Journal of Safety Research 71 (2019) 219–229

Contents lists available at ScienceDirect

Journal of Safety Research

journal homepage: [Link]/locate/jsr

Drivers’ visual scanning behavior at signalized and unsignalized

intersections: A naturalistic driving study in China
Guofa Li a,b, Ying Wang c,d, Fangping Zhu a, Xiaoxuan Sui a, Ning Wang d, Xingda Qu a,⇑, Paul Green e
a
Institute of Human Factors and Ergonomics, College of Mechatronics and Control Engineering, Shenzhen University, Shenzhen 518060, China
b
State Key Laboratory of Automotive Safety and Energy, Department of Automotive Engineering, Tsinghua University, Beijing 100084, China
c
Ipsos, User Experience, Chicago, IL 60601, USA
d
School of Transportation Science and Engineering, Beijing Key Laboratory for Cooperative Vehicle Infrastructure Systems and Safety Control, Beihang University, Beijing 100191,
China
e
University of Michigan Transportation Research Institute (UMTRI) & Department of Industrial and Operations Engineering, University of Michigan, Ann Arbor, MI 48109, USA

a r t i c l e i n f o a b s t r a c t

Article history: Introduction: Intersections are the most dangerous locations in urban traffic. The present study aims to
Received 25 February 2019 investigate drivers’ visual scanning behavior at signalized and unsignalized intersections. Method:
Received in revised form 15 July 2019 Naturalistic driving data at 318 green phase signalized intersections and 300 unsignalized ones were col-
Accepted 27 September 2019
lected. Drivers’ glance allocations were manually categorized into 10 areas of interest (AOIs), based on
Available online 20 November 2019
which three feature subsets were extracted including glance allocation frequencies, durations and AOI
transition probabilities. The extracted features at signalized and unsignalized intersections were com-
Keywords:
pared. Features with statistical significances were integrated to characterize drivers’ scanning patterns
Intersection safety
Visual scanning behavior
using the hierarchical clustering method. Andrews Curve was adopted to visually illustrate the clustering
Signalized intersection results of high-dimensional data. Results: Results showed that drivers going straight across signalized
Unsignalized intersection intersections had more often glances at the left view mirror and longer fixation on the near left area.
Hierarchical clustering When turning left, drivers near signalized intersections had more frequent glances at the left view mirror,
fixated much longer on the forward and rearview mirror area, and had higher transition probabilities
from near left to far left. Compared with drivers’ scanning patterns in left turning maneuver at signalized
intersections, drivers with higher situation awareness levels would divide more attention to the forward
and right areas than at unsignalized intersections. Conclusions: This study revealed that intersection types
made differences on drivers’ scanning behavior. Practical applications: These findings suggest that future
applications in advanced driver assistance systems and driver training programs should recommend dif-
ferent scanning strategies to drivers at different types of intersections.
Ó 2019 National Safety Council and Elsevier Ltd. All rights reserved.

1. Introduction Besides poor intersection design, intersection crashes may also

be caused by drivers’ inappropriate behaviors from multiple
Intersections have been identified as the most dangerous loca- aspects in the driver-vehicle-environment system. Drivers’ behav-
tions in urban traffic due to the higher situation awareness iors can be analyzed at different levels, including strategic behav-
requirement caused by the complex traffic environment (Richard, ior, tactical behavior, situation awareness behavior, and
Campbell, & Brown, 2006; Tay, 2015). Up to 60% of severe injuries operational behavior (Fig. 1). Typical strategic behaviors include
in traffic accidents occurred at intersections in European countries traffic light running and decision making in the dilemma zone
(Werneke & Vollrath, 2014; Lemonnier, Brémond, & Baccino, 2015). (Richard et al., 2006). Examples of dangerous tactical behaviors
In the United States, 47% of car crashes occurred at intersections or include tailgating, speeding, and not-yielding to pedestrians (Li,
intersection related junctions in 2015, and these intersection Li, Cheng, & Green, 2017). At the situation awareness level, behav-
crashes contributed to over 50% of crash injuries (NHTSA, 2017). iors of long glances away from the forward scene, not checking
traffic situations before entering intersections, and turning without
⇑ Corresponding author at: Institute of Human Factors and Ergonomics, 3688
checking was reported to compromise driving safety at intersec-
Nanhai Avenue, Shenzhen University, Shenzhen, Guangdong Province 518060, tions (Shinar, 2008). Examples of dangerous operational behaviors
China. include strong acceleration, hard braking, sharp steering, and turn-
E-mail address: quxd@[Link] (X. Qu). ing without signaling (Toledo, Musicant, & Lotan, 2008).

[Link]
0022-4375/Ó 2019 National Safety Council and Elsevier Ltd. All rights reserved.
220 G. Li et al. / Journal of Safety Research 71 (2019) 219–229

Fig. 1. A framework on the analysis of driver behavior at intersections.

