You are on page 1of 12

Journal of Safety Research 80 (2022) 416–427

Contents lists available at ScienceDirect

Journal of Safety Research


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

An improved automated braking system for rear-end collisions: A study


based on a driving simulator experiment
Junyu Hang a, Xuedong Yan a,⇑, Xiaomeng Li b, Ke Duan a, Jingsi Yang c, Qingwan Xue d
a
MOT Key Laboratory of Transport Industry of Big Data Application Technologies for Comprehensive Transport, School of Traffic and Transportation, Beijing Jiaotong University,
Beijing 100044, PR China
b
Queensland University of Technology (QUT), Centre for Accident Research and Road Safety-Queensland (CARRS-Q), Institute of Health and Biomedical Innovation (IHBI), Kelvin
Grove, Queensland 4059, Australia
c
CRSC Communication & Information Group Company Ltd., Beijing 100070, PR China
d
Beijing Key Laboratory of Urban Intelligent Traffic Control Technology, North China University of Technology, Beijing 100144, China

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

Article history: Introduction: To assist drivers in avoiding rear-end collisions, many early warning systems have been
Received 27 May 2021 developed up to date. Autonomous braking technology is also used as the last defense to ensure driver’s
Received in revised form 6 August 2021 safety. Method: By taking the accuracy and timeliness of automatic system control into account, this paper
Accepted 19 December 2021
proposes a rear-end Real-Time Autonomous Emergency Braking (RTAEB) system. The system inserts brake
Available online 29 December 2021
intervention based on drivers’ real-time conflict identification and collision avoidance performance. A
driving simulator-based experiment under different traffic conditions and deceleration scenarios were
Keywords:
conducted to test the different thresholds to trigger intervention and the intervention outcomes. The sys-
Rear-end collision
AEB
tem effectiveness is verified by four evaluation indexes, including collision avoidance rate, accuracy rate,
Driving simulator sensitivity rate, and precision rate. Results: The results showed that the system could help avoid all colli-
Collision avoidance sion events successfully and enlarge the final headway distance, and a TTC threshold of 1.5 s and a max-
imum deceleration threshold of 7.5 m/s2 could achieve the best collision avoidance effect. The paper
demonstrates the situations that are more inclined to trigger the RTAEB (i.e., a sudden brake of the leading
vehicle and a small car-following distance). Moreover, the study shows that driver characteristics (i.e.,
gender and profession) have no significant association with system trigger. Practical Applications: The
study suggests that development of collision avoidance systems design should pay attention to both
the real-time traffic situation and drivers’ collision avoidance capability under the present situation.
Ó 2021 National Safety Council and Elsevier Ltd. All rights reserved.

1. Introduction result in rear-end collisions. Statistical and experimental crash


data analysis showed that the common reasons related to rear-
Rear-end collision is one of the most frequent collisions on the end collisions are drivers’ inattentive driving (Li et al., 2016,
road compared with other types of collisions. According to the 2019), inadequate headways (Winsum & Heino, 1996; You et al.,
National Highway Traffic Safety Administration (NHTSA, US), there 2013), and late response to the leading vehicle’s deceleration
were 696,372 injured/fatal rear-end collisions in 2016, accounting (Sheu & Wu, 2015; Wood & Zhang, 2017).
for 45.87% of the total rear-end collisions (NHTSA, 2016). The rear- In recent years, autonomous driving or assistant driving tech-
end collision avoidance process is dynamic and composed of risk nologies have rapidly developed and they are expected to elimi-
recognition, decision-making, and vehicle operation. A complete nate human errors and, consequently, reduce traffic accidents.
rear-end collision avoidance process includes braking response Autonomous driving systems are divided into 6 levels (Level 0–5)
stage, deceleration adjustment stage, maximum deceleration and the highest level (Level 5) will achieve the ultimate goal of
maintaining stage, and recover stage (Li et al., 2016; Xue et al., self-driving without any control input from human drivers (Kala
2018). Any judgments or behavior errors in these stages could & Warwick, 2013; Wenda & Dolan, 2012; Ziegler et al., 2014). Prior
to the launch of fully autonomous vehicles, much effort has been
dedicated to developing and improving various driver assistance
⇑ Corresponding author.
technologies, such as the Collision Avoidance Systems (CASs)
E-mail addresses: 18114033@bjtu.edu.cn (J. Hang), xdyan@bjtu.edu.cn (X. Yan),
xiaomeng.li@qut.edu.au (X. Li), 17114231@bjtu.edu.cn (K. Duan), 15114234@bjtu. including Collision Warning Systems (CWSs), Autonomous Emer-
edu.cn (J. Yang), qwxue@ncut.edu.cn (Q. Xue). gency Braking Systems (AEBs), and Lane-departure Systems. These

https://doi.org/10.1016/j.jsr.2021.12.023
0022-4375/Ó 2021 National Safety Council and Elsevier Ltd. All rights reserved.
J. Hang, X. Yan, X. Li et al. Journal of Safety Research 80 (2022) 416–427

systems are assembled inside the vehicles and capable of providing 1.3. Key parameters of CWSs and AEBs
drivers with warnings in advance and even assist the vehicle con-
trol at safety–critical situations. Overall, the above-mentioned approaches of rear-end CASs
have been confirmed with significant effects in improving the dri-
1.1. Collision warning systems (CWSs) ver safety. However, when designing the algorithms of rear-end
CWSs and AEBs, most previous research tended to ignore the indi-
In the case of an impending rear-end situation, the key to avoid vidual difference among drivers and used fixed values for key
a collision is that the driver recognizes the leading vehicle’s brake parameters such as deceleration rate or TTC (Bella & Russo, 2011;
or the gap reduced between vehicles, and takes proper and timely Ho et al., 2006; Jeong et al., 2017; Scott & Gray, 2008; Wu et al.,
actions (i.e., braking or lane-changing). To remind drivers of the 2018a; Zavareh et al., 2017). For instance, Kusano and Gabler
imminent danger, Collision Warning Systems (CWSs) are designed (2012) set 1.2 s as the warning delivered time, 0.8 s as the brake
to provide precise warning information. Previous research showed assist delivered time, and 0.45 s as the autonomous braking time,
that the release of warning messages could effectively reduce the and limited the maximum deceleration rate to 0.8 g.
collision rate (Lee et al., 2002; Wu et al., 2018a; Xiang et al., A review on the related literature showed that TTC threshold as a
2016). This reduction was achieved by shortening driver’s reaction safety indicator, played a critical role in assessing rear-end collision
time, decreasing driving speed, and increasing the brake force risks and influencing the collision avoidance outcomes (Li et al.,
(Haas & Van Erp, 2014; Mohebbi et al., 2009; Wu et al., 2018b; 2018; Weng et al., 2015). Generally, a smaller TTC threshold repre-
Zhu et al., 2020). The effectiveness of CWSs has been verified in dif- sents a higher risk assessment accuracy but it also increases the
ferent scenarios, such as intersections and work-zone related area crash risks. On the contrary, a larger TTC value allows for a longer
(Chang et al., 2009; Qiao et al., 2017; Zhang et al., 2016, 2018). collision avoidance time but it may lead to an increase in false alarm
rate. Therefore, it is necessary to adjust the threshold range of risk
1.2. Autonomous emergency braking systems (AEBs) indicators according to different situations. It should be noted that
some studies did consider some of the key parameters as variables
In addition to the warning systems, some automobile manufac- in the process of vehicle braking. For example, Seacrist et al. (2020)
turers have developed collision avoidance systems to help drivers applied a deceleration rate varying with time, and the operating
take evasive actions (Doi, 1994; Fujita, 1995). Among the various principle of the AEB system was used to estimate the deceleration
collision avoidance systems, Rear-end Collision Avoidance Systems based on different road frictions. Nevertheless, drivers’ individual
(e.g., AEBs) is the most frequently used one (Braitman et al., 2010). differences and dynamic response that may lead to changes in the
According to Cicchino (2017), low-speed AEB systems could reduce parameters (such as their response abilities toward different situa-
rear-end crash rates and rear-end injury crash rates to 43% and tions) are still largely ignored in prior studies.
45%, respectively. Different from the CWSs, the AEB systems are Drivers’ reaction time to a rear-end collision is one of the key
applied to intervene in the vehicle control, which is usually acti- parameters that may change according to situations (Aust et al.,
vated when the risk threshold is less than a safety margin in critical 2013; Li et al., 2016; Liebermann et al., 1995; Wang et al., 2016;
situations. The millimeter-wave radar and/or video cameras are Xue et al., 2018). Liebermann et al. (1995) reported that drivers
usually used to track the moving vehicles in front of the AEB- reacted faster when the following distance got shorter. Especially
equipped vehicle (Brannstrom & Sjoberg, 2008; Fildes et al., in urgent situations (e.g., small TTC), drivers tended to release
2015; Kusano & Gabler, 2010). Once the driver does not respond the accelerator and braked more quickly. Hulst and Der (1999) fur-
to the emergent situations, the AEB system will automatically ther found that drivers could adjust their reaction time to the
apply a specific deceleration to avoid the collisions. deceleration level of the leading vehicle. Wang et al. (2016) tested
AEB system is used to reduce the risk of common rear-end col- the effect of situational urgency on driver’s rear-end collision
lision scenarios. However, due to the lack of vehicle dynamics data avoidance behaviors and reported that an increased situational
in real-world driving situations with and without the activation of urgency could lead to a faster deceleration. Given the varied driver
AEB system, it is challenging to evaluate the performance of AEB in response in different situations, the collision avoidance systems
collision avoidance. At present, the traditional AEB system is acti- should take driver’s real-time response into account when evaluat-
vated by comparing the real-time TTC with pre-designed TTC ing the situation and making the decisions of warning or autono-
threshold. For example, in the study of Kusano and Gabler (2010, mous control.
2012), the TTC threshold of 0.45 s was identified as an optimal
value for triggering AEB system. Duan et al. (2017) further intro- 1.4. Application of driving simulators
duced a response time (Tb) to measure whether the AEB was
needed based on driver’s behavior characteristics. Driving behavior research based on driving simulators has been
In these studies, the deceleration rate required by the driver a great contributor to the field of road safety. First, the driving sim-
was calculated using two methods. The first method calculated ulator can depict human behavior in real time to the greatest
the deceleration pulse at fixed jerk and maximum deceleration extent, and collect human driving behavior in a near real environ-
by using the momentum conservation theorem or kinematics rela- ment, such as car following (Li et al., 2019; Sheu & Wu, 2015), lane
tionship according to the relative speed and position relationship changing (Hess et al., 2020), and gap acceptance (Li et al., 2020).
of the drivers when specific TTC was reached (Duan et al., 2017; Secondly, driving simulators provide the opportunity to test differ-
Kusano & Gabler, 2010, 2012). This method used idealized AEB ent driving conditions in a controlled and safe environment. Some
deceleration profiles for analysis, which is lack of the driver’s real high-risk scenarios cannot be reproduced and tested in the real
reflection in the deceleration process. The second method used world. The driving simulator can test the risky scenarios without
the counterfactual algorithm to deduce the conditions that should causing actual collision risk, such as distraction (Li et al., 2019;
trigger the AEB system from the historical collision data (Bargman Muttart et al., 2007; Navarro et al., 2017), adverse weather (Li
et al., 2017; Seacrist et al., 2020). However, most of the above et al., 2015; Mcavoy et al., 2007; Yan et al., 2014; Zhao et al.,
methods are based on the historical accident data (such as reports 2019), and fatigue driving (Meng et al., 2019). Third, the driving
from the police or insurance companies) and due to the lack of real simulator can also provide a low-cost, high-efficiency, and repeat-
vehicle dynamics data, it is difficult to verify the results correlated able environment for researchers to develop and test new inter-
to the actual driving scenarios. ventions, such as the improvement of road auxiliaries (Hang
417
J. Hang, X. Yan, X. Li et al. Journal of Safety Research 80 (2022) 416–427

