stress condition. After the
exit toll, drivers followed the off-ramp to a turn around and re-entered the highway heading in
the opposite direction. After exiting the highway, the drivers returned through the city, down
the same busy main street and back to the starting point. The relative duration of these events
can be seen in Figure 3. The total duration of the drive, including rest periods, varied from
approximately fifty minutes to an hour and a half, depending on traffic conditions. Immediately
after each drive, subjects were asked to fill out the subjective ratings questionnaires.

2.1 Data Collection

Four types of physiological sensors were used during the experiment: electrocardiogram (EKG),
electromyogram (EMG), skin conductivity (also known as EDA, electro-dermal activation and
GSR galvanic skin response) and respiration (through chest cavity expansion). These sensors
were connected to a FlexComp[12] analog to digital converter which kept the subject optically
isolated from the power supply. The FlexComp unit was connected to an embedded computer
in a modified Volvo S70 series station wagon. The EKG electrodes were placed in a modified
lead II configuration to minimize motion artifacts and maximize the amplitude of the R-waves
since both the heart rate[13] and heart rate variability[14][15] algorithms used in this analysis
depend on R-wave peak detection. The EMG was placed on the trapezius (shoulder), which
has been used as an indicator of emotional stress[16]. The skin conductance was measured
in two locations: on the palm of the left hand using electrodes placed on the first and middle
finger and on the sole of the left foot using electrodes placed at each end of the arch of the foot.

3
Respiration was measured through chest cavity expansion using an elastic Hall effect sensor
strapped around the driver’s diaphragm. Figure 1 shows the general placement of sensors with
respect to the automotive system.
The physiologic monitoring sensors were chosen based on measures previously recorded in real
world driving and flight experiments. Helander (1978)[17] used an electrocardiogram (EKG),
skin conductivity and two EMG sensors to monitor drivers on rural roads. Heart rate and skin
conductance have been used to monitor task demand on pilots [18] [19] [20] [21] as have EMG
[20] and respiration[5] [20]. EMG [16] and skin conductivity [22] and heart rate variability[23]
have also been studied as a general indicators of stress.
Each signal was sampled at a rate appropriate for capturing the information contained in the
signal constrained by the sampling rates available on the FlexComp system. The EKG was
sampled at 496 Hz, the skin conductivity and respiration sensor were sampled at 31 Hz and
the EMG was sampled at 15.5 Hz after first passing through a 0.5 second averaging filter. The
signals were collected by an embedded computer in a modified car. The experimenter visually
monitored the physiological signals as they were collected using a laptop PC running a remote
display program. The video output from this laptop, displaying the physiological signals was
fed into a quad splitter to create a composite video record together with the video output from
three digital cameras: a small Elmo camera mounted on the steering wheel, a Sony digital video
camera with a wide angle (0.42) lens mounted on the dashboard and a third camera used for
event. This record was used to create the continuous stress metric. A sample frame from one of
the composite video records is shown in Figure 2.
Figure 3 shows an example of the signals collected on a typical day’s drive along with markings
showing driving periods and events. In total, 27 drives were completed, six by drivers who
completed the course only once and seven each from three drivers who repeated the course on
multiple days. In the first analysis, 24 complete data sets were used. Of the initial 27, one data
set was incomplete because the hand skin conductivity sensor fell off, one data set could not
be used because the EKG signal was too noisy to extract the R-R intervals necessary for the
heart rate and heart rate variability metrics and one data set was lost because it was accidentally
overwritten. In the second analysis all 16 drives were used for which video records were created
(see Section V).

3 Questionnaire Analysis
The questionnaire analysis was used to validate a perception of low, medium and high stress
during the rest, highway and city driving periods. Two kinds of ratings were used: a free scale
and a forced ranking of events. The free scale section asked drivers to rate driving events on a
scale of “1” to “5” where a rating of “1” was used to represent a feeling of “no stress” and a
“5” was used to represent a feeling of “high stress.” The forced scale section required drivers to
rank events on a scale of 1 to 7 where “1” was assigned to the least stressful driving event and
“7” to the most stressful driving event. Using this scale, drivers were asked to rate a number
of events including encountering toll booths, merging and exiting as well as the rest, city and
highway driving tasks. The extra categories were used to help drivers define the scale, but were
not used in the questionnaire analysis.

