Longitudinal driving behaviour on different roadway categories: an instrumented-vehicle experiment, data collection and case study in China

1

Longitudinal Driving Behaviour on Different Roadway Categories: An Instrumented-Vehicle Experiment, Data Collection, and Case Study in China

Jianqiang Wang1, Chenfeng Xiong2,*, Meng Lu3, Keqiang Li4

* Corresponding Author (Tel: +1(301)661-9635; Email: [email protected])

1. State Key Laboratory of Automotive Safety and Energy,

Tsinghua University, Beijing, 100084. P.R. China. Email: [email protected]

2. 1173 Glenn L. Martin Hall, Department of Civil Engineering, University of Maryland, College Park, 20742, MD, USA.

Email: [email protected] 3. Dutch Institute for Advanced Logistics,

Princenhagelaan 13, 4813 DA Breda, The Netherlands Email: [email protected]

4. State Key Laboratory of Automotive Safety and Energy, Tsinghua University, Beijing, 100084. P.R. China.

Email: [email protected] This paper is a post-print of a paper submitted to and accepted for publication in IET Intelligent Transport Systems and is subject to Institution of Engineering and

Technology Copyright. The copy of record is available at IET Digital Library Abstract A significant portion of the observed variability in roadway performance can be due to the difference and innate heterogeneity in drivers’ behaviour. Analytical models, stated preference data collection and studies, and laboratory-based simulator experiments are developed to understand the driver behaviour for years. However, little has been done to fill the important gap between the survey/laboratory observed behaviour and the field observed behaviour. This study investigates drivers’ actual behaviour by conducting real-world field experiments in Beijing’s roadway system. In the experiment platform developed, instrumented vehicles are employed for the advanced data collection and analysis in order to understand the impact of roadway category on drivers’ longitudinal behaviour, i.e. car-following and car-approaching. These behaviour dimensions are identified in this study and quantified by parameters including relative speed, leading vehicle speed, accelerator release, braking activation, distance headway, time headway, and time-to-collision. The analysis suggests that the drivers’ behaviour variation heavily depends on roadway characteristics, which supplements further theoretical and survey-based behavioural research. The research findings provide insight for theoretical advances, evaluating driving assistance systems (DAS), and roadway-specific incentive designs for traffic harmonization, speed reduction, collision warning/avoidance, safety enhancement, and energy consumption savings. Keywords: roadway category; driving behaviour; experiment; car following; car approaching; instrumented vehicle

2

1. Introduction Driver Assistance Systems (DAS) are extra electronic features in motor vehicles that support the drivers in certain driving situations. Primarily aimed at enhancing roadway safety, mitigating emissions and environmental impacts, and harmonizing traffic, DAS are often designed with the functionalities of dynamic speed assistance, horizontal flow controls (e.g. active gas pedal, collision warning/avoidance, distance control, adaptive cruise control, etc.), lateral controls (e.g. lane departure warning, lane keeping support), monitoring (drowsiness detection, alcohol interlocks), and cooperative systems. Examples can be drawn from a large number of studies [1-5]. It is noted that a major assistances that a typical DAS system offers is the monitoring and enhancement to drivers’ longitudinal performances (i.e. acceleration and deceleration) and lateral performances (i.e. driving behaviour that affects manoeuvring in a direction more or less perpendicular to the driving direction). The term “longitudinal and lateral” driving behaviour is thus broadly used for describing these important performances. DAS’ assistances are achieved by: (a) helping the driver in “recognition,” “judgment,” and “actions”; (b) displaying warning signs in case the driver’s incorrect actions are judged to be dangerous; and (c) taking over the control of the vehicle in case the driver is unable to avoid a collision [6, 7]. These advanced assistances can potentially be effective in accident prevention in the way that DAS help drivers maintain comfortable and safe vehicle following/approaching distances by assisting partially in speed and distance control. One prerequisite to that is the correct understanding of the “longitudinal driving behaviour”. A good behavioural model can bring together effective designs, evaluation, and implementation of DAS, integrated operations and demand management, and dynamic traffic control. But a key problem for the time being is the lack of understanding of longitudinal driving behaviour. Driving behavioural models often consider car-following and car-approaching as the two types of longitudinal behaviour. Over the past three decades, the behaviour has been extensively studied from the point of view of psychology (e.g. [8-10]), ergonomics (e.g. [11-13]), and traffic flow theory (e.g. [14-17]). Various measures have been employed to characterize the behaviour, including leading vehicle speed, headways [18], and time to collision (e.g. [19]). A number of variables from the perspective of automotive engineering could be added to the list, including elapsed time, steering angle, brake pedal deflection, etc. Nevertheless, a well-accepted specification of driving behavioural models is still lacking, especially concerning the heterogeneous behaviour and psychological factors [20, 21]. Limited data sources are available for a validity check against those models. Simulation and driving simulators have long been implemented to collect drivers’ behaviour data. For instance, earlier simulation-based models have employed vehicle velocity [14]; optimal velocity model [22, 23], and acceleration [24, 25] as the control variable. [26] has developed a driving simulation platform to collect behavioural data for the development of DAS wherein drivers’ psychological status (prudent v.s. aggressive) and driving habits (risk-seeking v.s. risk-averse; skilful v.s. non-skillful) were incorporated into the parameter set. However, these laboratory-based data sources are expensive and often criticized to be biased [27, 28]. Environment built in the simulation experiment may be perceived as artificial by the experiment subjects [29]. Driving simulators for driver behaviour analysis can be potentially more useful if field data collected from “actual behaviour” becomes available, since the “actual behaviour” can be used to calibrate