et al., 2018; Ma & Yan, 2021) and new driver assistance technolo- collect driver’s real time data and vehicle dynamic data. Combining
gies (Calvi et al., 2020; Zhang et al., 2021) the RTAEB system algorithm with experimental data, a counterfac-
The limitations of driving simulators should also be recognized. tual simulation method was used to evaluate the system effective-
First of all, driving simulators may cause motion sickness of partic- ness. The counterfactual simulation algorithm assumed that RTAEB
_
ipants (Almallah et al., 2021; Duzmańska et al., 2018); thus, it is system would intervene in the vehicle control when the risk
necessary to select drivers who are suitable to operate the driving increased to a certain level and the driver did not take timely
simulator. Moreover, the validation of driving simulators is impor- and appropriate action. Based on the simulation results, collision
tant, mainly to ensure the data collected from a driving simulator avoidance rate (CAR), accuracy rate (AR), sensitivity rate (SR),
is comparable to the real world. At present, many studies have ver- and precision rate (PR) of AEB system were assessed under various
ified the authenticity of driving simulators at both subjective and conditions. In addition, the vehicle dynamic data were compared
objective levels (e.g. Bella, 2005; Bham et al., 2014; Hussain between driver control and RTAEB system control.
et al., 2019). For example, Rossi et al. (2020) compared driver’s
gap acceptance behavior through both field observations and driv- 2.1. Framework design of the real-time autonomous Emergency
ing simulator, and found that the mean critical gaps estimated in braking (RTAEB) system
the field and in the virtual environment are not significantly differ-
ent, confirming the consistency of behaviors in a simulator and in The RTAEB system proposed in this study consists of risk
the real world. Hussain et al. (2019) validated a fixed-base driving recognition function and collision avoidance function in a car-
simulator by comparing drivers’ speed perception and actual following situation and the execution procedure is shown in
speed, and the results showed the effectiveness of a fixed-base Fig. 2. During the collision avoidance process, which starts from
driving simulator in speed assessment. Therefore, the commonly the leading vehicle’s (LV’s) deceleration and ends as the subject
recognized relative validity of driving simulators suggests that (following) vehicle (SV) completes deceleration, the system mon-
they can be used as a valuable tool to evaluate driver’s behavior, itors the speed and the headway distance of the two vehicles at a
especially in risky situations. frequency of 60 Hz.

2.1.1. Risk recognition


1.5. Objective of this study
After entering the work zone area, the system detects the vehi-
cles’ speed and position (including leading and subject vehicles) in
To take into account drivers’ dynamic response to a rear-end
real-time. The conflict is recognized to appear if the speed of the
collision situation, this study proposed a new framework of the
following vehicle is higher than that of the leading vehicle, and
real-time Autonomous Emergency Braking (RTAEB) System. The
the distance between the two vehicles is smaller than a normal
algorithm of the RTAEB was designed based on the principle of
car-following distance margin. In this study, we considered a time
existing ones. However, the proposed framework differentiated
headway of 4 s as the normal car-following distance margin
from previous research by validating the combination of different
(Lewis-Evans et al., 2010; Siebert et al., 2014).
TTC and maximum deceleration thresholds under various condi-
As the conflict appears, TTC (Time-to-Collision) is monitored as
tions, and considering the individual’s risk responding capabilities
the risk indicator. Smaller TTC values indicate higher collision risk
in real time. To obtain the complete vehicle dynamics data during
and vice versa. A threshold value of TTC is usually applied to distin-
the collision avoidance process, a driving simulator experiment
guish the safe zone and critical zone boundary. In the risk recogni-
was conducted and the collected data was used to verify the effec-
tion stage, hazards could be identified earlier if a larger TTC
tiveness of the system. As the proposed AEBs considers the driver’s
threshold is used, but it may also increase the likelihood of false
real-time response to an evolving risk situation, it is expected that
alarms. In contrast, a smaller TTC threshold corresponds to a more
the AEBs could, to a large extent, avoid unnecessary intervention
accurate risk recognition but it may increase the possibility of late
and reduce the discrepancy between drivers’ expectation and the
intervention and failed collision avoidance. Prior research has
system’s intervention in safety–critical situations. Specifically, this
applied TTC threshold in CASs within a range of 0.5 s  3 s (Bella
study aimed to address the following research questions:
& Russo, 2011; Benedetto et al., 2015; Lee et al., 2002). In this
study, a series of TTC thresholds (i.e., 0.5 s, 1 s, 1.5 s, 2 s, 2.5 s,
(1) What are the key situation indicators of rear-end collisions
3 s) will be used to test the trade-off between false alarm rate
and how do they influence the RTAEB system activation?
and successful intervention rate under different situations. Once
(2) How do the drivers’ real-time responses influence the RTAEB
the real-time TTC value is smaller than the TTC thresholds, it is
system activation?
regarded as the emergence of danger situation and the collision
(3) What are the effects of RTAEB system on the collision avoid-
avoidance process will be switched on.
ance process and outcomes?

