• Non ci sono risultati.

Driver’s behaviour changes with different LODs of Road scenarios

N/A
N/A
Protected

Academic year: 2021

Condividi "Driver’s behaviour changes with different LODs of Road scenarios"

Copied!
19
0
0

Testo completo

(1)

EUROPEAN TRANSPORT CONFERENCE 2020

© AET 2020 and contributors 1

DRIVER’S BEHAVIOUR CHANGES WITH DIFFERENT LODS OF ROAD SCENARIOS

Giandomenico Caruso

Politecnico di Milano, DMEC Yuan Shi

Politecnico di Milano, DMEC Islam Sayed Ahmed Politecnico di Milano, CERM

Alfonso Ferraioli Politecnico di Milano, iDrive

Barbara Piga

Politecnico di Milano, DASTU, Lorenzo Mussone

Politecnico di Milano, DABC ABSTRACT

This paper presents the outcomes of a research aiming at studying and assessing driver behaviours in relation to simulated driving environments with a different Level of Detail (LOD). The study was conducted through an experimental survey based on a data collection related to the functional variables of the vehicle and the psychological responses of the driver. The focus of this study relies on buildings and other artefacts visible to a driver from the vehicle in motion. The hypothesis is that a different LOD of these elements could affect driver’s behaviour and, consequently, the validity of driving experience assessment. The construction of scenarios could be complex and time consuming, especially when the research aims at evaluating the behaviour of a driver in different conditions. Consequently, using in an effective way a lower LOD scenario should lead to a reduction of time required for the implementation of the scenario. In this study, four scenarios with different LODs of its components of the same urban road path (1.12km long) have been developed following a protocol to optimize the 3D modelling activity. The experiment involved 26 participants, all of them drove the path twice with each LOD scenario. The simulation was developed by using CarMaker software (IGP). Different variables of vehicle, such as trajectory, speed, braking, acceleration, and steering angle, as well as some driver’s physiological data have been collected. Empirical approaches and statistical analysis have been adopted to investigate the relationship between LOD and collected data. Results show that effect of LOD is not the same for all variables: trajectory is almost unrelated to LOD, conversely Skin Conductance is strongly affected; besides, also interaction between variables may play a significant role in the relationship. Anyhow, in general terms, the needed level of realism of a simulation may change according to the specific goal of the study and the peculiar context taken into consideration.

(2)

EUROPEAN TRANSPORT CONFERENCE 2020

© AET 2020 and contributors 2

1. INTRODUCTION

Increasing safety of roads and improving the user experience within the vehicle are important objectives that the automotive sector is addressing over the decades even through the study of the driver’s behaviour. Driver’s behaviour is influenced by multiple variables (Ellison, 2015) including the structural elements constituting the driving environment. In (van der Horst, 2007), for instance, the authors observed that drivers tend to follow a different trajectory if safety barriers are present along the border of the road, but driving with different barriers, in terms of type and size, do not produce any trajectory alterations. Consequently, is not possible to identify a priori if a specific characteristic of structural elements can lead to a significant driver's behavioural change. Driving simulators can be effectively used to evaluate this aspect (Kaptein, 1996) not only because the same driving scenario can be subjected to multiple drivers, but also because data can be harvested in real-time along the testing session and it is very helpful for the further data processing and the reliability of those data, compared with the real-car driving, as it has been proved in several research (Robbins, 2019; Ruscio, 2017; Wang, 2010).

Usually, driving simulator validation studies tended to assess speed and speed adaptation (Godley, 2002), and lane keeping (Lee, 2003), to measure relatively low-level vehicle control, like perceptual-motor abilities. However, the driver’s behaviour involves also other elements associated to active controlling and decision-making situations, (e.g. speed adjustment, driving manoeuvre, overtaking, stopping), that requires the definition of the driver’s psychophysical states (Groeger, 2000). The identification of which type of state can be registered by psychological and physiological measures is related not only to individual variability, but also to the type of driving situation that the driver must face. Higher-level physiological measures, like frequency-domain measure of Heart Rate Variability (HRV) (Johnson, 2011) and Skin Conductance (SC), are valuable indexes to measure cognitive effort and workload (Lenneman, 2009). Each signal can be correlated with a specific psychological parameter which can inform about the emotional or cognitive state of the driver. For example, by measuring the SC, the emotional level related to a specific event during the experiment can be captured numerically, e.g. if the driving situation implies that the user is stressed, the increase of SC indicates the related event enhanced the stress level. Through the analysis of these data the emotional state of the user can be derived with statistical accuracy.