4
Table 1: The Overall and Comparative questionnaire rating results after using a z-score and
back transformation. The results of ANOVA analysis found the three states to be significantly
different at the 95% confidence level with p¿0.001 for both the ratings
Condition Overall Comparative
rating (1-5) rating (1-7)
µ σ µ σ
Rest 1.16 0.88 0.81 1.68
Highway 2.00 0.92 2.69 1.50
City 2.55 1.02 4.01 1.56

For each questionnaire, the values for the both stress ratings were normalized using a z-score
(z = x−µσ
)[24], then the average and standard deviation were calculated and back-transformed.
The results, see Table 3, show that subjects found the rest periods to be the least stressful, the
highway driving to be more stressful and the city driving to be the most stressful. ANOVA anal-
ysis on the z-score transformed variables determined that the means were significantly different
at the 95% confidence level with p > 0.001 for both the Overall and Comparative ratings. These
results support the assumptions of the experimental design.

4 Video Coding
The composite video record of the drives were coded to help assess driver stress levels. Two
video coders scored each video tape record based on a list of observable actions and events that
might correspond to an increase in driver stress. This list of potential stress indicators included
including stops, turning, bumps in the road, head turning and gaze changes. The coders were
also allowed to use their judgment and score any number of additional events in an “other”
column. The two coders analyzed the video tapes by advancing them at one second intervals
and recording the number of stress indicators in each frame. For each drive, an average of
over 25,000 frames were scored. Due to time limitations, this process was only completed for
16 of the 24 drives. The two coders were not involved in other aspects of the analysis. To
test the inter-coder reliability Cronbach’s alpha[25] was calculated for a drive that was scored
independently by both coders. These results were α = 1.0 for the highway segments, α =.91
for the city segments and α =.97 for the highway segments. Since all a coefficient of .80 is
considered acceptable for most applications, these scores show that the rating system yielded
consistent results between coders.
To create a stress metric, the number of stress indicators was first summed over each second
of the drive. For example, if the driver was turning the steering wheel, changing gaze and
turning his or her body during a frame, that frame would get a score of “3”. If the driver was
driving straight and only looking around for a turn the frame would get a score of “1”. If no
stress indicator was observed the score was entered as “0.” The sum of stress indicators at each
second n of the drive was recorded in a time series Id (n) for each drive, d.
To further validate the assumption of low, medium and high stress conditions during the rest,

5
Table 2: The average number of stress indicators per minute during each of the three driving
conditions: rest, highway and city
Condition Stress Indicator per Minute
Rest 13.6
Highway 61.4
City 87.7

highway and city segments, the time series Id (n), were averaged over each type of segment for
all 16 drives, d, and divided by the time of each segment time, T , to obtain an estimate of the
number of stressors per minute for each type of driving. The results, shown in Table 4, support
the assumption of the design by showing that the greatest concentration of stress indicators
occurred during the city driving condition, followed by fewer stress indicators during highway
driving and the least during the rest conditions. As shown by the results, the rest conditions were
not completely free of stress. During these periods some drivers would display restlessness by
moving around, shifting position and reacting to noises from a nearby road. Some fidgeting may
also have come from the initial discomfort of wearing the sensors, boredom, or anticipation of
either the beginning or end of the experiment. In one case the driver needed to use the rest room
during the end rest period. The rest periods were not designed to keep the subject entirely free
from stress, but to provide a lower stress situation just as city driving was designed to provide a
higher stress situation.

5 Creating a Continuous Stress Metric

A continuous stress metric was created to develop a finer grain picture of the stressors encoun-
tered throughout the drives on various days. Although each drive contained thirty minutes of
driving within the rest, city and highway conditions, it also contained approximately forty min-
utes of driving under other conditions that were not well defined by the experimental design.
Unlike laboratory experiments where repeatable stress conditions can be created and controlled,
the real world driving conditions encountered in this experiment were largely unpredictable and
uncontrollable. The stress metric was designed to give a rough approximation of driver task
load by counting the number of stress indicators at each second of the drive and smoothing the
signal to incorporate the effect of anticipation and past events.
The video code scores captured a continuous record of all stress indicators that occurred through-
out the drive, reflecting individual differences in driver reactions and varying traffic conditions.
A continuous stress metric was developed from these scores to be correlated with each of the
time series of physiological features calculated for that drive. To create this metric each stressor
was convolved with a simple model of its assumed stress effect. The stress effect was modeled
as having both anticipatory and persistence effects. In a model for pilot workload, Sheridan
and Simpson identified several types of mental workload tasks that preceded each observed
task: operating tasks, monitoring tasks, and planning tasks. They modeled the effect of each
of these as a continuous workload function spanning a period of time between when the pilot

