A Physiological investigation of Rest in Commercial Long-Haul Truck Drivers

(1)

——————————

Master thesis

A Physiological investigation of

Rest in Commercial Long-Haul

Truck Drivers

Author Mattias Axelson Supervisors Christer Ahlstr¨om Johannes Johansson Examiner Ingemar Fredriksson

Department of Biomedical enginering

University of Link¨

oping

(2)

(3)

Abstract

The development of automated vehicles is something most vehicle manufacturers are working on these days. Different levels of automation allow the driver to perform other tasks while traveling than focus on the dynamic driving tasks. For professional drivers where there are strict laws for the amount of driving hours that is allowed without stopping and taking a break, resting while the vehicle is in an automated driving mode can increase the transport efficiency and the comfort of the driver. With data collected from 11 professional long-haul truck drivers in the ADAS&ME project, the goal of this thesis is to investigate if it is possible to obtain rest during autonomous driving (simulated with a confederate driver).

Pre- and post-drive tests, KSS and SUS ratings, HRV features ob-tained from ECG data and blink features obob-tained from vertical EOG data was used in order to evaluate if rest could be obtained during sim-ulated autonomous driving compared to normal driving.

The results show that no clear trends or statistically significant differ-ences can be seen while comparing simulated autonomous driving with normal driving. However one of the participants showed indications in KSS and SUS ratings together with the HRV features that rest was ob-tained during the simulated autonomous driving.

The results indicate that it could be possible to obtain rest during autonomous driving, but a larger set of participants and a more demand-ing study setup is needed to verify the impact of autonomous drivdemand-ing as a substitute for regular rest breaks in terms of obtaining rest.

(4)

(5)

Acknowledgements

Firstly I would like to thank VTI for providing me with the opportunity to write this thesis and especially my supervisor Christer Ahlström for his en-couragement, advice, and support during this thesis. I would also like to thank Johannes Johansson for his valuable inputs during the project, and Ingemar Fredriksson for taking his time to be my examiner. Thank you to Anna Pers-son for all the interesting discussions regarding HRV. Furthermore, I would like to thank Göran Kecklund and Wessel van Leeuwen at Stockholm Univer-sity, Stas Krupenia and Jon Friström at Scania for providing all the data and answering all my questions of how it was obtained.

This thesis work was carried out within the ADAS&Me project which was funded by the European Union’s Horizon 2020 research and innovation pro-gramme under grant agreement No 688900.

(6)

(7)

List of abbreviations

ANS Autonomic Nervous System

ECG Electrocardiogram

EDA Electrodermal activity

EMG Electromyogram

EOG Electrooculogram

KSS Karolinska Sleepiness Scale

LCT Lane change task

PNS Parasympathetic Nervous System

PPG Photoplethysmogram

PSD Power Spectral Density

RMSSD Root Mean Square of the Successive Differences SNS Sympathetic Nervous System

SUS Stockholm University Stress Scale

TAP-M Test of Attentional Performance - Mobility

VTI The Swedish National Road and Transportaion Research In-stitute

(8)

1 Introduction and background

In this chapter the motivation and aim of this thesis work is presented, together with the question formulations and limitations of this project.

1.1 Motivation

The development of automated driving is a hot topic these days and something most vehicle manufacturers are developing. Audi now claims to have a system ready for production that supports SAE level 3 conditional automation [1]. There are six SAE levels of automation [2], a short description of these levels can be seen in table 1. During SAE level 3, it might be possible for the driver to rest, although sleep is not recommended due to sleep inertia since the system might request the driver to intervene and take control of the vehicle.

SAE level Name Description

0 No Automation The driver handles all dynamic driving tasks 1 Driver Assistance The system assists the driver with the execution

of steering or acceleration/deceleration

2 Partial Automation The system assists the driver with the execution of steering and acceleration/deceleration

3 Conditional Automation The driver can hand over all dynamic driving tasks to the automated system, but the driver is expected to be able to take control of the vehicle if requested by the system

4 High Automation The driver can hand over all dynamic driving tasks to the automated system, the system can also handle the vehicle if the driver does not respond to a request to intervene

5 Full Automation All aspects of driving are handled by the system on all types of road and environmental condi-tions

Table 1: Description of the different SAE levels of automation [2]

For professional drivers, there are laws for the amount of driving and rest hours. These are in place in order to ensure reasonable working conditions for the driver, but also to contribute to an increased traffic safety [3]. During one week, the maximum driving period for one driver is 56 hours (maximum 90 hours during two consecutive weeks). The maximum driving period of one day is 9 hours (10 hours are allowed two times each week). A continuous driving period of 4 hours and 30 minutes are allowed until the driver has to take a break for at least 45 minutes. The rest break can be split into two parts during

(11)

the driving period. If this is done, the first break has to be at least 15 minutes and the second break at least 30 minutes. After a total break of 45 minutes, a new driving period of 4 hours and 30 minutes can be started. [4]

If the driver can obtain rest during automated driving it might be possible to change the tachograph (a device that records if the vehicle is moving or standing still) mode from driving to rest, during the automated driving mode. This would lead to an increased transport efficiency but also increase the comfort for the driver.

To recover fully or to get enough rest, one assumption may be that measurable levels in the body should return to the baseline level, i.e. the same level as when work started. It is unlikely that a full recovery is possible during the automated driving. In order to claim that the body is recovering (resting), the most important thing is to show that the levels don’t increase, or preferably moves towards the baseline level during the resting period [5].

Rest breaks and the effects they have on performance and fatigue has been discussed in literature [6]. It was found that workers should be able to take relatively frequent and short rest breaks, but also one or two longer breaks are to prefer in order to sustain attention and focus while driving [6]. However, the effects of rest during automated driving is an area of research where very little work has been done.

ADAS&ME (an acronym for, Adaptive Advanced Driver-Assistance Systems to support incapacitated drivers Mitigate Effectively risks through tailor made Human-Machine Interface under automation), have conducted a study involv-ing eleven commercial long-haul truck drivers to better understand the effects of rest during simulated autonomous driving (the drivers are allowed to sleep during the experiment).

1.2 Aim

With data collected from eleven commercial long-haul truck drivers in the ADAS&ME project [7], the general aim of this master thesis is to investigate whether simulated autonomous driving can be a substitute for breaks (stop-ping and exiting the vehicle), where the main focus will be to investigate if and to what extent long-haul truck drivers can obtain rest in a moving vehicle during autonomous driving (in this study the autonomous driving is simulated by a confederate driver).

