You are on page 1of 10

Accident Analysis and Prevention 141 (2020) 105508

Contents lists available at ScienceDirect

Accident Analysis and Prevention


journal homepage: www.elsevier.com/locate/aap

Influence of traffic congestion on driver behavior in post-congestion driving T


a a a a a, a b
Guofa Li , Weijian Lai , Xiaoxuan Sui , Xiaohang Li , Xingda Qu *, Tingru Zhang , Yuezhi Li
a
Institute of Human Factors and Ergonomics, College of Mechatronics and Control Engineering, Shenzhen University, Shenzhen, 518060, China
b
Laboratory of Neural Engineering, College of Mechatronics and Control Engineering, Shenzhen University, Shenzhen, 518060, China

A R T I C LE I N FO A B S T R A C T

Keywords: Traffic congestion is more likely to lead to aggressive driving behavior that is associated with increased crash
Driving safety risks. Previous studies mainly focus on driving behavior during congestion when studying congestion effects.
Driver behavior However, the negative effects of congestion on driving behavior may also affect drivers’ post-congestion driving.
Traffic congestion To fill this research gap, this study examined the influence of traffic congestion on driver behavior on the post-
Hierarchical clustering
congestion roads (i.e., the roads travelled right after congestion). Twenty-five subjects participated in a driving
Advanced driver assistance systems
simulation study. They were asked to complete two trials corresponding to post-congestion and non-congestion
conditions, respectively. Driver behavior quantified by driving performance measures, eye movement measures,
and electroencephalogram (EEG) measures was compared between the two conditions. Ten features were se-
lected from the measures with statistical significance. The selected features were integrated to characterize
drivers’ response patterns using a hierarchical clustering method. The results showed that driver behavior in
post-congestion situations became more aggressive, more focused in the forward area but less focused in the
dashboard area, and was associated with lower power of the β-band in the temporal brain region. The clustering
results showed more aggressive and lack-of-aware response patterns while driving in post-congestion situations.
This study revealed that traffic congestion negatively affected driver behavior on the post-congestion roads.
Practical implications for driving safety education was discussed based on the findings from the present study.

1. Introduction caused by congestion could last on the non-congested roads following


the congestion (Khatri, 2015; Mekker et al., 2016; Huang et al., 2018;
By 2050, 70% of the world’s population is predicted to live in cities, Paranjothi et al., 2020), namely post-congestion roads. In fact, crashes
putting increased strain on urban infrastructure and transport systems happened even more often in post-congestion situations (Mekker et al.,
(Alvarez et al., 2017). Among the resulting problems, traffic congestion 2016; Huang et al., 2018; Paranjothi et al., 2020). However, driver
will continuously be one of the most imminent ones to be solved. Zheng behavior in post-congestion situations is still unclear. To fill this re-
(2012) investigated the relationship between traffic conditions and search gap, this study was conducted to examine driver behavior in
crash occurrence likelihood and found that crash occurrence likelihood post-congestion situations. A driving simulation experiment was con-
in congested situations was six times of that in free traffic flows. The ducted, in which traffic congestion effects on driver behavior were
number was even worse (25 times) according to a report by Mekker examined while driving on post-congestion roads. Driver behavior was
et al. (2016). Besides, existing literature proved that unsafe driver be- assessed by drivers’ driving performance, eye movements and electro-
havior (e.g., long time glance away from the road ahead) was the encephalogram (EEG) responses. Meanwhile, an innovative response
leading cause of traffic crashes (Stanton and Salmon, 2009; Li et al., pattern extraction method was proposed to facilitate a comprehensive
2017). Therefore, understanding the influence of traffic congestion on understanding of drivers’ post-congestion driver behavior.
driver behavior and design assistance systems for help would enhance
traffic safety in congestion related situations. 2. Literature review
To date, researchers have mainly focused on driver behavior during
congestion when studying congestion effects (Emo et al., 2016; Qi et al., Driver behavior can be analyzed from different levels, including
2017). However, the negative emotions and aggressive driving behavior strategic behavior, tactical behavior, situation awareness behavior,


Corresponding author at: Institute of Human Factors and Ergonomics, 3688 Nanhai Avenue, Shenzhen University, Shenzhen, Guangdong Province, 518060,
China.
E-mail addresses: guofali@szu.edu.cn (G. Li), 418320384@qq.com (W. Lai), nauyoaixius@163.com (X. Sui), 512099831@qq.com (X. Li),
quxd@szu.edu.cn (X. Qu), zhangtr@szu.edu.cn (T. Zhang), yzli@sina.com (Y. Li).

https://doi.org/10.1016/j.aap.2020.105508
Received 6 November 2019; Received in revised form 12 February 2020; Accepted 13 March 2020
0001-4575/ © 2020 Elsevier Ltd. All rights reserved.
G. Li, et al. Accident Analysis and Prevention 141 (2020) 105508