Secondary-task behaviors like phone-use, texting, eating, and ized and unsignalized intersections were compared to examine
smoking also contribute to increased risks of accidents while driv- the influence of intersection types on drivers’ scanning measures.
ing through intersections (Klauer et al., 2014). Among these Besides, an unsupervised machine learning approach (i.e., hierar-
aspects, understanding drivers’ visual scanning behavior is critical chical cluster analysis), was used to identify scanning patterns
for the improvement of intersection safety from the situation based on the examined measures with statistical significances.
awareness aspect. It was found in Owsley, Ball, Sloane, Roenker, The contributions of this study include: (1) Differences on drivers’
and Bruni (1991) that drivers with scanning problems had 4.2 visual scanning behavior between signalized and unsignalized
times more collisions and 15.6 more intersection accidents than intersections are examined; (2) A novel way to correlate drivers’
those without this problem. visual scanning measures is proposed for a more comprehensive
Existing studies have intensively analyzed drivers’ visual scan- understanding on their behavioral patterns.
ning behavior at intersections when considering the effects of
age and driving experience (Feng et al., 2018; Scott, Hall,
Litchfield, & Westwood, 2013). Although the effects of age and 2. Related work
driving experience are important topics for driving safety, we
noted that intersections could be either signalized or unsignalized. Driver behavior has been well studied at either signalized or
Drivers may adopt different visual scanning strategies under the unsignalized intersections. Dukic and Broberg (2012), for exam-
signalized and unsignalized intersection conditions. Previous stud- ple, conducted a field test to identify drivers’ visual scanning
ies mainly focus on drivers’ visual scanning behavior at signalized behavior while driving through signalized intersections. Their
or unsignalized intersections. Few have revealed the differences results showed that compared with younger drivers, older drivers
between signalized and unsignalized intersections. In fact, it was looked more at road lines and markings to position themselves in
reported that the number of crashes occurring near signalized the surrounding traffic. Carter, Hummer, Foyle, and Phillips
intersections was almost twice of that near unsignalized intersec- (2005) analyzed the following headways in exclusive left turn
tions (NHTSA, 2017; Yamani, Horrey, Liang, & Fisher, 2016). The lanes at 14 signalized intersections, and found that intersections
higher crash risk at signalized intersections could possibly be with double left-turn lanes would shorten the headways to
attributed to the differences of drivers’ scanning behavior at signal- involve more U-turn crashes. Wu and Xu (2017) examined driver
ized and unsignalized intersections (Tay, 2015). Therefore, there is operational behavior for right turn maneuvers at signalized inter-
an urgent need to identify such differences. sections using naturalistic driving data. They reported that drivers
The objective of this study was to investigate drivers’ visual had higher acceleration frequency when turning on red than on
scanning behavior at signalized and unsignalized intersections by green.
examining drivers’ visual scanning behavior using a multifaceted At unsignalized intersections, a study conducted by Bao and
representation of the measures (e.g., glance frequency and dura- Boyle (2009) found that old drivers focused on the straight ahead
tion in each area of interest (AOI)) both independently and jointly. region more intensively, but paid significantly less attention to
Naturalistic driving data were collected. Driver’s glance allocation the left and right regions. Similarly, Yamani et al. (2016) reported
frequencies, durations, and AOI transition probabilities at signal- that old drivers glanced less at the potential threats on roadsides
 G. Li et al. / Journal of Safety Research 71 (2019) 219–229 221

while turning at unsignalized intersections. Scott et al. (2013) con- 3. Methods

ducted a simulation experiment for right turn maneuver analysis
at unsignalized intersections. Two unique glance transition pat- 3.1. Apparatus
terns were observed for young novice drivers including from cen-
ter to near left area and from middle left to near left. Romoser, A driving recorder (GoSafe360, PAPAGO) was mounted on the
Pollatsek, Fisher, and Williams (2013) reported that at unsignalized windshield of a vehicle (Toyota Land Cruiser). The recording frame
intersections younger drivers spent more time looking at the cen- rate was 30 frames per second. Two cameras were included in the
tral area while turning left but less time while going straight. recorder. One was used to monitor the driving environment and
Lemonnier et al. (2015) found that while passing unsignalized help determine if the intersections were signalized or unsignalzied.
intersections, driver’s lane attracted the most attention from the The other one was used to record drivers’ scanning behavior. To
drivers, with dashboard ranking the second. Liu, Chen, Lu, and avoid possible interference, the sound option was turned off during
Wang (2017) modeled the relationship between drivers’ speed naturalistic driving data collection.
choices and their risk perception while going straight at unsignal-
ized intersections. Vehicle speed was precisely estimated based on 3.2. Participants
the proposed model and the crash risk with other vehicles was pre-
dicted. Mollu et al. (2018) investigated the effect of different prior- Ten licensed drivers (eight males and two females) aging from
ity rules on drivers’ speed choice at unsignalized intersections. 20 to 45 years old (Mean ± SD = 32.0 ± 6.1 years) recruited from
Results showed that speed was significantly decreased when the the university and local community participated in this study.
priority of right rules was indicated. Their driving experience ranged from 2 to 17 years (Mean ± -
Although the above-mentioned studies have analyzed driver SD = 6.9 ± 4.1 years). No accidental events were reported by the
behavior at either signalized or unsignalized intersections, few participants.
have revealed driver behavior differences between signalized
and unsignalized intersections. Intersection types affect driver 3.3. Experiment design and procedure
behavior. The number of crashes occurring near signalized inter-
sections was almost twice of that near unsignalized intersections Verbal instructions regarding the study protocol were given to
(NHTSA, 2017; Yamani et al., 2016). The higher crash risk at sig- the participants before data collection. After signing the consent
nalized intersections could possibly be attributed to the differ- form, the participants were told to drive anywhere they want in
ences of drivers’ scanning behavior at signalized and the city traffic of Beijing at their own preference. No assistant
unsignalized intersections (Tay, 2015). To our best knowledge, was present in the vehicle throughout the study to avoid interfer-
only two studies reported driver behavior differences between ence with naturalistic driving data collection. All the data were col-
signalized and unsignalized intersections. Specifically, Liu and lected on dry roads and non-peak hours in clear days to avoid
Ozguner (2007) proposed a general architecture of human driver possible interference by these factors.
model at both signalized and unsignalized intersections, which
can simulate drivers’ decision making and their operational 3.4. Data collection and reduction
responses. Their simulation results showed that driver behavior
at unsignalized intersections involved more decision making Drivers’ visual glance behavior while executing an intersection
and stopping behavior than at signalized intersection where fol- maneuver was assessed in 10 AOIs, including forward (FW), near
lowing the traffic light is the primary rule. Another study have left (NL), near right (NR), far left (FL), far right (FR), rearview mirror
shown that older drivers’ over-involvement in either left turn, (RM), left view mirror (LVM), right view mirror (RVM), clusters
right turn or straight across maneuvers at unsignalized intersec- (CL), and others (OT). See Fig. 2 for a general overview of the AOIs.
tions (Lococo & Staplin, 2018). Both studies did not examine dri- Timestamps were used to extract the duration of each glance by
vers’ scanning patterns, which is the most important for risk computing the absolute time difference between the current allo-
perception (Owsley et al., 1991; Li & Wang, 2018). Therefore, cation and the next. Glance frequencies, durations, and glance tran-
we conducted this study to fill the knowledge gap. The identified sitions were computed based on the extracted glance episodes, as
differences would help better understand the inter-relationship described in Section 3.5.
between visual scanning measures (i.e., glance allocation frequen- Videos recorded by the cameras were reviewed for the data
cies, durations and AOI transition probabilities) and further con- about the orientation and movement of drivers’ eyes and head.
tribute to the knowledge on drivers’ visual scanning behavior Glance allocations (location and timestamp) were then manually
for a better design of intersection safety enhancement systems. assessed and extracted by two coders and mediated by a third,
Table 1 summarizes the relevant studies on driver behavior at and binned to the 10 AOIs, following the naturalistic driving data
intersections. coding scheme proposed by researchers from MIT (Fridman,

