IET Intelligent Transport Systems

Research Article

Curve speed model for driver assistance ISSN 1751-956X

Received on 14th November 2016
Revised 11th May 2017
based on driving style classification Accepted on 12th July 2017
E-First on 10th August 2017
doi: 10.1049/iet-its.2016.0294
www.ietdl.org

Duanfeng Chu1, Zejian Deng1, Yi He1, Chaozhong Wu1 , Chuan Sun1, Zhenji Lu2
1Intelligent Transport Systems Research Center, Wuhan University of Technology, Engineering Research Center of Transportation Safety,

Ministry of Education, Heping Avenue #1040, Wuchang District, Wuhan 430063, People's Republic of China
2Faculty of Mechanical, Maritime and Material Engineering, Delft University of Technology, Mekelweg 2, 2628 CD Delft, The Netherlands

E-mail: [email protected]

Abstract: Inappropriate speed in negotiating curves is the primary cause of rollovers and sideslips. In this study, the authors
proposed an improved curve speed model considering driving styles, as well as vehicle and road factors. On the basis of a
vehicle–road interaction model, the driver behaviour factor was introduced to quantify driving styles of curve speed choices.
Firstly, the fuzzy synthetic evaluation method was utilised to classify the driving styles of 30 professional drivers into three
different types (i.e. cautious, moderate and aggressive). Secondly, the classification results using fuzzy synthetic evaluation
were compared to and verified with the K-means clustering method resulting over 60% the similarities. Finally, the proposed
curve speed model was built and compared with four existing models. The authors’ model has the following promising
advantages: (i) it reflects the speed preferences of three different types of drivers on the premise of driving safety on curves;
and (ii) it shows a stationary speed transition when the road adhesion coefficient exceeds 0.8, which indicates that rollover,
instead of sideslip, becomes the primary cause for lateral instability crashes on curves. Therefore, this proposed curve speed
model could be applied in a curve speed warning system to improve both driving safety and comfort.

1 Introduction reflecting driving style. The simulation studies show that the
vehicle would be out of sideslip and rollover risks if a driver drives
Over the past few years, both the amount and the severity of curve at a speed within the improved curve speed model proposed.
accidents in China have been maintaining a high rate. There were Furthermore, it is substantiated that the improved curve speed
35,000 curve crashes in 2014, accounting for 17.8% in all crashes. model not only meets the safety requirements, but also adapt to the
Moreover, the fatality rate in curve accidents reached up to 0.8, i.e. drivers’ expected speed on a curve.
0.8 people were killed in each curve incident, which was much
higher than 0.3 the average fatality rate among all crashes [1].
Curve crashes are caused by many factors, including road 2 Literature review
conditions, vehicle dynamics and driver behaviours. Inappropriate The vehicle lateral instability crashes, such as rollovers and
choice of driving speed is a major cause of curve accidents [2]. sideslips, are always related to road surface conditions, road
Vehicles in high speed are exposed to lateral instability hazards like geometry features, vehicle dynamics and so on. Some previous
rollovers and sideslips. Naturally, the adjustment of the vehicle studies have focused on road and environment impacts on the
speed during curve negotiation becomes one of the effective driving safety on curves [3–5].
solutions. A common method that has always been used for drivers As drivers characteristics vary, the curve negotiation speed
to adjust curve speed is to set roadside speed limit signs along preferences would be distributed around a range of values.
curves. The speed limits on roadside are mainly obtained using Therefore, the analysis of driver behaviours is indispensable to
standards and empirical data from the highway engineering. This avoid inaccurate predictions of curve speeds. MacAdam [6]
method, however, cannot rectify drivers’ inappropriate speed identified human driver as the primary control element within the
effectively. First, the roadside speed limits cannot fully take into long-established driver–vehicle system. Driver models including
account some dynamic and individual factors, such as vehicle human traits were proved to be useful in predicting the
structural parameters and driver behaviors. Moreover, drivers may performance of the combined driver–vehicle system. Salvucci [7]
ignore the speed limit signs intentionally because their values are developed a rigorous computational model of driver behaviour in a
derived from the conservative conditions which are probably below cognitive architecture that incorporate basic properties and
drivers’ expectations. limitations of human, which could be applied to predict and
Curve speed warning (CSW) is an on-board driver assistance recognise driver behaviour and distraction.
system that can calculate appropriate speed based on the driver– Most existing curve speed models were established based on the
vehicle–road interactions and send warning signals if the driver analysis of vehicle–road interactions. Glaser et al. [8, 9] proposed a
drives over the calculated speed limit. However, in a situation more comprehensive model than previous ones. The road geometry
where drivers lose the trust to the warning system, inappropriate was fully used. Chen et al. [10] proposed a BP neural network-
speed selection can occur and lead to more dangerous situation. based model to represent the motion states of vehicles on curves.
Therefore, it is important to improve the system to reduce wrong Unfortunately, these models neglected driver behaviours which
warnings. could largely influence the curve negotiating process.
This paper proposes an improved curve speed model for Some researchers intended to introduce human factors to
calculating the safe speed during curve negotiation by combining conventional curve speed models. Bosetti et al. [11] concluded that
multiple factors from drivers, vehicles and roads. Specifically, the one important aspect of the curve driving was the driver speed
influence of driving behaviour is described by driver behaviour choices. Zhang et al. [12] concluded that the curve speed was
factor which can vary as driving styles differ. Several indices are related to the drivers’ preferred velocity and the initial velocity on
selected to make the classification according to their properties of a curved road. Lee et al. [13, 14] added three different gain factors