Furthermore, this thesis will investigate algorithms for detecting rest via min-imally invasive sensors. The data collected includes both subjective and ob-jective experimental data.

(12)

1.3 Question formulations

• Is there a difference in subjective alertness levels when comparing simu-lated autonomous driving with normal driving?

• Are there differences in the results from the pre- and post-drive tests when comparing simulated autonomous driving with normal driving? • Is it possible to distinguish any differences in obtained rest from the

physiological measurements utilizing electrodes if the driver is allowed to use the simulated autonomous driving instead of normal driving?

1.4 Limitations

This thesis will base its results on the data collected in the ADAS&ME project where the autonomous drive is simulated with a confederate driver.

(13)

2 Theory

The following sections cover the relevant theory that relates to rest and how to find rest indicators in the different signals used in this study.

2.1 Autonomic nervous system

The autonomic nervous system, (ANS), controls the muscles of internal organs in the body, such as regulating breathing, blood pressure, the beating of the heart and metabolic processes in the body. [8]

ANS is composed of two main sub-divisions, the sympathetic nervous system, (SNS), and the parasympathetic nervous system, (PNS). The SNS is associ-ated with emergency situations and more demanding physical activity and is sometimes referred to as the ”fight or flight” response. On the opposite end, the ”rest and digest” response, occurs during more relaxed conditions and is associated with more dominant PNS activity. [9]

2.2 ECG - Heart rate variability

A person at rest with no external physical or mental distractions has essentially a regular heart rate. However, the small variations that occur are known as heart rate variability (HRV). HRV refers to the change in beat-to-beat intervals, i.e. the time between two consecutive R-peaks in the QRS complex obtained from the ECG and referred to as RR-intervals. Changes in PNS and SNS activity can be observed by analyzing HRV, where increased PNS activity leads to a decreased heart rate and an increased HRV, whereas an increased SNS activity leads to increased heart rate and decreased HRV. [10]

RR-intervals that originates from wrongfully detected R-peaks or R-peaks gen-erated from ectopic beats (disturbances in the cardiac rhythm) are considered outliers in HRV analysis. These outliers have a very large influence on HRV parameters. There are different ways of managing outliers, they can either be removed from the set, or replaced with an interpolated value based on the neighboring RR-intervals. The two methods for managing outliers have very similar effects on the resulting HRV parameters [11, 12], the removal method will be used in this thesis. An example of the effect three outliers have on HRV analysis covering a 5-minute epoch can be seen in figure 1.

(14)

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 Frequency [Hz] 0 0.05 0.1 0.15 0.2 PSD [s 2 /Hz] PSD of HRV, outliers excluded 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 Frequency [Hz] 0 0.05 0.1 0.15 0.2 PSD [s 2 /Hz] PSD of HRV, outliers included

Figure 1: Example of the effect of three outliers on the PSD during a 5 min epoch with 384 heartbeats i.e. 0.78% wrongfully detected heartbeats. The red area cor-responds to the HF band and blue area to the LF band. The top graph shows the PSD with outliers removed and the bottom graph shows the same PSD with outliers included

By analyzing the power spectral density, (PSD), in the frequency domain and looking at the low- and high-frequency bands, it has been shown that the low-frequency band, LF (0.04-0.15 Hz) is influenced by SNS activity. Dur-ing restDur-ing conditions it is possible that PNS activity can influence the LF band, especially during periods of slow respiration rates (∼ 3 to 9 breaths per minute), this is due to the fact that the breathing frequency moves from the HF band to the LF band.[13, 14]

The high-frequency band, HF (0.15-0.4 Hz) is influenced by PNS activity and is often referred to as the respiratory band. The reason for this is because changes in heart rate correspond to changes in the respiration. During inhalation, the heart rate is accelerated and during exhalation, the heart rate is decelerated. A low power in the HF region has been shown to have a correlation to stress and worry. During periods of slow respiration rates (∼ 3 to 9 breaths per minute) a lower power in the HF band will be observed (due to breathing frequency in the LF band). [13, 14]

The ratio between LF and HF is another commonly used measurement in HRV analysis. Where the main goal is to estimate the ratio between SNS

(15)

and PNS activity. This assumption suggests that a high LF/HF ratio is an indication of high SNS activity and a low LF/HF ratio is an indication of a high PNS activity. As mentioned earlier the LF region is not solely composed of frequencies generated by the SNS and should, therefore, be interpreted carefully during resting conditions when very slow respiration rates are more likely to occur. [14, 15]

Another way to analyze HRV is in the time domain using the root mean square of successive differences (RMSSD), equation 1. RMSSD has been shown to have a high correlation to the HF band [16]. The HF band as mentioned earlier, is correlated to PNS activity and therefore also rest and recovery.

RM SSD = v u u t 1 N − 1 N −1 X i=1 (RRi+1− RRi)2 ! (1)

Where RRi is the time interval between two successive R-peaks in the QRS

complex.

2.3 EOG

The eye can be described as a dipole, where the cornea is the positive pole and the retina the negative pole. It is this property that is the basis of the electrooculogram, (EOG). For a vertical EOG, which is the signal that will be used in this study, electrodes are placed above and below the eye (preferably as close to the eye as possible to maximize the signal) and the potential is measured differentially, the magnitude of the recorded signal is correlated to the magnitude of the corneo-retinal potential that ranges between 0.4 and 1.0 mV. [17, 18]

Utilizing EOG to observe changes in blink parameters have been shown to be a successful method to estimate sleepiness [19]. Parameters related to blink duration has been shown to have the strongest correlation to sleepiness, where an increased blink duration and increased blink frequency can be indications of sleepiness. [20, 21]

2.4 Rest

Defining rest is tricky since there is no established gold standard to measure or quantify rest. However as described in section 2.1, the activity in the PNS and SNS gives an indication whether the body is in a more relaxed state (dominant PNS activity) that can be associated with sleepiness. Whilst emergency situations (dominant SNS activity) can be associated with a more

(16)

stressed and alert state. Identifying indicators of stress and sleepiness in the objective data that correlates with the subjective data are of importance in order to verify and compare the data. The subjective and self-reported data covers sleepiness and stress in form of the KSS- and the SUS-scale. In the objective data, measurable levels in the body (features that correlate to PNS and SNS activity) can be indicators of sleepiness and stress. [22, 15, 23] Other methods to measure rest (although not applicable while in a vehicle) could be done by analyzing the activity in different parts of the brain such as the amygdala or prefrontal cortex. This could be done with imaging techniques such as positron-emission tomography, PET, or functional magnetic resonance imaging, fMRI. The activity in these areas also showed to be correlated with HRV. [24]

This thesis will go under the assumption that rest is obtained during decreasing subjective ratings. For the objective measurements, activity in the PNS is the main measurable signal that correlates to rest and recovery. To claim that rest has been obtained, this signal level needs to be more dominant, i.e. higher [15, 23], compared to the signal level measured before the rest period. A summary of how each feature used in this report can be an indication of rest is shown in table 2.