6
anticipated the task and when the task was completed[26]. This model implies that before a
stressor is observed there is an increase in driver stress due to anticipatory, monitoring and plan-
ning effects. In addition, the expected physiological effect of a stressor occurs slightly after the
stimulus and may take several seconds or several minutes to recover depending on the type of
stimulus event[27]. It is also known that physiological reactions add non-linearly and depend
on habituation effects and components of the individual’s physiology[27].
To precisely model the effect of each observed stressor, the anticipatory components of mental
workload and the expected persistence of the physiological effect would have to be individually
modeled for each observation, taking into account all previous and concurrent events and a
model of each driver’s physiology. Such a model would have been too complex for this analysis.
Instead, each observed event was modeled by using a 100 second Hanning window, H, centered
on the observation to approximate these effects.
The 100 second window was chosen for several reasons: it approximates the time needed for
autonomic signals such as the skin conductivity to extinguish, it is the same window as the
shortest window used for heart rate variability and it provides a level of smoothing that allows
the essentially discrete stressor metric to approximate a continuous signal.
This window was convolved with the metric of events for each drive I d to create a signal Vd that
represented the modeled effect of the stressors as stated in Equation 1.

Vd (n) = Id (n) ⊗ H (1)

For each of the 16 drives, d, the stress effect signal Vd (n) was correlated with each of the
physiological time series. The results are shown in Table 6.2 and discussed in the section
Analysis II.

6 Data Analysis
The collected data were subject to two types of analysis. Analysis I used five minute intervals
of data from well defined segments of the drive where drivers experienced low, medium and
high stress situations to train an automatic recognition algorithm. Analysis II investigated how
continuous physiological features, calculated at one second intervals throughout the entire drive,
correlated with a metric of driver stress derived from video tape records.

6.1 Analysis I: Recognizing General Stress Levels

The algorithm for general level stress recognition was developed using features derived from
five minute non-overlapping segments of data taken from each of the rest, city and highway
driving periods. Each of these segments was designed to represent a period of low, medium or
high stress. To ensure consistency in the stress conditions, the data segments were taken from
specific parts of the drive. The segments for the low stress condition were taken from the last
five minutes of the rest periods, giving subjects enough time to relax from the previous task. The
segments for the medium stress condition were taken from a stretch of uninterrupted highway
driving between two toll booths, after the driver had completed a merge onto the highway and

7
was safely in the right hand lane. The segments for the high stress condition were taken after
the driver turned onto a busy main street in the city.
Nine statistical features were calculated for each segment: the normalized mean of the EMG and
the normalized mean and variance for respiration, heart rate and skin conductivity on the hand
and on the foot. The EMG, respiration and heart rate signals were normalized by subtracting the
mean of the first rest period before each drive. The skin conductivity signals were normalized
by subtracting the baseline minimum and dividing by the baseline range[16]. Heart rate was
uniformly sampled and smoothed using a heart rate tachometer[13][28].
Four spectral power features were calculated from the respiration signal representing the en-
ergy in each of four bands. The power spectrum was calculated using 2048 data points from
the middle of each segment. A Hanning window was applied and an implementation of Welch’s
averaged, modified periodogram method[29] was used to calculated the normalized power spec-
trum. Four spectral power density features were calculated by summing the energy in the bands
0-0.1Hz, 0.1-0.2Hz, 0.2-0.3Hz and 0.3-0.4Hz. These features were found useful for discrimi-
nating emotion in previous work [30].
Eight additional skin conductivity features were calculated to characterize orienting responses.
An orienting response is a sudden rise in the skin conductance due to ionic filling of the skin’s
sweat glands in response to sympathetic nervous activation. A series of three orienting responses
is shown in Figure 4, along with the marks indicating the onset and peak of the response and
the measurements of the magnitude, OM , and duration, OD , of the response. The algorithm
detected the onsets and peaks of the orienting responses by first detecting slopes exceeding
a critical threshold and then finding the local minimum preceding that point (onset) and the
local maximum following that point (peak)[31]. Using this algorithm four orienting response
features were calculated: the total number of such responses in the segment, the sum of the
startle magnitudes ΣOM , the sum of the response durations ΣOD and a sum of the estimated
areas under the responses Σ( 21 OM ∗ OD ). These four features were calculated for both the hand
skin conductance and the foot skin conductance signals.
The final feature was a heart rate variability (HRV) feature which has been used to representing
sympathetic tone. The parasympathetic nervous system is able to modulate heart rate effectively
at all frequencies between 0 and 0.5 Hz, whereas the sympathetic system modulates heart rate
with significant gain only below 0.1Hz[32]. By taking the ratio of the low frequency heart
rate spectral energy to the high frequency heart rate spectral energy we derive a feature that
represents the ratio of the sympathetic to parasympathetic influence on the heart. Our hypothesis
is that increased stress will lead to an increase in sympathetic nervous activity and an increase
in this ratio.
To calculate the HRV feature, we used the instantaneous heart rate time series derived from
the EKG. A Lomb periodogram[15] was used to calculate the power spectrum[33][34] of the
heart rate time series because it can directly use unevenly sampled inter-beat interval data and
because it is robust to missed beats[35]. The total energy in the low frequency (LF ) band (0-
0.08 Hz) and in the high frequency (HF ) band (0.15Hz-0.5Hz) were calculated and the ratio
LF
HF
was used as the final feature. In Analysis II, another suggested sympatho-vagal balance
ratio, LFHF
+M F
, using the mid-frequency (MF) range (0.08Hz-0.15Hz) was also used along with
a shorter window size.