3

the analyses or models developed using simulator data. Real-world driving behaviour data is one type of “ground truth” which is not readily available yet in existing research. An increasingly popular research concern has been raised that megacities in developing countries such as Beijing (a city with over 20 million population, 5 million automobiles, and 6 beltways) suffer from much more severe traffic and transportation safety conditions when compared with cities of the similar scale in the western world. For instance, Beijing’s total number of annual fatal traffic incidents is 4.5 times larger than that of Tokyo, a city with 12 million population and over 4 million vehicles. While many factors contribute to this discrepancy, it is our primary goal and long-term objective to work on DAS systems with well-calibrated behavioural model parameters for Beijing. In order to achieve this goal, this study aims at designing a real-world experiment and employing an experimental platform [30] to observe and record the “actual behaviour” of drivers in Beijing. This research will further adapt the two main components of the platform in order for the longitudinal driving behavioural data collection and analysis: an instrumented vehicle test-bed wherein vehicle’s surroundings and interior are captured, and an advanced data collection and analysis tool that serves as the data warehouse. A limited number of other vehicle test-bed systems have been constructed for research in intelligent transport systems. Sato and Akamatsu [31] have employed instrumented-vehicles to analyse drivers’ preparatory behaviour before turning at intersections. Larue et al.’s most recent research has studied actual driving behaviour and fuel consumption/emissions on urban roads [32]. Among the existing studies, the contradictory findings regarding the difference of driving behaviour among different roadway categories draw our attention. Dijker et al. have found that car-following and approaching behavior varies with the traffic flow regimes using trajectory data [33]. Among several studies with opposite conclusions, Brackstone, et al. studied instrumented-vehicle data in congested traffic on two high-speed (speed limits: 80 km/h and 120 km/h) roadway categories [34]. The paper found no effect of roadway type in influencing the behaviour through a thorough data analysis, in which various combinations of explanatory variables are tested. Brackstone et al.’s seminal work has setup a standard for designing and conducting instrumented-vehicle research to understand driving behaviour. Based on their study, the authors design an experiment platform and aim at collecting data and conducting a case study in Beijing. There is still room to refine their experiment design in order to achieve this paper’s research objectives. Different roadway categories can be entirely different in Beijing’s roadway system with different layout, capacity, speed limits, traffic volume, and density. When designing experiment, it is important to make sure that all these types of roadways are covered by the study. Moreover, if possible, experiment should include multiple experimental segments for each roadway category so that small sample bias can be mitigated to some extent. Another possible enhancement to Brackstone et al.’s experiment design is to incorporate multiple scenarios when collecting the data. Drivers’ performance under various circumstances such as stable car-following, unstable car-following, and car-approaching can provide sufficient information for researchers and practitioners to understand heterogeneous driving behaviour. These aforementioned considerations will be elaborated in the Experiment Design section. Under an experimental framework, the following empirical questions still need to be carefully answered: how to empirically test whether drivers tend to behave differently on different

4

roadway categories? What are the data information pieces that need to be collected in an experiment? To answer those questions, existing studies are reviewed, especially for the variables considered and for the settings of the control objectives [35]. In order to obtain “actual behaviour” data, an instrumented-vehicle experiment is designed and conducted by using and adapting an experimental platform. Over thirty drivers have been recruited to participate in the experiment during which each subject was asked to drive extensively on six different roadway segments (three urban access roads, two distributor roads, and one freeway segment) in order to avoid potential selection bias. The study selected the aforementioned roadway types by following a typical roadway categorization: freeway, distributor (or arterial) road, and local access road (see e.g. [36]). The distinction is meaningful as different road categories involved different types and levels of traffic risk [37, 38]. The paper emphasizes the experimental design and behavioural analysis. It is worth reiterating the importance of including multiple roadway categories and multiple segments for each category to receive sufficient variation in the observations. In general, a specific countermeasure often relates to and is designed for one specific road category. Therefore, designing carefully the experiment would benefit the data collection of a more credible and behaviourally rich dataset, deepen our understanding about the actual behaviour, and potentially enhance the robustness of the estimated behaviour models. Longitudinal behaviours including car following and car-approaching are analysed in this research. Various results of the analysis are visualized and statistics suggest that drivers tend to behave differently on different roadway categories. Roadway category actually is an influencing factor to be considered in car-following and approaching models. If treating these behaviour dimensions differently for different roadway categories both in modelling and in implementations, more reasonable and effective outcomes would be expected. The remainder of the paper is organized as follows. Section 2 describes the experimental platform and details the design of the instrumented-vehicle experiment. Data collection and results of the comparative analysis are presented in Section 3. Conclusions and future research scope are offered at the end of the paper. 2. The Experiment Design 2.1. Experiment design and data collection While a large amount of data shall be collected, the major control variables and the longitudinal driving behaviour terminologies are summarized and listed as follows:

• Car following/approaching behaviour • Accelerator release behaviour: When approaching the leading vehicle, the driver releases

the gas pedal and the vehicle reaches a fully closed throttle state. This behaviour is recorded once the accelerator position data is lower than a predefined lower bound threshold.

• Braking behaviour: When approaching the leading vehicle, the driver activates the brake. This is identified by the fact that the brake lamp signal becomes on.

• Gap closing: car-following with positive relative speed.