operational behavior, and non-driving related behavior (Li et al., 3. Methods


2019b). Typical strategic behaviors include time-saving or short-dis-
tance route selection, traffic light running decision-making, etc. (Li 3.1. Participants
et al., 2017). Examples of tactical behaviors include tailgating, lane
changing, and approaching a lead vehicle (Toledo et al., 2008). Drivers’ Twenty-five male participants aged 20–30 years old (Mean = 24.2
visual searching behavior and physiological responses are included in years, SD = 1.9 years) were recruited from the local community. They
the situation awareness level. Typical dangerous behaviors in this level all held a valid driving license with a minimum of two-year driving
include long glances away from the forward road, not checking before experience and had normal or corrected-to-normal vision. All partici-
turning at intersections, etc. (Shinar and Compton, 2004). Drivers’ pants completed an informed consent procedure approved by the
throttle and brake pedal operations, steering behavior, and the gener- Institutional Review Board of Shenzhen University.
ated vehicle speed and acceleration responses are included in the op-
erational behaviors (Toledo et al., 2008; Li et al., 2015). Examples of 3.2. Apparatus
non-driving related behavior include cellphone use (calling or texting),
smoking, chatting with passengers while driving (Liao et al., 2016). A fixed-base driving simulator was used for the experiments. The
This paper mainly focuses on driver behavior analysis from the aspects simulated roadway and traffic scenarios were designed using UC-Win/
of situation awareness and driver operation while driving in post-con- Road Ver.8 Standard (FORUM8, Tokyo, Japan) with a refreshing rate of
gestion situations. 20 Hz. The simulator can record measures accounting for vehicle lo-
As for the driver operation aspect, studies have been conducted to cations, vehicle movement (e.g., speed, acceleration/deceleration) and
investigate how traffic congestion affects drivers’ driving performance. driver operations (e.g., steering wheel angle). Tobii Pro X3-120 (Tobii
Researchers have found that traffic congestion is able to induce ag- Technology Inc., US.) was adopted to collect drivers’ eye movement
gressive driving behavior, such as higher speeds and greater accelera- data (e.g., fixation, blink, and saccade) with a logging frequency of
tions (Hennessy and Wiesenthal, 1999; Schroeder et al., 2013; Li et al., 60 Hz. A 32-channel electrode cap and a bioelectric amplifier (SYMTOP
2019a). Emo et al. (2016) conducted a simulation experiment to ex- UEA-BZ, China) were used to collect drivers’ EEG signals with a sam-
amine the congestion effect on drivers’ risk-taking behavior. Their re- pling rate of 1000 Hz. Fig. 1(a) shows a participant driving on the si-
sults showed that traffic congestions led to more frequent tailgating and mulator wearing the EEG cap, and Fig. 1(b) presents the EEG channels.
overtaking responses. Qi et al. (2017) studied the effect of traffic con- MATLAB 2016a and EEGLab were used for data analysis. EEGLab is an
gestion on drivers’ lane change behavior and found that congestion open source toolbox widely accepted for advanced EEG signal proces-
would lead to more risky lane changes. Similarly, Huang et al. (2018) sing (Mognon et al., 2011).
analyzed how traffic congestion affected drivers’ speed choice based on
the naturalistic driving data from taxi drivers in Shanghai and reported 3.3. Simulated driving scenarios
that aggressive drivers accelerated faster in congestion situations.
Eye movement data are also frequently used to assess driver beha- Nine connected signalized intersections were constructed in the
vior from the aspect of situation awareness. It was reported that drivers’
visual searching performance would be degraded by a secondary tasks
(e.g., using cell phone for texting) while engaged in traffic congestions
(Stavrinos et al., 2013; Lansdown et al., 2015). Distraction by the sec-
ondary tasks would result in longer eyes-off-the-road durations (Zheng
et al., 2016) which would increase the crash risk (Liang et al., 2012). Qi
et al. (2013) examined drivers’ visual behavior in traffic congestion and
found that their fixations were more concentrated in the central dash-
board area. They also reported that drivers’ average fixation duration
decreased and the saccade speed increased with traffic congestion, in-
dicating that drivers had to process more traffic information for better
situation awareness in congestion situations.
As for the congestion impact on drivers’ physiological responses,
previous studies reported that traffic congestion would easily make
drivers irritable (Shinar and Compton, 2004; Huang et al., 2018) and
stressful (Hennessy and Wiesenthal, 1999; Lansdown et al., 2015),
which could compromise driving safety. Specifically, Yan et al. (2017)
proposed an emotion index based on EEG signals and found that traffic
congestion could more easily irritate drivers’ negative emotions than
normal traffic density. Further, Yan et al. (2018) used blood volume
pulse, skin conductance, and EEG features (the power percentages of δ
and β bands) to distinguish angry driving caused by traffic congestion
from normal driving. Their results showed that these physiological
features were effective indicators of congestion induced anger.
In summary, previous studies mainly focused on the congestion ef-
fects on drivers’ driving performance, eye movement, and even phy-
siological responses while driving in congestion. However, the negative
influences of congestion on driver behavior could continue on the fol-
lowing post-congestion roads (Khatri, 2015; Mekker et al., 2016;
Paranjothi et al., 2020), which has not been analyzed in previous stu-
dies. This study designed a simulator experiment to fill this research
gap.

Fig. 1. The driving apparatus.

2
G. Li, et al. Accident Analysis and Prevention 141 (2020) 105508

Fig. 2. Driving scenarios.

driving simulator with a random distance from 800 m to 1200 m be- 9.


tween two consecutive intersections. The total driving distance was Measures accounting for driving performance were steering wheel
9.8 km. The road connecting intersections had three lanes in each of the angle (S), lane offset (O), lateral acceleration (AX), velocity (V), long-
two driving directions. Participants were required to drive straight itudinal acceleration (AY), throttle (T) and brake (B) pedal opening.
across all the intersections, complying with traffic rules. The speed limit The first three measures describe the lateral performance, while the
was 60 km/h. Two traffic flow conditions, i.e., non-congestion and others describes the longitudinal performance. All these measures were
congestion, were set. Among the ten road segments illustrated in Fig. 2, frequently used in previous studies to evaluate driving performance
the second, third, sixth, and seventh segments (marked in red) were from different aspects (Li et al., 2019d; Chang et al., 2019; Bichicchi
designed with congestions in the post-congestion trial. The congestion et al., 2020). The mean value, standard deviation (std), maximum value
was defined as a failure in passing a single intersection within three (max) and Shannon entropy (sn) were computed for each measure (Li
traffic light periods. Driver behavior on the post-congestion road seg- et al., 2015; Huang et al., 2018).
ments (marked within the green boxes) was analyzed and compared As for the eye movement features, the areas of interest (AOIs) were
with that on the same road segments in the non-congestion trial. defined as illustrated in Fig. 3. The corresponding frequency and time of
The first congestion started from the beginning of the second road fixations in each AOI, and the transition frequencies between the AOIs
segment to the end of the third road segment, and the second conges- were extracted from the original eye tracking data. Other eye move-
tion started from the beginning of the sixth road segment to the end of ment features included fixation point coordinate, saccade speed, sac-
the seventh road segment (please see Fig. 2 for the illustration of cade acceleration, and saccade range. Among the eye movement mea-
driving scenarios). Each congestion lasted about 15 and 20 min. The sures, frequency, time of fixation, fixation point coordinate have been
traffic density on the non-congested roads (the blue road segments in widely used to examine the fixation of a driver’s attention while driving
Fig. 2) varied from 200 to 300 vehicles per hour. The participants had (Muñoz et al., 2016; Zhang et al., 2019), and the transition frequencies
full control of the vehicle (e.g., accelerate, decelerate, keep a constant describe the trajectories of attention across space (Muñoz et al., 2016;
speed, or change lanes) to drive as in their usual real-world driving Li et al., 2019b). The saccade measures also have been widely accepted
following the traffic rules (e.g., speed limit). to describe drivers’ rapid eyeball movements during transitions be-
tween fixations (Di Stasi et al., 2012).
3.4. Experiment procedure A band-pass filter was used to remove the noises lower than 1 Hz
and higher than 30 Hz from the raw EEG data. After that, independent
Before the experiment, each participant was given a verbal in- component analysis (ICA) was adopted to find and remove the noise
troduction of the experiment including the equipment they needed to caused by eye movements, and a z-score standardization was used to
wear, the tasks they needed to complete, the operations they could do, minimize the effect of individual differences on EEG signals. The pro-
the road speed limit, etc. After reading and signing the consent form, cessed EEG data were then transformed into the frequency domain
each participant was equipped with the EEG cap by the experiment using fast Fourier transform (FFT). The average power in each of the
assistant. Then, they were given a 10-minute practice trial to get fa- four waves (i.e., δ-band: 1–3 Hz, θ-band: 4–7 Hz, α-band: 8–13 Hz, and
miliar with the simulator operations. To minimize the order effect, 13 β-band: 14–30 Hz) was calculated. These average powers of the bands
participants started with the non-congestion trial and the rest drove the are useful indicators of drivers’ brain responses (Chen et al., 2010).
post-congestion trial first. Each simulated driving trial was required to Each participant drove through the four target road segments (road
be completed within 25 min. There was a 5-minute break between the segments 4, 5, 8, 9 in Fig. 2) in each trial. Therefore, 100 data samples
two trials. Drivers’ irritable emotion levels, assessed by a 1–5 rating (25 participants × 4 road segments) were collected for the non-con-
scale with 1 corresponding to a very low irritable emotion level and 5 gestion and post-congestion conditions, respectively. Drivers’ driving
corresponding to a very high irritable emotion level, were manually
recorded at the end of road segments (ERS) 2, 3, 6, 7 (Fig. 2) and at the
very beginning and end of each trial, respectively. Each participant was
paid 150 Chinese Yuan for their participation.

3.5. Data analysis

Driver behavior was quantitatively assessed by driving performance


measures, eye movement measures, and EEG measures. To avoid the
influence of traffic light signals on driver behaviors (e.g., acceleration
and braking), the driving performance, eye movement, and EEG data on
each road segment were collected from 100 m after passing the previous
intersection to 100 m before arriving at the next intersection. All the
examined measures were calculated covering road segment 4, 5, 8, and Fig. 3. Areas of interest (AOIs).