Signal Feature Indication of rest

Subjetive/Self report KSS Decreasing

SUS Decreasing

ECG LF/HF ratio Decreasing

Relative power LF Decreasing Relative power HF Increasing

RMSSD Increasing

EOG Mean blink duration Decreasing

Blink frequency Decreasing

(17)

3 Method

This section describes the design of the ADAS&ME study and the different types of data collected. Followed by the signal processing steps used to obtain each feature, it also covers how the features are analyzed in order to obtain the results.

3.1 Participants

Scania Transport Laboratory distributed information to about 100 professional long-haul truck drivers with an invitation to participate in the study. All participation was voluntary and the drivers were paid their normal working salary. 11 drivers participated in the study (8 male and 3 female), with an average age of 45 years (SD = 17) and an average of 15.8 years of professional truck driving experience (SD = 19.3). Written consent was obtained on the first day of testing.

3.2 Experimental design

Participants drove for approximately three hours followed by a rest period of four hours (controls were handed over to the confederate driver), followed by a final 15 minutes of driving. This test condition will be referred to as the rest condition. The participants also drove the same route on another occasion, however this time they drove the entire route by themselves and there was no confederate driver, this test condition will be referred to as the drive condition. The total route for each test condition was 450km. Both test conditions were carried out during the daytime. The participants drove from the Scania Trans-port Laboratory (located in Södertälje) on the E4 motorway to Ödeshög rest stop before turning back to Scania Transport Laboratory. A more detailed summary of the driving route and rest stops is described in table 3, although the arrival times are preliminary since they are affected by traffic conditions. Before and after each test condition the participants performed two tests, these will hereby be referred to as the pre-and post-tests. These tests were conducted in a test room within the Scania Transport Laboratory. Together with the pre-and post-tests several objective pre-and subjective measures were obtained during the drive and rest conditions, these will be described in sections 3.3 and 3.4.

(18)

Time Drive condition Rest Condition 08:30 – 09:30 Depart Scania Transport

Labora-tory and drive to Nyk¨opingsbro

Depart Scania Transport Labora-tory and drive to Nyköpingsbro 09:30 – 09:45 Rest Break at Nyköpingsbro Rest Break at Nyköpingsbro 09:45 – 10:45 Depart Nyköpingsbro and drive to

Herrbeta

Depart Nyk¨opingsbro and drive to Herrbeta

10:45 – 11:15 Lunch break at Herrbeta Lunch break at Herrbeta, hand con-trol to confederate driver

11:15 – 12:20 Depart Herrbeta and drive to rest stop ¨Odesh¨og (exit 106)

No driving

12:20 – 15:00 Return to highway and drive to J¨arna

No driving

15:00 – 15:05 Rest break at J¨arna Take back control from confederate driver

15:05 – 15:15 Depart J¨arna and drive to Scania Transport Laboratory

Depart J¨arna and drive to Scania Transport Laboratory

Table 3: Summary of the driving routes and rest breaks during the drive and rest conditions

3.3 Subjective and self-reported measures

The subjective and self-reported measures consist of the Karolinska Sleepiness Scale (KSS) and the Stockholm University Stress Scale (SUS). During the driving part of the test, the participants were required to grade their sleepiness and stress every fifth minute.

KSS is a 9-point Likert scale with verbal anchors to each scale [25, 26]. The KSS is shown in table 4. Value Description 1 Extremely alert 2 Very alert 3 Alert 4 Rather alert

5 Neither alert nor sleepy 6 Some signs of sleepiness

7 Sleepy, but no difficulty remaining awake 8 Sleepy, some effort to keep alert

9 Extremely sleepy, fighting sleep

(19)

SUS is a 9-point Likert scale. Five of the nine categories includes verbal anchors to the corresponding scales. The SUS can be seen in table 5 [27].

Value Description

1 Very low stress (very calm and relaxed) 2

3 Low stress (calm and relaxed) 4

5 Neither low nor high stress 6

7 High stress (high tension and pressure) 8

9 High stress (Very high tension and pressure)

Table 5: The Stockholm University Stress Scale

3.4 Objective measures

The objective measures are divided into two parts. The first part covers the measures obtained during the rest and drive condition. The second part, pre-and post-tests, covers the measures obtained before pre-and after the participants started and ended the rest and drive conditions.

3.4.1 Measures collected during the rest and drive condition

The following measures were collected during the drive and rest conditions. Vitaport II

Data were recorded with a medical reference system from Temec Instruments, called Vitaport II. Data collected by the Vitaport includes EOG (horizontal and vertical), Electrodermal activity (EDA), Electromyogram (EMG, trape-zoid) and Electrocardiogram (ECG). This data was recorded at 256Hz during the entire drive and rest condition.

Empatica E4

An Empatica E4 wristband collected the following data, Photoplethysmogram (PPG), EDA, Inter beat intervals (IBI), 3-axis accelerometer (ACC) and tem-perature. Data was recorded at 1Hz during the entire drive and rest condition. CAN

Data from the Controller Area Network, (CAN), was logged both during the entire drive and rest condition.

(20)

Eye Tracker

A three-camera SmartEye eye tracking system recorded during the drive period in the rest condition and during the entire drive condition.

Optical Cameras

Recorded the participant’s overall movement inside the cabin, the positions of the hands, and the forward field of view. Data were recorded during the entire drive and rest condition.

3.4.2 Pre- and post-tests

The participants went through two performance tests before and after they started and ended the rest and drive condition.