IET Intell. Transp. Syst., 2017, Vol. 11 Iss. 8, pp. 501-510 501
© The Institution of Engineering and Technology 2017
Table 1 General information of 30 studied drivers
Driver number Gender Age Driving years Driver number Gender Age Driving years
D01 male 52 23 D16 male 42 16
D02 male 53 16 D17 male 48 22
D03 male 33 6 D18 male 48 18
D04 male 50 16 D19 male 57 21
D05 male 36 15 D20 male 34 3
D06 male 55 21 D21 male 35 8
D07 male 52 16 D22 male 50 12
D08 male 47 20 D23 male 53 9
D09 male 47 16 D24 female 51 17
D10 male 52 27 D25 male 37 6
D11 male 46 9 D26 male 49 17
D12 female 49 17 D27 male 49 20
D13 male 42 10 D28 male 43 17
D14 female 49 25 D29 male 48 26
D15 female 36 10 D30 male 47 14

associated with vehicles, roads and drivers to the curve speed In (3), kd describes drivers’ impact on speed choices
model to represent how these factors affect the curve speed. quantitatively and is related to driving skills, speed preferences and
However, the process of determining the appropriate gain values even driving mood. Though driver behaviours are difficult to
was carried out without any on-field experiment for further describe through a mathematical model, their driving styles
verification. normally stay steady for a long time, which provides a novel idea
Currently, curve speed models mainly consider the coupling by adding a constant coefficient to the theoretical speed. As
effects of vehicles and roads with little efforts on the quantitative changing the coefficient for each driver behaviour is difficult to
analysis of the driver behaviour characteristics. As drivers have implement, it is more practical to classify the drivers according to
different expectations and preferences of the curve speed, it is their driving styles. In this way, the appropriate values of kd to
necessary to obtain appropriate curve safety speed with the indicate the accurate curve speed that fits drivers themselves can be
consideration of individual differences, otherwise drivers would derived. Moreover, the results of driving style classifications
lose the trust to CSW system, which might increase the risks of directly influence the accuracy of curve safety speed.
traffic crashes.
3.2 Classifications of driving styles
3 Curve speed model based on driving style
The study divides the driving styles into three typical types:
clustering
cautious driving, moderate driving and aggressive driving. A
3.1 Improved curve speed model considering driver cautious driver would drive more carefully and avoid high speed
behaviours and hard acceleration. In contrast, aggressive drivers prefer
exciting driving experiences. Moderate drivers would drive
For light vehicles, the critical value of the curve safety speed
vehicles with relative steady motions that are neither too cautious
should focus on the prevention of sideslips because they are more
nor too aggressive.
likely to skid than overturn on the same curve conditions due to
their lower centre of gravity compared to heavy trucks. However,
rollover risks of heavy trucks should not be neglected in the 3.2.1 On-field test for driving style classifications: A driving
analysis of the curve safety speed. Therefore, the curve safety behaviour experiment was carried out. Before recruitment of
speed is defined as the minimum value vsr between the critical participants, a sample of taxi drivers were surveyed through
interview and questionnaires. It was found that more than 85% of
sideslip speed vs and the critical rollover speed vr, which is stated
them were male and their ages ranged from 30 to 60 years old.
in [15] Based on the survey, 30 professional drivers were recruited to
complete the experiment. Their driving experiences ranged from 3
vsr = min vs, vr (1) to 27 years (mean = 15.8, STD = 6.2) with ages ranging from 34 to
55 years old (mean = 46.3, STD = 6.6). Also, the gender ratio was
where controlled around 85% (see Table 1).
The experiment was conducted on the Hanshi Freeway from
μ + iy B + 2hiy Wuhan to Xiangyang and the total distance was about 600 km, as
vs = ⋅ gR, vr = ⋅ gR (2)
1 − μiy 2h − Biy shown in Fig. 1. It took each driver 5–7 h to finish the whole
experiment. However, there is a service area in Suizhou
where B and h are the vehicle track width and the height of centre (approximately in the half section of Hanshi Freeway), as shown in
of gravity, respectively. g denotes the gravity, R, μ and iy are the Fig. 1. Before the experiment, each participant took a trial drive
curve radius, the road adhesion coefficient and the superelevation, through the experimental route in order to get used to the road
respectively. conditions. Besides, all subjects were not allowed to start
The above model contains the factors of road conditions and experimenting until they were reported to get used to on-board
vehicle states. However, driver behaviours, which can significantly driving assistance devices so that they could drive in a naturalistic
influence the adaptation of the curve speed to different drivers, way. During the experiment, all participants were asked to drive
have not been included. Therefore, the curve speed model should with a safe and comfortable speed that reflected their real driving
be improved through the introduction of the influence of driver style.
behaviours which is represented by the following factor kd. Data acquisition equipment includes vehicle-mounted CAN
bus, Mobileye C2–270 system, smartphones, angle sensors and HD
vsafe = kd ⋅ vsr (3) video cameras. The information about the collected experimental
data is listed in Table 2. Steering wheel angle data was acquired
where kd represents the influence factor of driver behaviours.