3
G. Li, et al. Accident Analysis and Prevention 141 (2020) 105508

performance, eye movement, and EEG measures were compared be- To visually illustrate the clustering results of high-dimensional data,
tween the two conditions using paired t-tests when the normality re- Andrews Curve was adopted to reveal the differences between the
quirement of data was satisfied. Otherwise, a nonparametric test clustered patterns. For each clustered pattern, Andrews Curve defines a
(Wann-Whiteney U-test) was used to test whether the data originated finite Fourier series as follows,
from the same distribution. The statistical significance level was se- n
Oi,1
lected to be 0.05. fi (t ) = + ∑ Oi,k ((1 − mod (k, 2)) sin ( ⌊k /2⌋∙2πt )
2 k=2

3.6. Hierarchical clustering + mod (k , 2) cos ( ⌊k /2⌋∙2πt )) (1)


where t ∈ [0, 1] and mod (k , 2) is an operation returning the remainder
The hierarchical clustering method was used to specify drivers’ re-
after division of k by 2 . Thus, each data point can be viewed as a curve
sponse patterns. Selected features used in hierarchical clustering were
line between 0 and 1. If the Andrews Curve of one clustered pattern is
from the examined measures with statistical significance in congestion
obviously different from the curve of another clustered pattern, it in-
condition comparisons. The sklearn.feature_selection.SelectKBest
dicates that the clustered patterns are effective to describe drivers’ re-
method in the scikit-learn (Pedregosa et al., 2011) library (v0.21.2) in
sponses by jointly considering multiple behavioral descriptors.
Python (v3.5) was used for the implementation of feature selection. The
available codes can be found at: https://github.com/WillKoehrsen/
feature-selector. This feature selection method uses ANOVA (analysis of 4. Results
variance) F-values to rank the features and the top ten features were
selected for the next step (Bisong, 2019). The selected features were 4.1. Driving performance, eye movement, and drivers’ physiological
used as an integrated input to characterize drivers’ response patterns response
using the hierarchical clustering method. Algorithm 1 shows the pseudo
code for the hierarchical clustering algorithm, which works by se- 4.1.1. Driving performance
quentially merging similar clusters to find the best number of clustered The changes of driving performance measures were compared be-
patterns, and returns the corresponding clustering results afterwards. tween the first and second post-congestion conditions using paired t-
As shown in Algorithm 1, Oi = {Oi,1, Oi,2, ⋯, Oi, n} is a multivariate ob- tests. No significant differences were found between these two condi-
servation n , where n is the number of selected features (n=10) and Oi, j tions, which indicates that these two conditions could be combined
represents the standardized value of the selected features j for ob- together for further analysis. Table 1 presents the results of post-con-
servation i , 1 ≤ j ≤ n . Therefore, O = {O1, O2, ⋯, Om} is the dataset in- gestion effects on driving performance. Most of the examined measures
cluding all the samples to be clustered, and m is the total number of showed statistical significance between the non-congestion and post-
samples. M (i, j ) is the calculated distance between cluster Ci and cluster congestion trials. Comparatively, both the speed and longitudinal ac-
Cj . M is the distance matrix between each two clusters. q is the number celeration were higher on the post-congestion roads than on the same
of clusters in each iteration. IC is the inconsistency coefficient of the
Table 1
clustering result in each iteration. The corresponding index k with the
Comparisons of driving performance measures between the examined trials. *
max variation between each adjacent IC values is selected to cluster the
represents statistical significance at 0.05 level (p ≤ 0.05). ** represents statis-
observations into m − k clusters. C = {C1, C2, ⋯, Cm − k } is the final tical significance at 0.01 level (p ≤ 0.01).
output of the cluster results.
Feature Non-congestion Post-congestion

Steering (10-2 rad) Smean** 0.51(0.43) 0.91(0.65)


Algorithm 1 Hierarchical clustering algorithm
Smax** 2.40(2.33) 4.92(4.13)
Sstd** 0.51(0.52) 0.95(0.71)
Input: Dataset O = {O1, O2, …, Om}
Ssn** 29.20(46.80) 67.81(84.19)
Output: Clusters C = {C1, C2, …, Cm − k }
1: for j = 1, 2, …, m do
Velocity (km/h) Vmean* 58.15(11.63) 63.40(14.50)
2: Cj = {Oj}; Vmax* 71.81(14.00) 78.64(16.50)
3: end for Vstd 8.97(5.67) 10.12(4.51)
4: for i = 1, 2, …, m do Vsn (107) −1.33(0.52) −1.35(0.44)
5: for j = 1, 2, …, m do
6: M (i, j ) = CalculateDistances(Ci, Cj ); Throttle Tmean* 0.29(0.15) 0.37(0.21)
7: M (j, i) = M (i, j ); Tmax** 0.82(0.18) 0.96(0.09)
8: end for Tstd** 0.28(0.10) 0.35(0.09)
9: end for Tsn** 56.41(31.28) 30.62(23.24)
10: Set number of current clusters: q = m;
11: z = 1; Brake Bmean* 0.017(0.024) 0.029(0.032)
12: while q > 1 do Bmax* 0.40(0.45) 0.58(0.45)
13: Find the two closest clusters Ci* and Cj* ; Bstd* 0.07(0.09) 0.11(0.10)
14: Merge Ci* and Cj : Ci = Ci ∪ Cj ; Bsn 1.65(2.61) 2.51(3.65)
* * * *
15: for j = j * + 1, j * + 2, …, q do
16: Remember Cj as Cj − 1 ; Lateral acceleration (m/s2) AXmean** 0.15(0.13) 0.30(0.24)
17: end for AXmax* 0.94(1.22) 2.43(5.06)
AXstd** 0.16(0.15) 0.34(0.31)
18: Delete row j* and column j* of M;
19: for j = 1, 2, …, q − 1 do AXsn 7.84(110.56) −277.43(1600.73)
20: M (i*, j ) = CalculateDistances(Ci , Cj );
* Longitudinal acceleration AYmean** 0.59(0.39) 0.86(0.44)
21: M (j, i*) = M (i*, j ) ; (m/s2) AYmax** 4.35(3.26) 5.92(3.25)
22: end for AYstd** 0.79(0.61) 1.12(0.63)
23: q=q−1; AYsn* -1965.6(3005.9) -3470.7(4245.1)
24: IC (z ) = CalculateInconsistencyCoefficient(C );
25: z=z+1; Offset (m) Omean 0.39(0.18) 0.43(0.17)
26: end while Omax** 1.15(0.60) 1.42(0.49)
27: k = index of the max variation of IC; Ostd** 0.26(0.12) 0.32(0.12)
28: Restore C : Use the (k − 1)th result of the merger as the final output; Osn 49.66(60.06) 22.33(70.82)
29: Return Clusters C = C1, C2, …, Cm − k

4
G. Li, et al. Accident Analysis and Prevention 141 (2020) 105508

Table 2
Dependent variables of driving performance with a Pearson correlation higher than 0.85. ** represents statistical significance at 0.01 level (p ≤ 0.01).
DV pair Correlation DV pair Correlation DV pair Correlation