5

• Relative speed. • Leading vehicle speed. • Distance headway (DHW): the distance between the front bumpers of subject vehicle and

its leading vehicle. • Time headway (THW): distance headway divided by leading vehicle speed. • Time-to-collision (TTC): distance headway divided by relative speed. • Time-to-collision inverse (TTCi): the inverse of TTC. This statistic is meaningful in

analysis because under certain circumstances the relative speed of the two vehicles is zero and the corresponding TTC approaches infinity.

Three scenarios are designed for studying the plausible influence of roadway category on longitudinal driving behaviour. These definitions are defined based on DAS’s predefined criteria (e.g. speed range is based on the specifications of DAS’s Adaptive Cruise Control system).

• Scenario A: stable car-following. This stable car-following scenario represents the case when the relative speed of the two vehicles is low. The absolute TTC (time-to-collision) is longer than 20 seconds. And the speed is within a predefined range for each roadway category.

• Scenario B: unstable car-following. As a comparison, Scenario B is designated as a scenario where TTC is less than or equal to 20 seconds or the speed exceeds or is below the predefined range for each roadway category.

• Scenario C: car-approaching. The car-approaching scenario represents the case when the following vehicle is approaching the leading vehicle (i.e. gap closing lasts longer than a certain period of time). Specifically, the following conditions are defined: (1) the duration of gap closing is longer than 5 s; and (2) TTC is smaller than 50 s.

Based on the observation and analysis to the common road systems in Beijing, 3 types of roads have been selected for this experiment: urban distributor roads (road section 1 and 2), urban access roads (road section 3, 4, and 5), and freeways (road section 6). The route map is illustrated in Figure 1. The southeast-bound (the red arrow) and the northwest-bound (the green arrow) of the 4th Ring Road (speed limit: 70 km/h) have been chosen as the urban distributor roads. Another three urban road segments (with two/three lanes, speed limit: 40 km/h) were included as the urban access roads. A part of the four-lane freeway between Beijing and Shijiazhuang (the capital of Hebei Province) has been chosen as the freeway section (speed limit: 90-120 km/h). In order to incorporate all the differences in the traffic density and the road characteristics and to avoid selection bias, 2 urban distributor road segments and 3 urban access road segments were included in the test sections. The data used for this study was collected using the instrumented vehicle experimental platform described in the following section. The research has recruited thirty-three drivers to drive on the six experimental routes. Subjects’ age ranges from 30 to 69 with mean equals 44.9 and standard deviation equals 10.8. Their driving experience is inferred from the number of years of driver’s license holding: from 1 year to 47 years (mean equals 12.0, standard deviation equals 11.3). Before the experiment, each driver was directed to take a 30-min trial to get used to the vehicle and the roadway segments used in the research in order to mitigate the learning effects to the minimum. Therefore, the data collected reflect the normal driving behaviour of the driver. Drivers strictly used all six experimental roadway sections sequentially. When in the experiment, the subjects were instructed to drive on each roadway section starting from the same locations

6

(identified in Figure 1). After completing the driving experiment on one roadway, drivers are instructed to head to the starting location of the next section. All the experiments were conducted during the same time of day with clear meteorology condition.

Figure 1. The route map of the instrumented-vehicle experiment

2.2. Experimental platform for comprehensive longitudinal driving behaviour study Driving simulators and stated preference surveys have long been implemented to collect driver behaviour data. These laboratory-based data sources are often criticized to be biased [27, 28]. Environment built in the simulation experiment may be perceived as artificial [29]. Driving simulators for driver behaviour analysis can be more useful if field data collected from “actual behaviour” becomes available. In order to observe and record this “actual behaviour” of drivers in Beijing and to calibrate model parameters for in-vehicle driving assistance systems (DAS) and other modelling/operations applications, the research team adopts an experimental platform [30] and designs the experiment. Two main components of the platform are further developed: the instrumented vehicle test-bed, and the data warehousing tool, as illustrated by Figure 2(a) and 2(b). The instrumented vehicle test-bed (Figure 2(a)) is built to obtain complete coverage of data on drivers’ behaviour, vehicle kinematics, and vehicle surroundings. The system incorporates a number of detectors, including laser radar (used to detect the distance headway and relative speed), global positioning system (GPS, employed to obtain the longitude and latitude), cameras (forward camera, drivers’ hand camera, and foot camera), and internal vehicle kinematics sensors (used to accurately measure wheel speeds, acceleration, etc.). Images of the frontal road environment are recorded by a charge-coupled device (CCD) camera installed behind the windshield. In order to observe the driver steering and speed control operation, another two CCD cameras have been installed in the vehicle to capture the images of the hands and the feet movements of the driver. The signals of pedal position and steering angle are also obtained. The

7

data warehouse collects, synchronizes, processes, and stores a large amount of information from the instrumented vehicle. Its interface is illustrated as Figure 2(b). The real-time information being collected include: vehicle run time, brake lamp signal, accelerator-pedal position, distance to destination, speed (vehicle speed, horizontal/vertical speed, relative speed), acceleration (lateral/longitudinal acceleration), steering angle, yaw rate, direct angle, latitude, longitude, and altitude.