502 IET Intell. Transp. Syst., 2017, Vol. 11 Iss. 8, pp. 501-510
© The Institution of Engineering and Technology 2017
Fig. 1 Driving route for the experiment

Table 2 Some data collected during the field experiment

Parameter types On-board equipment Sampling frequency, Hz
speed, km h−1 Smartphone 16
steering wheel angle, deg. Angular transducer 77
acceleration, ms−2 Smartphone 16
lane departure, m Mobileye C2–270 5–15

using angular transducer. Smartphones were used to collect Standard deviation of speed (vstd) represents the stability of
dynamic data such as vehicle speeds and accelerations. driving speed, which is able to reflect the fluctuation of driver's
Mobileye C2–270 is an image processing product provided by operations. Aggressive drivers generally keep their speed at a high
the company Mobileye. It could sense moving or fixed objects and level and reduce it under complex traffic environment, which,
determine positions between host vehicles and surrounding traffic inevitably, leads to a wide range of driving speed. Thus, standard
signs. A camera and a display screen are included in the system. deviation of speed differs between aggressive and cautious drivers.
The camera is used to collect the traffic information, like lane lines Driving style reflects comprehensive performance of
and relative time headway. In this experiment, Mobileye C2–270 longitudinal and lateral control to the vehicle. Steering angle and
system was used to collect time headways (within 2.5 s) and lane driving speed are important parameters in two directions.
departure displacements. This device has been used to collect Generally, drivers will slow down the vehicle when making a turn
vehicle motion states for the research of driving behaviour in many in case of unstable lateral accidents. However, aggressive drivers
researches [16, 17]. are likely to take risks pursuing high speed. Based on this
assumption, we define maximum product of the steering wheel
3.2.2 Driving style classification using fuzzy synthetic angle and vehicle speed (kmax) as a new parameter.
evaluation: Fuzzy-logic based mechanism has been utilised to Besides, many literatures are based on the parameters
identify driving styles for evaluating energy-saving performance mentioned above to study the modelling and classification of
[18] and driver profiling [19]. In this paper, fuzzy synthetic personal driving style [20–22].
evaluation is applied to classify the driving styles by analysing the The following four steps are used to determine a driver's driving
evaluation indexes reflecting each driver's characteristic. The style:
chosen indices include the following driving parameters. Step 1: Define appropriate membership functions for the five
Mean speed (v̄) differs when driving style changes. Aggressive chosen indices. In fuzzy mathematics, membership function of a
drivers prefer to drive faster to reach the destination while the research is first roughly obtained from subjective knowledge and
cautious ones would choose lower speed with more safety then determined according to distributions of the experiment data.
concerns. In this field experiment, participants drove their vehicles In this study, drivers’ temperament test which psychologically
in freeway, where driving style was hardly influenced by traffic reflects their driving styles is combined with on-field tests to
flow, therefore, driving speed could reflect their driving styles in a determine membership functions.
large extent. In [23], Chen proposed a ‘60 questions’ temperament test to
Speeding is a kind of common driving behaviour especially help people to understand their temperament types. The
when driving in wide road. Aggressive drivers are inclined to drive temperament test consisting of 60 questions is very popular, and it
over the speed limits to shorten travel time, leading to higher has been widely used in various fields in China, such as education,
speeding frequency than cautious drivers. The number of speeding medicine and sports. The research divided the personality into four
(ns) is selected as an indicator for classification of driving style. In types: choleric, sanguineous, phlegmatic and melancholic
this experiment, speeding is defined when driving speed is over temperaments.
120 km/h which is the speed limit in Chinese freeway. Human behaviours are directly influenced by their
Acceleration is an indirect measure of the acceleration pedal psychological characteristics. Drivers’ driving styles are
position and it can reflect the driving smoothness. Aggressive psychologically related to the intrinsic temperament types of their
drivers tend to depress acceleration pedal harder to get higher own. For example, as displayed in Fig. 2, the average speed
speed during a short period, resulting in higher acceleration. profiles show apparent differences as personal characteristics
Therefore, maximum positive acceleration (amax +
) is selected as a change of recruited drivers who did temperament test [23].
parameter for fuzzy evaluation.

IET Intell. Transp. Syst., 2017, Vol. 11 Iss. 8, pp. 501-510 503
© The Institution of Engineering and Technology 2017
Fig. 2 Probability and cumulative probability distributions of vehicle speeds for the divers with different temperaments