AXmax & AXsn −0.931** Bmean & AYstd 0.909** Vmean & Vmax 0.879**
AYmax & AYstd 0.921** Bmean & AYsn −0.913** AYstd & AYmean 0.877**
AXmean & Smean 0.957** Bmax & AYmax 0.969** AYstd & AYsn −0.886**
Smax & AXstd 0.901** Bmax & AYstd 0.908** Smean & AXstd 0.861**
Sstd & AXstd 0.921** Bstd & AYstd 0.970** Ssn & AXstd 0.854**
Smean & Sstd 0.903** Bstd & Bmean 0.949** Ssn & Smean 0.894**
Sstd & Ssn 0.915** Bstd & Bmax 0.917** Ssn & Smax 0.890**
Tmean & Vmax 0.856** Bmean&AYmean 0.900** Bstd & AYmean 0.859**
Bstd & AYmax 0.887** Bstd & AYsn −0.895** Omax & Ostd 0.884**

roads in the non-congestion trial. The higher throttle and brake pedal respectively. The standard deviations of vertical coordinates were 0.137
opening values together with the higher maximum and standard de- and 0.113, respectively. Statistical significance on the standard devia-
viation values on the post-congestion roads indicate that the long- tion of vertical coordinates was found between the two trials
itudinal performance was more aggressive (Huang et al., 2018; Li et al., (p = 0.012).
2019d). By comparing the lateral performance measures (steering, As shown in Table 3, fixations in R3 predominated in both the non-
lateral acceleration, and offset) between the two trials, more aggres- congestion and post-congestion trials. Congestion led to 50% less fixa-
siveness can also be found in post-congestion driving with more fre- tion frequency in R4 (p < 0.001). The average fixation time in each
quent steering operations, faster lateral accelerations, and greater lane AOI in one minute is also presented in Table 3. The average fixation
offset. Three times more lane changes were observed in the post-con- time in R4 in the non-congestion trial was about two times of that in the
gestion trial than in the non-congestion trial. post-congestion trial with statistical significance (p < 0.001). The
The correlations between dependent variables (DVs) of driving average fixation time in R3 in the non-congestion trial was significantly
performance were examined using Pearson correlation coefficients. The shorter than that in the post-congestion trial (p = 0.018).
value of 0.85 was suggested as the correlation threshold when testing The comparison results of saccade measures between the two trials
the existence of multicollinearity between variables (Schroeder, 1990). are also presented in Table 3. A significantly lower saccade speed was
All the DV pairs with a Pearson correlation higher than 0.85 were found in the post-congestion trial (p = 0.014). No statistical sig-
shown in Table 2. These correlations were all with statistical sig- nificance was found on the other examined saccade measures. As for the
nificances (p ≤ 0.01). The presented results in Table 2 show that lateral transition frequencies between the AOIs, statistical significance was
acceleration measures were closely related to steering wheel angle, and found in the following transition patterns: (1) from R3 to R3
longitudinal acceleration measures were closely related to brake pedal (p = 0.015), (2) from R3 to R4 (p < 0.001), (3) from R4 to R3
operation. These results are reasonable as steering wheel and brake are (p < 0.001), (4) from R4 to R4 (p = 0.005).
the controls of lateral and longitudinal accelerations in vehicle opera- The correlations between the examined eye movement measures
tion. The correlation results could be useful for future DV selection in were also examined. All the DV pairs with a Pearson correlation higher
driving performance studies to avoid unnecessary information re- than 0.85 were shown in Table 4. These correlations were all with
dundancy (Li et al., 2017). There are few strong correlations between statistical significances (p ≤ 0.01). The presented results indicate that
throttle features and other performance variables. The correlations fixation frequency and fixation time in the peripheral vision regions
between Tmean & Vmean, Tmax & Vmax, Tstd & Vstd, Tsn & Vsn were 0.743 (R1, R2, R4, and R5) were related. The back and forth transitions be-
(p ≤ 0.01), 0.325 (p ≤ 0.01), 0.483 (p ≤ 0.01), and -0.132 (p > 0.05), tween R1 and R3 were also related as well as the round transitions
respectively. between R3 and R4, and R3 and R5.

4.1.2. Eye movement 4.1.3. EEG response


Drivers’ fixation points are plotted in Fig. 4. The origin in Fig. 4 The average powers of β-band at T3, T4 and Tp7 in the post-con-
corresponds to the upper left corner of the middle screen in the simu- gestion trial were significantly lower than the corresponding powers in
lator, and the length and width of each screen were normalized to [0, the non-congestion trial (Fig. 5). All these three electrodes belong to the
1]. Intuitively, the fixation points in the post-congestion trial were more temporal brain region. No statistical significances were found on the
concentrated. The standard deviations of the horizontal coordinates in other EEG measures. The correlation analysis between these three EEG
the non-congestion and post-congestion trials were 0.133 and 0.143, measures show that they were not related.

4.1.4. Driver emotion


Fig. 6 presents the results of drivers’ subjective ratings on their ir-
ritable emotion level at different phases during the experiment. Results
show that traffic congestion would lead to an accumulative negative
effect on drivers’ irritable emotions. When the trials ended, drivers’
subjective irritable emotion level slightly decreased.

4.2. Drivers’ response patterns

The selected features included Tmax (maximum throttle), Tstd


(standard deviation of throttle), Tsn (Shannon entropy of throttle),
AXmean (mean of lateral acceleration), Smax (maximum steering wheel
angle), Smean (mean of steering wheel angle), fixation time in R4 (R4t),
fixation transition from R4 to R3 (4to3) and from R3 to R4 (3to4), and
Fig. 4. Distribution of drivers’ fixation points. saccade speed (SCV). Six of them were from the driving performance

5
G. Li, et al. Accident Analysis and Prevention 141 (2020) 105508

Table 3
Comparisons of eye movement measures between the examined trials. The transition patterns that did not occur during the experiments were not listed in this table. *
represents statistical significance at 0.05 level (p ≤ 0.05). ** represents statistical significance at 0.01 level (p ≤ 0.01).
Feature Non-congestion Post-congestion Feature Non-congestion Post-congestion

Fixation measures Frequency in R1 5.97(7.13) 6.75(6.29) Time in R1 1.16(1.65) 1.33(1.50)


Frequency in R2 1.23(2.11) 1.17(1.91) Time in R2 0.22(0.39) 0.18(0.31)
Frequency in R3 86.94(26.10) 93.81(34.44) Time in R3* 38.07(9.05) 41.09(7.96)
Frequency in R4** 22.03(16.95) 12.26(12.12) Time in R4** 5.29(4.41) 2.57(2.46)
Frequency in R5 1.28(2.96) 1.92(3.69) Time in R5 0.26(0.57) 0.38(0.76)
Saccade measures Distance (normalized) 0.28(0.18) 0.28(0.20) Frequency (times/min) 220.63(140.85) 304.09(664.21)
Speed (normalized distance/s)* 1.43(0.84) 1.15(0.71) Time (s/min) 7.27(3.88) 8.69(14.41)
Transition frequencies From R1 to R1 2.21(3.31) 2.77(3.26) From R1 to R3 3.63(4.31) 3.95(3.62)
From R1 to R4 0.11(0.46) 0.02(0.19) From R2 to R1 0.02(0.17) 0.06(0.40)
From R2 to R2 0.58(1.37) 0.53(1.11) From R2 to R3 0.56(0.94) 0.53(0.93)
From R2 to R4 0.03(0.20) 0.05(0.34) From R3 to R1 3.31(3.88) 3.50(3.46)
From R3 to R2 0.62(0.98) 0.63(1.08) From R3 to R3* 68.15(29.42) 80.17(35.92)
From R3 to R4** 13.82(8.00) 8.17(7.11) From R3 to R5 0.92(1.77) 1.32(2.45)
From R4 to R1 0.26(0.99) 0.26(0.76) From R4 to R2 0.04(0.24) 0.02(0.19)
From R4 to R3** 13.74(7.83) 7.96(6.89) From R4 to R4** 7.94(11.78) 3.99(6.40)
From R4 to R5 0.02(0.24) 0.01(0.12) From R5 to R3 0.90(1.88) 1.38(2.53)
From R5 to R4 0.07(0.37) 0.00(0.00) From R5 to R5 0.30(1.31) 0.55(1.52)