Table 1
A brief summary of the driver behavior studies at intersections.

Authors Publication year Driver behavior Intersection type Research topic

Carter et al. 2005 Following headway signalized Driver operation behavior in U-turns
Dukic & Broberg 2012 Scanning signalized Older drivers’ scanning behavior
Wu and Xu 2017 Operation signalized Driver operational behavior for right turns
Bao & Boyle 2009 Scanning unsignalized Age-related differences in scanning
Scott et al. 2013 Scanning unsignalized Glance transitions for drivers with different age and experience
Romoser et al. 2013 Scanning unsignalized Age-related differences in scanning
Lemonnier et al. 2015 Scanning unsignalized Visual attention
Liu et al. 2017 Operation unsignalized Drivers’ crossing behavior models
Mollu et al. 2018 Operation unsignalized Speed performance under different priority rules
Liu & Ozguner 2007 Operation both Driver behavior model
Lococo & Staplin 2018 Scanning both Scanning behavior of older drivers
222 G. Li et al. / Journal of Safety Research 71 (2019) 219–229

Fig. 2. Areas of interest (AOIs) and the corresponding snapshot.

Langhans, Lee, Reimer, 2016) and Virginia Tech (Kusano, Chen, glances transit from Ax to Ay . The glance allocations were defined
Montgomery, & Gabler, 2015). The type of intersections (signalized as follows:
or unsignalized) and passing maneuvers (i.e., left turn, right turn
and going straight) were also recorded. The two coders had rich
experience in subjective evaluation of driving behavior. They all
had a valid driver license and had been involved in driving behav-
ior analysis and related projects for more than two years. Accord- ð1Þ
ing to Fridman, Langhans, et al. (2016), where drivers were looking
at could be identified by their head and eye orientations and their The transition probability from Ax to Ay for a maneuver was
operations. Therefore, the coders classified drivers’ glance alloca- denoted by:


tions in each frame segment mainly according to the observed dri-
 x Ay jAx
vers’ head and eye orientations and driving operations. After axy ¼ p Ay jAx ¼ PN
 ð2Þ
finishing their initial evaluation, the coders checked their results i¼1 x Ay jAx

with each other. When there was a conflict between the results, where x was the number of transition events, axy 2 ½0; 1, N was the
a third colder who had more than 10 years research experience number of AOIs (N = 10), 1  x; y  N. The transition matrix was
in driving behavior analysis would be involved to finally determine  
defined as A ¼ axy . Smaller axy suggests lower probability of the
the AOI.
transition from Ax to Ay : The transition probabilities indicate the
As most of the intersection crashes occur while a vehicle is
glance transition preferences while passing signalized or unsignal-
moving, this study mainly examined drivers’ visual scanning
ized intersections.
behavior in the green light phase. The data collected from 15 s
The above-mentioned scanning measures were compared
before arriving at the stop line to 5 s after entering the target
between signalized and unsignalized intersections using two-
branch were used for analysis.
sample t-tests. Such comparisons were conducted for the three
driving maneuvers (i.e., straight across, turning left, and turning
3.5. Visual scanning measures right), separately. The statistical significance level was selected to
be 0.05.
Features employed to describe drivers’ visual scanning behavior
include glance frequency and duration in each AOI, and glance 3.6. Hierarchical clustering
transition probabilities between AOIs. By building gaze concentra-
tion profiles per maneuver, the glance frequency in each AOI was The above-mentioned scanning measures with statistical signif-
counted for each collected intersection episode. As per SAE J2396 icances were selected as an integrated input to characterize dri-
(2000), the duration of each glance was computed by taking the vers’ scanning patterns using hierarchical clustering method. See
absolute time difference between its timestamp and that of the Algorithm 1 for the workflow of the hierarchical clustering. It
next glance. works by sequentially merging similar clusters to find the best
When calculating glance transition probabilities between AOIs, number of clustered patterns, and returns the corresponding clus-
we used Ax and Ay to represent the former and the latter glance tering results afterwards. As shown in Algorithm 1,
 