Fig. 3 Membership functions for five indices of three driving styles

The cumulative probability distributions of vehicle speeds were different personal characteristics and could help to determine the
presented to indicate the speed preferences of drivers with different interval points in the membership functions of the fuzzy synthetic
personal characteristics and then determine the interval points in evaluation. Interval points of other four indices have been
the membership functions of the fuzzy synthetic evaluation, i.e. ai, determined in the same way. Based on the analysis of the
bi and ci in Fig. 3. The conclusions could be drawn from the relationship between driving style and driver characteristics, we
changes of slopes. can get the membership functions for five selected driving
For example, as speed points of melancholic temperament parameters of three driving types as presented in Fig. 3.
drivers are mainly distributed in the low speed interval, the driver There is a reason for combining choleric and sanguineous
is more likely to be melancholic temperament if the driving speed temperaments as one group. In [23], Chen made a correlation
is <83 km/h. For phlegmatic drivers, their speed points are almost analysis of the number of people with four different temperaments.
linearly distributed from 80 to 95 km/h, so it is reasonable to set The result showed that there was no significant variation ONLY
the starting and terminal points as 80 and 95, respectively, in the between people of choleric and sanguineous temperaments, which
speed membership function. Those driving at higher speed (i.e. indicates that these two kinds of people are very similar in
>89 km/h) are more likely to be choleric or sanguineous, so the personality. Besides, driving styles are usually divided into three
membership function of aggressive drivers could be determined. categories in many previous literatures [24, 25], which is also an
These turning points indicate the speed preferences of drivers with important reason for our decisions.

504 IET Intell. Transp. Syst., 2017, Vol. 11 Iss. 8, pp. 501-510
© The Institution of Engineering and Technology 2017
Table 3 Statistics of driving styles indices
Driver no. Indices
v̄, km·h−1 vstd, km·h−1 ns +
amax, m·s−2 kmax, °·km·h−1
D01 93.6 25.64 80 1.58 1050.4
D02 82.74 37.28 33 2.09 3274.8
… … … … … …
D30 91.49 20.34 1 1.78 2017.2

Table 4 Membership degree of 30 studied drivers

Membership degree Cautious Moderate Aggressive Membership degree Cautious Moderate Aggressive
B*D01 0.147 0.463 0.39 B*D16 0.37 0.329 0.301
B*D02 0.098 0.35 0.552 B*D17 0.375 0.522 0.103
B*D03 0.314 0.344 0.342 B*D18 0.116 0.382 0.502
B*D04 0.15 0 0.85 B*D19 0.345 0.504 0.151
B*D05 0.238 0.452 0.31 B*D20 0.508 0.283 0.209
B*D06 0.098 0.506 0.396 B*D21 0.174 0.598 0.228
B*D07 0.036 0.52 0.444 B*D22 0.471 0.412 0.117
B*D08 0.07 0.477 0.453 B*D23 0.254 0.447 0.299
B*D09 0.053 0.497 0.45 B*D24 0.313 0.456 0.232
B*D10 0.253 0.395 0.352 B*D25 0.261 0.516 0.223
B*D11 0.292 0.409 0.299 B*D26 0.584 0.324 0.092
B*D12 0.305 0.249 0.299 B*D27 0.54 0.404 0.056
B*D13 0.544 0.222 0.525 B*D28 0.641 0.31 0.049
B*D14 0.243 0.232 0.525 B*D29 0.359 0.567 0.074
B*D15 0.367 0.451 0.182 B*D30 0.374 0.213 0.413

+
Step 2: Determine weighing values of assessment indices. If the indices v̄, vstd, ns, amax and kmax in turn could also be determined,
five indices equally affected by driving behaviours, we can set the where j = 1, 2, …, 5 and k = 1, 2, 3. The evaluation matrix Q is then
weight values for all of them as 0.2 averagely. However, the impact defined to describe the fuzzy relationship consisting of qjk, where
of driving behaviour on these indicators is comparatively different.
The reasons for the weighting values are based on their relations q11 ⋯ q13
with driving style. If the index is affected more by driving style
than traffic environment, the weighting will be higher than the Q = ⋮ ⋱ ⋮ (4)
average one. For example, average speed v̄ and standard deviation q51 ⋯ q53
of speed vstd are more vulnerable to the traffic condition because
drivers have to keep speed in a reasonable range considering Then the relation matrix can be calculated in terms of Q. Finally,
driving safety. Other indicators such as maximum positive each driver's membership degree BDi is calculated through the
+
acceleration amax , as well as the maximum product of the steering following equation:
wheel angle and vehicle speed kmax are easier to be affected by
driver behaviours because drivers are free to depress the BDi = w ⋅ Q, i = 01, 02, … , 30 (5)
accelerator pedal intensely or gently. Therefore, we decrease the
5
weight values of v̄ and vstd by 25%, then conversely increase those where ∑ j = 1 w j = 1, and w = [0.15, 0.15, 0.2, 0.25, 0.25], which is
+
of amax and kmax by 25% from the average. Finally, based on described in Step 2.
comprehensive comparisons of the five indices, we determine the After the normalised processing of membership values, the
+
weighting value of v̄, vstd, ns, amax and kmax as 0.15, 0.15, 0.2, 0.25 result of membership degree of 30 drivers described by BDi* is
and 0.25, respectively. shown in Table 4.
If the weighting is made differently, the membership degree BDi Step 4: Determine driving styles according to the principle of
the maximum membership degree. As the principle describes,
in Step 3 will change and the classification results will also vary.
though each object gets three membership degrees, the maximum
However, it is impossible to determine the exact weighting value of
one determines which group the object belongs to. Therefore, the
each index because fuzzy evaluation method is a kind of subjective
final classification results are shown in Table 5.
system based on human experience in most cases. All we can do is
setting appropriate range of weighting values through reasonable
demonstrations. 3.2.3 Driving style verification using K-means
Step 3: Calculate membership degree values of each driver. clustering: Fuzzy synthetic evaluation is often used to describe the
First, we calculate the values of five indices for each driver during ambiguous things in nature, like driving styles in this case. In fact,
the whole drive cycle on the freeway, as presented in Table 3. it is hard to accurately distinguish the boundaries of different
As the membership function is described in the form of graphs, driving styles. Therefore, we use K-means clustering with different
the membership degree could be determined by substituting the evaluation indexes to classify drivers’ driving styles. This method
average values of evaluation indices into the abscissa. For example, is further used to verify the above classification results of fuzzy
if one index value of a driver is x, the corresponding membership synthetic evaluation.
degree will be qj1, qj2 and qj3 which reflect the driver's tendencies K-means clustering is often used in clustering analysis of large
of cautious, moderate and aggressive styles, respectively, as shown amounts of data. In this study, we choose three driving parameters
in Fig. 3. Therefore, the membership degree qjk of the above five for clustering analysis which reflect the differences of driving
styles.