category and the other four were from the eye movement category.
Andrews curve, an approach of dimension-reducing visualization,
was employed to qualitatively describe the degree to which the clus-
tered patterns differed with each other. The upper and lower bound-
aries for each clustered pattern indicate the 75% and 25% data range.
Based on the ten selected features, drivers’ response patterns were
clustered into two patterns in the non-congestion situation (Fig. 7(a))
and three patterns in the post-congestion situation (Fig. 7(b)).
As shown in Fig. 8(a), the two clustered patterns for non-congestion
can be described as non-congestion-cautious-aware (NCCA) and non-
congestion-aggressive-unaware (NCAU). The ratio of sample numbers
in NCCA and NCAU was about 2:1 (62:34). The NCCA pattern shows
more cautious driving operations (lower Tstd, Tmax, Smean, etc.) and
higher situation awareness (more frequent fixations transitions between
R4 and R3, etc.). The three clustered patterns for the post-congestion Fig. 5. EEG measures with statistical significance between the non-congestion
trial in Fig. 8(b) can be described as post-congestion-cautious-aware and post-congestion trials.
(CCA), post-congestion-cautious-unaware (CCU), and post-congestion-
aggressive-unaware (CAU). The ratio of sample numbers in CCA, CCU
and CAU was about 1:5:2 (12:52:22). Similar with response patterns in
non-congestion situations, the CCA pattern also shows more cautious
driving operations and higher situation awareness, and the CAU pattern
shows more aggressive driving operations and lower situation aware-
ness. The CCU pattern shows more cautious driving operations but
lower situation awareness.

5. Discussion

5.1. Driving performance

Significant increases were observed in most of the examined per-


formance measures while driving on post-congestion roads, probably Fig. 6. Drivers’ irritable emotion ratings during the experiment.
because of the increase of irritable emotion level caused by congestions

Table 4
Dependent variables of eye movement with a Pearson correlation higher than 0.85. ** represents statistical significance at 0.01 level (p ≤ 0.01).
DV pair Correlation DV pair Correlation

saccade time & saccade frequency 0.984** transition from R1 to R3 & transition from R3 to R1 0.942**
fixation frequency in R1 & fixation time in R1 0.908** transition from R3 to R3 & fixation frequency in R3 0.969**
fixation frequency in R2 & fixation time in R2 0.906** transition from R3 to R5 & fixation frequency in R5 0.922**
fixation frequency in R5 & fixation time in R5 0.918** transition from R4 to R3 & transition from R3 to R4 0.987**
fixation frequency in R4 & fixation time in R4 0.867** transition from R5 to R3 & fixation frequency in R5 0.938**
transition from R1 to R1 & fixation frequency in R1 0.907** transition from R5 to R3 & transition from R3 to R5 0.980**
transition from R1 to R3 & fixation time in R1 0.912** transition from R3 to R1 & fixation time in R1 0.885**
transition from R1 to R3 & fixation frequency in R1 0.932** transition from R3 to R1 & fixation frequency in R1 0.895**
transition from R4 to R4 & fixation frequency in R4 0.893**

6
G. Li, et al. Accident Analysis and Prevention 141 (2020) 105508

Fig. 8. The interrelationships between the selected features in the clustered


response patterns. The size of each dot indicates the standard deviation of the
feature.

show that “from R3 to R3” is the main transition pattern in the post-
Fig. 7. Illustrated clustering performance for the non-congestion and post- congestion trial, which suggests that drivers’ attention focus more on
congestion trials using Andrews Curve.
the road ahead (R3) in post-congestion than in non-congestion situa-
tions.
(Huang et al., 2018). As frequently reported in previous studies (Dula As indicated by previous studies (Ma and Kaber, 2005; Naweed,
and Geller, 2003; Wickens et al., 2011; Beck et al., 2013; Kovácsová 2013; Kim et al., 2014; Teh et al., 2014; Deng et al., 2019; Du et al.,
et al., 2016; Precht et al., 2017; Suhr and Dula, 2017; Huang et al., 2019), the frequent and greater throttle-brake pedal operations, more
2018), negative emotions led to aggressive driving behaviors (e.g., lane changes, as well as time pressure, increased drivers’ workload in
more frequent throttle-brake pedal operations, more lane changes). the post-congestion trial. The higher levels of workload caused drivers’
Similarly, these aggressive behaviors were also observed in this study, tunnel vision (Rantanen and Goldberg, 1999; Recarte and Nunes, 2003;
leading to the increased speed variation, higher speed, greater long- Young, 2012; Marquart et al., 2015), which means that drivers would
itudinal and lateral accelerations, which were associated with higher pay more attention to the forward area but less attention to the per-
driving risks and reduced driving safety (Wang et al., 2016; Li et al., iphery environment including the dashboard area. This explains why
2017). These results are in line with Huang et al. (2018) who reported the fixation points were more vertically concentrated and more allo-
that drivers’ post-congestion driving performance became aggressive cated in R3 but less in R4 in the post-congestion situations. The in-
and was more likely to be involved in crashes. Besides, Huang et al. creased transition patterns (from R3 to R3) were also related with the
(2018) also presented that higher irritable emotion level led to more R3 area. The more attention to the R3 area indicates that drivers may
aggressive behavior. Thus, the observed changes of driving perfor- not carefully examine the traffic situations in the other areas for better
mance in the post-congestion trial could be attributed to the arousal environment perception (Li and Wang, 2018). The decreased attention
negative emotion caused by traffic congestion. To improve safety in to the periphery traffic environment degraded drivers’ situation
post-congestion driving, speed monitoring and alert when necessary awareness in the post-congestion situations because vigilant drivers
could be considered on recurrent post-congestion roads (Li et al., 2020), should keep their attention on both the front and the periphery traffic
and driver emotion intervention strategies could also be useful to avoid environments (Jainta et al., 2015; Chouinard et al., 2017; Li et al.,
aggressive driving (Huang et al., 2018). 2018). Therefore, future safety enhancement strategies should consider
roadside infrastructures on recurrent post-congestion roads to alert
5.2. Eye movement drivers when there is a pedestrian.