8
Table 3: The confusion matrix for the recognition algorithm. Correctly recognized segments
are found along the diagonal. This classifier mistakenly classified two medium stress segments
as high stress and one high stress segment as medium stress.
Recognition Results
Labeled As
Recognized as: Low Medium High Recognition Rate
Low 36 0 0 100%
Medium 0 36 1 94.7%
High 0 2 37 97.4%

These 22 features were used to create a single vector representing each of the segments used
in the recognition analysis. A total of 112 segments were used: 36 from rest periods, 38 from
highway driving and 38 from city driving. The resulting 112 feature vectors were then used
to train and test the recognition algorithm. Each vector was sequentially excluded from the
training set and the recognition algorithm was trained using the remaining 111 vectors. The
training vectors were used to create a Fisher projection matrix and a linear discriminant. The
Fisher projection was determined by solving a factorization for the generalized eigenvectors of
the covariance matrices for the between class scatter and the within class scatter of the labeled
training vectors[36]. The generalized eigenvectors corresponding to the two greatest eigenval-
ues were used to project the 22 dimensional feature vectors onto a two dimensional space where
the between class scatter was maximized and the within class scatter was minimized. Using
the projection determined by the training data the test vector y was projected into a two di-
mensional vector ŷ. In the two dimensional space, a linear discriminant function g c (ŷ) was
determined using the sample mean (mc ) and the a priori probability P r[wc ] for each class c and
pooled covariance K of the training vectors. The test vector was classified as belonging to the
class for which gc (ŷ) was the greatest.

−1 T −1
gc (ŷ) = 2mT
c K ŷ − mc K mc + 2ln(P r[wc ]); (2)

Table 6.1 is a confusion matrix for the recognition algorithm in which all correctly classified
segments are shown along the diagonal and all incorrectly classified segments are off diagonal.
As this table shows, all low stress segments were correctly recognized but two periods that were
labeled as medium stress were recognized as being high stress and one period labeled as high
stress was classified as medium stress. The results thus show very good discrimination between
the classes. These physiologically based results also show a perfect discrimination between the
low stress rest period and the two driving periods which agrees with both the perception of stress
as evaluated by the questionnaire and the scoring of observed stressors obtained from the video
tape analysis, suggesting that these features accurately represent a driver’s general stress level.

9
6.2 Analysis II: Continuous Correlations
The recognition algorithm gives good separation between three general types of driving stress,
but it does not account for variations in the drives and it does not give a fine grain assessment of
stressors. An ideal indicator of stress would be a physiological variable that continuously varied
proportional to every driver’s internal stress. To determine which features might be the best
candidates for such a variable, continuous calculations were made on each of the physiological
sensor signals at one second intervals throughout the entire drive for each of the sixteen drives
for which the video was scored. These calculations included the mean and variance of the EMG
(µE , σE2 ), hand skin conductivity (µS , σS2 ), respiration (µR , σR2 ) and the mean of the tachometer
heart rate (µH ) over one second intervals throughout the drive.
For this analysis four metrics of heart rate variability were calculated. In addition to the 300
LF
second window HF used in Analysis I, a 100 second window and a LFHF +M F
were also calculated
LF
for comparison. These time series are denoted: L100, M 100, L300 and M 300 for the HF
(L) and LFHF +M F
(M) power ratios in the 100 and 300 second periodograms respectively. To
create a continuous time series, Lomb periodograms were calculated using both 100 and 300
second windows (Hanning) of instantaneous heart rate data, centered on the second of interest,
advanced by one second for each second of the drive. The 150 seconds at the beginning and end
of the drive were excluded because there would not have been enough data for the periodogram.
For each of the drives d, the video stress metric Vd (n) was correlated with each of the feature
time series and a correlation coefficient rd was calculated:

KV P
rd = (3)
σV V ∗ σ P P
where KV P is the covariance of the time series Vd (n) with one of the physiological time series
for the same drive d and σV V and σP P are the standard deviations for Vd (n) and physiological
time series, respectively.
If the feature time series were independent of the stress metric, the correlation coefficient would
be zero. To test this null hypothesis, each of the stress metrics was also correlated with a white
noise signal w. Table 6.2 shows the results for each time series for all sixteen drives. As
expected, the correlation coefficients with white noise, w, were all close to zero. The variance
of the EMG, σE2 , and the mean of the respiration, µR , were also close to zero. This was also
expected, since the EMG signal was pre-processed with a smoothing filter and the respiration
mean primarily represents the baseline stretch of the sensor which varies mostly with sensor
2
movement (slippage) with respect to the chest cavity. The variance of the respiration, σ R , and
2
the variance of the skin conductivity, σG , also did not correlate well with the stress metric,
most likely because the variance over one second intervals in these signals has a large noise
component.
To determine which sensors might be most useful for use as a real time indicator of stress, the
averages of the correlation coefficients were calculated in two ways, first by calculating a z-
score for each day’s scores, averaging and then back transforming to get the result shown in row
“µ-zs” and second by using the normalizing z-transform, zd = 0.5 ∗ (ln(1 + rd ) − ln(1 − rd )),
and averaging to get the result shown in row “µ-zt.” The z-score transformed data is more likely
to be robust against a poor stress metric on a given drive and the z-transformed data creates

10
Table 4: Correlation coefficients “rd ” between the stress metric created from the video and
variables from the sensors indicating how closely the sensor feature varies with the stress metric.
As a null hypothesis, a set of random numbers, “w” was also correlated with the video metric
for each drive. The last rows show the mean over all days as calculated by using the z-score and
z-transform methods respectively.
Day L100 L300 M100 M300 HR µE σE2 µG σG2 µR σR 2
w
S1-2 .53 .61 .53 .64 .34 .22 .01 .75 .09 -.53 .04 .01
S1-3 .45 .45 .44 .42 .35 .04 .01 .77 .08 -.49 .04 .00
S1-4 .45 .58 .47 .60 .53 .14 .06 .71 .18 -.33 .26 .01
S1-5 .41 .35 .22 .09 .46 .30 .08 .85 .22 -.22 .15 .01
S1-6 .62 .62 .59 .62 .31 .32 .09 .74 .00 -.56 .16 .01
S1-7 .46 .36 .41 .31 .52 .28 .04 .77 .23 -.23 .16 .01
S2-2 .49 .66 .55 .69 .49 .02 .03 .13 .00 -.24 .15 -.01
S2-4 .22 .29 .13 .17 .41 .27 .01 .59 .12 .12 .18 .00
S3-2 .74 .73 .75 .74 .44 .20 .06 .78 .20 .17 .25 -.01
S3-4 .46 .41 .48 .48 .38 .16 .06 .77 .15 .59 .19 .01
S3-5 .41 .51 .44 .50 .35 .09 .00 .81 .20 .21 .01 -.02
S3-6 .44 .53 .44 .51 .40 .20 .04 .73 .14 .67 .24 .03
S3-7 .35 .35 .39 .35 .29 .22 .08 .78 .16 .44 .12 -.01
R2-1 .41 .58 .39 .54 .30 .20 .06 .47 .06 .10 .03 .00
R3-1 .32 .42 .35 .41 .30 .16 .13 .45 .08 .03 .10 .01
R4-1 .49 .55 -.08 -.19 .76 .37 .09 -.07 .03 -.28 .22 -.03
µ-zs .52 .60 .49 .57 .48 .17 .03 .99 .08 -.42 .10 -.01
µ-zt .50 .56 .45 .49 .45 .20 .05 .81 .12 -.03 .15 .00

11
a more normal distribution of the data, which may give a better estimate of the true mean.
Both transformations yield similar results suggesting that skin conductivity is the best real-time
correlate of stress followed by the heart rate variability and heart rate measures. In general, the
skin conductance performed well (with the notable exceptions of drives S2-2 and R4-1) and the
heart rate variability measures performed similarly to each other, with the exception of drive R4-
1, where the two metrics using LFHF +M F
ratio correlated differently than the two metrics using the
LF
HF
ratio. The 100 second and 300 second windows for HRV performed similarly, suggesting
that it is possible to use the shorter 100 second window to derive features for HRV, although this
window excludes some of the low frequency power typically used in HRV calculations. The
mean heart rate, µR , was the best correlated measure for only one of the drives.
There were individual differences in how drivers responded. In Drive S3-2, there were very high
correlations for both the mean skin conductivity and for the average HRV measures, however
Drive S2-4 showed a much stronger correlation with skin conductivity and heart rate than with
HRV measures and Drive S2-2 showed a weak correlation with skin conductance and stronger
correlations heart rate and HRV measures. For all drivers studied, the lowest correlation be-
tween either the heart rate or skin conductance metrics was 0.49, suggesting that between these
two sensors a reliable metric can be obtained. These correlations were performed over ap-
proximately 25,000 sample points per drive. It is not clear from these results if individuals
consistently respond to stress with similar physiological reactions. For S1 and S3 there was
less variance in mean skin conductance response for the same subject over many drives than
for all subjects over all drives and for S1 the same was also true of heart rate variability. We
performed ANOVA anaylsis on the correlation coefficients and found significant individual dif-
ferences in the mean of the skin conductance µG (p = 0.0007) and the mean of the respiration
µR (p = 0.0001). The difference in the mean skin conductance is most observable for subject
S3. This may be due to a physiological difference in the number of sweat glands on the palm or
from a different in electrode contact due to the way the subject gripped the steering wheel. The
differences in the respiration means are most likely due to physical differences in chest size.
Figure 5 shows an example of the stress metric plotted against signals from drive R2-1. For
this drive, the best correlating signal shown is the mean of the skin conductivity (.47) followed
by L100 (.41) and heart rate (.30). This graph shows qualitatively how well each of the signals
reflects the stress metric. During this drive, the subject was unusually agitated during the second
rest period due to a need to use the restroom. This agitation is reflected in the stress metric, but
would not have been taken into account by using the task based categorization.