IET Intell. Transp. Syst., 2017, Vol. 11 Iss. 8, pp. 501-510 505
© The Institution of Engineering and Technology 2017
Table 5 Comparisons of fuzzy synthetic evaluation and K-means clustering
Fuzzy synthetic evaluation K-means clustering analysis Similarity, %
Driving style Driver no. Cluster Driver no.
cautious D13, D16, D17, D20, D22, D26, D27, D28, D30 I D06, D13, D14, D16, D20, D22, D26, D30 66.7
moderate D01, D03, D05, D06, D07, D08, D09, D10, D11, II D01, D03, D05, D07, D08, D10, D11, D12, D15, D17, 87.5
D15, D19, D21, D23, D24, D25, D29 D19, D21, D23, D24, D25, D27,D28, D29
aggressive D02, D04, D12, D14, D18 III D02, D04, D9, D18 60

Fig. 4 Driving styles classification based on K-means clustering

The first parameter, represented by η, is the proportion of time two methods. Hence, the similarity is calculated on 66.7% (i.e. six
when the driving speed is over 80% of the limited speed. drivers in common divided by nine drivers in total). Similarly, the
Aggressive drivers usually tend to drive faster. In contrast, the ratio similarities are 87.5 and 60% for the other two types. The
when the speed is at high level is lower for cautious drivers. Hence, resemblances are all over 60% through the verification of K-means
the proportion η reflects the speeding tendency of drivers with clustering analysis indicating that the fuzzy synthetic evaluation
different driving style. Due to the choice of highway traffic can be used for the driving styles classification.
situation, 80% of the limited speed is 96 km/h.
Standard deviation of the positive acceleration as+ is the second 3.2.4 Driving style impact on curve speed model: To calibrate
chosen parameter. Acceleration reflects the control of the the influence factor of driver behaviors kd for different drivers in
accelerator pedal and the brake pedal, while standard deviation of the curve speed model, vehicle speeds of entering the exit ramps on
the acceleration reflects the discrete degree of acceleration. the freeway are sampled because curve radius in those conditions
Positive acceleration results from the control of the accelerator are small enough to reveal drivers’ styles classified by the above
pedal and is closely related to the traffic and driving behaviour fuzzy synthetic evaluation. The statistical results are listed in
characteristics. For aggressive drivers, positive acceleration often Table 6.
shows the driving characteristics that include opening or closing The scatter diagram in Fig. 5 shows the actual speed vsafe
level of the accelerator and sudden changes of acceleration and chosen by drivers and the theoretical velocity vsr calculated by (1).
deceleration. The driver behaviours influence factor can be obtained by
The third parameter is minimum time headway thead. Time calculating the slopes of each driver presented in (3). In this way,
headway refers to the time interval that two consecutive vehicles in we obtain the mean value of kd in three driving styles, where
the same lane need to go through a section. It reflects the risk of kd_cautious = 0.475 (STD = 0.065), kd_moderate = 0.554 (STD = 0.123)
collision between two vehicles. Under car-following scenarios,
and kd_aggressive = 0.636 (STD = 0.152).
aggressive drivers tend to maintain high-speed state, and look for
the chances for overtaking which usually result in small values of Obviously, the more aggressive a driver is, the greater the value
the time headway. Therefore, time headway could be the index of of kd would be. This kind of tendency suggests that aggressive
driving style characterisation. drivers would drive faster than those who drive cautiously when
As shown in Fig. 4, 30 participants are divided into three types. entering segment of a curve.
For each type, there is a centroid which represents the average
characteristics of the corresponding cluster. Compared with Cluster 3.3 Numerical study of the improved curve speed model
II (56.9, 1.41, 0.77) and Cluster III (75.1, 2.02, 0.23), the centroid
To evaluate the proposed curve speed model, we designed a series
of Cluster I (42.5, 1.21, 1.6) has much lower η ( = 42.5) and as+ ( =
of simulation tests in TruckSim and MATLAB/Simulink, and we
1.21), as well as higher thead ( = 1.6), which conforms to the detected the lateral stability indices that could reflect the safety
characteristics of the group of cautious drivers. Similarly, it could performance of the vehicle. Lateral load transfer ratio (LTR) is an
be seen that Clusters II and III represent moderate and aggressive evaluating indicator that estimates the risk of rollover. Generally,
drivers separately according to the locations of their centroids. the vehicle could be in a safe state if its LTR is under 0.6 [26].
Comparison of driving style results classified by fuzzy synthetic Another indicator sideslip gradient (SSG) is used to evaluate the
evaluation and K-means clustering is shown in Table 5. When risk of sideslips. If SSG > 0, the vehicle is in the understeering
comparing the cautious type with Cluster I, six drivers (i.e. D13, state; and if SSG < 0, the vehicle is in the oversteering state [8].
D16, D20, D22, D26, D30 which are the underlined bold texts in SSGs of the left and right side of the vehicle are represented as
Table 5) always belong to the same classification analysed by both SSGL and SSGR, separately.