(a) The architecture of the experimental platform for driver behaviour study

(b) the interface of the data warehouse Figure 2: The information capture system of the instrumented-vehicle experiment

Vehicle surround

Laser Radar

GPS

Forward camera

Driver behaviour

Driver hand camera

Foot camera

Steering angle

Pedal sensor

Vehicle kinematics

Wheel speeds

Acceleration

Yaw rate

In-car capture and synchronization

CAN/AD card

Serial port

Image card

Data processing

Image capture and compression

Data warehouse

Monitor hard drive

Data integration

Visualization/analysis

8

3. Data Analysis and Results The longitudinal driving scenarios have been analysed by using the developed experimental platform. The objects of the statistical analysis are the assembled data sample points from data sets collected on different roadway categories. Descriptive statistics concerning longitudinal driving behaviour for each scenario (i.e. A. stable car-following, B. unstable car-following and C. car-approaching) on different roadway categories for all drivers are presented in Table 1. Table 1. Descriptive statistics of the scenarios on different roadway categories Control parameters speed

(km/h) DHW (m)

DHW accelerator release (m)

DHW brake

activation (m)

THW (s)

TTCi (s-1)

TTC accelerator release (s)

TTC brake

activation (s) Urban Access Roads: Scenario A mean 47.14 24.08 - - 1.83 0.012 - - std. dev. 3.82 4.48 - - 0.32 0.030 - - Urban Access Roads: Scenario B mean 45.68 22.60 - - 1.78 0.039 - - std. dev. 3.95 3.92 - - 0.29 0.011 - - Urban Access Roads: Scenario C mean - - 25.29 22.24 - - 19.05 14.20 std. dev. - - 3.93 3.78 - - 3.15 3.80 Urban Distributor Roads: Scenario A mean 68.98 33.46 - - 1.72 0.024 - - std. dev. 4.30 6.50 - - 0.27 0.035 - - Urban Distributor Roads: Scenario B mean 68.18 31.67 - - 1.64 0.085 - - std. dev. 6.20 6.17 - - 0.24 0.012 - - Urban Distributor Roads: Scenario C mean - - 33.50 27.37 - - 21.96 14.68 std. dev. - - 6.80 8.02 - - 3.46 3.35 Freeway: Scenario A mean 94.65 47.86 - - 1.80 0.025 - - std. dev. 8.25 9.75 - - 0.36 0.033 - - Freeway: Scenario B mean 91.49 45.19 - - 1.77 0.093 - - std. dev. 13.16 9.41 - - 0.32 0.087 - - Freeway: Scenario C mean - - 44.60 34.16 - - 20.04 14.64 std. dev. - - 7.65 9.16 - - 2.59 5.71 Statistics presented in Table 1 are obtained using our entire database, wherein 38% data samples are in Scenario A, 34% in Scenario B, and 28% in Scenario C. In the following text, the correlations between leading vehicle speed (V) and other parameters (DHW, TTC and TTCi) on each road category for each scenario are analysed. The result shows distinct differences of the correlations between speed and other parameters on different roadway categories. Firstly, the results of Scenario A, B, and C reveal that the distributions of distance headway (DHW) differ with different vehicle speed on each road category. It is proved by a subsequent regression analysis. For urban access roads, a linear fit yields the relationship of DHW=0.56V-2.06 (with R2 equals 0.52). For urban distributor roads and freeways, the relationships are DHW=0.97V-33.68 (R2 = 0.46), and DHW=0.68V-17.52 (R2 = 0.40) respectively. A series of T-Tests suggests that

9

these estimates are significant at 99% confidence level and indicates that with the increase of the leading vehicle speed, the values of DHW increase accordingly on all three roadway categories. However, the minimum restricted vehicle speed on different roads is not the same. For example, the leading vehicle speed on urban access roads is always larger than 20 km/h while this number for freeways is 60 km/h. Therefore, the minimum value of DHW will be different for these two different roadway types based on the estimated equations. It can be seen from the regression results of Scenario B (unstable car-following) that the correlations between vehicle speed and DHW are as follows: DHW=1.53V-1.39 (urban access roads), DHW=0.90V-30.14 (urban distributor roads), and DHW=0.69V-19.02 (freeways). The results for Scenario B are in contrast with the results of Scenario A (stable car-following). The results of Scenario C (car-approaching) show the correlations between DHW and vehicle speed for (1) accelerator release: DHW=0.58V-2.03 (urban access roads), DHW=0.69V-15.02 (urban distributor roads), and DHW=0.56V-7.13 (freeways); and (2) brake activation: DHW=0.51V+0.25 (urban access roads), DHW=0.46V-2.41 (urban distributor roads), and DHW=0.43V+0.003 (freeways). Figure 3 compares various statistics on different roadway categories.

Urban Access Roads Urban Distributor Roads Freew ays

0

10

20

30

40

50

60

Roadway Categories

Dis

tanc

e H

eadw

ay (m

)

Scenario AScenario BAccelerator releaseBrake activation

Figure 3.Comparison of longitudinal driving behaviour on different roadway categories For different road categories, the mean values of DHW (distance headway) simultaneously increase with vehicle speed (speed limit increases from urban access roads to urban distributor roads and to freeways, see Figure 3. Meanwhile, the values of the standard deviations of these parameters also synchronously increase as shown in Figure 3, which suggests that drivers tend to choose more variable speed and distance headway when they are driving on roadways with higher speed limit. Other longitudinal driving behaviour (measured by THW, TTCi, and TTC) varies for different road categories except for TTCi in Scenario A and TTC in accelerator release (see Table 1 for detailed statistics).