7 Discussion
In the future, we may want vehicles to be more intelligent and responsive, managing information
delivery in the context of the driver’s situation. Physiological sensing is one method of accom-
plishing this goal. This study tested the applicability of physiological sensing for determining a
driver’s overall stress level in a real environment using a set of sensors that do not interfere with
the driver’s perception of the road. The results showed that three stress levels could be recog-
nized with an overall accuracy of 97.4% using five minute intervals of data and that heart rate
and skin conductivity metrics provided the highest overall correlations with continuous driver

12
stress levels.
Using a continuously updated record of the last five minutes of a driver’s physiology, the stress
recognition algorithm might be used to manage real-time, non-critical applications such as mu-
sic selection and distraction management (cell phones and navigation aids, etc.), which could
tolerate a delay in updating the user’s state precisely. The original five minute time window
was chosen because it was the interval recommended for calculating heart rate variability using
the spectrograms[23] and because the limiting time factor for the driving segments, the unin-
terrupted highway segment between the two toll booths, was just over five minutes long. In a
similar study, Wilson et al. [5] trained an artificial neural network to recognize three levels of
pilot task demand using five minute intervals of rest and low and high levels of difficulty on
the NASA Multiple Attribute Task Battery during a simulation. For this experiment, heart rate,
EEG, electrooculographic (EOG) and respiration data were used. The algorithm was first tested
on the five minute training segments, then it was run continuously to detect stress in real time.
When a high stress level was detected, the simulation adapted by turning off two of the sub-
tasks, enabling a 33% reduction in errors. A similar test could be performed with the algorithm
developed in Analysis I if road conditions could be made constant and drivers could be allowed
to make safe errors while talking on the cell phone or using visual navigation aids. If a high
stress condition were detected using the algorithm on the last five minutes of data, the driver
distractions could be turned off until the driver recovered to a medium stress level. The level of
driver error for drivers using this adaptive aid could then be compared to a set of control drivers
who did not have this feedback.
Although the original experiment was not designed to test how the five minute algorithm would
perform in a real time scenario, the second analysis compared near real-time features to a con-
tinuous stress metric to determine how well these signals reflected driver stress on a continuous
basis. Driver’s reaction time to specific stressors was not measured because the latencies in-
volved fall beneath the resolution of the coding metric. For example, the skin conductivity
latency is on the order of 1.4 seconds [37] and anticipatory EMG has been measured in the lab-
oratory at 30 milliseconds[38]. In this experiment, the video was scored at one second intervals
and the video clock and sensor clock were not synchronized to be sensitive to time differences
within a few seconds. The latency measurements would also be confounded by the open road
conditions where many stressors occurred concurrently and before the effects of previous stres-
sors had extinguished.
Despite these limitations, these experiments show that physiological signals provide a viable
method of measuring a driver’s stress level. Although physiological sensing systems have not
yet developed to the point where they are as inexpensive and convenient to use as on board
cameras, sensors are becoming smaller and researchers are developing new ways to integrate
them into existing devices. The results of the second analysis suggest that the first sensors
that should be integrated into a car, or a mobile wearable device that communicates with a car,
should be skin conductance and heart rate sensors. These measures could be used in future in-
telligent transportation systems to improve safety and to manage in-vehicle information systems
cooperatively with the driver.
Additionally, future computer vision algorithms and car sensors might be able to automatically
calculate a stress metric similar to the one created by video coding analysis. Such methods

13
might provide an automatic non-contact method for predicting or otherwise anticipating chang-
ing levels of driver stress related to cognitive or emotional load.