TAP-M

The first test was an alertness and sustained attention test which included two components of the Test of Attentional Performance-Mobility (TAP-M) [28]. The first was the Alertness-TAP-M, designed to assess tonic alertness, i.e. how ready an individual is to respond, where the participants are to react and press a button as fast as possible in response to an ”x” appearing on a screen. The second was the Go/NoGo-TAP-M, designed to assess the specific ability to suppress undesired responses, which is performed the same way as the Alertness-TAP-M with the difference that a ”+” can appear on the screen, which should yield no response from the participants.

LCT

The second pre- and post-test was the Lane Change Task (LCT). LCT is a simulation- and analysis software designed to asses driver distractions that are caused by in-vehicle-tasks [29]. Performance is evaluated by the participant’s deviation from the optimal route. During the LCT, participants complete a 1-back task. The principle of the 1-back task is that the driver responds to a sequence of numbers being read out loud and continuously repeating the second last number in the sequence [30]. Performance is evaluated by the total amount of correct responses.

3.5 Data acquisition

As described in the section above, there was a lot of different data obtained in the ADAS&ME study. The data with the strongest correlation to rest was included in this thesis, ECG and EOG from the Vitaport II, and the pre-and post-tests were selected from the objective data together with KSS pre-and SUS from the subjective data. Unfortunately, the data recorded with the Empatica E4 turned out to be of poor quality with a lot of connection losses, therefore it was excluded from this study. The data from the eye tracker and

(21)

optical cameras only contained data from four participants and therefore it was decided not to include this data in the study.

From the eleven participants, only five complete data sets were acquired, the other six data sets had some missing data or artifacts, see table 6. KSS and SUS values were obtained from nine of the eleven participants and used to check whether there are any differences in subjective rest while the driver is allowed to use simulated autonomous driving compared to normal driving. For the investigation whether rest can be distinguished from physiological measurements, 6 of the eleven data sets are used for ECG data and 6 data sets for the EOG data, with the difference that data from TS3 was used in the EOG analysis and TS11 in the ECG analysis.

The Vitaport collected data continuously during the experiment, even when the driver left the vehicle (during scheduled breaks). This can cause a lot of noise in the signals due to electrode movement. During these times the driver did not answer any KSS- or SUS-prompts. Therefore the parts of the signal where the driver is not in the vehicle were removed from the data sets. The KSS and SUS data registers new values with five-minute intervals. The physiological data is sampled at 256 Hz, and are divided into 5-minute epochs.

Test-subject ECG EOG KSS, SUS Comment

TS1 Test aborted

TS2 √ √ √

TS3 √ √ Premature ventricular

con-traction, interferes with HRV analysis

TS4 √ √ √

TS5 Test aborted

TS6 √ Missing Vitaport data

TS7 √ Missing Vitaport data

TS8 √ √ √

TS9 √ √ √

TS10 √ √ √

TS11 √ √ Missing Vitaport data

Table 6: The available data from each participant in this study is marked with the symbol;√

(22)

3.6 Signal processing

All signal processing were done in MATLAB (Version R2017b). An overview of the different processing steps used to obtain each feature can be seen in figure 2.

Figure 2: Flowchart, feature extraction

3.6.1 ECG

To obtain the R-peaks from the QRS complex in the ECG signals the MAT-LAB toolbox, Complete Pan Tompkins Implementation ECG QRS detector [31] was used. This algorithm is based on the QRS detection algorithm devel-oped by Pan J and Tompkins, W.J [32].

Algorithm overview - Preprocessing

The Pan-Tomkins QRS detection algorithm is based on ECG data sampled at 200 Hz. Due to this, the data is downsampled from 256 Hz (which is the sampling frequency of the data in this study) to 200 Hz. This is followed by

(23)

a bandpass filter (5-15 Hz) to maximize the QRS complex, remove baseline wander and muscle noise.

The signal is then differentiated and squared. The final step of the prepro-cessing is a moving window integration, where the signal is averaged with a window length of 30 samples (150 ms). The output from the preprocessing is a pulse-shaped waveform.

Algorithm overview - Decision rule

Following the preprocessing, it has to be decided if the pulses correspond to a true QRS complex and not originating from e.g. noise or a prominent T-wave. Identifying the pulse to a single point of time is done by locating the highest peak of the pulse, furthermore to dismiss peaks that do not originate from the QRS complex a distance of at least 40 samples (200 ms) has to pass before a new peak can be registered. 200 ms is based on the physiological constraint that a new QRS complex cannot be generated within 200 ms since ventricular depolarization cannot occur during the refractory period.

There are two thresholds in the decision part of the algorithm, one noise threshold, and one signal threshold. These thresholds are initialized during a short training period of two seconds. Where the signal threshold is initialized as 25% of the maximum amplitude and the noise threshold is initialized as 50% of the mean signal. If a detected peak is larger than the signal threshold it is considered a QRS complex candidate, if the detected peak is between the two thresholds it is considered a noise peak. The thresholds are constantly updated depending on the signal strength and the amount of noise.

If a long time passes between two detected R-peaks, the algorithm triggers a search within the time period of the last two detected R-peaks. The highest peak during this period is assumed to be the missing R-peak. This peak has to be between the signal and noise threshold to be considered a missed R-peak (note, this would be considered to be a noise peak if a search had not been triggered). The output of the algorithm is the locations and amplitudes of the R-peaks. An example of the detected R-peaks can be seen in figure 3.

Outlier detection

To detect outliers in the obtained RR-interval a method based on standard deviation was used. Values that differ more than 4 standard deviations from the mean value of all RR-intervals in the current epoch are considered outliers. To verify that the outliers removed are in fact outliers, visual inspection was done on the ECG for smaller parts of the data sets.[11, 12] It should be noted that not all ECG data were inspected for outliers since this would be extremely time-consuming, the visual inspection was done in order to confirm that the outlier algorithm was working in a correct way. The majority of the removed outliers originated from wrongfully detected R-peaks, there were, however, a small number of wrongfully removed RR-intervals (or missed outliers) originat-ing from parts with a high HRV. Experiments with another outlier detection

(24)

method called percentage change were done. This method is based on the mean of the four previously accepted RR-intervals in the epoch. If the cur-rent RR-interval differs more than 30% of the mean of the four previously detected RR-intervals it is considered an outlier [11]. The standard deviation performed better than the percentage change method and was therefore used in this study. 2.3 2.35 2.4 2.45 2.5 2.55 2.6 104 0 0.5 1 Raw Signal 2.3 2.35 2.4 2.45 2.5 2.55 2.6 104 -0.4 -0.2 0 0.2 0.4 0.6

Band Pass Filtered

2.3 2.35 2.4 2.45 2.5 2.55 2.6

104

-0.5 0 0.5

Filtered with the derivative filter

2.3 2.35 2.4 2.45 2.5 2.55 2.6 104 0.1 0.2 0.3 0.4 0.5 Squared 2.3 2.35 2.4 2.45 2.5 2.55 2.6 104 0.02 0.04 0.06 0.08 0.1 0.12 0.14

Averaged with 30 samples length,Black noise,Green Adaptive Threshold, RED Sig Level,Red circles QRS adaptive threshold

(25)

LF/HF-ratio

HRV was obtained by calculating the time difference between each successive RR-interval. The PSD was obtained by the MATLAB function plomb [33] which calculate the Lomb-Scagle periodogram. The advantage of plomb is that it can calculate the PSD on unevenly sampled data (such as HRV) with-out interpolation. Following this, the relative power of the LF and HF band was calculated. An example of the obtained PSD can be seen in figure 4.