2.1.2. Collision avoidance


The paper is organized in the following sequence: Section 2
The collision avoidance function will be triggered depending
describes the technical framework design of the RTAEB and the
on whether the drivers’ real-time response is sufficient or not.
design of driving simulator experiment. Section 3 demonstrates
Once the actual TTC between the two vehicles drops below the
the experimental results and the system effectiveness. Section 4
preset TTC threshold, the system calculates the required deceler-
presents the interpretations and discussions of the results. Sec-
ation rate (amf(ti)) based on the following algorithm using the
tion 5 concludes the implications of the proposed RTAEB frame-
real-time speed, headway distance, and leading vehicle’s deceler-
work and future study directions.
ation rate.
The driving distance of leading and following vehicle within the
2. Methodology time interval (ti, tf) could be calculated through Eqs. (1) and (2).
R tf  R tf     
The methodology framework of this study was shown in Fig. 1. xl ðu; tÞ ¼ ti
v l ðt i Þ þ ti
al ðt i Þdu dt u 2 ti ; tf ; t 2 ti ; tf ð1Þ
The framework of RTAEB system was designed considering differ-
R tf  R tf     
ent key parameters, such as TTCth, amf(ti), amax. In light of the defect xf ðu; t Þ ¼ ti
v f ðt i Þ þ ti
amf ðt i Þdu dt u 2 ti ; tf ; t 2 ti ; tf ð2Þ
of historical data, a driving simulator experiment was conducted to
418
J. Hang, X. Yan, X. Li et al. Journal of Safety Research 80 (2022) 416–427

Fig. 1. Methodology framework.

Fig. 2. The procedure of RTAEB system.

 2
where xl(u,t) is the driving distance of the leading vehicle; xf(u,t) is amf ðti Þ ¼ al ðt i Þ þ 0:5  v f ðti Þ  v l ðti Þ =ðLv  Dðti ÞÞ ð7Þ
the driving distance of the following vehicle; ti is the start time of
the collision avoidance process; tf is the end time of the process; vl(- where D(ti) is the headway distance between leading and following
ti) is the leading vehicle’s speed at time ti; vf(ti) is the following vehicle at time ti.
vehicle’s speed at time ti; amf(ti) is the following vehicle’s required Through comparing the value of actual deceleration apf(ti) and
deceleration rate at time ti; al(ti) is the leading vehicle’s decelera- required deceleration amf(ti) calculated from Eqs. (1)–(7), the con-
tion rate at time ti. trol mode to be triggered can be determined. However, as a sudden
It should be noted that tf is an unknown variable in the calcula- hard brake could lead to driver discomfort and even cause driver
tion of xl(u,t) and xf(u,t) as it will be determined by the system injury, a maximum acceptable deceleration value was considered
algorithm according to the definition of the end of collision avoid- to avoid an unrealistic amf(ti), which was represented as amax.
ance. It was defined that a successful collision avoidance ends as Therefore, the modified value of amf(ti) could be obtained in the fol-
speed of the following vehicle reduces to the same as that of the lowing equation:
leading vehicle, and the two vehicles maintain the required safety  
clearance distance, as shown in Eqs. (3) and (4). amf ðti Þ ¼ min amf ðt i Þ; amax ð8Þ
 
D t f ¼ Lv ð3Þ It should be noted that the maximum deceleration rate varies
    according to road surface conditions, with the value higher on
vl tf ¼ v f tf ð4Þ dry roads and lower on icy roads (Chen et al., 2017; Seacrist
et al., 2020). The braking capacity of vehicles equipped with Antil-
where D(tf) is the headway distance between leading and following
ock Brake System (ABS) was also enhanced by increasing road
vehicle at time tf; vl(tf) is the leading vehicle’s speed at time tf; vf(tf)
adhesion (Kudarauskas, 2007; Vangi & Virga, 2007). Since the max-
is the following vehicle’s speed at time tf; Lv is the safety clearance
imum acceptable deceleration correlates to the system triggering, a
distance set in the algorithm.
range of values (i.e., 7, 7.5, 8, 8.5 and 9m/s2) were chosen
Then, the value of time interval Dt could be calculated through
according to the empirical values applied in the above studies.
the following Eq. (5):
Based on Eqs. (1)–(8), the required deceleration rate amf(ti) is
   
Dt ¼ t f  t i ¼ v l ðti Þ  v f ðti Þ= amf ðt i Þ  al ðt i Þ ð5Þ calculated in real-time once a conflict is identified. In the cases
when the situation is under control by the driver, which means
Based on the positional relationship between the leading and the driver takes an equivalent or larger deceleration rate than
following vehicle during the collision avoidance process, amf(ti) the calculated value (apf(ti)  amf(ti)), the collision avoidance func-
can be obtained through Eqs. (6) and (7). tion will not be activated. In the other cases (apf(ti) > amf(ti)), the
Lv  Dðt i Þ ¼ xl ðu; t Þ  xf ðu; tÞ ð6Þ collision avoidance function will be activated to help driver achieve
the modified required deceleration amf(ti) and avoid the collision.
419
J. Hang, X. Yan, X. Li et al. Journal of Safety Research 80 (2022) 416–427

2.2. Driving simulator experiment drivers were other occupations who often drove for commute. Each
participant held a valid driver’s license and had at least 3 years of
2.2.1. Scenario design driving experience. Each participant held a valid driver’s license
The study designed a typical car-following scenario on a two- and had at least 3 years of driving experience. The age of them ran-
way, two-lane sub-urban road with the speed limit of 70 km/h. ged from 30 to 40 years old with an average of 33 years old and
As drivers’ behavior and safety in a car-following scenario were 3 years of standard deviation. As age was identified as an impor-
correlated with the traffic situation, especially the leading vehicle’s tant factor of driving behavior (Lambert-Belanger et al., 2012;
maneuver, different levels of traffic density and deceleration rates Pitta et al., 2021; Zafian et al., 2021), the study used a strict age
of the leading vehicle were designed. The traffic density was repre- range to collect a homogeneous sample with relatively mature
sented by the operating speed of the surrounding vehicles and it and stable driving performance.
included three levels: low density/speed (5–30 km/h), mid den- The study used the BJTU driving simulator located at the Beijing
sity/speed (30–50 km/h), and high density/speed (40–70 km/h), Jiaotong University (see Fig. 4). It is a high-performance, high-
as shown in Fig. 3. Fig. 3 also shows the speed profile of the leading fidelity driving simulator produced by the Real-time Technologies
vehicle in the three density/speed conditions. Specifically, in each Inc. in the United States. It has 300° horizontal field of view with
drive the leading vehicle had three acceleration stages (with the scenarios displayed in 8 channels (five forward views and three
acceleration rate of 0.5 m/s2, 1 m/s2 and 1.5 m/s2), three decelera- rear view mirrors) at a resolution of 1400  1050 pixels. A full-
tion stages (with the deceleration rate of 2 m/s2, 4 m/s2 and 6 m/ size fixed-based vehicle cabin (Ford Focus) is equipped with com-
s2) and five constant speed driving stages. plete operational interface as a real-vehicle, including steering
wheel, brake pedal, gas pedal, dashboard, gear, rear-view mirrors,
and so forth. Meanwhile, an environmental noise simulation sys-
2.2.2. Participants and apparatus
tem and a vehicle vibration simulation system are also included
This study recruited 46 participants including 24 males (12 pro-
in this simulator to simulate the real driving environment. The
fessional drivers and 12 non-professional drivers) and 22 females
SimCreator and SimVista software is used for designing the sce-
(10 professional drivers and 12 non-professional drivers), and all
nario and vehicle dynamic model. The simulator collects data at
of them completed the experiment successfully. The professional
a sampling rate of 60 Hz during the experiment.
drivers were full-time taxi drivers, and the non-professional

 (a) Low driving speed (5-30 km/h)


Deceleration Deceleration
sensor 3 Deceleration
 Deceleration sensor 2:
sensor 3:
Vl (km / h)

sensor 1: al=-4m/s 2
al=-6m/s 2
al=-2m/s 2


L 



    W V
(b) Mid driving speed (30-50 km/h)
 Deceleration
Deceleration Deceleration sensor 2: Deceleration
sensor 2 sensor 1: al=-2m/s 2 sensor 3:
 al=-6m/s 2
Vl (km / h)

al=-4m/s 2


L 


 W
   W V
Deceleration
sensor 1
Leading
(c) High driving speed (40-70 km/h)

L vehicle  Deceleration
 sensor 3:
Vl (km / h)

al=-6m/s 2
Subject 
S vehicle Deceleration
Deceleration
sensor 2:
sensor 1:
 al=-2m/s 2
al=-4m/s 2
 W
   W V

Fig. 3. Experimental scenario and speed and deceleration rate of leading vehicle.

420
J. Hang, X. Yan, X. Li et al. Journal of Safety Research 80 (2022) 416–427

Fig. 4. Illustration of the driving simulator.