allocations, respectively. AOI transition counts were extracted Oi ¼ Oi;1 ; Oi;2 ;    ; Oi;n is a multivariate observation Rn , where n
from the annotated maneuvers, illustrating how often driver indicates the number of selected measures and Oi;j indicates the
 G. Li et al. / Journal of Safety Research 71 (2019) 219–229 223

standardized value of the selected measure j for observation i, information from multiple aspects of their scanning behavior
1  j  n. Therefore, O ¼ fO1 ; O2 ;    ; Om g is the dataset including descriptions.
all the samples to be clustered, and m is the total number of sam-
ples. Mði; jÞ is the calculated distance between cluster C i and cluster
4. Results
C j . M is the distance matrix between each two clusters. q is the
number of clusters in each iteration. IC is the inconsistency coeffi- The numbers of the collected episodes for left turn, straight
cient of the clustering result in each iteration. The corresponding across, and right turn maneuvers through the signalized intersec-
index k with the max variation between each adjacent IC values tions in the green light phase were 72, 150 and 96, respectively.
is selected to cluster the observations into m  k clusters. For unsignalized intersections, the corresponding numbers were
C ¼ fC 1 ; C 2 ;    ; C mk g is the final output of the cluster results. For 64, 156 and 80, respectively.
more details about hierarchical clustering, see Bouguettaya, Yu,
Liu, Zhou, and Song (2015).
4.1. Visual scanning measures
Algorithm 1 Hierarchical clustering algorithm
As illustrated in Fig. 3, glances to FW predominated (about 50%)
Input: Dataset O ¼ fO1 ; O2 ;    ; Om g for both signalized and unsignalized intersections. When driving
Output: Clusters C ¼ fC 1 ; C 2 ;    ; C mk g straight across signalized intersections, drivers would check LVM
1:for j ¼ 1; 2;    ; m do 78% more frequently than at unsignalized intersections (t = 2.164,
 
2: C j ¼ Oj ; df (degrees of freedom) = 304, p = 0.031). When turning left at sig-
3: end for nalized intersections, drivers significantly checked LVM 158% more
4: for i ¼ 1; 2;    ; m do frequently than at unsignalized intersections (t = 2.459, df = 134,
5: for j ¼ 1; 2;    ; m do p = 0.015). Interestingly, drivers checked FR 82% less frequently


6: Mði; jÞ ¼CalculateDistances C i ; C j ; at signalized intersections than at unsignalized intersections while
7: Mðj; iÞ ¼ Mði; jÞ; turning left (t = 3.371, df = 134, p = 0.001). No statistical signifi-
8: end for cances were found between signalized and unsignalized intersec-
9: end for tions while turning right.
10: Set number of current clusters: q = m; As illustrated in Fig. 4, durations for glances in FW were signif-
11: z = 1; icantly longer than that in the other AOIs. When driving straight
12: while q > 1 do across signalized intersections, the durations in NL were 95%
13: Find the two closest clusters C i and C j ; longer than that at unsignalized intersections (t = 1.985, df = 304,
14: Merge C i and C j : C i ¼ C i [ C j ; p = 0.048). When turning left at signalized intersections, drivers’
average durations in FW and RM were 2.97 s and 1.24 s, respec-
15: for j ¼ j þ 1; j þ 2;    ; q do
tively. The corresponding durations at unsignalized intersections
16: Renumber C j as C j1 ;
were 2.19 s and 0.65 s. The differences between signalized and
17: end for
unsignalized intersections for durations in both FW (t = 2.064,
18: Delete row j and column j of M;
df = 134, p = 0.041) and RM (t = 2.649, df = 134, p = 0.009) were sta-
19: for j ¼ 1; 2;    ; q  1 do

 tistically significant. No statistical significances were found
20: M ði ; jÞ= Calculate Distances C i ; C j ; between signalized and unsignalized intersections when turning
 
21: Mðj; i Þ ¼ M ði; j Þ; right.
22: end for In Fig. 5, the row AOIs and column AOIs correspond to the for-
23: q ¼ q  1; mer and latter allocations of two consecutive glances, respectively.
24: IC ðzÞ ¼ Calculate Inconsistency Coefficient (C); The color indicates the transition probability value, describing a
25: z ¼ z þ 1; driver’s visual scanning preference. A warmer color suggests a
26: end while greater transition probability. As expected, drivers’ glance transi-
27: k ¼ index of the max variation of IC; tions at signalized intersections were different from those at
th
28: Restore C: Use the ðk  1Þ result of the merger as the unsignalized intersections. Statistical significances were found in
final output; the following glance transitions when turning left: (1) from FW
29: Return Clusters C ¼ fC 1 ; C 2 ;    ; C mk g to FR (t = 3.343, df = 134, p = 0.001), (2) from FW to LVM
(t = 2.236, df = 134, p = 0.027), (3) from NL to FL (t = 2.131,
df = 134, p = 0.035), (4) from FR to FW (t = 2.033, df = 134,
p = 0.044), (5) from LVM to FW (t = 2.545, df = 134, p = 0.012), and
(6) from FR to FL (t = 2.682, df = 134, p = 0.008). Table 2 presents
To visually illustrate the clustering results of high-dimensional
the transition probability values of these six typical transitions.
data, Andrews Curve was adopted to reveal the differences between
As for the right turn maneuvers, statistical significance was found
the clustered patterns. For each clustered pattern, Andrews Curve
in the glance transitions from RVM to FW (t = 2.037, df = 174,
defines a finite Fourier series as follows,
p = 0.043). No statistical significances were found in the straight
Oi;1 X n
across maneuvers.
f i ðt Þ ¼ pffiffiffi þ Oi;k ðð1  mod ðk; 2ÞÞsinðbk=2c  2pt Þ
2 k¼2
þ mod ðk; 2Þcosðbk=2c  2ptÞÞ ð3Þ 4.2. Results from cluster analysis