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Frequency [Hz] 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 PSD [s 2/Hz] PSD of HRV

Figure 4: Power spectral density of the HRV. Blue area corresponds to the LF band (0.04-0.15 Hz), the red area corresponds to the HF band (0.15-0.4 Hz)

RMSSD

RMSSD was obtained by applying the built in MATLAB function rms [34] on the successive RR intervals.

(26)

3.6.2 EOG

Features from the EOG signal was extracted using an automatic blink de-tection algorithm developed by Jammes, B et al. [35]. The algorithm was provided by VTI.

Algorithm overview - calibration

To find the thresholds needed for a specific individual to identify a blink event, the algorithm needs to be calibrated. During this calibration, a short sample of the EOG-signal is used, in this case, the first five minutes of the signal. From this signal, the algorithm calculates two outputs, velocity of a normal blink and amplitude of a normal blink.

Algorithm overview - blink detection

The two parameters obtained from the calibration are used as input in the blink detection algorithm. The EOG-signal is filtered with a low-pass filter to remove frequencies over 10 Hz, following this the derivative of the signal is computed. The derivative is used in order to search for specific events in the signal that exceeds a threshold (eyelid closing velocity) and falls below another threshold (eyelid opening velocity). These thresholds were selected empirically by the authors of the algorithm.

The parameters obtained from the calibration are used in relation to the orig-inal signal if the signal exceeds the calibration parameters it is considered to be a blink event.

3.7 Mean and standard deviation

The mean and standard deviation are calculated from the participants for each signal in order to see if any observable trends can be seen. For the rest condition (simulated autonomous drive), the data is split into three parts, pre rest, during rest and post rest, an illustration of this can be seen in figure 5. The mean and standard deviation is calculated separately for these three time periods.

For the drive condition where the participants drove the entire route them-selves, mean and standard deviation are calculated during the entire period.

(27)

Figure 5: Illustration of what parts the mean and standard deviation are calculated on during the rest condition

3.8 Pre- and post-drive tests

The pre and post-drive tests have been analyzed by Stockholm University (project partner with ADAS&ME) and no significant differences were found. The first and last KSS is added to the results in order to see if there is a correlation between KSS and the results obtained in the pre- and post-tests.

3.9 Statistics

In order to analyze if the results are statistically significant, they are evaluated with IBM SPSS Software. Different versions of ANOVA (analysis of variance) are used to determine if there are any significant differences in variance and mean between the different groups. A probability level of p ≤ 0.05 was set for statistically significant results. Together with the probability level, the cor-responding F-statistic is also presented, F(df_between,df_error). Where df_between is the degrees of freedom between, calculated as k-1, where k is the number of levels in the measurement, i.e. how many times (observations) a value is measured in each TestCondition. The degrees of freedom error (df_error), is cal-culated as df_within− df_subjects. df_within is calculated as N-k, N is the number of samples available. df_subjects is calculated as S-1, where S is the number of participants. The degrees of freedom shows information about the number of independent pieces of information in the analysis.

(28)

To check if the rest condition has any differences compared to the drive condi-tion, a two way repeated measures ANOVA was used, where several different effects are evaluated. TestCondition is the drive and rest condition, Time is the first and last sample registered in each test condition. TS is the dif-ference between the participants and used as a random factor. Time and TestCondition are the two independent variables used (together with KSS in the 3-way repeated measures ANOVA used for the pre- and post-tests). Time*TestCondition is the interaction between the test condition and time. The dependent variable is the feature looked upon (KSS, RMSSD, etc.). All these effects and if there is any significant difference in variance in relation to the feature looked upon is evaluated in the repeated measures ANOVA. The pre- and post-tests analyzed by the Stockholm University was done with a two-way repeated measures ANOVA, where Time and TestCondition was used as independent variables. A 3-way repeated measures ANOVA was used when the first and last KSS value was added as an independent variable. The dependent variable for both of these methods was the feature looked upon (Reaction time, Errors, etc.).

(29)

4 Results

This section is divided into two major parts, one covering the results obtained from all available data, and one covering the results from a smaller sub-sample. This sub-sample contains data from three participants that showed a consid-erable increase in subjective KSS ratings during the drive condition, while the KSS ratings during the rest conditions were much lower. One of these par-ticipants also showed considerable changes in HRV features during the rest condition. The reason for selecting this sub-sample is to highlight the need for further analysis to determine the effects of rest during autonomous driving.

4.1 Pre- and post-tests

The results from the pre- and post-test without KSS added as a factor are shown in table 7. Table 8 shows the results where KSS has been added as a factor, it should be noted that there was one less test subject in table 8 compared to table 7 due to missing KSS values. No statistically significant results were found in either of the cases.