Acknowledgments
The authors would like to acknowledge the comments of Dr. Roger Mark, Dr. Ary Goldberger,
Joe Mietus, George Moody, Isaac Henry, Dr. Paul Viola, Dr. John Hansman. We would also
like to thank George Moody for his EKG software and advice and our students Kelly Koskelin,
Justin Seger and Susan Mosher for their assistance with the experiment. We would like to thank
Media Lab’s Things That Think and Digital Life Consortia for sponsoring this research.

References
[1] Y. I. Noy. (2001) International harmonized research activities report of working group
on intelligent transportation systems (ITS). Proceedings of the 17th International
Technical Conference on the Enhanced Safety of Vehicles. [Online]. Available:
http://www-nrd.nhtsa.dot.gov/pdf/nrd-01/esv/esv17/proceed/00134.pdf

[2] P. Burns and T. C. Lansdown. (2000) E-distraction: The challenges for safe and usable in-
ternet services in vehicles. NHTSA Driver Distraction Internet Forum. [Online]. Available:
http://www-nrd.nhtsa.dot.gov/departments/nrd-13/driver-distraction/Papers.htm

[3] National Highway Traffic Safety Administration (NHTSA). Proposed driver

workload metrics and methods project. [Online]. Available: http://www-
nrd.nhtsa.dot.gov/departments/nrd-13/driver-distraction/PDF/32.PDF

[4] M. J. Skinner and P. A. Simpson, “Workload issues in military tactical aircraft,” Int. J. of
Aviat. Psych., vol. 12, no. 1, pp. 79–93, 2002.

[5] G. F. Wilson, J. D. Lambert, and C. A. Russell, “Performance enhancement with real-

time physiologically controlled adaptive aiding,” Proceedings of the Human Factors and
Ergonomics Society 46th Annual Meeting, vol. 3, pp. 61–64, 1999.

[6] H. Selye, Selye’s Guide to Stress Research. Van Nostrand Reinhold Company, 1980.

[7] I. J. K. G. Eisenhofer and D. Goldstien, “Sympathoadrenal medullary system and stress,”

in Mechanisms of Physical and Emotional Stress. Plenum Press, 1988.

[8] A. Baddeley, “Selective attention and performance in dangerous environments,” British

Journal of Psychology, vol. 63, pp. 537–546, 1972.

[9] M. Vidulich, M. Stratton, and G. Wilson, “Performance-based and physiological measures

of situational awareness,” Aviation, Space and Environmental Medicine, pp. 7–12, May
1994.

14
[10] R. Helmreich, T. Chidster, H. Foushee, S. Gregorich, and J. Wilhelm, “How effective is
cockpit resource management training? issues in evaluating the impact of programs to
enhance crew coordination,” Flight Safety Digest, vol. 9, no. 5, pp. 1–17, 1990.

[11] S. Newbury, The Car Design Yearbook. Merrell Publishers, 2002.

[12] ProComp Software Version 1.41 User’s Manual, Thought Technology Ltd., 2180 Belgrave
Ave. Montreal Quebec Canada H4A 2L8, 1994.

[13] G. Moody. tach [c-language software]. [Online]. Available:

http://www.physionet.org/physiotools/wag/tach-1.htm [1985]

[14] ——. ihr [c-language software]. [Online]. Available:

http://www.physionet.org/physiotools/wag/ihr-1.htm [2002]

[15] N. R. Lomb, “Least-squares frequency analysis of unequally spaced data,” Astrophysics

and Space Science, vol. 39, pp. 447–462, 1976.

[16] J. T. Cacioppo and L. G. Tassinary, “Inferring psychological significance from physiolog-

ical signals,” American Psychologist, vol. 45, no. 1, pp. 16–28, Jan. 1990.

[17] M. Helander, “Applicability of drivers’ electrodermal response to the design of the traffic
environment,” Journal of Applied Psychology, vol. 63, no. 4, pp. 481–488, 1978.

[18] T. C. Hankins and G. F. Wilson, “A comparison of heart rate, eye activity, eeg and subjec-
tive measures of pilot mental workload during flight,” Aviation, Space and Environmental
Medicine, vol. 69, pp. 360–367, 1998.

[19] J. A. Veltman and A. W. K. Gaillard, “Physiological indicies of workload in a simulated

flight task,” Biological Psychology, vol. 42, pp. 323–342, 1996.

[20] G. F. Wilson, “An analysis of mental workload in pilots during flight using multiple psy-
chophysiologic measures,” The International Journal of Aviation Psychology, vol. 12,
no. 1, pp. 3–18, 2001.

[21] M. A. Bonner and G. F. Wilson, “Heart rate measures of flight test and evaluation,” The
International Journal of Aviation Psychology, vol. 12, no. 1, pp. 63–77, 2001.