where t 2 ½0; 1 and mod ðk; 2Þ is an operation returning the remain- Not many statistical significances between signalized and
der after division of k by 2. Thus each data point can be viewed as a unsignalized intersections were found on drivers’ visual scanning
line between 0 and 1. If the Andrews Curve of one clustered pattern measures in either straight across or right turn maneuvers. Thus,
obviously differentiates from the curve of another clustered pattern, drivers’ visual scanning patterns were only clustered for the left
it indicates that the clustered patterns are effective to describe dri- turn maneuvers at the examined intersection types. See Fig. 6 for
vers’ scanning behavior at intersections by jointly considering the the illustrated performance of the clustering method. Andrews
224 G. Li et al. / Journal of Safety Research 71 (2019) 219–229

Fig. 4. Durations in each AOI at signalized and unsignalized intersections.

Results show that drivers’ visual scanning behavior can be clus-

tered into two patterns for left turn maneuvers at either signalized
or unsignalized intersections.
Fig. 3. Glance frequency percentages in each AOI at signalized and unsignalized
intersections. The error bar represents standard error. * represents significance at
As shown in Fig. 7(a), the two clustered scanning patterns when
0.05 level (p  0.05). ** represents significance at 0.01 level (p  0.01). turning left at signalized intersections can be described as left
areas aware (LAA) and left areas not aware (LANA). The number
of events for LAA and LANA are 39 (54.2%) and 33 (45.8%), respec-
Curve, an approach of dimension-reducing visualization, was tively. The LAA pattern shows significantly increased attention in
employed to qualitatively describe how significantly the clustered the left areas than the LANA pattern, while their scanning behavior
patterns differ with each other. The upper and lower boundaries to the right areas does not differentiate much from each other.
for each clustered pattern indicate the 75% and 25% data range. Similarly, the clustered visual scanning patterns when turning left
 G. Li et al. / Journal of Safety Research 71 (2019) 219–229 225

Fig. 5. Visualization of glance transition matrices at signalized and unsignalized intersections.

at unsignalized intersections can be described as side areas aware at signalized intersections. However, drivers scanning with pattern
(SAA) and side areas not aware (SANA). The corresponding event SAA also transfer more frequently between FR and FW. The com-
numbers for SAA and SANA are 34 (53.1%) and 30 (46.9%), respec- parison of scanning patterns for drivers turning left at signalized
tively. See Fig. 7(b). Drivers scanning with pattern SAA focuses and unsignalized intersections shows that drivers scanning with
more on the left areas, which is similar with the clustering results pattern SAA focuses more on FW and RM than drivers scanning
226 G. Li et al. / Journal of Safety Research 71 (2019) 219–229

Table 2
Transition probability values of the six typical transition patterns.

FW ? FR FW ? LVM NL ? FL FR ? FW LVM ? FW FR ? FL
Signalized 0.023 0.057 0.100 0.200 0.167 0.000
Unsignalized 0.069 0.214 0.033 0.153 0.527 0.015

Fig. 7. Interrelationships between the selected measures in the clustered scanning

Fig. 6. Illustrated clustering performance for left turn maneuvers using Andrews patterns. The size of each dot indicates the standard deviation of the corresponding
Curve. measure. (F for frequency, D for duration, TP for transition probability.)

quency, glance duration, and AOI transition probabilities were

with pattern LAA. The LANA and SANA patterns do not differentiate used to assess visual scanning behavior. Glance frequency and
much from each other. The scanning patterns can be considered as glance duration have been widely used to examine the fixation of
one of drivers’ attributes indicating their driving safety levels or a driver’s attention while driving (Muñoz, Reimer, Lee, Mehler, &
driving styles, which can be verified by drivers’ crash involvement Fridman, 2016). Higher frequency and longer duration suggests a
statistics in years. higher task demand (Owsley et al., 1991; Wang, Reimer, Dobres,
& Mehler, 2014). As presented in this study, drivers would check
LVM 78% more frequently when driving straight across signalized
5. Discussion than unsignalized intersections (p = 0.031). This indicates that the
cognitive demand for receiving and processing the information
This study examined drivers’ visual scanning behavior at differ- presented in LVM is higher near signalized than unsignalized
ent types of intersections. Ten AOIs were assessed. Glance fre- intersections.
 G. Li et al. / Journal of Safety Research 71 (2019) 219–229 227