2.2.3. Experiment procedure Based on the classification listed in Table 1, four indexes includ-
Upon arrival, each participant read and signed an informed con- ing collision avoidance rate (CAR), accuracy rate (AR), sensitivity
sent form (per IRB). They were briefly introduced about the exper- rate (SR) and precision rate (PR) were calculated to evaluate the
iment and advised to drive as they normally would in a real-life system effectiveness in different conditions. Definitions of the four
situation. Before the formal test, each participant drove the simu- indexes are listed as below. The indexes will be compared under
lator for about 5 min to get familiar with the operations such as different TTC and amax thresholds.
acceleration, deceleration, and turning. Moreover, participants
were advised that they could quit the experiment at any time in (1) Collision avoidance rate (CAR): ratio of successful collision
case of motion sickness or any discomfort. avoidance samples (STS) in all true activation samples. Note
During the formal test, each participant drove three scenarios that not all samples under correct intervention successfully
corresponding to the three speed profiles of the leading vehicle avoided the collisions due to the smaller TTC thresholds or
as shown in Fig. 3. The order of three scenarios were counterbal- other reasons. Thus, CAR reflects whether the system could
anced among all the participants. As the leading vehicle deceler- help avoid collisions successfully.
ated three times at different deceleration rates in each scenario,
each participant would encounter 3  3 times of rear-end collision CAR ¼ STS=TA  100%
avoidance events in total. Between each two scenarios, the partic-
ipant took a rest for about 5 min, and the entire experiment took
about 1 h to complete. Each participant was asked to finish a ques-
tionnaire regarding the validity of the driving simulator and driv- (2) Accuracy rate (AR): ratio of true activation samples and true
ing scenario after all the formal tests. Six questions were inactivation samples in all samples, reflecting the system
included in the questionnaire (i.e., four ratings of the driving sim- accuracy in judging the safety–critical situations.
ulator operation: overall performance, accelerator, brake, and
steering wheel control) and two ratings of the driving environment AR ¼ ðTA þ TN Þ=ðTA þ FN þ FA þ TNÞ  100%
(i.e., other vehicles and buildings) on a five-point Likert scale. After
completing the experiment, each participant received 100 RMB
(about 15 USD) as a compensation for their participation.
(3) Sensitivity rate (SR): ratio of true activation samples in all
2.3. System evaluation indexes and key variables activation-required samples (i.e. collision samples).

The proposed system algorithm were performed on the col- SR ¼ TA=ðTA þ FNÞ  100%
lected samples data. If the system determines that driver’s deceler-
ation is insufficient, the sample data was divided into trigger
group. Otherwise, the sample data was divided into non-trigger
group. The relationships between the system control states and (4) Precision rate (PR): ratio of true activation samples in all
collision outcomes is shown in Table 1. triggered samples.

PR ¼ TA=ðTA þ FAÞ  100%


Table 1
System effectiveness indexes. Apart from the system evaluation indexes, the situational indi-
cators and driver behavioral indicators were also extracted from
System control Collision outcome
the raw experiment data to investigate their effects on system acti-
Collision Non-collision vation pattern. Meanwhile, through comparing the safety indica-
Trigger True activation False activation tors such as reaction time, maximum deceleration rate, minimum
(TA) (Inaccurate judgment) (FA) distance, and collision number between system activation and
Non-trigger False inactivation True inactivation
(Missing judgment) (FN) (TN)
inactivation conditions, the system effectiveness was validated
and analyzed. The variables and the definitions were shown in
Table 2.
421
J. Hang, X. Yan, X. Li et al. Journal of Safety Research 80 (2022) 416–427

Table 2
Variables and their definitions under real driving.

Variable Variable Definition


type
System mode Categorical 0-driver control mode (DCM, no
system-triggered condition), 1-system
control mode (SCM, system-triggered
condition).
Driver behavioral indicators
Gender Categorical 0-male, 1-female.
Profession Categorical 0-professional driver, 1-non-
professional driver.
Situational indicators
Traffic condition Categorical 0-low speed, 1-mid speed, 2-high
speed.
Leading vehicle’s Categorical 0- 2m/s2, 1- 4m/s2, 2- 6m/s2.
deceleration rate
(al)
Initial speed of Continuous The subject vehicle’s speed at the start
following vhielce of leading vehicle’s deceleration.
(vf(t0)) (m/s)
Fig. 5. The validity of driving simulator (1 = not effective, 5 = highly effective).
Initial speed of Continuous The leading vehicle’s speed at the start
leading vhielce of leading vehicle’s deceleration.
(vl(t0)) (m/s)
increase of TTC threshold, the collision avoidance rate also
Initial distance (D Continuous The distance from the rear of leading
(t0)) (m) vehicle to the front of subject vehicle increased. When the TTC threshold was greater than 2 s, all the col-
at the start of leading vehicle’s lisions could be successfully avoided. Meanwhile, the increase of
deceleration. amax threshold also improved the successful rate of collision avoid-
Safety indicators ance. An appropriate combination of TTC threshold and amax
Reaction time (tr) (s) Continuous Driver’s brake reaction time, i.e. the threshold could increase the system’s collision avoidance rate to
time from when the leading vehicle
100%, e.g. TTC = 1.5 s and amax = 7.5 m/s2.
starts to decelerate to the time that the
driver starts to decelerate. The sensitivity rate (SR) represents the system’s sensitivity to
Maximum Continuous Driver’s maximum deceleration rate rear-end risks and low SR means that the system will miss some
deceleration rate from the subject vehicle begun to dangerous cases in the judgement process. As shown in Fig. 6(b),
(afmax) (m/s2) decelerate to the deceleration finished. the SR increased gradually with the increase of TTC thresholds.
Minimum distance Continuous The minimum distance from the rear
When the TTC threshold increased to 1.5 s, all the missing cases
(Dmin) (m) of leading vehicle to the front of
subject vehicle during the subject could be eliminated and the SR reached to 100%. Besides, the
vehicle begun to decelerate to the results also showed that for small TTC thresholds (e.g., 0.5 s or
deceleration finished. 1 s), a larger amax threshold could help improve the system
sensitivity.
The precision rate (PR) reflects the extent of system’s inaccurate
3. Results judgments. Low PR means that a large number of safe cases are
regarded as requiring system control. The results showed that
Among the 414 (46 participants  3 driving speed  3 deceler- the PR remarkably decreased when the TTC threshold increased
ation rate) data samples, 61 cases were excluded as outliers as from 0.5 s to 3 s as seen in Fig. 6(c), which means higher TTC
the driver had a large car-following distance (60 m) or a long thresholds could lead to more inaccurate judgment of the system
reaction time (3 s), remaining 353 data samples for further on the situations. As for different amax thresholds, changes in PR
analysis. were not obvious.
The post-experiment questionnaire was first analyzed as a sub- The accuracy rate (AR) measures the accuracy of the system in
jective evaluation of the simulator validity. As shown in Fig. 5, all implementing interventions, which integrated the information of
the six questions received higher scores than a middle value (i.e., PR and SR. Similar to the tendency of PR, the AR increased with
3), indicating that drivers were overall satisfied with the validity the decrease of TTC threshold as shown in Fig. 6(d). When the
of the simulator. The item with highest score was accelerator TTC threshold reduced to 0.5 s, the system could intervene with
(Mean = 4.43, S.D. = 0.65), while the lowest was brake pedal an accuracy rate of over 97%. However, there was no significant
(Mean = 3.23, S.D. = 1.20). As for the driving environment, driver’s changes in AR under different amax thresholds.
subjective ratings were almost the same for other vehicles In the premise of 100% collision avoidance rate, an optimal
(Mean = 4.02, S.D. = 1.07) and surrounding buildings (Mean = 4.06, combination of TTC and amax thresholds could be obtained given
S.D. = 1.01). the alternative values. According to the results of AR, it seemed
that smaller TTC thresholds are more conducive to improve the
accuracy of the system. However, small TTC thresholds also needs
3.1. System evaluation larger amax to ensure the CAR of the system control to reach 100%,
which may increase driver’s discomfort during the collision avoid-
Based on the experimental data, the system evaluation indexes ance process. More importantly, a small TTC threshold is likely to
were calculated, and changes of the indexes under different TTC omit some risky cases. As for a large TTC threshold, although lower
and amax thresholds were shown in Fig. 6. PR may bring distrust of some drivers, it can ensure more safety.
The collision avoidance rate (CAR) is a critical safety index that Given all the consideration, a TTC threshold of 1.5 s and an amax
reflects the RTAEB’s capability in avoiding collisions. A primary threshold of 7.5 m/s2 were recommended in this study, and the
principle of the system is that a 100% successful rate in collision combination was used in the following system performance
avoidance should be a priority. Fig. 6(a) showed that with the analysis.