506 IET Intell. Transp. Syst., 2017, Vol. 11 Iss. 8, pp. 501-510
© The Institution of Engineering and Technology 2017
Table 6 Driver behaviours influence factors of 30 studied drivers
Driver no. Style R, m vsafe, km/h vsr, km/h kd
D01 moderate 172 78 106.7 0.731
D02 aggressive 172 89.6 106.7 0.84
D03 moderate 196 75 113.9 0.67
D04 aggressive 662 90 209.3 0.43
D05 moderate 196 77.5 113.9 0.681
D06 moderate 172 82.8 106.7 0.776
D07 moderate 196 80.7 113.9 0.709
D08 moderate 172 62 106.7 0.581
D09 moderate 662 85.4 209.3 0.408
D10 moderate 303 62.2 141.6 0.439
D11 moderate 303 61.8 141.6 0.437
D12 aggressive 196 78.8 113.9 0.692
D13 cautious 196 55.1 113.9 0.484
D14 aggressive 172 60 106.7 0.566
D15 moderate 303 62.2 141.6 0.439
D16 cautious 196 61.7 113.9 0.542
D17 cautious 662 90.3 209.3 0.432
D18 aggressive 303 92 141.6 0.65
D19 moderate 303 64.9 141.6 0.458
D20 cautious 303 47.1 141.6 0.333
D21 moderate 303 81.1 141.6 0.573
D22 cautious 303 76.3 141.6 0.539
D23 moderate 303 73.6 141.6 0.52
D24 moderate 303 72.1 141.6 0.509
D25 moderate 303 74 141.6 0.523
D26 cautious 303 66.3 141.6 0.468
D27 cautious 303 72.4 141.6 0.511
D28 cautious 303 72.2 141.6 0.51
D29 moderate 172 43.9 106.7 0.412
D30 cautious 303 65 141.6 0.459

Fig. 5 Calibration result of kd for drivers with different driving styles

Four different scenarios were considered with the variations of rises from 0 to the peak value; then it is steady during most of the
the road adhesion coefficient and the curve radius. The route was following time; it come down in the end. All peak values are found
set in the order of straight road, curve road and straight road. All to be lower than 0.6, indicating that the vehicle is always out of the
variables were kept the same except the road adhesion coefficient risk of rollovers on the curve. Similarly, SSGL and SSGR are
and the curve radius in each simulation scenario. The data was between −0.05 to 0.05 rad most of the time. In accordance with
extracted during the driving process on curves and they are plotted previous finding, this implies that the yaw motion of the vehicle is
(see Fig. 6). under control and out of the danger of sideslips [8].
As shown in Fig. 6, LTR varies when the road condition
changes due to the influence of the vehicle speed to the lateral
stability. The curve shapes are similar. In the early period, LTR

IET Intell. Transp. Syst., 2017, Vol. 11 Iss. 8, pp. 501-510 507
© The Institution of Engineering and Technology 2017
Fig. 6 Data was extracted during the driving process on curves
(a) LTR, (b) SSGL and, (c) SSGR analysis in four scenarios with the corresponding curve speeds

4 Comparisons with existing curve speed models 4.2 Impact analysis of coefficient of road adhesion on curve
speed threshold
In this part, simulation experiments were carried out under
different road conditions to analyze the influences of the road Fig. 8 shows the variation tendency of safety speed with respect to
adhesion coefficient and the curve radius on the curve speed the road friction coefficient. The value of speed is positively
calculated by the proposed model. Thereafter, the results were correlated to the friction coefficient under the fixed radius in most
compared between different speed models. To make the simulation models, except Glaser's model whose variation is not completely
environment similar to the field test, we chose the road parameters monotonically increasing. When the friction coefficient is small,
that were common in highways. The superelevation iy = 0.04 rad the curve safety speed would rise quickly with the increasing
and four kinds of common road adhesion coefficients are 0.4 friction coefficient. But when the friction coefficient increases, the
(gravel), 0.6(wet asphalt), 0.8(dry asphalt) and 0.85(dry concrete) curve safety remains steady.
[27]. The curve radiuses are 100 200, 400 and 1000 m. Based on When the friction coefficient comes to 0.8, the curve speed
these conditions, four existing curve speed models including the calculated by the improved model reaches its saturation point due
simplified model [27], Glaser's model [8, 9], Chen's model [10] to rollover's role in our model. When the friction coefficient is over
and Lee's model [13, 14] were compared with the proposed model. 0.8, rollover becomes the leading factor causing speed growth
stagnation that could affect the traffic safety. On the contrary, other
4.1 Impact analysis of road curvature on curve speed speed models have not shown the similar property, which could be
threshold the limitations in practical application.