Glance transitions between AOIs can assess visual scanning tra- ous and needs to be considered in the safety improvement strate-
jectories, which reflects the flow of attention across space (Liang, gies for unsignalized intersections. For left turn maneuvers at
Lee, & Yekhshatyan, 2012; Muñoz et al., 2016). A more frequent signalized intersections, checking the left areas (including LVM)
glance transition between two AOIs indicates a higher correlation is a critical scanning strategy drivers should execute for safety
between the areas (Wong & Huang, 2013; Fridman, Toyoda, (Richard et al., 2006). However, drivers with LANA pattern show
et al., 2016). As shown in Table 1, the transition probability from much less attention in the left areas compared with the LAA dri-
LVM to FW near unsignalized intersections is 0.527, while the vers. This low situation awareness level of LANA drivers will
number near signalized intersections is only 0.167. The higher cor- greatly challenge traffic safety at signalized intersections.
relation between LVM and FW near unsignalized intersections To avoid crashes caused by drivers with LASA or SANA scanning
indicates that drivers attended more to the traffic situation and patterns, new strategies can be proposed to improve intersection
information presented in the function-related areas (e.g., LVM, safety, both offline and online, based on the results reported in this
FW). It demonstrates that function-related areas play an important paper. In offline training and improving programs, driver’s scan-
role for safe driving near unsignalized intersections (Peng & Boyle, ning measures at all the recorded intersections can be computed.
2015). Therefore, to enhance driving safety, Advanced Driver Assis- If the driver is clustered as a LANA or SANA driver in the left turn-
tance Systems (ADAS) should be designed to provide drivers with ing maneuver at intersections, deficiencies in the driver’s scanning
effective automatic feedback concerning the information pre- behavior can be summarized (e.g., not sufficiently checking the left
sented in the function-related areas when driving near unsignal- areas at signalized intersections, not paying enough attention to
ized intersections. the right areas at unsignalized intersections). A feedback report
The results from the present study show that when driving on the monthly scanning performance at intersections will be
straight across or turning right, Chinese drivers do not perform given to the driver for continuous training and improving educa-
much differently at signalized and unsignalized intersections. As tion, which has been convinced by Toledo et al. (2008) that offline
reported earlier, risky driving behavior is engrained in Chinese feedback could effectively reduce crash rates and driving risk.
traffic culture (Yang, Du, Qu, Gong, & Sun, 2013). Chinese drivers Besides offline programs, online driver assistance could also be
tend NOT to decelerate to a full stop when passing an unsignalized proposed. In particular, if a LANA driver turns left at a signalized
intersection. As we only observed the scanning behavior in the intersection in green phase, the driver should check LVM and tran-
green light phase near intersections, risky driving tendencies sit between LVM and FW for safety reasons. If the driver do not
among Chinese drivers could help explain our finding that visual check the LVM while turning, a visual reminder from the LVM
scanning performance was similar between signalized and can be triggered. Otherwise, the visual reminder should not be trig-
unsignalized intersections in the straight across and right turning gered to avoid annoyance. Besides the left areas, drivers’ attention
maneuvers. should also be given to the FR for left turn safety at unsignalized
At unsignalized intersections, drivers have to keep aware of the intersections. Similar assistance strategy can be applied. To avoid
current traffic situations, make a decision, and behave properly to crashes caused by ‘looked-but-not-see’ problems, an emergency
pass the intersection safely without conflicting with other vehicles braking can still be applied just in case two vehicles get very close
(Liao et al., 2016). In China, driving rules require drivers to drive on but the driver does not react to this danger.
the right hand side of the road. This indicates that task demands by A limitation of the present study is that drivers’ scanning behav-
left turn maneuvers would be higher than those by right turn and ior was manually coded, which is less precise compared to eye-
straight across maneuvers at the unsiganlized intersections, as dri- tracker-based coding. Eye-trackers will be employed in our further
vers have to attend more to the traffic in the opposite direction and studies to automatically and precisely record drivers’ scanning
the other surrounding traffic environment at the same time when behavior, and more participants are needed to be involved in
turning left (Werneke & Vollrath, 2014). Thus, when drivers are future studies. Besides, previous studies showed that turning
turning left at unsignalized intersections, they attend more to the maneuvers (left, straight, right) would influence drivers’ visual
far areas and transfer more frequently between the non-adjacent scanning behavior (Wu & Xu, 2017). In the present study, we only
areas for a better understanding of the traffic environment, so as focused on the visual scanning differences near signalized and
to avoid traffic incidents. unsignalized intersections. No comparisons between the turning
At signalized intersections, however, traffic lights organize maneuvers were examined. Future efforts would concentrate on
vehicles to pass the intersections orderly in the green light phase. driver behavior distinction between different turning maneuvers
Drivers turning left do not need to pay much attention to the far to provide more evidences for better design of ADAS and driver
areas, which has been suggested to be particularly important for training programs. Further, the drivers involved in the present
driving safety near unsignalized intersections (Yamani et al., study were all experienced (>2 years). Drivers’ experience makes
2016). Thus, the findings presented in this study show that drivers significance effects on their driving performance (Romoser et al.,
attend more to FW, RM and LVM without paying much attention to 2013). Knowledge about driving experience effects could con-
the far areas at signalized intersections. Meanwhile, although the tribute to driving safety enhancement programs (Li, Li, & Cheng,
task demand near signalized intersection is relatively lower, we 2015). Therefore, future studies should examine experience effects
have to keep in mind that there are far more vehicles and pedestri- on drivers’ scanning behavior when passing through signalized and
ans near signalized than near unsignalized intersections (NHTSA, unsignalized intersections.
2017). This explains why the number of crashes near signalized
intersections was almost twice of the number for unsignalized
intersections (NHTSA, 2017). 6. Conclusions
The cluster analysis results show that the LANA and SANA pat-
terns do not differentiate much from each other, which means that This study examines the differences of drivers’ visual scanning
some drivers adopt the same scanning patterns at either signalized behavior between signalized and unsignalized intersections not
or unsignalized intersections. However, safety consideration only in terms of the independent measures (frequency, duration
requires drivers at unsignalized intersections to check both the left and transition probability), but also the correlation between them
and right areas before entering (Richard et al., 2006). The SANA to see how they are inter-related with typical scanning patterns.
scanning pattern neglect the right areas, indicating a low situation We found that intersection types made differences on drivers’
awareness level. Therefore, the SANA scanning pattern is danger- visual scanning behavior. These findings suggest that future appli-
228 G. Li et al. / Journal of Safety Research 71 (2019) 219–229