422
J. Hang, X. Yan, X. Li et al. Journal of Safety Research 80 (2022) 416–427

Fig. 6. Rate of collision avoidance, accuracy, sensitivity and precision under different TTC and amax thresholds.

3.2. System performance In this section, the differences between the key variables in the
two sets of raw data were analyzed (i.e., samples requiring system
3.2.1. System intervention intervention vs. samples that did not require system intervention).
According to the 353 valid samples, 94.62% of them safely As seen in Table 3, the differences were checked by T-test and Chi-
avoided the collision and the rest 5.38% had a collision with the square test. The results showed that leading vehicle’s deceleration
leading vehicle. After applying the RTAEB algorithm on the 353 rate al (Chi-square = 101.851, P-value < 0.001), initial distance D
samples, 84 samples (23.80%) were determined to require a system (t0) (t = 7.557, P-value < 0.001), reaction time tr (t = 2.257, P-
intervention, while the rest of the samples could complete the value = 0.025), maximum deceleration rate afmax (t = 16.126, P-
deceleration by driver control. value < 0.001), and minimum headway distance Dmin (t = 19.119,

Table 3
T-test analysis and Chi-square test results of initial condition.

Parameter No intervention Intervention Chi-square P-value


Count N% Count N%
Gender Male 137 38.81 40 11.33 0.281 0.596
Female 132 37.39 44 12.46
Profession P 125 35.41 37 10.48 0.151 0.698
NP 144 40.79 47 13.31
Traffic condition Low 80 22.66 29 8.22 0.987 0.610
Mid 94 26.63 25 7.08
High 95 26.91 30 8.50
al 2m/s2 113 32.01 1 0.28 101.851 <0.001
4m/s2 101 28.61 17 4.82
6m/s2 55 15.58 66 18.70
Collision numbers 0 0.00 19 5.38
Total numbers 269 76.20 84 23.80
Mean S.D. Mean S.D. t P-value
D(t0) (m) 24.04 12.99 15.30 7.73 7.557 <0.001
vl(t0) (m/s) 12.41 4.81 11.56 4.27 1.558 0.121
vf(t0) (m/s) 12.63 5.04 11.61 4.26 1.822 0.070
tr (s) 1.27 0.91 1.10 0.47 2.257 0.025
afmax (m/s2) 4.48 1.90 7.31 1.22 16.126 <0.001
Dmin (m) 15.13 10.08 2.57 2.13 19.119 <0.001

423
J. Hang, X. Yan, X. Li et al. Journal of Safety Research 80 (2022) 416–427

P-value < 0.001) between the system intervention and non- that triggered the system were different. Paired-sample T-tests
intervention groups showed significant difference. were performed and the results were shown in Table 4.
The leading vehicle’s deceleration rate significantly influenced For pattern 1 (late response group), the reaction time under dri-
whether the system was activated or not. As the leading vehicle’s ver control and reaction time after the modification of system
deceleration rate increased from 2 m/s2 to 4 m/s2, the number of intervention showed significant difference (t = 5.547, P-
samples that required system intervention increased from 1 value < 0.001). The reaction time in SCM group (Mean = 0.98 s, S.
(0.28%) to 66 (18.70%). The initial headway distance in the inter- D. = 0.60 s) was smaller than that in DCM group (Mean = 1.18 s, S.
vention group (Mean = 15.30 m, S.D. = 7.73 m) was significantly D. = 0.60 s), which illustrated that the system could compensate
smaller than that of the non-intervention group (Mean = 24.04 m, some drivers’ late reaction. The mean value of minimum distance
S.D. = 12.99 m), which indicates that drivers who followed the was significantly different (t = 5.139, P-value < 0.001) between
leading vehicle closely were more likely to trigger the system. driver control and system control, and the system (Mean = 2.98 m,
Regarding the safety indicators, the reaction time of the sam- S.D. = 1.30 m) enlarged the minimum distance by nearly 2 m com-
ples who needed system intervention (Mean = 1.10 s, S.D. pared to the human drivers (Mean = 1.01 m, S.D. = 1.20 m). Fur-
= 0.47 s) was smaller than those who did not (Mean = 1.27 s, S.D. thermore, under system control, all eight of the collisions that
= 0.91 s). The maximum deceleration rate of samples who needed occurred in this pattern group could be avoided successfully. How-
system intervention (Mean = 7.31 m/s2, S.D. = 1.22 m/s2) was lar- ever, the average deceleration rate showed no significant differ-
ger than those who did not (Mean = 4.48 m/s2, S.D. = 1.90 m/s2). ence between driver control and system control (t = 1.999, P-
Moreover, samples who needed system intervention also ended value = 0.061).
up with smaller minimum distance with the leading vehicle For pattern 2 (insufficient deceleration group), there was no sig-
(Mean = 2.57 m, S.D. = 2.13 m) compared with the non- nificant difference on the reaction time. However, the maximum
intervention group (Mean = 15.13 m, S.D. = 10.08 m). The experi- deceleration rate (t = 2.675, P-value = 0.009) and minimum dis-
ment collected 19 collisions and they all occurred on samples tance (t = 2.034, P-value = 0.046) showed significant differences
who were identified to require a system intervention. Overall, it between driver control and system control. Samples under system
can be inferred that samples under the group of system interven- control showed a smaller maximum deceleration rate
tion always encounter more dangerous situations. The combina- (Mean = 6.79 m/s2, S.D. = 1.79 m/s2) compared to the drivers
tion of the TTC threshold and amax threshold selected in this (Mean = 7.28 m/s2, S.D. = 1.24 m/s2). The minimum distance
research could effectively identify the dangerous cases. under system (Mean = 3.52 m, S.D. = 2.09 m) was also larger than
that under human control (Mean = 3.03 m, S.D. = 2.13 m). Mean-
while, the system control could successfully avoid all the 11 colli-
3.2.2. System effectiveness
sions that occurred in driver control group.
According to the observation of driver’s deceleration behavior,
the cases requiring the system intervention could be divided into
two categories: drivers with late response and/or drivers with 4. Discussion
insufficient deceleration. Typical deceleration patterns of the two
type of drivers were shown in Fig. 7. When the driver did not This paper designed a real-time Autonomous Emergency Brak-
respond timely to the leading vehicle’s deceleration, the system ing (RTAEB) system based on the traditional conflict judgment
would intervene before the driver’s actual deceleration started and risk perception process, which was used to judge whether to
(Fig. 7(a)). For drivers who were aware of the risk but failed to take implement autonomous braking according to driver’s different
sufficient deceleration, the system would enlarge the deceleration reactions. In order to preliminarily verify the effect of the system,
rate to a required value (Fig. 7(b)). a driving simulator-based experiment was conducted to explore
To examine the effectiveness of the system, the differences of the system effectiveness under different risk thresholds and com-
key variables before and after the application of system control fort levels, and to compare which conditions and driver character-
were compared. Behavior under driver control mode (DCM) repre- istics were more likely to trigger autonomous control. Further
sented the results of raw data set, which required the system inter- comprehensive discussions will be illustrated in this section.
vention, and behavior under system control mode (SCM) To make up for driver mistakes, previous studies have
represented the results of the same data set after being modified attempted to deliver warning massages in advance under emer-
by RTAEB. The comparison was conducted based on the two pat- gent situations, which help drivers to avoid collisions by shorten-
tern groups depicted in Fig. 7, respectively, since the conditions ing their reaction time and smoothing the braking process (Ho

Fig. 7. Different behavior patterns.

424
J. Hang, X. Yan, X. Li et al. Journal of Safety Research 80 (2022) 416–427

Table 4
Descriptive analysis and T-test results.