Drivers’ eye movements show that their fixation points in the post- 5.3. Physiological responses
congestion trial were more vertically concentrated and more focused in
the R3 area but less focused in the R4 area, and the saccade speed also Significant changes of EEG measures were only observed in the
declined in post-congestion driving. The transition frequencies also average power of β-band in the temporal brain region. The temporal

7
G. Li, et al. Accident Analysis and Prevention 141 (2020) 105508

brain region is involved in perception of driving speed (Yang et al., prone to traffic crashes compared to young male drivers (Bose et al.,
2018), and the decrease of β-band power indicates a lower activation 2011; Lombardi et al., 2017). Future efforts should be made to examine
level of drivers’ brain activities (Gourab and Schmit, 2010). Usually gender-related and age-related differences in post-congestion effects.
when people are nervous or excited, the β-band power will increase The responses of drivers with different personalities would also be an
(Stevenson et al., 2011). Therefore, the decrease of β-band power in the interesting research topic (AbuAli and Abou-zeid, 2016). Besides, as
temporal brain region probably indicates that drivers’ perception of reported in previous studies (Ünal et al., 2012; van der Zwaag et al.,
speed declined when driving was affected by congestion. This partially 2012), music affects drivers’ driving performance and their physiolo-
explains why more aggressive behavior were found in the post-con- gical responses. Music intervention strategies can be considered in the
gestion trial. Similar with our results, Fairclough and Spiridon (2012) design of assistance systems to help drivers mitigate the negative effect
found a reduction of the EEG power while driving in congestion. of congestion (Zhu et al., 2016). The personality effects and effective-
It was well reported that drivers’ emotional states may only change ness of music intervention strategies are both worth investigating in the
when sudden stimulus occurred while driving on each road segment or relevant future work.
across adjacent segments (Vuilleumier et al., 2003; Gunes and Piccardi,
2007), and experimentally induced emotions could retain or only 6. Conclusion
slightly degrade with time without further emotion induction stimulus
(Tipples, 2008; Jack et al., 2014; Zhu et al., 2016). In our experiment, This study examined driver behavior in post-congestion situations
there existed no sudden emotional stimulus, and the continuous con- from the aspects of driving performance, eye movements, and EEG re-
gestion stimulus while driving on the congested roads could enhance sponses. The results showed that drivers became more aggressive (e.g.,
the congestion-induced emotion. Therefore, the emotion states results higher speed, greater accelerations, more lane changes) with degraded
could effectively reveal drivers’ emotion change during the test driving. situation awareness level (e.g., less focused in the periphery areas) and
As shown in Fig. 6, drivers’ subjective irritable emotion level con- declined speed perception (indicated by the lower power of the β-band
tinuously increased in the post-congestion trial. Similar negative effect in the temporal brain region) while driving in post-congestion than in
of congestion on driver emotion was reported in Hennessy and non-congestion situations. A more comprehensive method to extract
Wiesenthal (1997) and Morris and Hirsch (2016). Accumulative irri- drivers’ response patterns using hierarchical clustering was proposed.
table emotions in the post-congestion trial would be more likely to be The response pattern results suggested that drivers tend to adopt a more
triggered as road rage which is dangerous for driving safety (Kennedy, aggressive and lack-of-aware response patterns while driving in post-
1992; Zheng, 2012). Therefore, how to decrease drivers’ irritable congestion situations.
emotion level in congestion-related situations needs to be considered in To the best of our knowledge, the present study is one of the first
future studies for safety enhancement. Besides, as shown in Fig. 6, attempts to specifically examine driver behavior in post-congestion si-
drivers’ irritable emotion level decreased at the end of the trials. This is tuations. Our findings provide empirical evidence that congestion ad-
probably because drivers felt released from congestion at the end of the versely affects driver behavior on the post-congestion road segments,
trials. thus special attention should be paid to safe driving in post-congestion
situations. In addition, unlike previous studies, we employed drivers’
5.4. Response patterns response measures from different aspects (driving performance, eye
movement, and EEG) to comprehensively assess driver behavior, which
The response patterns illustrated in Fig. 8 show that drivers’ beha- can help better reveal the change of drivers’ behavioral patterns in-
vioral patterns in post-congestion situations were more complex than in fluenced by congestion.
non-congestion situations. About 2/3 of the collected samples in non-
congestion situations were clustered as NCCA whereas the ratio of CRediT authorship contribution statement
cautious-aware samples in post-congestion situations was only about 1/
8. This indicates that drivers are more likely to perform aggressive or Guofa Li: Conceptualization, Methodology, Writing - original draft,
unaware responses when affected by congestion. The cautious-unaware Funding acquisition. Weijian Lai: Validation, Visualization, Software,
and aggressive-unaware patterns have higher crash risks because they Formal analysis. Xiaoxuan Sui: Formal analysis, Data curation,
are associated with more risky driving behavior (e.g., speeding, faster Investigation. Xiaohang Li: Data curation. Xingda Qu: Supervision,
acceleration) (Shi et al., 2010; Li et al., 2017). For safety driving, the Writing - review & editing. Tingru Zhang: Writing - review & editing.
NCCA and CCA patterns are preferred for strategies development on Yuezhi Li: Resources.
driver assistance and autonomous driving systems (Li et al., 2019c).
Based on the above arguments, we understand that even if the Declaration of Competing Interest
congestion ended, drivers were still negatively affected. This requires
actions to be taken to minimize the adverse effects of congestion on None.
driving on the post-congestion roads. Some practical implications could
be derived from the findings from the present study. Firstly, more traffic Acknowledgements
signs (e.g., speed limit) and surveillance cameras could be deployed to
restrain drivers’ aggressive driving patterns on the daily recurrent post- This study is supported by the National Natural Science Foundation
congestion roads. Secondly, driver training programs can be designed to of China (grant number 51805332), the Natural Science Foundation of
help drivers understand the negative effects of congestion, and learn Guangdong Province (grant number 2018A030310532), the Shenzhen
and practice the safe driving patterns (NCCA and CCA patterns) in post- Fundamental Research Fund (grant number
congestion situations. Offline feedback, which has been proved to be JCYJ20190808142613246), and the Young Elite Scientists Sponsorship
effective in reducing crash rates and driving risk (Toledo et al., 2008), Program by China Society of Automotive Engineers.
could be incorporated into the training programs for continuous safe
driving pattern training and education. References

5.5. Limitations and future work AbuAli, N., Abou-zeid, H., 2016. Driver behavior modeling: developments and future
directions. Int. J. Veh. Technol. 2016, 1–12.
A limitation of this study is that only young (18–30 years old) male Alvarez, P., Lerga, I., Serrano, A., Faulin, J., 2017. Considering congestion costs and
driver behaviour into route optimisation algorithms in smart cities. The Second
drivers were recruited. Young female drivers and older drivers are more