cations in advanced driver assistance systems and driver training Driving simulator study on speed and lateral position. Transportation Research
Part F: Traffic Psychology and Behaviour, 58, 156–166.
programs should recommend different scanning strategies to dri-
Muñoz, M., Reimer, B., Lee, J., Mehler, B., & Fridman, L. (2016). Distinguishing
vers at different types of intersections. patterns in drivers’ visual attention allocation using Hidden Markov Models.
Transportation Rescanning Part F: Traffic Psychology and Behaviour, 43, 90–103.
NHTSA (2017). Traffic Safety Facts 2015 Data (DOT HS 812 353, updated March
Declaration of Competing Interest 2017). Washington, DC: National Highway Traffic Safety Administration.
Owsley, C., Ball, K., Sloane, M. E., Roenker, D. L., & Bruni, J. R. (1991). Visual/cognitive
The authors declare that they have no known competing finan- correlates of vehicle accidents in older drivers. Psychology and Aging, 6(3),
403–415.
cial interests or personal relationships that could have appeared
Peng, Y., & Boyle, L. N. (2015). Driver’s adaptive glance behavior to in-vehicle
to influence the work reported in this paper. information systems. Accident Analysis & Prevention, 85, 93–101.
Richard C. M., Campbell J. L., & Brown J. L. (2006). Task Analysis of Intersection
Driving Scenarios: Information Processing Bottlenecks (Technical Report
Acknowledgement FHWA-HRT-06-033). Federal Highway Administration, McLean, VA, USA.
Romoser, M. R. E., Pollatsek, A., Fisher, D. L., & Williams, C. C. (2013). Comparing the
This study is supported by the National Natural Science Foundation glance patterns of older versus younger experienced drivers: Scanning for
hazards while approaching and entering the intersection. Transportation
of China with 51805332, 71401004, Natural Science Foundation of Rescanning Part F: Traffic Psychology and Behaviour, 16, 104–116.
Guangdong Province with 2018A030310532, the State Key Labora- Scott, H., Hall, L., Litchfield, D., & Westwood, D. (2013). Visual information scanning
tory of Automotive Safety and Energy under Project No. KF1801, in simulated junction negotiation: Gaze transitions of young novice, young
experienced and older experienced drivers. Journal of Safety Research, 45,
and the Natural Science Foundation of SZU (grant no. 2017033).
111–116.
The authors also would like to thank Liming Liang for his help on Shinar, D. (2008). Looks are (almost) everything: Where drivers look to get
data collection and glance allocation extraction from videos, and information. Human Factors, 50(3), 380–384.
Society of Automotive Engineers (2000). Definitions and Experimental Measures
Weijian Lai for his help on data processing and figure plotting.
Related to the Specification of Driver Visual Behavior Using Video Based
Techniques (Surface Vehicle Recommended Practice J2396). Society of
References Automotive Engineers, Warrendale, PA.
Tay, R. (2015). A random parameters probit model of urban and rural intersection
crashes. Accident Analysis & Prevention, 84, 38–40.
Bao, S., & Boyle, L. N. (2009). Age-related differences in visual scanning at median-
Toledo, T., Musicant, O., & Lotan, T. (2008). In-vehicle data recorders for monitoring
divided highway intersections in rural areas. Accident Analysis & Prevention, 41
and feedback on drivers’ behavior. Transportation Research Part C: Emerging
(1), 146–152.
Technologies, 16(3), 320–331.
Bouguettaya, A., Yu, Q., Liu, X., Zhou, X., & Song, A. (2015). Efficient agglomerative
Wang, Y., Reimer, B., Dobres, J., & Mehler, B. (2014). The sensitivity of different
hierarchical clustering. Expert Systems with Applications, 42(5), 2785–2797.
methodologies for characterizing drivers’ gaze concentration under increased
Carter, D., Hummer, J. E., Foyle, R. S., & Phillips, S. (2005). Operational and safety
cognitive demand. Transportation Rescanning Part F: Traffic Psychology and
effects of U-turns at signalized intersections. Transportation Research Record,
Behaviour, 26, 227–237.
1912(1), 11–18.
Werneke, J., & Vollrath, M. (2014). How do environmental characteristics at
Dukic, T., & Broberg, T. (2012). Older drivers’ visual scanning behaviour at
intersections change in their relevance for drivers before entering an
intersections. Transportation Rescanning Part F: Traffic Psychology and
intersection: Analysis of drivers’ gaze and driving behavior in a driving
Behaviour, 15(4), 462–470.
simulator study. Cognition, Technology & Work, 16(2), 157–169.
Feng, J., Choi, H., Craik, F. I. M., Levine, B., Moreno, S., Naglie, G., & Zhu, M. (2018).
Wong, J.-T., & Huang, S.-H. (2013). Attention allocation patterns in naturalistic
Adaptive response criteria in road hazard detection among older drivers. Traffic
driving. Accident Analysis & Prevention, 58, 140–147.
Injury Prevention, 19(2), 141–146.
Wu, J., & Xu, H. (2017). Driver behavior analysis for right-turn drivers at signalized
Fridman, L., Langhans, P., Lee, J., & Reimer, B. (2016). Driver gaze region estimation
intersections using SHRP 2 naturalistic driving study data. Journal of Safety
without use of eye movement. IEEE Intelligent Systems, 31(3), 49–56.
Research, 63, 177–185.
Fridman, L., Toyoda, H., Seaman, S., Seppelt, B., Angell, L., Lee, J., Mehler, B., & Reimer,
Yamani, Y., Horrey, W. J., Liang, Y., & Fisher, D. L. (2016). Age-related differences in
B. (2016). What can be predicted from six seconds of driver glances?. arXiv
vehicle control and eye movement patterns at intersections: Older and middle-
preprint arXiv:1611.08754.
aged drivers. PLoS One, 11(10) e0164124.
Klauer, S. G., Guo, F., Simons-Morton, B. G., Ouimet, M. C., Lee, S. E., & Dingus, T. A.