Pattern Parameter Count Behavior under DCM Behavior under SCM T-test results
Mean S.D. Mean S.D. t P-value
1 tr (s) 19 1.18 0.60 0.98 0.60 5.547 <0.001
afmax (m/s2) 19 7.43 1.16 6.47 1.84 1.999 0.061
Dmin (m) 19 1.01 1.20 2.98 1.30 5.139 <0.001
Collision numbers 19 8 (2.27%) under DCM, 0 (0.00%) under SCM
2 tr (s) 65 1.08 0.42 1.08 0.42 – –
afmax (m/s2) 65 7.28 1.24 6.79 1.79 2.675 0.009
Dmin (m) 65 3.03 2.13 3.52 2.09 2.034 0.046
Collision numbers 65 11 (3.12%) under DCM, 0 (0.00%) under SCM

et al., 2006; Wu et al., 2018a; Zhu et al., 2020). However, if the cle braked with larger deceleration rate and when the initial dis-
warning system was triggered too late, or drivers failed to comply tance between two vehicles was small. This finding is consisted
with the warning information, it would still lead to accidents and with previous studies that reported that rear-end collision risk
even cause property damage and casualties. Different from the would increase due to the leading vehicle’s sudden brake or close
warning system, the autonomous collision avoidance system, as car-following distance (Shangguan et al., 2020; Xue, et al., 2018).
the last defensive line, needed to interfere with driver’s behavior Besides, from the following vehicle drivers’ perspective, the cases
when the danger came, and thus effectively reduce the crash rates requiring system control often corresponded to drivers with a
(Cicchino, 2017). In the previous studies of autonomous braking short reaction time, which means they probably needed to apply
system’s algorithm, researchers usually used the speed and posi- an emergent brake (Li et al., 2019; Liebermann et al., 1995;
tion information of the leading vehicle to infer the safe distance Wang et al., 2016). Even so, the rear-end collisions might still be
(Bella & Russo, 2011; Doi, 1994; Wang et al., 2016). However, the inevitable. Therefore, when the effectiveness of warning messages
changes in driver behavior were often unpredictable, and the vari- cannot be guaranteed, autonomous control intervention is neces-
ation of deceleration and reaction time would also reduce the accu- sary to be adopted as once the cases are judged to require system
racy of the autonomous emergency braking system (Xiong et al., control, any form of warnings would be invalid due to the insuffi-
2019). Based on the speed and position information in each time cient response time.
interval, a real-time based Autonomous Emergency Braking algo- The study verified the system effectiveness through comparing
rithm was established in this paper, which judged driver’s risk the differences of key variables before and after the application of
level according to their real-time reaction. system control. It was found that the RTAEB could solve the inap-
The speed and accuracy of drivers in identifying and responding propriate deceleration cause by late response or insufficient decel-
to the impending rear-end risks determine if they can avoid the eration. For the late response drivers, the application of the system
collisions successfully or not. For the collision avoidance system, could shorten the reaction time of taking brake action and enlarge
overly-early intervention normally sacrifices the accuracy to some the minimum distance. For the insufficient deceleration drivers,
extent and may confuse the drivers when the risk cues have not the system could reduce hard braking and increase the minimum
been perceived yet (Wang et al., 2016; Xiong et al., 2019). On the distance through intervening the braking process early. All the col-
other hand, overly-late system interventions could result in the lision cases in the raw data sets were eliminated after applying the
failure of collision avoidance (Brannstrom & Sjoberg, 2008; RTAEB.
Vahidi & Eskandarian, 2004). This study proposed a RTAEB frame- In addition, the research also found that there was no signifi-
work to help driver avoid rear-end collisions and examined a series cant difference caused by driver’s characteristics in terms of trig-
of TTC thresholds (from 0.5 s to 3 s) and maximum deceleration gering or not triggering the RTAEB. To some extent, the system
thresholds (from 7.5 m/s2 to 9m/s2) to seek an optimal balance did not need to distinguish driver’s gender and profession factors.
between safety and accuracy. Four system evaluation indexes were The finding implies that for the rear-end collision avoidance sys-
analyzed including collision avoidance rate, accuracy rate, preci- tem, more attention should be paid to the real-time risk situation
sion rate, and sensitivity rate. The study demonstrated the changes of external environment and driver’s risk avoidance performance,
of the indexes under different TTC and amax thresholds. It was such as reaction ability and deceleration ability.
found that when the TTC threshold was larger than 2 s, all the col-
lisions were successfully avoided. However, larger TTC thresholds 5. Conclusion and limitation
also brought lower precision rate since there might be some safe
cases divided into the system intervention group. Although smaller This study proposed a new framework of Autonomous Emer-
TTC thresholds could achieve a higher accuracy rate and precision gency Braking System (RTAEB). According to vehicles’ real-time
rate, it needed a larger amax to ensure the collision avoidance rate information and driver’s real-time response, the system could dis-
of the system could reach 100%, which may increase driver’s dis- tinguish whether there was a risky situation in the car-following
comfort during the deceleration process. Moreover, smaller TTC process. If the situation became critical, the system could assist dri-
thresholds had another serious defect (i.e., the missing judgments vers to brake safely. A driving simulator experiment was con-
due to the low sensitivity rate of the system), which may omit ducted to collect the drivers’ collision avoidance behavioral data
some risky situations. Therefore, after comprehensive considera- and thus provided the test database for the RTAEB system. The test
tion, this research recommended a TTC threshold of 1.5 s and an results generated several implications:
amax threshold of 7.5 m/s2 as the parameters of RTAEB, and the
combination of the mentioned thresholds were used as a case  The system performance changed with the thresholds of TTC
study to further examine the performance of the system. and maximum deceleration rate. TTC threshold directly affected
In the case study, the paper explored which driving conditions the accuracy of the system. Larger TTC thresholds could result in
were more inclined to trigger the system. The results indicate that inaccurate judgments, and smaller TTCs could cause missed
the system was more likely to be triggered when the leading vehi- judgments. The maximum deceleration threshold directly