10

0 2 4 6 8 10 120

20

40

60

80

100

120

140

160

Relative Speed [m/s]

Dis

tanc

e H

eadw

ay [m

]

0 20 40 60 80 100 120 140 1600

5

10

15

20

25

30

35

40

45

50

Host Vehicle Speed [km/h]

Tim

e to

Col

lisio

n [s

ec]

0 2 4 6 8 10 120

20

40

60

80

100

120

140

160

Relative Speed [m/s]

Dis

tanc

e H

eadw

ay [m

]

0 20 40 60 80 100 120 140 1600

5

10

15

20

25

30

35

40

45

50

Host Vehicle Speed [km/h]

Tim

e to

Col

lisio

n [s

ec]

(a) data at the first accelerator release

(Urban access roads) (b) data at the first braking behaviour

(Urban access roads)

0 2 4 6 8 10 120

20

40

60

80

100

120

140

160

Relative Speed [m/s]

Dis

tanc

e H

eadw

ay [m

]

0 20 40 60 80 100 120 140 1600

5

10

15

20

25

30

35

40

45

50

Host Vehicle Speed [km/h]

Tim

e to

Col

lisio

n [s

ec]

0 2 4 6 8 10 120

20

40

60

80

100

120

140

160

Relative Speed [m/s]

Dis

tanc

e H

eadw

ay [m

]

0 20 40 60 80 100 120 140 1600

5

10

15

20

25

30

35

40

45

50

Host Vehicle Speed [km/h]

Tim

e to

Col

lisio

n [s

ec]

(c) data at the first accelerator release

(Urban distributor roads) (d) data at the first braking behaviour

(Urban distributor roads)

Free

way

s

0 2 4 6 8 10 120

20

40

60

80

100

120

140

160

Relative Speed [m/s]

Dis

tanc

e H

eadw

ay [m

]

0 20 40 60 80 100 120 140 1600

5

10

15

20

25

30

35

40

45

50

Host Vehicle Speed [km/h]

Tim

e to

Col

lisio

n [s

ec]

0 2 4 6 8 10 120

20

40

60

80

100

120

140

160

Relative Speed [m/s]

Dis

tanc

e H

eadw

ay [m

]

0 20 40 60 80 100 120 140 1600

5

10

15

20

25

30

35

40

45

50

Host Vehicle Speed [km/h]

Tim

e to

Col

lisio

n [s

ec]

(e) data at first accelerator release

(Freeways) (f) data at the first braking behaviour (Freeways)

Figure 4. Results of the analysis of Scenario C on different road categories: absolute relative speed vs. DHW, and leading vehicle

speed vs. TTC

11

The relationships between leading vehicle speed and other factors such as THW and TTC are plotted in Figure 4. In the interest of paper length, Scenario C is plotted as an example. By taking a closer look at Figure 4, we can judge qualitatively that the relationships between the behaviours and the leading vehicle speed vary on different roadway categories. The paper further examines the frequency contour of time headway and time to collision inverse for all the data collected. This is typically used to indicate drivers’ car-following characteristics. Taking Scenario B as an example (performances on different roadways are presented in Figure 5a, b, and c), the percentile numbers on area borders (25, 50, 75, 95, and 99) indicate the percentages of the data samples, which fall inside the associated areas. In Figure 5(a), for instance, the “99” contour curve indicates that on urban access roads, 99% of the TTCi-THW samples fall in that contour area. Firstly, the triangle-shaped contour suggests that the variation of TTCi is much higher when the time headway is relatively smaller. It makes sense since when the time headways are high, times-to-collision data samples are more concentrated to higher values (i.e. TTCi concentrate to [-0.1, 0.1] when THW reaches 5 sec as shown in Figure 5a). This concentrated region is smaller on urban distributor roads and freeways. Another interesting finding is that the figure evidently shows that the data for urban distributor roads and freeways are relatively more concentrated in the contour, while the data for urban access roads are more dispersed. For instance, in urban access road cases, TTCi ranges from about -0.38 s-1 to about 0.35 s-1 (that is, vehicles’ time-to-collision is controlled to be higher than 2.86 sec), while TTCi data for urban distributor roads and freeways ranges from -0.25 s-1 to about 0.25 s-1 (time-to-collision is higher than 4 sec). This phenomenon suggests that drivers prefer to keep THW and TTCi more cautiously when driving on urban distributor roads and freeways (could be due to the higher speed limit). On the other hand, drivers tend to behave more heterogeneously when driving on urban access roads. These findings indicate that there is an effect of roadway category on longitudinal driving behaviour. The impact of specific characteristics of roadway categories needs to be measured through estimation and validation of statistical models, which is subject for immediate future analysis.

(a) Urban access roads

12

(b) Urban distributor roads

(c) Freeways

Figure 5. The frequency contour of THW vs. TTCi: results for unstable car-following The Mann-Whitney U-test method is used for statistical testing of the above results for validation. When the population distribution of the scenario occurrences (mean and variance) is unknown and the number of samples is relatively large (more than 30 samples in this study), the U-test is a ubiquitous test method [39-41]. A null hypothesis is made that the expectations of the two groups of samples (such as data collected on urban distributor roads v.s. data collected on freeways) are equal. 95% confidence level is specified, from which a rejection region for the null hypothesis is derived. The absolute value of the testing variable u is calculated with the formula of the U-test. Finally, it is determined whether the calculated value of u is located within the rejection region or not in order to determine whether to accept the null hypothesis or not (i.e. whether the two samples are identical or not). Scenario B is set as control group. Data points collected in Scenario A and C are included in the statistical test. Two important longitudinal behaviours, THW and TTC, are tested using this

13

Mann-Whitney U-test. After several trials, we have replaced TTC with TTCi because under certain circumstances, the relative speed of the two vehicles is zero and thus makes TTC close to infinity. Therefore, THW and TTCi are selected for the U-test. The testing variable u is defined as follows:

2 21 2

1 2

X YuS Sn n

−=

+

(1)

Where X and Y are the expectation values of THW or TTCi for the two sample groups being tested. Variables S1 and S2 are the standard deviations of THW or TTCi for the two sample groups. Variables n1 and n2 are the sample sizes of the two sample groups. We make a null hypothesis H0 and an alternative hypothesis H1: ( ) ( ) ( ) ( )0 1: 0, : 0H E X E Y H E X E Y− = − ≠ (2) Where E(X) and E(Y) are the expectation values of the two samples. Consider the result for α = 0.05 (90% confidence level) and 0.025 (95% confidence level), two typical levels of significance used in the U-test [42]. The rejection regions for the two confidence levels are |u|>1.65 and 1.96, respectively. We have conducted U-test for each subject and for each pair of roadway categories. The U-test statistics for THW and TTCi are shown in Table 2. For a majority of the subjects in the experiment, we can conclude, on a 95% confidence level, that H0 is rejected and most drivers’ choices of THW and TTCi differ substantially and are influenced by roadway categories. 4. Closing Remarks In this research, an instrumented-vehicle experiment has been reported to study the possible influence of roadway categories on longitudinal driving behaviour, which is measured by a number of variables, including speed (leading vehicle speed and relative speed), distance headway, time headway, and time-to-collision. Behavioural data has been collected from a carefully designed experiment wherein instrumented vehicles, in-car data collection/processing, and data warehousing have been employed and integrated. The results show that the longitudinal driving behaviours are influenced by roadway categories. The instrumented-vehicle data exhibits different correlation between the leading vehicle speed and DHW on different roadway sections, while the correlation between the relative speed and DHW largely remain unchanged. Another set of validation data has been collected from the same research and data procedure to verify the results. Only the data obtained during clear weather conditions and same time of day are chosen and analysed. The validation significantly matches the patterns observed in the instrumented-vehicle experiment. In our study, driving behaviour under three different driving behaviour scenarios, including stable following, disturbed following, and car-approaching, has been studied. The analysis uses information collected from a relatively larger sample (33 drivers). At the same time, 6 different road segments falling in three roadway categories (3 urban access road segments, 2 distributor segments, and 1 freeway segment) are chosen as the test field to collect the behaviourally rich dataset.

14

Table 2. Absolute values of U-test variables based on different road categories Driver number

U-test for THW U-test for TTCi

Access roads & distributor roads

Access roads & Freeways

Distributor roads & Freeways

Access roads & distributor roads

Access roads & Freeways

Distributor roads & Freeways

1 7.9507** 7.96** 0.3186 6.9561** 1.3958 13.4147** 2 9.7178** 4.8134** 11.6548** 36.4094** 41.3795** 15.9786** 3 14.0946** 9.1887** 13.3268** 44.4871** 1.7786* 99.2302** 4 11.3276** 19.4584** 15.747** 10.5652** 14.2633** 12.4475** 5 4.776** 9.4072** 14.5693** 11.294** 8.5239** 4.1502** 6 3.6912** 1.1318 6.6802** 44.7998** 67.3504** 50.1572** 7 18.3322** 10.1289** 21.5564** 32.2364** 8.4551** 103.6063** 8 5.1799** 0.3604 15.4514** 8.3393** 27.8315** 62.1012** 9 11.3713** 3.0731** 37.5491** 14.0531** 19.2531** 13.6999** 10 3.1255** 9.301** 23.1081** 10.6438** 14.4555** 10.1004** 11 7.211** 5.3371** 5.5613** 11.4161** 39.8083** 97.0732** 12 2.2928** 6.7915** 11.5126** 13.0054** 23.1922** 23.6187** 13 7.7454** 11.1861** 11.0772** 5.2334** 50.7283** 58.7001** 14 3.5249** 7.4738** 10.6219** 5.707** 14.4148** 34.1414** 15 34.544** 10.4058** 53.0379** 14.2043** 15.5245** 1.851* 16 5.5331** 0.8049 18.7247** 25.6831** 10.9557** 40.5881** 17 14.8603** 13.1136** 3.6537** 12.2499** 5.3268** 34.3462** 18 1.6867* 8.5749** 20.2412** 31.1119** 7.6987** 41.78** 19 16.8363** 15.5033** 0.4244 32.3644** 20.7317** 24.6619** 20 14.2633** 10.0043** 11.7663** 16.7769** 8.023** 13.7928** 21 5.8154** 9.3964** 38.9002** 27.9054** 10.3097** 50.9016** 22 5.3174** 11.2824** 44.772** 3.3306** 11.2747** 21.9681** 23 6.2247** 5.6265** 2.032** 14.5222** 14.7068** 0.9013 24 17.8596** 15.3124** 5.8275** 11.4765** 34.4025** 56.5802** 25 12.6612** 10.7626** 8.5319** 39.2769** 48.6192** 55.1755** 26 31.5498** 32.3261** 3.3505** 39.491** 16.9235** 32.1954** 27 11.3669** 33.4324** 39.7057** 16.7955** 1.5894 19.3263** 28 3.3338** 3.5822** 24.4785** 28.4096** 30.7226** 8.2991** 29 1.1243 1.5782 7.3341** 4.4246** 4.1819** 16.3476** 30 9.7426** 18.5558** 39.0048** 7.2276** 14.5314** 23.6253** 31 7.3964** 0.1126 21.7154** 10.2368** 5.4859** 8.7816** 32 5.5081** 0.1503 22.7224** 0.9347 9.3137** 26.3953** 33 1.9863** 4.7303** 15.5774** 25.5432** 19.3219** 11.8391** * - Significant at 90% confidence level; ** - significant at 95% confidence level.