Test Feature Effect F(1,9) p

TAP-M Alertness Reaction time TestCondition 0.655 0.439

Time 0.858 0.378

Time*TestCondition 2.194 0.173

TAP-M GoNogo Errors TestCondition 0.000 1.000

Time 0.488 0.502

Reaction time TestCondition 1.572 0.242

Time 0.203 0.663

LCT Deviation from optimal line TestCondition 0.431 0.528

Time 2.953 0.120

1-back task Correct responses TestCondition 0.003 0.960

Time 1.899 0.201

Table 7: Results from the pre- and post-tests, analyzed with a 2-way repeated measures ANOVA. Data from 10 participants

(30)

Test Feature Effect F(2,21) p

TAP-M Alertness Reaction time TestCondition*KSS 1.542 0.237

Time*KSS 0.289 0.752

Time*TestCondition*KSS 0.249 0.782

TAP-M GoNogo Errors TestCondition*KSS 0.780 0.471

Time*KSS 0.721 0.498

Reaction time TestCondition*KSS 0.052 0.931

Time*KSS 0.062 0.940

LCT Deviation from optimal line TestCondition*KSS 0.667 0.524

Time*KSS 0.172 0.843

1-back task Correct responses TestCondition*KSS 1.515 0.243

Time*KSS 2.177 0.138

Table 8: Results from the pre- and post-tests with KSS added as a factor, analyzed with a 3-way repeated measures ANOVA. Data from 9 participants

4.2 Mean and standard deviation

The results from plotting the mean and standard deviation from each feature are shown in figures 6 - 13. The gray lines in these figures are the results from each individual participant whereas the red lines are the mean and standard deviation calculated from the results of all participants. On the x-axis, time is calculated backward, so at the first time zero (in the intersection between ”Pre rest” and ”During rest”) all participants arrived at the time where they change to the confederate driver. The same goes for the second zero time (intersection between ”During rest” and ”Post rest”) all participants arrived at the time where they take over control from the confederate driver.

The aim of these figures was to see trends in the mean that indicate rest according to table 2, but also to observe if there are any differences between the drive and rest conditions. However, trends in the drive conditions are similar to the rest conditions and no indications of rest can be seen. The reason for the gaps in the subjective data during the rest conditions is as mentioned in section 3.3, due to the fact that KSS and SUS values were only obtained during the driving parts of the rest condition.

(31)

Figure 6: Mean KSS and standard deviation during the rest and drive condition. Data from 9 participants

Figure 7: Mean SUS and standard deviation during the rest and drive condition. Data from 9 participants

(32)

Figure 8: Mean LF/HF ratio and standard deviation during the rest and drive condition. Data from 6 participants

Figure 9: Mean relative power LF and standard deviation during the rest and drive condition. Data from 6 participants

(33)

Figure 10: Mean relative power HF and standard deviation during the rest and drive condition. Data from 6 participants

Figure 11: Mean RMSSD and standard deviation during the rest and drive condi-tion. Data from 6 participants

(34)

Figure 12: Mean blink duration and standard deviation during the rest and drive condition. Data from 6 participants

Figure 13: Mean blink frequency and standard deviation during the rest and drive condition. Data from 6 participants

(35)

4.3 Statistical analysis

Table 9 show the results from the two way repeated measures ANOVA. The relative LF and HF power shows statistically significant change with time, although by a small margin.

Feature Effect F(1,8) p KSS TestCondition 2.992 0.122 Time 1.818 0.214 TS 1.022 0.486 Time*TestCondition 2.286 0.169 SUS TestCondition 1.730 0.225 Time 0.640 0.447 TS 1.946 0.265 Time*TestCondition 0.640 0.447 Feature Effect F(1,5) p LF/HF TestCondition 0.030 0.869 Time 3.025 0.143 TS 2.237 0.499 Time*TestCondition 0.078 0.791

Relative power LF TestCondition 0.003 0.961

Time 7.258 0.043

TS 2.955 0.408

Relative power HF TestCondition 0.003 0.961

Time 7.258 0.043 TS 2.955 0.408 Time*TestCondition 0.155 0.710 RMSSD TestCondition 4.216 0.095 Time 0.252 0.637 TS 0.514 0.769 Time*TestCondition 1.530 0.271

Blink duration TestCondition 2.314 0.189

Time 0.570 0.484

Blink frequency TestCondition 0.060 0.817

Time 0.000 0.989

TS 1.163 0.619

(36)

4.4 Sub-sample

Three participants, TS4, TS7 and TS11, showed a large increase in KSS ratings during the drive condition, KSS 3 → 6, 2 → 5, 3 → 7. The corresponding ratings during the rest condition are KSS 1 → 2, 3 → 3, 2 → 2.

Figures 14 and 15 shows the results from TS11, where the effects of the study are very clear and reflect the need for further analysis of the effects of au-tonomous driving as a supplement for regular breaks. No EOG data was recorded for this participant and are therefore not included in figure 15. The self-reported results shown in figure 14 shows considerably lower KSS and SUS ratings after the rest period in the rest condition compared to the drive condition. In the ECG features shown in figure 15 no clear difference can be seen after the rest period compared to the drive condition. There is, however, a large change in all objective features towards the end of the rest period. During this time the logs showed that the driver was sleeping or at least the driver’s eyes were closed.

In figure 15 there are several data gaps in the drive condition. These originate from a removal of data when the drive is not inside the vehicle. There are three scheduled breaks during the drive condition, this participant has however done two unscheduled stops during the drive condition, the reason for these unscheduled stops are unknown.

EOG and ECG data were available for TS4, but there were no clear indications of rest in any of the objective features. No objective data was available for TS7.

(37)

Figure 14: KSS and SUS results from TS11

(38)

5 Discussion

This section will discuss the results obtained in this study and highlight the results from the smaller sub-sample, but also future work that can be done to investigate whether drivers can obtain rest during autonomous driving.

5.1 Results

Since there were only 5 complete data sets available the results are based on the amount of data that were available for each feature. Nine drivers had subjective measurements, while 6 had ECG data and 6 had EOG data (not the same 6 as for the ECG data). Comparing the data when it originates from different sources is not to prefer and can introduce inaccuracies in the results. The pre- and post-test showed no statistically significant results with the ad-dition of the first and last KSS value, table 8. No clear trend can be seen compared to the results without KSS values added, table 7. However this is to be expected, in table 9, it can be seen that KSS have no significant impact on any of the effects. That is, the drivers were still alert enough during the post-test to complete the tests in a similar way as the pre-tests, both during the rest and drive condition.

From figures 6 - 13, a clear trend cannot be seen between the drive and rest condition. From table 9, the only features that showed any statistically sig-nificant difference was the relative LF and HF power, these features showed a difference in time (p = 0.043). The reason for this remains unknown, and could be due to random effects, the LF/HF ratio in the same table is not close to being statistically significant with Time which could be expected since the relative LF and HF power were. TS in the same table shows no significant difference, that is there were no significant differences between the individual drivers. From these results, the conclusion should be that the rest condition had little to no effect compared to the drive condition, in terms of obtaining rest.