425
J. Hang, X. Yan, X. Li et al. Journal of Safety Research 80 (2022) 416–427

affected the comfort level of drivers. Considering the two Bella, F. . (2005). Validation of a driving simulator for work zone design.
Transportation Research Record(1937), p.136-144.
parameters, it was suggested that 1.5 s TTC and 7.5 m/s2 max-
Benedetto, F., Ca Lvi, A., D’Amico, F., & Giunta, G. (2015). Applying
imum deceleration rate were suitable thresholds for risk assess- telecommunications methodology to road safety for rear-end collision
ment, which also provides a theoretical reference for the future avoidance. Transportation Research Part C, 50(jan.), 150–159.
autonomous control system. Bham, G. H., Leu, M. C., Vallati, M., & Mathur, D. R. (2014). Driving simulator
validation of driver behavior with limited safe vantage points for data collection
 The system was more likely to be triggered when the situation in work zones. J Safety Res, 49.
was urgent or dangerous, such as leading vehicle’s sudden Braitman, K. A., Mccartt, A. T., Zuby, D. S., & Singer, J. (2010). Volvo and infiniti
brake, or too close car-following. In these cases, the reaction drivers’ experiences with select crash avoidance technologies. Traffic Injury
Prevention, 11(3), 270–278.
time left for the driver was very short. If the driver did not Brannstrom, M. , Sjoberg, J. , & E C oelingh. (2008). A situation and threat assessment
respond in time or brake improperly, the collision risks rose. algorithm for a rear-end collision avoidance system. Intelligent Vehicles
This also explains why the autonomous control system is Symposium, 2008 IEEE. IEEE.
Calvi, A., D’Amico, F., Ferrante, C., & Ciampoli, L. B. (2020). A driving simulator study
needed as a final defense to protect the drivers. to assess driver performance during a car-following maneuver after switching
 The driver’s characteristics, such as gender and profession, did from automated control to manual control. Transportation Research Part F Traffic
not show significant differences in this study. This suggests that Psychology and Behaviour, 70, 58–67.
Chang, S. H., Lin, C. Y., Hsu, C. C., Fung, C. P., & Hwang, J. R. (2009). The effect of a
in the process of collision avoidance, driver’s present capability collision warning system on the driving performance of young drivers at
to avoid danger is more important. In the future, driver’s reac- intersections. Transportation Research Part F Traffic Psychology & Behaviour, 12
tion and braking ability should be highlighted in driver train- (5), 371–380.
Chen, C., Xiang, H., Qiu, T., Cong, W., Yang, Z., & Chang, V. (2017). A rear-end collision
ing/education process.
prediction scheme based on deep learning in the internet of vehicles. Journal of
Parallel and Distributed Computing, 117(jul.), 192–204.
Finally, some limitations of the study should be mentioned. The Cicchino, J. B. (2017). Effectiveness of forward collision warning and autonomous
study only examined limited TTC and maximum deceleration emergency braking systems in reducing front-to-rear crash rates. Accident;
analysis and prevention, 99(Pt A), 142–152.
thresholds because of the limitation of driving simulator experi- Doi, A. (1994). Development of a rear-end collision avoidance system with
ment. Although we believe the investigated options are sufficient automatic brake control. JSAE Review, 15(4), 335–340.
to reflect the underlying patterns, more values should be tested Duan, J., Li, R., Hou, L., Wang, W., Li, G., Li, S. E., et al. (2017). Driver braking
behavior analysis to improve autonomous emergency braking systems in
using different traffic simulation methods to validate the results typical chinese vehicle-bicycle conflicts. Accident Analysis & Prevention, 108
observed in the study. Moreover, a restricted age group of drivers (nov.), 74–82.
were examined to avoid the potential confounding effect of drivers’ _
Duzmańska, N., Strojny, P., & Strojny, A. (2018). Can simulator sickness be avoided?
a review on temporal aspects of simulator sickness. Frontiers in Psychology, 9.
age. To verify the generalizability of the study results, a larger sam- Fildes, B., Keall, M., Bos, N., Lie, A., Page, Y., Pastor, C., et al. (2015). Effectiveness of
ple size with broad age distribution should be tested in the future. low speed autonomous emergency braking in real-world rear-end crashes.
In addition, this paper uses the car-following behavior data derived Accident Analysis & Prevention, 81(aug.), 24–29.
Fujita, Y. (1995). Radar brake system. JSAE Review, 16(1), 113.
from a driving simulator to preliminarily verify the rear-end colli- Haas, E. C., & Van Erp, J. B. F. (2014). Multimodal warnings to enhance risk
sion avoidance algorithm of RTAEB, which provides theoretical communication and safety. Safety ence, 61, 29–35.
basis for the application of RTAEB. In the future, the algorithm Hang, J., Yan, X., Ma, L., Duan, K., & Zhang, Y. (2018). Exploring the effects of the
location of the lane-end sign and traffic volume on multistage lane-changing
should be integrated into the driving simulator to conduct an
behaviors in work zone areas: A driving simulator-based study. Transportation
online test. Research Part F Traffic Psychology and Behaviour, 58(OCT.), 980–993.
Hess, S., Choudhury, C. F., & Bliemer, M. C. J. (2020). Modelling lane changing
behaviour in approaches to roadworks: Contrasting and combining driving
CRediT authorship contribution statement simulator data with stated choice data. Transportation Research Part C: Emerging
Technologies, 112, 282–294.
Ho, C., Reed, N., & Spence, C. (2006). Assessing the effectiveness of ‘‘intuitive”
Junyu Hang: Conceptualization, Methodology, Software, Formal vibrotactile warning signals in preventing front-to-rear-end collisions in a
analysis, Investigation, Visualization, Writing – original draft. Xue- driving simulator. Accid Anal Prev, 38(5), 988–996.
dong Yan: Supervision, Conceptualization, Methodology, Writing – Hulst & Der, M. V. (1999). Anticipation and the adaptive control of safety margins in
driving. Ergonomics, 42(2), 336–345.
review & editing. Xiaomeng Li: Methodology, Writing – review & Hussain, Q., Alhajyaseen, W., Pirdavani, A., Reinolsmann, N., Brijs, K., & Brijs, T.
editing. Ke Duan: Methodology, Writing – review & editing. Jingsi (2019). Speed perception and actual speed in a driving simulator and real-
Yang: Writing – review & editing. Qingwan Xue: Writing – review world: A validation study. Transportation research, 62F(APR.), 637–650.
Jeong, E., Oh, C., & Lee, S. (2017). Is vehicle automation enough to prevent crashes?
& editing. role of traffic operations in automated driving environments for traffic safety.
Accident Analysis & Prevention, 104, 115–124.
Kala, R., & Warwick, K. (2013). Multi-level planning for semi-autonomous vehicles
Acknowledgements in traffic scenarios based on separation maximization. Journal of Intelligent &
Robotic Systems, 72(3–4), 559–590.
Kudarauskas, N. (2007). Analysis of emergency braking of a vehicle. Transport, 22(3),
This research was supported by the BJTU Basic Scientific
154–159.
Research (2019YJS105) and National Natural Science Foundation Kusano, K. D., & Gabler, H. C. (2010). Potential occupant injury reduction in pre-
of China (72171017, 71621001). crash system equipped vehicles in the striking vehicle of rear-end crashes.
Annals of advances in automotive medicine, 54, 203–214.
Kusano, K., & Gabler, H. C. (2012). Safety benefits of forward collision warning,
References brake assist, and autonomous braking systems in rear-end collisions. IEEE
Transactions on Intelligent Transportation Systems.
Lambert-Belanger, A., Dubois, S., Weaver, B., Mullen, N., & Bedard, M. (2012).
Almallah, M., Hussain, Q., Reinolsmann, N., & Alhajyaseen, W. (2021). Driving
Aggressive driving behaviour in young drivers (aged 16 through 25) involved in
simulation sickness and the sense of presence: Correlation and contributing
fatal crashes. J Safety Res, 43(5–6), 333–338.
factors. Transportation Research Part F Traffic Psychology and Behaviour, 78(7),
Lee, J. D., Mcgehee, D. V., Brown, T. L., & Reyes, M. L. (2002). Collision warning
180–193.
timing, driver distraction, and driver response to imminent rear-end collisions
Aust, M. L., Engstr, M. J., & Vistr, M. M. (2013). Effects of forward collision warning
in a high-fidelity driving simulator. Human Factors: The Journal of the Human
and repeated event exposure on emergency braking. Transportation Research
Factors and Ergonomics Society, 44(2), 314–334.
Part F: Traffic Psychology and Behaviour, 18, 34–46.
Lewis-Evans, B., W Aa Rd, D. D. , & Brookhuis, K. A. . (2010). That’s close enough-a
Bargman, J., Boda, C. N., & Dozza, M. (2017). Counterfactual simulations applied to
threshold effect of time headway on the experience of risk, task difficulty, effort,
shrp2 crashes: The effect of driver behavior models on safety benefit
and comfort. Accident; Analysis and Prevention, 42(6), 1926-1933.
estimations of intelligent safety systems. Accident Analysis & Prevention, 102
Li, X., Oviedo-Trespalacios, O., & Rakotonirainy, A. (2020). Drivers’ gap acceptance
(MAY), 165–180.
behaviours at intersections: A driving simulator study to understand the impact
Bella, F., & Russo, R. (2011). A collision warning system for rear-end collision: A
of mobile phone visual-manual interactions. Accident Analysis & Prevention, 138
driving simulator study. Procedia - Social and Behavioral Sciences, 20(none),
105486.
676–686.