Fig. 7 shows that the value of speed positively correlates to the 5 Conclusion and recommendations
radius under the fixed friction coefficient. When the radius is small,
the curve safety speed would rise quickly with the increasing A curve speed model is proposed with the consideration of vehicle
radius. When the radius becomes bigger, the curve safety speed parameters, road conditions and driving styles. The driver behavior
becomes steady. factor kd is introduced to reflect the driving styles. The
The simplified model does not take the driver behaviors and improvements of driving safety and comfort are the objectives of
vehicle parameters into consideration, resulting in the apparent the curve speed modeling. Driving comfort can be guaranteed
larger value of curve safety speed. For example, when the vehicle when calculated curve speeds are compatible with drivers’
drives on the gravel road (μ = 0.4) entering a curve of 400 m, the psychological expectations of driving speeds when negotiating
curve safety speed calculated by simplified model is close to 150 road curves ahead. On the premise of driving safety, the proposed
km/h, which is too high. Instead, the results calculated by the curve speed model could meet the psychological expectations of
proposed curve speed model mostly fall between the result of Lee's drivers in different (cautious, moderate and aggressive) driving
model and Glaser's model. The reason could be that Lee's model styles by quantitatively evaluating the driving features. In this
adds three different influence factors to the vehicle-road based regard, the proposed model could improve both driving safety and
calculation. comfort on curve negotiations.
Specifically, through our analysis and the results shown in A calibration test with 30 participants is conducted to determine
Fig. 7, our curve speed model stays the same when the friction the values of kd. First, the driving style is defined by analyzing five
coefficient are 0.8 and 0.85 as the curve radius changes. In the evaluation indexes with fuzzy synthetic evaluation model. The
meantime, the curve safety speed calculated by other models still classification is verified by comparing with K-means clustering
changes, although the friction coefficient is big enough. Thus, it model. Thereafter, the entry speeds chosen by drivers on the curve
can be seen that the improved curve speed model takes the risk of of the exit ramp are collected to linear fit the curve speed
rollover into consideration when the friction coefficient of road coefficient of drivers in different driving styles. Through the joint
surface is high. From this perspective, our improved model simulation of TruckSim and MATLAB/Simulink, LTR and SSG are
performs better. chosen to evaluate risks of vehicle lateral instabilities including
rollovers and sideslips. The results show that the vehicles are out of

508 IET Intell. Transp. Syst., 2017, Vol. 11 Iss. 8, pp. 501-510
© The Institution of Engineering and Technology 2017
Fig. 7 Variation of curve speed with respect to the curve radius

Fig. 8 Variation of curve speed with respect to the road friction coefficient

risks of rollover and sideslip after entering the curve roads with the tasks under all traffic scenarios in the foreseeable future [28, 29].
calculated speeds based on the proposed curve speed model. One important aspect in the human–vehicle cooperation is the
Although the number of the experiment drivers is 30, which is accurate evaluation of driver behaviour made by driver assistance
not enough to draw generalisable conclusions, this study lays the system when responding to driving tasks because of the variety of
foundation for the speed profiling of the automated driving on drivers’ adaptability to it. Therefore, personalised automated
curve roads. Nowadays, there is a common view that the human– driving systems considering drivers’ driving styles are more likely
vehicle cooperation will still be the dominating form in automated to be accepted by different types of drivers or passengers, which
driving before autopilot system takes over all aspects of the driving