Yang, J., Du, F., Qu, W., Gong, Z., & Sun, X. (2013). Effects of personality on risky
(2014). Distracted driving and risk of road crashes among novice and
driving behavior and accident involvement for Chinese drivers. Traffic Injury
experienced drivers. New England Journal of Medicine, 370(1), 54–59.
Prevention, 14(6), 565–571.
Kusano, K. D., Chen, R., Montgomery, J., & Gabler, H. C. (2015). Population
distributions of time to collision at brake application during car following
from naturalistic driving data. Journal of Safety Research, 54, 95–104. Dr. Guofa Li is an assistant professor at the Institute of Human Factors and Ergo-
Lemonnier, S., Brémond, R., & Baccino, T. (2015). Gaze behavior when approaching nomics, Shenzhen University, China. He received the B.E. degree from Beijing
an intersection: Dwell time distribution and comparison with a quantitative Institute of Technology in 2010, and the Ph.D. degree from Tsinghua University in
prediction. Transportation Rescanning Part F: Traffic Psychology and Behaviour, 35, 2016. From 2012 to 2013, he was a visiting scholar with University of Michigan
60–74. Transportation Research Institute (UMTRI). His research interests include driving
Li, G., Li, S. E., & Cheng, B. (2015). Field operational test of advanced driver assistance behavior analysis, driving safety applications, machine learning techniques, and
systems in typical Chinese road conditions: The influence of driver gender, age human factors in automotive and transportation engineering.
and aggression. International Journal of Automotive Technology, 16(5), 739–750.
Li, G., Li, S. E., Cheng, B., & Green, P. (2017). Estimation of driving style in naturalistic Dr. Ying Wang is a human factors and traffic safety expert. She holds a doctorate
highway traffic using maneuver transition probabilities. Transportation Research degree from Tsinghua University, Beijing, China (2009). She is currently the Prin-
Part C: Emerging Technologies, 74, 113–125. cipal of United Human Factors, a consulting service based in Illinois, USA, and used
Li, G., & Wang, Y. (2018). Surface transportation. In S. N. Samman (Ed.), Human to be an Associate Professor in the School of Transportation Science and Engineering
Factors and Ergonomics for the Gulf Cooperation Council: Processes, Technologies, at Beihang University, Beijing, China (2013–2016) and a Postdoctoral Research
and Practices (pp. 191–210). CRC Press. Associate at MIT AgeLab (2009–2012). Her research seeks to build safety and well-
Liang, Y., Lee, J. D., & Yekhshatyan, L. (2012). How dangerous is looking away from being in surface transportation systems from a human factors perspective.
the road? Algorithms predict crash risk from glance patterns in naturalistic
driving. Human Factors, 54(6), 1104–1116. Mr. Fangping Zhu is a graduate research assistant at the Institute of Human Factors
Liao, Y., Li, S. E., Wang, W., Wang, Y., Li, G., & Cheng, B. (2016). Detection of driver and Ergonomics, Shenzhen University, China. He enrolled in this institute as a
cognitive distraction: A comparison study of stop-controlled intersection and master student in 2017. His main research areas include driving behavior analysis,
speed-limited highway. IEEE Transactions on Intelligent Transportation Systems, and driving style estimation using machine learning techniques.
17(6), 1628–1637.
Liu, M., Chen, Y., Lu, G., & Wang, Y. (2017). Modeling crossing behavior of drivers at Ms. Xiaoxuan Sui is a graduate research assistant at the Institute of Human Factors
unsignalized intersections with consideration of risk perception. Transportation and Ergonomics, Shenzhen University, China. She enrolled in this institute as a
Research Part F: Traffic Psychology and Behaviour, 45, 14–26. master student in 2016. Her main research areas include driving behavior analysis,
Liu, Y., & Ozguner, U. (2007). Human driver model and driver decision making for signal processing, and traffic congestion strategies development.
intersection driving. In Proceedings of the 2007 IEEE Intelligent Vehicles
Symposium (pp. 642–647). Ms. Ning Wang is an algorithm engineer at Aptiv, USA. She was an undergraduate
Lococo, K. H., & Staplin, L. (2018). Visual Scanning Training for Older Drivers: A research assistant at Beihang University from 2012 to 2016. Her main research
Literature Review (Technical Report No. DOT HS 812 514). Washington, DC: areas include driving behavior analysis, computer vision and autonomous driving.
National Highway Traffic Safety Administration.
Mollu, K., Biesbrouck, M., Van Broeckhoven, L., Daniëls, S., Pirdavani, A., Declercq, K., Dr. Xingda Qu is a professor with the Institute of Human Factors and Ergonomics,
... Brijs, T. (2018). Priority rule signalization under two visibility conditions: Shenzhen University, China. He received his PhD degree in Human Factors and
 G. Li et al. / Journal of Safety Research 71 (2019) 219–229 229

Ergonomics from Virginia Tech in 2008. His research interests include transporta- University’s Human Factors Engineering Short Course, the flagship continuing
tion safety, occupational safety and health, and human computer interaction. education course in the profession, now in its 60th year. His research interests
include driver interfaces and driver workload, and the development of standards to
Dr. Paul Green is a Research Professor with the University of Michigan Trans- get research into practice. Prof. Green is a past president of the Human Factors and
portation Research Institute, and an Adjunct Professor with the Department of Ergonomics Society.
Industrial and Operations Engineering at Umich, USA. He is the Leader of the