There are many factors that might have an influence on the results. The way that the test is designed, it should represent a regular work day for the drivers. This can have a big influence on the test in the sense that not all drivers have exactly the same layout of the test. They are still driving the same route, but there are some unscheduled breaks affecting some drivers, one example of this can be to stop and fill up fuel. Another example of the differences between different drivers is during the scheduled breaks. During this time some driver decided to exit the vehicle and some decided to stay inside the vehicle during the break. The effect of taking a short walk during a break compared to sitting still is not covered in this thesis. But since it involves differences in

(39)

physiological activity during the breaks, it will in some way have an impact on the features extracted.

The time period in the rest condition where the driver takes back controls after the simulated autonomous drive is very short, roughly 15 minutes. This makes it hard to see any trend in the data, or what the longer effects are from the resting period.

Not having subjective measurements during the simulated autonomous drive makes it hard to verify the objective measurements during this time period. The reason for not having subjective measurements during this period is to let the participant be as relaxed and free to rest as possible and give the partici-pants the option to sleep. One solution to this could be to have longer prompts between each subjective response during this period, e.g. 15 min. This would make the participants unable to sleep, but the advantages to compare the sub-jective measurements with the obsub-jective measurements weighs higher in my opinion.

The study is also performed during conditions that might not be very demand-ing from the participants. The study was conducted on professional long-haul truck drivers during daytime on a regular workday. This fact might raise the question of whether there should be an observable change in rest. How to design a test where the effects of a resting period during autonomous driving can be seen more clearly is discussed more in section 5.3.2.

5.2 Sub-sample

The high KSS values reported by the three drivers in the sub-sample are interesting. The last KSS values reported during the drive condition are 6, 5 and 7. During the rest condition, the corresponding KSS values are 2, 3 and 2. This indicates that the resting period has had a positive effect on these drivers.

Looking at the results in figures 14 - 15 from one of the drivers strengthens the theory that at least this driver has obtained some sort of rest. Comparing the KSS and SUS results from the rest condition with the drive condition shows considerably lower KSS and SUS values. The ECG features in figure 15, are about the same in the final part of both conditions. In the rest condition, during the second part of the simulated autonomous drive (during rest ), there is, however, a large change in the registered signals from all features. During this time the driver’s eyes were closed and it is possible that the driver was sleeping. The LF/HF ratio and LF power decrease, while HF and RMSSD increase. All these features indicate dominant PNS activity and fulfill the requirements for rest shown in table 2. The reason that all these features return to ”normal” before the driver is going to take control of the vehicle

(40)

again might be due to some sort of anticipation or mental preparation to start working again. In some sense, the body might be preparing for a more demanding task. The period ”post rest” only covers about 15 minutes and the effect that the rest might have had is hard to tell during such a short time frame.

Unfortunately, this driver did not have any EOG data and therefore it is not possible to make any conclusions whether the EOG features support the claim that this particular driver has obtained rest, but since the driver had their eyes closed, the EOG data would not add much information.

5.3 Future work

This section is divided into two parts, one covering the signal processing used and one covering the design of the study.

5.3.1 Signal processing

The objective measurements are generated from two major algorithms. HRV features are based on the RR-intervals obtained from the QRS detection al-gorithm developed by Pan J and Tompkins, W.J [32] and the EOG features from the blink detection algorithm developed by Jammes, B et al. [35]. The HRV features are dependent on correctly identified R-peaks and the impact of outliers are as mentioned in section 2.2 very big. Though the QRS detection algorithm has steps to make sure that all detected R-peaks are in fact R-peaks, it is a very strong possibility that there are some missed or wrongly classified R-peaks in the output. The best way to detect R-peaks is to visually inspect the ECG, but this is not doable with the amount of data in this study. The missed or wrongly detected R-peaks from the QRS detection algorithm will impact the outlier detection. Even though outliers are removed from the RR-intervals, the outlier detection is based on standard deviation from the mean of the epoch. If there are many outliers, the standard deviation will increase and the risk of including RR-intervals that originate from wrongly detected R-peaks increases.

The output from the blink detection algorithm is not post-processed in any way. In the same way as the QRS detection algorithm, there is a very strong possibility that there are missed or wrongly detected blinks. Although the majority of the detected features used in this study originates from correctly detected R-peaks and blinks, more steps should be taken to reduce sources of error and make sure that all features originates from correctly detected events.

(41)

5.3.2 Study design

To more clearly see the effects that autonomous driving have on the driver’s ability to obtain rest more work has to be done. A larger set of data is preferred to avoid smaller factors and sources of error to influence the results to a large extent. A longer period of data collection after the driver takes back controls is needed in order to more clearly see the effects of the resting period. The data in this study only covers about 15 minutes.

The study is also performed as a regular work day during daytime and this fact might make it harder to observe the effects of the rest condition. To get a better understanding of how and to what extent the rest period affects the drivers it might be suitable with a more demanding study design. One example of this could be to perform the study during nighttime or not allowing the participants to get a full night sleep before the study. By doing this the participants might be in more need of rest and the rest period could have a stronger impact on the drivers.

(42)

6 Conclusions

Since the number of complete data sets available in this study was fairly small (5 complete data sets), it is hard to make any claims whether rest can be obtained during autonomous driving. The results show that the majority of the drivers were fairly unaffected by the simulated autonomous drive. There were, however, three drivers that showed indications in the subjective ratings that the autonomous driving had a positive effect compared to normal driving. One driver with clear indications in both the ECG and subjective features, that rest was obtained during the rest condition (figure 15). The pre- and post-tests showed no differences between the drive and rest condition.

Further studies need to be done with a larger set of complete data sets. With this set of data, it is not possible to determine the effects of autonomous driving as a substitute for regular breaks in terms of obtaining rest. Further studies should also involve a more demanding design. Such as a study should be performed during nighttime or possibly during daytime but not allowing the participants to get a full night sleep before the day of the

(43)

(44)

study-References

[1] Audi MediaCenter. Piloted Driving [Internet]. Audi MediaCenter; [cited 18 January 2018]. Available from: https://www.audi-mediacenter. com/en/piloted-driving-3651.

[2] SAE International. Automated driving [Internet]. SAE International; [cited 18 January 2018]. Available from: https://www.smmt.co.uk/ wp-content/uploads/sites/2/automated_driving.pdf.

[3] Transportstyrelsen. K¨or- och vilotider [Internet]. Transport-styrelsen; [cited 17 January 2018]. Available from: https: //www.transportstyrelsen.se/sv/vagtrafik/Yrkestrafik/

Kor--och-vilotider/.