15

The scale of the study, the experiment design, and real-world data collection efforts highlight the significance of this research. Also, the presented platform and case study in Beijing provides a unique application reference in a developing country for researchers and practitioners. In this paper, we have reached conclusion that driving behaviour tends to be different on different types of roadways. A Mann-Whitney U-Test has been conducted to statistically test the conclusion. In particular, it has been found that the lower the leading vehicle speed, the shorter the distance headway (DHW). One convincing explanation is that drivers operate vehicles with an increasing difference on roadways with lower speed limit, narrower width, and/or more locally complicated geometry. And thus they are likely to choose a substantially different car-following distance. However, we cannot conclude that roadway category is a significant determinant for longitudinal driving behaviour before we conduct rigorous modelling and test its statistical significance. Thus, this paper only answers the empirical questions raised in the Introduction partly. Modelling longitudinal driving behaviour and empirically test whether the roadway type is a significant factor for driving are subjects for immediate future research. The results of this research provide supports for algorithms development for driving assistance systems (DAS), especially for the dynamic speed assistance with different feedback models that can adapt to the variety of driving behaviour. Future research is necessary to figure out drivers’ behaviour under a number of “what-if” scenarios in order to completely understand the behaviour response (or adapts) to DAS. The dissimilarities of longitudinal driving behaviour will be further studied for a more accurate behavioural modelling, as the behavioural model is seen as one of the key components that needs further work to enhance the effective design, implementation, and evaluation of DAS in order to potentially achieve traffic safety, fuel consumption, and traffic flow harmonisation goals. The results also provide profound insight for a potential application in traffic simulation modelling. The heterogeneous car-following behaviour observed from this instrumented-vehicle experiment study suggests another round of calibration or re-estimation of those car-following model parameters embedded in microscopic traffic simulation models. Ideally, as an immediate next step, we will specify and estimate different driving behaviour models for different roadway categories and explore their applications with microscopic models. Our mature instrumented-vehicle experimental platform, the fine-tuned data collection and data warehousing tool, and the stimulating environment of a fast-growing vehicle ownership/usage in Beijing (already over 5 million private vehicles) ensure that this line of research is a promising future direction. Acknowledgments The research was supported by the National Natural Science Foundation of China, No.: 51175290. and the joint project of NISSAN Motor, Co., Ltd. and Tsinghua University. Special thanks go to Tomohiro Yamamura (Nissan Motor Co., Ltd.), Nobuyuki Kuge (Nissan Motor Co., Ltd.), Tsunehiko Nakagawa (Nissan (China) Investment Co., Ltd.), Lei Zhang (Tsinghua university), Xiaojia Lu (China Agricultural University), Qing Xiao (China Agricultural University), Xiaoyun Lu (University of California), and Kees Wevers (NAVTEQ). The conclusions do not necessarily reflect the opinions of the sponsors. The authors are solely responsible for all the statements presented in this paper.

16

References [1] Shladover, S.E. Potential contributions of Intelligent Vehicle/Highway Systems (IVHS) to reducing

transportation's greenhouse gas production. PATH, University of California, Berkeley. 1991. [2] Minderhoud, M.M. Supported driving: impacts on motorway traffic flow. Doctoral Thesis, Delft

University of Technology, Delft. 1999. [3] Broge, J.L. Safety and comfort on the road. Tech Briefs, AEI (Automotive Engineering International),

SAE International (SAE - Society of Automotive Engineers). 2000. [4] Carsten, O.M.J., & Tate, F.N. Intelligent speed adaptation: accident savings and cost-benefit analysis.

Accident Analysis and Prevention, 2005, 37(3), pp. 407-416. [5] Lu, M., Wevers, K., & Van der Heijden, R. Technical feasibility of advanced driver assistance

systems (ADAS) for road traffic safety. Transportation Planning and Technology, 2005, 28(3), 167-187.

[6] Aufrere, R., J. Gowdy, C. Mertz, C. Thorpe, C. Wang, and T. Yata. Perception for collision avoidance and autonomous driving. Mechatronics, 2003, 13(10), 149-1161.

[7] Yoshida, T., Kuroda, H, & Nishigaito, T. Adaptive Driver-assistance Systems. Hitachi Review, 2004, 53(4), 212-216.

[8] Boer, E.R. Car following from the driver's perspective. Transportation Research Part F: Traffic Psychology and Behaviour, 1999, 2(4), 201-206.

[9] Vogel, K. A comparison of headway and time-to-collision as safety indicators. Accident Analysis and Prevention, 2003, 35(33), 427-433.

[10] Adell, E., Várhelyi1, A., Alonso, M., & Plaza, J. Developing human-machine interaction components for a driver assistance system for safe speed and safe distance. IET Intelligent Transport Systems, 2008, 2(1), 1-14.

[11] Van Winsum, W., & Heino, A. Choice of time-headway in car-following and the role of time-to-collision information in braking. Ergonomics, 1996, 39(4), 579-592.

[12] Van der Hulst, M., Meijman, T., & Rothengatter, T. Anticipation and the adaptive control of safety margins in driving. Ergonomics, 1998, 42(2), 336-345.

[13] Brackstone, M. An examination of the use of fuzzy sets in describing relative speed perception. Ergonomics, 2000, 43(4), 528-542.

[14] Gipps, P.G. A behavioural car-following model for computer simulation. Transportation Research Part B: Methodological, 1981, 15(2), 105-111.