[22] W. Boucsein, Electrodermal Activity. New York, NY: Plenum Press, 1992.

[23] C. M. A. vanRavenswaaij et al., “Heart rate variability,” Annals of Internal Medicine, vol.
118, pp. 436–447, 1993.

[24] E. Crow, F. A. Davis, and M. W. Maxfield, Statistics Manual. Toronto, Canada: General
Publishing Company, 2002.

[25] G. G. Berntson and J. T. B. J. et al., “Coefficient alpha and the internal structure of tests.”
Psychometrika, vol. 16, pp. 297–333, 1951.

15
[26] T. B. Sheridan and R. W. Simpson, “Toward the definition and measurement of the men-
tal workload of transport pilots,” Flight Transportation Laboratory, MIT Department of
Aeronautics and Astronautics, FLT Report R79-4, Jan. 1979.

[27] R. A. Sternbach, Principles of Psychophysiology. Academic Press, 1966.

[28] G. B. Moody and R. G. Mark, “Development and evaluation of a 2-lead ecg analysis
program,” Computers in Cardiology, vol. 9, pp. 39–44, 1982.

[29] Matlab Version 5.3.0.10183 (R11) On-line User’s Manual, The Mathworks, Inc., 3 Apple
Hill Drive,Natick, MA 01760-2098, 1999.

[30] R. Picard, E. Vyzas, and J. Healey, “Toward machine emotional intelligence,” IEEE Trans-
actions on Pattern Analysis and Machine Intelligence, vol. 23, no. 10, pp. 1175–1191,
2001.

[31] J. Healey, “Wearable and automotive systems for affect recognition from physiology,”
Ph.D. dissertation, MIT, 2000.

[32] G. G. Berntson and J. T. B. J. et al., “Heart rate variability: Origins, methods and interpre-
tive caveats,” Psychophysiology, vol. 34, pp. 623–647, 1997.

[33] G. Moody. lomb [c-language software]. [Online]. Available:

http://www.physionet.org/physiotools/wag/lomb-1.htm [1992].

[34] G. B. Moody, “Spectral analysis of heart rate without resampling,” Computers in Cardiol-
ogy, vol. 20, pp. 715–718, 1993.

[35] P. Laguna, G. B. Moody, and R. G. Mark, “Power spectral density of unevenly sampled
data by least-square analysis: performance and application to heart rate signals,” IEEE
Trans Biomed Eng, vol. 45, pp. 698–715, 1998.

[36] R. Duda and P. Hart, Pattern Classification and Scene Analysis. John Wiley & Sons,
1973.

[37] R. Lockhart, “Interrelations between amplitude, latency, rise time and the edelberg recov-
ery measure of the galvanic skin response,” Psychophysiology, vol. 9, no. 4, pp. 437–442,
1967.

[38] Y. Barnif. (2001) Using electromyography to predict head motion for virtual reality.
Research Report, NASA Ames Research Center, Human Factors Research and
Technology Division, Human Information Processing Branch. [Online]. Available:
http://human-factors.arc.nasa.gov/ihh/web/library/rt reports.php

16
Figure 1: The subject wore five physiological sensors, an electrocardiogram (EKG) on the chest,
an electromyogram (EMG) on the left shoulder, a chest cavity expansion respiration sensor
(Resp.) around the diaphragm and two skin conductivity sensors (SC), one on he left hand and
one on the left foot. The sensors were attached to a computer in the rear of the vehicle.

17
Figure 2: A sample frame from the quad split video collected during the experiment. The upper
left panel shows the driver facial expression, collected from a camera mounted on the steering
column. The upper right panel shows the camera used for experimenter annotations where a
“stop” annotation is shown. The lower left panel shows road conditions and the lower right
panel shows a visual trace of the physiological signals as they were being recorded.

18
Figure 3: This figure shows an illustration of the physiological data collected from the respi-
ration, heart rate, L100 spectral ratio, the skin conductivity (SC) from the hand and the elec-
tromyogram (EMG). This figure does not show vertical units because each signal is scaled and
offset to be shown with an illustrative amount of detail.

19
Figure 4: An example of three orienting responses occurring in a one minute segment of the
skin conductance signal. The onset as marked by the detection algorithm is marked with an “x”
and peak is marked with an “o”. The magnitude OM and duration OD features are measured as
shown.

20
Figure 5: This figure shows an illustration of the physiological data collected from the respi-
ration, heart rate, L100 spectral ratio, the skin conductivity (SC) from the hand and the elec-
tromyogram (EMG) along with the stress metric derived from the video tapes for this drive.
This figure does not show vertical units because each signal is scaled and offset to be shown
with an illustrative amount of detail.

21