426
J. Hang, X. Yan, X. Li et al. Journal of Safety Research 80 (2022) 416–427

Li, X., Oviedo-Trespalacios, O., Rakotonirainy, A., & Yan, X. (2018). Collision risk Wu, Y., Abdel-Aty, M., Park, J., & Zhu, J. (2018). Effects of connected-vehicle warning
management of cognitively distracted drivers in a car-following situation. systems on rear-end crash avoidance behavior under fog conditions.
Transportation Research Part F Traffic Psychology and Behaviour, 60, 288–298. Transportation research, 95(OCT.), 481–492.
Li, X., Oviedo-Trespalacios, O., Rakotonirainy, A., & Yan, X. (2019). Collision risk Wu, X., Boyle, L. N., Marshall, D., & O’Brien, W. (2018). The effectiveness of auditory
management of cognitively distracted drivers in a car-following situation. forward collision warning alerts. Transportation research, 59F(PT.A), 164–178.
Transportation Research Part F Traffic Psychology & Behaviour, 60, 288–298. Xiang, W., Yan, X., Weng, J., & Li, X. (2016). Effect of auditory in-vehicle warning
Li, X., Yan, X., & Wong, S. C. (2015). Effects of fog, driver experience and gender on information on drivers’ brake response time to red-light running vehicles
driving behavior on S-curved road segments. Accident Analysis & Prevention, 77, during collision avoidance. Transportation Research Part F Traffic Psychology and
91–104. Behaviour, 40(Jul.), 56–67.
Li, X., Yan, X., Wu, J., Radwan, E., & Zhang, Y. (2016). A rear-end collision risk Xiong, X., Wang, M., Cai, Y., Chen, L., Farah, H., & Hagenzieker, M. (2019). A forward
assessment model based on drivers’ collision avoidance process under collision avoidance algorithm based on driver braking behavior. Accident
influences of cell phone use and gender—a driving simulator based study. Analysis & Prevention, 129(AUG.), 30–43.
Accident Analysis & Prevention, 97, 1–18. Xue, Q., Markkula, G., Yan, X., & Merat, N. (2018). Using perceptual cues for brake
Liebermann, D. G., Ben-David, G., Schweitzer, N., Apter, Y., & Parush, A. (1995). A response to a lead vehicle: Comparing threshold and accumulator models of
field study on braking responses during driving. i. triggering and modulation. visual looming. Accident; analysis and prevention, 118(SEP.), 114.
Ergonomics, 38(9), 9. Yan, X., Li, X., Liu, Y., & Zhao, J. (2014). Effects of foggy conditions on drivers’ speed
Ma, S., & Yan, X. (2021). Examining the efficacy of improved traffic signs and control behaviors at different risk levels. Safety Science, 68, 275–287.
markings at flashing-light-controlled grade crossings based on driving You, S., Wang, X., & Chen, M. (2013). Driver collision avoidance behavior after
simulation and eye tracking systems. Transportation Research Part F Traffic braking. In ICTE 2013-Proceedings of the 4th International Conference on
Psychology and Behaviour, 81(1), 173–189. Transportation Engineering (pp. 2327–2334).
Mcavoy, D., Schattler, K., & Datta, T. (2007). Driving simulator validation for Zafian, T., Ryan, A., Agrawal, R., Samuel, S., & Knodler, M. (2021). Using shrp2 nds
nighttime construction work zone devices. Transportation Research Record data to examine infrastructure and other factors contributing to older driver
Journal of the Transportation Research Board, 2015(2015), 55–63. crashes during left turns at signalized intersections. Accident Analysis &
Meng, F., Wong, S. C., Yan, W., Li, Y. C., & Yang, L. (2019). Temporal patterns of Prevention, 156(106141).
driving fatigue and driving performance among male taxi drivers in Hong Kong: Zavareh, M. F., Mamdoohi, A. R., Nordfj?Rn, & T. (2017). The effects of indicating
A driving simulator approach. Accident Analysis & Prevention, 125(APR.), 7–13. rear-end collision risk via variable message signs on traffic behaviour.
Mohebbi, R., Gray, R., Tan & Hong (2009). Driver reaction time to tactile and Transportation Research Part F Traffic Psychology & Behaviour,
auditory rear-end collision warnings while talking on a cell phone. Human S1369847816303291.
Factors. Zhang, Y., Yan, X., & Li, X. (2021). Effect of warning message on driver’s stop/go
Muttart, J., Fisher, D., Knodler, M., & Pollatsek, A. (2007). Driving without a clue: decision and red-light-running behaviors under fog condition. Accident Analysis
Evaluation of driver simulator performance during hands-free cell phone & Prevention, 150 105906.
operation in a work zone. Transportation Research Record Journal of the Zhang, Y., Yan, X., Li, X., & Wu, J. (2018). Changes of drivers’ visual performances
Transportation Research Board, 43, 9–14. when approaching a signalized intersection under different collision avoidance
National Highway Traffic Safety Administration (2016). Visual-manual NHTSA warning conditions. Transportation Research Part F Traffic Psychology and
driver distraction guidelines for portable and aftermarket devices. Washington, Behaviour, S1369847816303850.
DC: National Highway Traffic Safety Administration (NHTSA), Department of Zhang, Y., Yan, X., Li, X., & Xue, Q. (2016). Drivers’ eye movements as a function of
Transportation (DOT). collision avoidance warning conditions in red light running scenarios. Accident
Navarro, J., Deniel, J., Yousfi, E., Jallais, C., Bueno, M., & Fort, A. (2017). Influence of Analysis & Prevention, 96(nov.), 185–197.
lane departure warnings onset and reliability on car drivers’ behaviors. Applied Zhao, X., Xu, W., Ma, J., Li, H., Chen, Y., & Rong, J. (2019). Effects of connected vehicle-
Ergonomics, 59, 123–131. based variable speed limit under different foggy conditions based on simulated
Pitta, L., Quintas, J. L., Trindade, I., Belchior, P., & Camargos, E. F. (2021). Older drivers driving. Accident Analysis & Prevention, 128(JUL.), 206–216.
are at increased risk of fatal crash involvement: Results of a systematic review Zhu, M., Wang, X., & Hu, J. (2020). Impact on car following behavior of a forward
and meta-analysis. Archives of Gerontology and, Geriatrics(2) 104414. collision warning system with headway monitoring. Transportation Research
Qiao, F., Rahman, R., Li, Q., & Yu, L. (2017). Safe and environment-friendly forward Part C: Emerging Technologies, 111, 226–244.
collision warning messages in the advance warning area of a construction zone. Ziegler, J., Bender, P., Dang, T., & Stiller, C. (2014). Trajectory planning for Bertha — A
International Journal of Intelligent Transportation Systems Research, 15(3), 1–14. local, continuous method. IEEE: Intelligent Vehicles Symposium.
Rossi, R., Meneguzzer, C., & Orsini, F. (2020). Gap-acceptance behavior at
roundabouts: Validation of a driving simulator environment using field Junyu Hang received the B.E. degree from the School of Automobile, Chang’an
observations - sciencedirect. Transportation Research Procedia, 47, 27–34. University, in 2016, where she is currently pursuing the Ph.D. degree with the MOT
Scott, J. J., & Gray, R. (2008). A comparison of tactile, visual, and auditory warnings Key Laboratory of Transport Industry of Big Data Application Technologies for
for rear-end collision prevention in simulated driving. Human Factors: The Comprehensive Transport, Beijing Jiaotong University. Her research interests
Journal of the Human Factors and Ergonomics Society, 50(2), 264–275. include intelligent transportation systems, driving behavior and human factor, and
Seacrist, T., Sahani, R., Chingas, G., Douglas, E. C., & Loeb, H. (2020). Efficacy of traffic safety.
automatic emergency braking among risky drivers using counterfactual
simulations from the shrp 2 naturalistic driving study. Safety Science, 128 Xuedong Yan is a professor from the MOT Key Laboratory of Transport Industry of
104746. Big Data Application Technologies for Comprehensive Transport, Beijing Jiaotong
Shangguan, Q., Fu, T., & Liu, S. (2020). Investigating rear-end collision avoidance University. His research expertise includes advanced driving simulator research,
behavior under varied foggy weather conditions: A study using advanced driving behavior and human factor, GIS application on traffic engineering, traffic
driving simulator and survival analysis. Accident Analysis & Prevention, 139 safety and security, statistical analysis, traffic simulation, highway geometric
105499. design, highway-rail grade crossing, ITS, and urban design and planning.
Sheu, J. B., & Wu, H. J. (2015). Driver perception uncertainty in perceived relative
speed and reaction time in car following – a quantum optical flow perspective. Xiaomeng Li is a research fellow from the Centre for Accident Research and Road
Transportation Research Part B, 80(OCT.), 257–274. Safety – Queensland (CARRS-Q), Queensland University of Technology. She was
Siebert, F. W., Oehl, M., & Pfister, H. R. (2014). The influence of time headway on awarded a PhD degree in January, 2017 from Beijing Jiaotong University. Her
subjective driver states in adaptive cruise control. Transportation Research Part F research mainly focuses on road traffic safety, driving performance, ITS and traffic
Traffic Psychology & Behaviour, 25, 65–73. engineering.
Vahidi, A., & Eskandarian, A. (2004). Research advances in intelligent collision
avoidance and adaptive cruise control. IEEE Transactions on Intelligent Ke Duan graduated from Shijiazhuang Tiedao University, in 2016. He is currently
Transportation Systems, 4(3), 143–153. pursuing the Ph.D. degree with the MOT Key Laboratory of Transport Industry of Big
Vangi, D., & Virga, A. (2007). Evaluation of emergency braking deceleration for Data Application Technologies for Comprehensive Transport, Beijing Jiaotong
accident reconstruction. Vehicle System Dynamics, 45(10), 895–910. University. His research interests include traffic safety research, driving behavior
Wang, X., Zhu, M., Chen, M., & Tremont, P. (2016). Drivers’ rear end collision analysis and intelligent transportation system.
avoidance behaviors under different levels of situational urgency.
Transportation Research Part C: Emerging Technologies, 71, 419–433. Jingsi Yang is a postdoctor from the CRSC Communication & Information Group
Wenda, X. u., & Dolan, J. M. (2012). A Real-Time Motion Planner with Trajectory Company Ltd. She was awarded a PhD degree in October, 2020 from Beijing Jiaotong
Optimization for Autonomous Vehicles. Advances in Electronic Commerce. Berlin University. Her research mainly focuses on road traffic safety, driving performance,
Heidelberg: Web Application and Communication. Springer. ITS, and road traffic signs and markings.
Weng, J., Xue, S., Yang, Y., Yan, X., & Qu, X. (2015). In-depth analysis of drivers’
merging behavior and rear-end crash risks in work zone merging areas. Accident Qingwan Xue is a lecturer from Beijing Key Laboratory of Urban Intelligent Traffic
Analysis & Prevention, 77(apr.), 51–61. Control Technology, North China University of Technology. She was awarded a PhD
Winsum, W. V., & Heino, A. (1996). Choice of time-headway in car-following and degree in June, 2019 from Beijing Jiaotong University. Her research interests include
the role of time-to-collision information in braking. Ergonomics, 39(4), 579–592. the modelling of driving behavior, and the assessment of the impacts of intelligent
Wood, J., & Zhang, S. (2017). Evaluating Relationships Between Perception-Reaction technologies on drivers.
Times, Emergency Deceleration Rates, and Crash Outcomes Using Naturalistic
Driving Data (No. Fargo, ND: Mountain-Plains Consortium, North Dakota State
University.

427

You might also like