[15] Kerner, B., & Klenov, S. Spatial-temporal patterns in heterogeneous traffic flow with a variety of driver behavioural characteristics and vehicle parameters. Journal of Physics A: Mathematical and General, 2004, 37(37), 8753-8788.

[16] Kesting, A. Microscopic modeling of human and automated driving: towards traffic adaptive cruise control. Doctoral Thesis, Technische Universität Dresden, Dresden. 2008.

[17] Ossen, S.J.L. Longitudinal driving behavior: theory and empirics. Doctoral Thesis, Delft University of Technology, Delft. 2008.

[18] Taieb-Maimon, M., & Shinar, D. Minimum and comfortable driving headways: reality versus perception. Human Factors, 2001, 43(1), 159-172.

[19] Wang, J., Lu, M., & Li, K. Characterization of longitudinal driving behaviour by measureable parameters. Transportation Research Record. 2010, 2185, 15-23.

[20] Tampère, C.M.J. Human-kinetic multiclass traffic flow theory and modelling. Doctoral Thesis, Delft University of Technology, Delft. 2004.

[21] Toledo. Driving Behaviour: Models and Challenges. Transport Reviews, 2007, 27(1), 65-84. [22] Bando, M., Hasebe, K., Nakanishi, K., Nakayama, A., Shibata, A., & Sugiyama, Y.

Phenomenological study of dynamical model of traffic flow. Journal de Physique I, 1995, 5(11), 1389-1399.

[23] Bando, M., Hasebe, K., Nakayama, A., Shibata, A., & Sugiyama, Y. Dynamical model of traffic congestion and numerical simulation. Physical Review E, 1995, 51(2), 1035-1042.

17

[24] Addison, P.S., & Low, D.J. A novel nonlinear car-following model. Chaos, 1998, 8(4), 791-799. [25] Treiber, M., Hennecke, & A., Helbing, D. Congested traffic states in empirical observations and

microscopic simulations. Physical Review E, 2000, 62(2), 1805-1824. [26] Wang, J., Li, S., Hang, X. and Li, K. Driving simulation platform applied to develop driving

assistance systems. IET Intelligent Transport Systems. 2009, 4(2), 121-127. [27] Bonsall, P. and Parry, T. Using an interactive route-choice simulator to investigate drivers'

compliance with route guidance advice, Transportation Research Record. 1991, 1306, 59-68. [28] Koutsopoulos, H. N., Lotan, T. and Yang, Q. A driving simulator and its application for modeling

route choice in the presence of information, Transportation Research Part C: Emerging Technologies. 1994, 2(2), 91-107.

[29] Xiong, C. & Zhang, L. A descriptive Bayesian approach to modelling and calibrating drivers’ en-route diversion behaviour. IEEE Transactions on Intelligent Transportation Systems. 2013, 15(4), pp. 1817-1824.

[30] Zhang, L., Wang, J., Li, K., Yamamura, T., Kuge, N., & Nakagawa, T. An instrumented vehicle test-bed and analysis methodology for investigating driver behaviour. In Proceedings of the 14th World Congress on Intelligent Transport Systems, Beijing. China, 2007.

[31] Sato, T. & Akamatsu, M. Analysis of drivers’ preparatory behaviour before turning at intersections. IET Intelligent Transport Systems. 2009, 3(4), pp. 379-389.

[32] Larue, G.S., Malik, H., Rakotonirainy, A., & Demmel, S. Fuel consumption and gas emissions of an automatic transmission vehicle following simple eco-driving instructions on urban roads. IET Intelligent Transport Systems. 2013, In Press.

[33] Dijker, T., Bovy, P.H.L., & Vermijs, R. Car-following under congested conditions: empirical findings. Transportation Research Record, 1998, 1644, pp. 20-28.

[34] Brackstone, M., Waterson, B., & McDonald, M. Determinants of following headway in congested traffic. Transportation Research Part F: Traffic Psychology and Behaviour, 2009, 12(2), pp. 131-142.

[35] Lu, M., Wang, J., Li, K., Zhang, L., Yamamura, T., Kuge, N., and Nakagawa, T. Road category dependence of variations in longitudinal driving behavior. Presentation at the 22nd ICTCT (International Cooperation on Theories and Concepts in Traffic safety) Workshop, Leeds, UK. Oct. 22, 2009.

[36] Lu, M., Wevers, K., & Bekiaris, E. Traffic safety principles and physical road infrastructure measures. Presentation at the 19th ICTCT (International Cooperation on Theories and Concepts in Traffic safety) Workshop, Minsk. 2006.

[37] Todorova, M., R. Donceva, J. Bunevska. Role of functional classification of highways in road traffic safety. TRANSPORT PROBLEMS, 2009, 4(3), 97-104.

[38] RWG. Assessing the Feasibility of a National Road Classification. ICSM Bi-Annual Meeting Paper, Version 2.0. 2006.

[39] Mann, H.B. & Whitney, D.R. On a test of whether one of two random variables is stochastically larger than the other. Annals of Mathematical Statistics, 1947, 18(1), pp. 50-60.

[40] Wilcoxon, F. Some uses of statistics in plant pathology. Biometrics Bulletin 1945, pp. 41-45. [41] Zhuang, C., & He, C. Application of mathematical statistics. South China University of Technology

Press, Guangzhou. (in Chinese). 2006. [42] Liu, C., & Wan, J. Probability and statistics. Higher Education Press, Beijing (PRC). (in Chinese).

2005.