8
G. Li, et al. Accident Analysis and Prevention 141 (2020) 105508

International Conferences on Smart Cities (Smart-CT) 39–50. Li, G., Zhu, F., Qu, X., Cheng, B., Li, S., Green, P., 2019d. Driving style classification based
Beck, K.H., Daughters, S.B., Ali, B., 2013. Hurried driving: relationship to distress toler- on driving operational pictures. IEEE Access 7, 90180–90189.
ance, driver anger, aggressive and risky driving in college students. Accid. Anal. Prev. Li, S., Li, G., Cheng, Y., Ran, B., 2020. Urban arterial traffic status detection using cellular
51, 51–55. data without cellphone GPS information. Transp. Res. Part C Emerg. Technol. 114,
Bichicchi, A., Belaroussi, R., Simone, A., Vignali, V., Lantieri, C., Li, X., 2020. Analysis of 446–462.
road-user interaction by extraction of driver behavior features using deep learning. Li, S.E., Li, G., Yu, J., Liu, C., Cheng, B., Wang, J., Li, K., 2018. Kalman filter-based
IEEE Access 8, 19638–19645. tracking of moving objects using linear ultrasonic sensor array for road vehicles.
Bisong, E., 2019. More supervised machine learning techniques with scikit-learn. Building Mech. Syst. Signal Process. 98, 173–189.
Machine Learning and Deep Learning Models on Google Cloud Platform. Apress, Liang, Y., Lee, J.D., Yekhshatyan, L., 2012. How dangerous is looking away from the
Berkeley, CA, pp. 287–308. road? Algorithms predict crash risk from glance patterns in naturalistic driving. Hum.
Bose, D., Segui-Gomez, M., Crandall, J.R., 2011. Vulnerability of female drivers involved Factors 54 (6), 1104–1116.
in motor vehicle crashes: an analysis of US population at risk. Am. J. Public Health Liao, Y., Li, S.E., Wang, W., Wang, Y., Li, G., Cheng, B., 2016. Detection of driver cog-
101 (12), 2368–2373. nitive distraction: a comparison study of stop-controlled intersection and speed-
Chang, F., Xu, P., Zhou, H., Lee, J., Huang, H., 2019. Identifying motorcycle high-risk limited highway. IEEE Trans. Intell. Transp. Syst. 17 (6), 1628–1637.
traffic scenarios through interactive analysis of driver behavior and traffic char- Lombardi, D.A., Horrey, W.J., Courtney, T.K., 2017. Age-related differences in fatal in-
acteristics. Transp. Res. Part F Traffic Psychol. Behav. 62, 844–854. tersection crashes in the United States. Accid. Anal. Prev. 99, 20–29.
Chen, Y.-C., Duann, J.-R., Chuang, S.-W., Lin, C.-L., Ko, L.-W., Jung, T.-P., Lin, C.-T., Ma, R., Kaber, D.B., 2005. Situation awareness and workload in driving while using
2010. Spatial and temporal EEG dynamics of motion sickness. NeuroImage 49 (3), adaptive cruise control and a cell phone. Int. J. Ind. Ergon. 35, 939–953.
2862–2870. Marquart, G., Cabrall, C., de Winter, J., 2015. Review of eye-related measures of drivers’
Chouinard, P.A., Peel, H.J., Landry, O., 2017. Eye-tracking reveals that the strength of the mental workload. Procedia Manuf. 3, 2854–2861.
vertical-horizontal illusion increases as the retinal image becomes more stable with Mekker, M.M., Remias, S.M., McNamara, M.L., Bullock, D.M., 2016. Characterizing in-
fixation. Front. Hum. Neurosci. 11 (143), 1–14. terstate crash rates based on traffic congestion using probe vehicle data (paper No.
Deng, T.-M., Fu, J., Shao, Y.-M., Peng, J., Xu, J., 2019. Pedal operation characteristics and 16-1194). The 95th Annual Meeting of the Transportation Research Board.
driving workload on slopes of mountainous road based on naturalistic driving tests. Mognon, A., Jovicich, J., Bruzzone, L., Buiatti, M., 2011. ADJUST: an automatic EEG
Saf. Sci. 119, 40–49. artifact detector based on the joint use of spatial and temporal features: automatic
Di Stasi, L.L., Renner, R., Catena, A., Cañas, J.J., Velichkovsky, B.M., Pannasch, S., 2012. spatio-temporal EEG artifact detection. Psychophysiology 48 (2), 229–240.
Towards a driver fatigue test based on the saccadic main sequence: a partial vali- Morris, E.A., Hirsch, J.A., 2016. Does rush hour see a rush of emotions? Driver mood in
dation by subjective report data. Transp. Res. Part C Emerg. Technol. 21 (1), conditions likely to exhibit congestion. Travel Behav. Soc. 5, 5–13.
122–133. Muñoz, M., Reimer, B., Lee, J., Mehler, B., Fridman, L., 2016. Distinguishing patterns in
Du, N., Haspiel, J., Zhang, Q., Tilbury, D., Pradhan, A.K., Yang, X.J., Robert, L.P., 2019. drivers’ visual attention allocation using Hidden Markov Models. Transportation
Look who’s talking now: implications of AV’s explanations on driver’s trust, AV Rescanning Part F: Traffic Psychology and Behaviour 43, 90–103.
preference, anxiety and mental workload. Transp. Res. Part C Emerg. Technol. 104, Naweed, A., 2013. Psychological factors for driver distraction and inattention in the
428–442. Australian and New Zealand rail industry. Accident Analysis & Prevention 60,
Dula, C.S., Geller, E.S., 2003. Risky, aggressive, or emotional driving: addressing the need 193–204.
for consistent communication in research. J. Safety Res. 34 (5), 559–566. Paranjothi, A., Khan, M.S., Patan, R., Parizi, R.M., Atiquzzaman, M., 2020. VANETomo: a
Emo, A.K., Matthews, G., Funke, G.J., 2016. The slow and the furious: anger, stress and congestion identification and control scheme in connected vehicles using network
risky passing in simulated traffic congestion. Transp. Res. Part F Traffic Psychol. tomography. Comput. Commun. 151, 275–289.
Behav. 42, 1–14. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., et al., 2011.
Fairclough, S.H., Spiridon, E., 2012. Cardiovascular and electrocortical markers of anger Scikit-learn: machine learning in Python. J. Mach. Learn. Res. 12, 2825–2830.
and motivation during a simulated driving task. Int. J. Psychophysiol. 84 (2), Precht, L., Keinath, A., Krems, J.F., 2017. Effects of driving anger on driver behavior –
188–193. results from naturalistic driving data. Trans. Res. Part F Traf. Psychol. Behav. 45,
Gourab, K., Schmit, B.D., 2010. Changes in movement-related β-band EEG signals in 75–92.
human spinal cord injury. Clin. Neurophysiol. 121 (12), 2017–2023. Qi, W., Pei, Y., Song, M., Bie, Y., 2013. Pattern analysis of driver’s “pressure-state-re-
Gunes, H., Piccardi, M., 2007. Bi-modal emotion recognition from expressive face and sponse” in traffic congestion. Discrete Dyn. Nat. Soc. 2013, 1–11.
body gestures. J. Netw. Comput. Appl. 30 (4), 1334–1345. Qi, W., Wen, H., Wu, Y., Qin, L., 2017. Effect model of urban traffic congestion on driver’s
Hennessy, D.A., Wiesenthal, D.L., 1997. The relationship between traffic congestion, lane-changing behavior. Adv. Mech. Engin. 9 (9), 1–12.
driver stress and direct versus indirect coping behaviours. Ergonomics 40 (3), Rantanen, E.M., Goldberg, J.H., 1999. The effect of mental workload on the visual field
348–361. size and shape. Ergonomics 42 (6), 816–834.
Hennessy, D.A., Wiesenthal, D.L., 1999. Traffic congestion, driver stress, and driver ag- Recarte, M.A., Nunes, L.M., 2003. Mental workload while driving: effects on visual
gression. Aggress. Behav. 25 (6), 409–423. search, discrimination, and decision making. J. Exp. Psychol. Appl. 9 (2), 119–137.
Huang, Y., Sun, D.(Jian), Zhang, L.-H., 2018. Effects of congestion on drivers’ speed Schroeder, M.A., 1990. Diagnosing and dealing with multicollinearity. Western Journal of
choice: assessing the mediating role of state aggressiveness based on taxi floating car Nursing Research 12 (2), 175–187.
data. Accid. Anal. Prev. 117, 318–327. Schroeder, P., Kostyniuk, L., Mack, M., 2013. National Survey of Speeding Attitudes and
Jack, R.E., Garrod, O.G.B., Schyns, P.G., 2014. Dynamic facial expressions of emotion Behaviors. Report No. DOT HS-811-865. Department of Transportation, NHTSA, U.S.
transmit an evolving hierarchy of signals over time. Curr. Biol. 24 (2), 187–192. Shi, J., Bai, Y., Ying, X., Atchley, P., 2010. Aberrant driving behaviors: a study of drivers
Jainta, S., Blythe, H.I., Nikolova, M., Jones, M.O., Liversedge, S.P., 2015. A comparative in Beijing. Accid. Anal. Prev. 42, 1031–1040.
analysis of vertical and horizontal fixation disparity in sentence reading. Vision Res. Shinar, D., Compton, R., 2004. Aggressive driving: an observational study of driver, ve-
110, 118–127. hicle, and situational variables. Accid. Anal. Prev. 36, 429–437.
Kennedy, H.G., 1992. Anger and irritability. Br. J. Psychiatry 161 (2), 145–153. Stanton, N.A., Salmon, P.M., 2009. Human error taxonomies applied to driving: a generic
Khatri, C.P., 2015. Real-time Road Traffic Information Detection Through Social Media driver error taxonomy and its implications for intelligent transport systems. Saf. Sci.
(master Thesis). Georgia Institute of Technology, Atlanta, GA, United States. 47 (2), 227–237.
Kim, H.S., Hwang, Y., Yoon, D., Choi, W., Park, C.H., 2014. Driver workload character- Stavrinos, D., Jones, J.L., Garner, A.A., Griffin, R., Franklin, C.A., Ball, D., Welburn, S.C.,
istics analysis using EEG data from an urban road. IEEE Trans. Intell. Transp. Syst. 15 Ball, K.K., Sisiopiku, V.P., Fine, P.R., 2013. Impact of distracted driving on safety and
(4), 1844–1849. traffic flow. Accid. Anal. Prev. 61, 63–70.
Kovácsová, N., Lajunen, T., Rošková, E., 2016. Aggression on the road: relationships Stevenson, C.M., Brookes, M.J., Morris, P.G., 2011. β-band correlates of the fMRI BOLD
between dysfunctional impulsivity, forgiveness, negative emotions, and aggressive response. Hum. Brain Mapp. 32 (2), 182–197.
driving. Transp. Res. Part F Traffic Psychol. Behav. 42, 286–298. Suhr, K.A., Dula, C.S., 2017. The dangers of rumination on the road: predictors of risky
Lansdown, T.C., Stephens, A.N., Walker, G.H., 2015. Multiple driver distractions: a sys- driving. Accid. Anal. Prev. 99, 153–160.
temic transport problem. Accid. Anal. Prev. 74, 360–367. Teh, E., Jamson, S., Carsten, O., Jamson, H., 2014. Temporal fluctuations in driving
Li, G., Li, S.E., Cheng, B., Green, P., 2017. Estimation of driving style in naturalistic demand: the effect of traffic complexity on subjective measures of workload and
highway traffic using maneuver transition probabilities. Transp. Res. Part C Emerg. driving performance. Trans. Res. Part F Traf. Psychol. Behav. 22, 207–217.
Technol. 74, 113–125. Tipples, J., 2008. Negative emotionality influences the effects of emotion on time per-
Li, G., Li, S.E., Liao, Y., Wang, W., Cheng, B., Chen, F., 2015. Lane change maneuver ception. Emotion 8 (1), 127–131.
recognition via vehicle state and driver operation signals—Results from naturalistic Toledo, T., Musicant, O., Lotan, T., 2008. In-vehicle data recorders for monitoring and
driving data. The 2015 IEEE Intelligent Vehicles Symposium (IV) 865–870. feedback on drivers’ behavior. Trans. Res. Part C Emerg. Technol. 16 (3), 320–331.
Li, G., Li, S.E., Zou, R., Liao, Y., Cheng, B., 2019a. Detection of road traffic participants Ünal, A.B., Steg, L., Epstude, K., 2012. The influence of music on mental effort and driving
using cost-effective arrayed ultrasonic sensors in low-speed traffic situations. Mech. performance. Accid. Anal. Prev. 48, 271–278.
Syst. Signal Process. 132, 535–545. van der Zwaag, M.D., Dijksterhuis, C., de Waard, D., Mulder, B.L.J.M., Westerink,
Li, G., Wang, Y., 2018. Surface transportation. In: Samman, S.N. (Ed.), Human Factors and J.H.D.M., Brookhuis, K.A., 2012. The influence of music on mood and performance
Ergonomics for the Gulf Cooperation Council: Processes, Technologies, and Practices. while driving. Ergonomics 55 (1), 12–22.
CRC Press, pp. 191–210. Vuilleumier, P., Armony, J.L., Driver, J., Dolan, R.J., 2003. Distinct spatial frequency
Li, G., Wang, Y., Zhu, F., Sui, X., Wang, N., Qu, X., Green, P., 2019b. Drivers’ visual sensitivities for processing faces and emotional expressions. Nat. Neurosci. 6 (6),
scanning behavior at signalized and unsignalized intersections: a naturalistic driving 624–631.
study in China. J. Safety Res. 71, 219–229. Wang, X., Fan, T., Li, W., Yu, R., Bullock, D., Wu, B., Tremont, P., 2016. Speed variation
Li, G., Yang, Y., Qu, X., 2019c. Deep learning approaches on pedestrian detection in hazy during peak and off-peak hours on urban arterials in Shanghai. Trans. Res. Part C
weather. IEEE Trans. Ind. Electron. https://doi.org/10.1109/TIE.2019.2945295. Emerg. Technol. 67, 84–94.