[4] K¨or-och vilotider Regler och v¨agledning [Internet]. Transport-styrelsen; [cited 23 May 2018]. Available from: https://www. transportstyrelsen.se/sv/Publikationer/Vag/Yrkestrafik/ Kor--och-vilotider---regler-och-vagledning/.

[5] Geurts SAE, Sonnentag S. Recovery as an explanatory mechanism in the relation between acute stress and chronic health impairment. Scand J Work Environ Health. 2006;32:482–92.

[6] Tucker P. The impact of rest breaks upon accident risk, fatigue and preformance: a review. Work Stress. 2003;17:123–37.

[7] ADAS&ME. Homepage [Internet]; 2018 [cited 2 February 2018]. Available from: http://www.adasandme.com/.

[8] Health P. Autonomic Nervous System (Involuntary Nervous System) [Internet]. PubMed; [cited 26 April 2018]. Available from: https://www. ncbi.nlm.nih.gov/pubmedhealth/PMHT0025455/.

[9] McCorry LK. Physiology of the Autonomic Nervous System. Am J Pharm Educ. 2007;71:78.

[10] Sornmo L, Laguna P. Bioelectrical Signal Processing in Cardiac and Neurological Applications. 1st ed. Amsterdam: Elsevier Academic Press; 2006.

[11] Kemper KJ, Hamilton C, Atkinson M. Heart Rate Variability: Impact of Differences in Outlier Identification and Management Strategies on Com-mon Measures in Three Clinical Populations. Pediatr Res. 2007;62:337– 42.

[12] Forcolin F, Buendia R, Karlsson J, Sj¨oqvist BA, Anund A. Compari-son of outlier heartbeat identidentification and spectral transformation strategies for deriving heart rate variability indices for drivers at different stages of sleepiness. Traffic Inj Prev. 2018;19:112–9.

(45)

[13] Rajendra Acharya U, Paul Joseph K, Kannathal N, Lim CM, Suri JS. Heart rate variability: a review. Med Biol Eng Comput. 2006;44:1031–51. [14] Schaffer F, Ginsberg JP. An Overview of Heart Rate Variability Metrics

and Norms. Front Public Health. 2017;5.

[15] Schaffer F, McCraty R, Zerr CL. A healthy heart is not a metronome: an integrative review of the heart’s anatomy and heart rate variability. Front Psyhcol. 2014;5.

[16] Kaida K, ˚Akerstedt T, Kecklund G, Nilsson JP, Axelsson J. Use of sub-jective and physiological indicators of sleepiness to predict performance during a vigilance task. Ind Health. 2007;45:520–6.

[17] Malmivuo J, Plonsey R. Bioelectrmagnetism - Principles and Applica-tions of Bioelectric and Biomagnetic Fields. New York, N.Y.: Oxford University Press; 1995.

[18] Heide W, Koenig E, Trillenberg P, K¨ompf D, Zee DS. Electroencephalog-raphy and clinical neurophysiology. Electroencephalogr Clin Neurophysiol Suppl. 1999;52:223–40.

[19] Schmidt J, Laarousi W R amd Stolzmann, Karrer-Gauß K. Eye blink detection for different driver states in conditionally automated driving and manual driving using EOG and a driver camera. Behav Res Methods. 2017;.

[20] ˚Akerstedt T, Ingre M, Kecklund G, Anund A, Sandberg D, Wahde Mea. Reaction of sleepiness indicators to partial sleep deprivation, time of day and time on task in a driving simulator – the DROWSI project. J Sleep Res. 2010;19:298–309.

[21] Schreicher R, Galley N, Briest S, Galley l. Blinks and saccades as in-dicators of fatigue in sleepiness warnings: looking tired? Ergonomics. 2008;51:982–1010.

[22] Vicente J, Laguna P, Bartra A, Bail´on R. Drowsiness detection using heart rate variability. Med Biol Eng Comput. 2016;54:927–37.

[23] Wiklund U, Olofsson BO. Hj¨artfrekvensvariabilitet. In: Pahlm O, S¨ornmo L, editors. Elektrokardiologi: klinik och teknik. Lund: Studentlitteratur; 2006. p. 223–52.

[24] Thayer JF, ˚Ahs F, Fredriksson M, Sollers JJ, Wager TD;.

[25] ˚Akerstedt T, Gillberg M. Subjective and Objective Sleepiness in the Active Individual. Int J Neurosci. 1990;52(1-2):29–37.

[26] Horne JA, Burley CV. We know when we are sleepy: Subjective versus objective measurements of moderate sleepiness in healthy adults. Biol Psychol. 2010;83(1-2):266–8.

(46)

[27] Dahlgren A, Kecklund G, ˚Akerstedt T. Different levels of work-related stress and the effects on sleep, fatigue and cortisol. Scand J Work Environ Health. 2005;31:277–85.

[28] Psytest. The mobility version of TAP [Internet]; 2017 [cited 02 February 2018]. Available from: https://www.psytest.net/index.php?page= TAP-M&hl=en_US.

[29] DaimlerChrystler AG Research and Technology. Lane Change Test 1.2 - User Guide [Internet]; 2005 [cited 31 January 2018]. Available from: http://apps.usd.edu/coglab/schieber/pdf/LCT-UserGuide.pdf. [30] Mehler B, Reimer B, Dusek JA. MIT AgeLab Delayed Digit Recall Task

(n-back). MIT AgeLab Working paper 2011-3B;.

[31] Sedghamiz H. Matlab Implementation of Pan Tompkins ECG QRS de-tector. 2014;.

[32] Pan J, Tompkins WJ. A Real-Time QRS Detection Algorithm. IEEE Trans Biomed Eng. 1985;32.

[33] Plomb [Internet]. Mathworks; 2016 [cited 13 Mars 2018]. Available from: https://se.mathworks.com/help/signal/ref/plomb.html.

[34] rms [Internet]. Mathworks; 2016 [cited 13 Mars 2018]. Available from: https://se.mathworks.com/help/signal/ref/rms.html.

[35] Jammes B, Sharabty H, Esteve D. Automatic EOG Analysis: A first step toward automatic scoring during wake-sleep transitions. Somnologie (Berl). 2008;12:227–32.

A Physiological investigation of Rest in Commercial Long-Haul Truck Drivers

Master thesis