IET Intell. Transp. Syst., 2017, Vol. 11 Iss. 8, pp. 501-510 509
© The Institution of Engineering and Technology 2017
will lead to lower driving risks and higher efficiency even in [12] Zhang, D., Xiao, Q., Wang, J., et al.: ‘Driver curve speed model and its
application to ACC speed control in curved roads’, Int. J. Autom. Technol.,
complicated traffic scenarios. 2013, 14, (2), pp. 241–247
Future studies can be conducted on optimising the driving style [13] Lee, Y.H., Deng, W.: ‘Speed control method for vehicle approaching and
classification and its coefficient calibration under more different traveling on a curve’. U.S. Patent 7,400,963, July 2008
driving scenarios. More importantly, real-world experiments with [14] Lee, Y.H.: ‘Automatic speed control system for vehicle approaching and
driving on a curve’. ASME 2008 Int. Mechanical Engineering Congress and
CSW systems based on the proposed model could be carried out Exposition, January 2008, pp. 345–353
under real-world traffic scenarios. [15] Sun, C., Wu, C., Chu, D., et al.: ‘Risk prediction for curve speed warning by
considering human, vehicle, and road factors’, Transp. Res. Rec., J. Transp.
Res. Board, 2016, 2581, pp. 18–26
6 Acknowledgments [16] Dukic, T., Ahlstrom, C., Patten, C., et al.: ‘Effects of electronic billboards on
driver distraction’, Traffic Inj. Prev., 2013, 14, (5), pp. 469–476
This study was sponsored by the National Natural Science [17] Son, J., Park, M., Park, B.B.: ‘The effect of age, gender and roadway
Foundation of China (51675390, 51605350), Hubei Provincial environment on the acceptance and effectiveness of advanced driver
Natural Science Foundation (2016CFB468), and Fundamental assistance systems’, Transp. Res. F, Traffic Psychol. Behav., 2015, 31, pp. 12–
Research Funds for the Central Universities (2017II50GX). 24
[18] Vangi, D., Virga, A.: ‘Evaluation of energy-saving driving styles for bus
drivers’, Proc. Inst. Mech. Eng. D, J. Autom. Eng., 2003, 217, (4), pp. 299–
7 References 305
[19] Castignani, G., Frank, R., Engel, T.: ‘An evaluation study of driver profiling
[1] Statistics Annals of Road Traffic Accident of People's Republic of China fuzzy algorithms using smartphones’. 21st IEEE Int. Conf. Network Protocols
(2014) (Traffic Administration Bureau, Chinese Ministry of Public Security, (ICNP), October 2013, pp. 1–6
2015) [20] Constantinescu, Z., Marinoiu, C., Vladoiu, M.: ‘Driving style analysis using
[2] Jiménez, F., Liang, Y., Aparicio, F.: ‘Adapting ISA system warnings to data mining techniques’, Int. J. Comput. Commun. Control, 2010, 5, (5), pp.
enhance user acceptance’, Accid. Anal. Prev., 2012, 48, pp. 37–48 654–663
[3] Pratt, M.P., Geedipally, S.R.: ‘Developing a framework for evaluating and [21] Enev, M., Takakuwa, A., Koscher, K., et al.: ‘Automobile driver
selecting curve safety treatments’. Transportation Research Board 95th fingerprinting’, Proc. Privacy Enhancing Technol., 2016, 2016, (1), pp. 34–50
Annual Meeting, Washington, the USA, 2016 [22] Miyajima, C., Nishiwaki, Y., Ozawa, K., et al.: ‘Driver modeling based on
[4] Funk, J., Wirth, J., Bonugli, E.,, et al.: ‘An integrated model of rolling and driving behavior and its evaluation in driver identification’, Proc. IEEE, 2007,
sliding in rollover crashes’ (SAE, 2012) 95, (2), pp. 427–437
[5] Yang, X.: ‘Balance of static and dynamic rollover thresholds for a three-axle [23] Zhang, T., Chen, H.: ‘Report about the development and preliminary trial of
vehicle’, SAE Int. J. Commer. Veh., 2011, 4, (1), pp. 22–30 the temperament test scale’, J. Shanxi Univ. (Philos. Soc. Sci. Ed.) (in
[6] MacAdam, C.C.: ‘Understanding and modeling the human driver’, Veh. Syst. Chinese), 1985, 1985, (4), pp. 73–77
Dyn., 2003, 40, (1-3), pp. 101–134 [24] Murphey, Y.L., Milton, R., Kiliaris, L.: ‘Driver's style classification using jerk
[7] Salvucci, D.D.: ‘Modeling driver behavior in a cognitive architecture’, Hum. analysis’. IEEE Workshop on Computational Intelligence in Vehicles and
Factors, J. Hum. Factors Ergon. Soc., 2006, 48, (2), pp. 362–380 Vehicular Systems, March 2009, pp. 23–28
[8] Glaser, S., Mammar, S., Sentouh, C.: ‘Integrated driver–vehicle–infrastructure [25] Li, G., Li, S.E., Cheng, B., et al.: ‘Estimation of driving style in naturalistic
road departure warning unit’, IEEE Trans. Veh. Technol., 2010, 59, (6), pp. highway traffic using maneuver transition probabilities’, Transp. Res. C,
2757–2771 Emerg. Technol., 2017, 74, pp. 113–125
[9] Sentouh, C., Glaser, S., Mammar, S.: ‘Advanced vehicle–infrastructure–driver [26] Larish, C., Piyabongkarn, D., Tsourapas, V., et al.: ‘A new predictive lateral
speed profile for road departure accident prevention’, Veh. Syst. Dyn., 2006, load transfer ratio for rollover prevention systems’, IEEE Trans. Veh.
44, (Sup1), pp. 612–623 Technol., 2013, 62, (7), pp. 2928–2936
[10] Chen, X.L., Lu, M., Zhang, D.Z.,, et al.: ‘Integrated safety speed model for [27] Wong, J.Y.: ‘Theory of ground vehicles’ (John Wiley & Sons, 2001, 3rd edn.)
curved roads’. Proc. of the 2010 World Automotive Congress, London, the [28] SAE J3016: ‘Taxonomy and definitions for terms related to on-road motor
UK, 2010, pp. 1–7 vehicle automated driving systems’, 2014
[11] Bosetti, P., Da Lio, M., Saroldi, A.: ‘On curve negotiation: from driver [29] Russell, H.E., Harbott, L.K., Nisky, I., et al.: ‘Motor learning affects car-to-
support to automation’, IEEE Trans. Intell. Transp. Syst., 2015, 16, (4), pp. driver handover in automated vehicles’, Sci. Robot., 2016, 1, (1), p. eaah5682
2082–2093

510 IET Intell. Transp. Syst., 2017, Vol. 11 Iss. 8, pp. 501-510
© The Institution of Engineering and Technology 2017