9
G. Li, et al. Accident Analysis and Prevention 141 (2020) 105508

Wickens, C.M., Wiesenthal, D.L., Flora, D.B., Flett, G.L., 2011. Understanding driver anger 326–342.
and aggression: attributional theory in the driving environment. J. Exp. Psychol. Zhang, X., Li, G., Xue, H., Zhao, H., 2019. Pilots’ scanning behavior between different
Appl. 17 (4), 354–370. airport intersection maneuvers in a simulated taxiing task. IEEE Access 7,
Yan, J., Tao, P., Shao, P., Xiu, W., Wang, L., 2017. Research on the dynamic changes of 150395–150402.
drivers’ emotion in long tunnel base on EEG. 2017 Chinese Automation Congress Zheng, R., Nakano, K., Ishiko, H., Hagita, K., Kihira, M., Yokozeki, T., 2016. Eye-gaze
(CAC) 6943–6946. tracking analysis of driver behavior while interacting with navigation systems in an
Yan, L., Wan, P., Qin, L., Zhu, D., 2018. The induction and detection method of angry urban area. IEEE Trans. Hum.-Mach. Syst. 46 (4), 546–556.
driving: evidences from EEG and physiological signals. Discrete Dynamics in Nature Zheng, Z., 2012. Empirical analysis on relationship between traffic conditions and crash
and Society 2018, 1–16. occurrences. Proc. Soc. Behav. Sci. 43, 302–312.
Yang, L., He, Z., Guan, W., Jiang, S., 2018. Exploring the relationship between electro- Zhu, Y., Wang, Y., Li, G., Guo, X., 2016. Recognizing and releasing drivers’ negative
encephalography (EEG) and ordinary driving behavior: a simulated driving study. emotions by using music: evidence from driver anger. The 8th International
Transp. Res. Rec. 2672 (37), 172–180. Conference on Automotive User Interfaces and Interactive Vehicular Applications
Young, R., 2012. Cognitive distraction while driving: a critical review of definitions and (AutomotiveUI’ 16) 173–178.
prevalence in crashes. SAE Int. J. Passenger Cars Electron. Electr. Syst. 5 (1),

10

You might also like