Besides the difficulties related to the acquisition and synchronisation of these heterogeneous data, setting up a driving simulation test implies the development of road scenarios that should include all the elements visible from the vehicle (e.g. buildings, road signs, lamppost, barriers, etc.) to provide a realistic driving experience. Therefore, making such scenarios is an intensive and time-consuming task, and researchers generally face with these questionings:

• How much realistic the simulation should be to reach the ecological validity (simulation vs the real driving environment)?

• What kind of elements must be modelled and rendered?

(3)

EUROPEAN TRANSPORT CONFERENCE 2020

© AET 2020 and contributors 3

• Which is the minimum render quality?

Consequently, the identification of the needed minimum requirements, in terms of quality and quantity of the elements that characterise the driving scenario, should lead to a reduction of time needed for the implementation.

This paper presents a study aimed at evaluating the impact of the Level Of Detail (LOD) of the road scenarios on drivers’ behaviour (see also Piga, 2020). Four driving scenarios, representing the same area have been modelled with different LODs in terms of graphical and presence of physical elements. The hypothesis is that a different LOD could affect driver’s behaviour and, consequently, the validity of the driving experience assessment. The assessment is based on the statistical evaluation of data collected during a survey on 29 samples. Data refer both to vehicular features (speed, trajectory, acceleration pedal) and driver’s physiological signals (SC, HRV).

The paper has other 6 sections: (2) Road Scenarios (3) Description of the experiment, (4) Data processing, (5) Results, (6) Discussion, (7) the Conclusion section synthesises the work and the results, and proposes future research activities.

2. ROAD SCENARIOS

A general driving scenario is composed by static elements (roads, buildings, road marking, vertical signing, perspective scenes), interactive elements (vehicles, pedestrians, unexpected obstacles), and dynamic elements (traffic lights). This study focuses on static elements, which are the ones that mainly characterise the driving environment.

The scenarios elaborated for the test relate to the area of the Politecnico di Milano “Campus Bovisa”, which located in the North-West of Milan. The area is about 350m x 350m wide, and the path is 1.12 km long, which is the public one-way ring road that runs along and in between the campus area. The area is mainly composed of isolated buildings (the university ones), surrounded by open spaces and fences (Figure. 1).

The 3D models, used in this study, were built by using the 3D design software SketchUp (SketchUp, 2020). Starting from 2.5D volumetric extrusions of the building, it was then possible to add other details, e.g. the roofs, calculated from the 2D map and the images provided by Google Earth©. A site survey was conducted and photos of visible facades from the road were taken with a digital camera. The images were then organized and post-processed to be used in the 3D model. Non-accessible facades were reconstructed starting from pictures imported from Google Street View©. Each part of the model was organized in layers, to easily recall the different LODs of each scenario.

(4)

EUROPEAN TRANSPORT CONFERENCE 2020

© AET 2020 and contributors 4

Figure 1: Map of the area elaborated for the test.

In this research, a different way to define the LOD of the scenarios has been elaborated starting from the CityGML (City Geography Markup Language) approach (Kolbe, 2009). Some elements (e.g. the roads, traffic signs) appear in all scenarios while others have been incrementally added, as listed in the following:

• LOD 0 includes roads, vertical signs, and road markings.

• LOD 1 extends LOD 0 with buildings as simple volumes with the same colour.

• LOD 2 extends LOD 1 with roofs as simple volumes with the same colour.

• LOD 3 is the scenario with the highest level of detail. Buildings and roofs have their photorealistic materials. Moreover, fences, walls and streetlamps have been added. Table 1 summarises the elements included in each LOD and shows images of the scenarios from the driver's perspective.

(5)

EUROPEAN TRANSPORT CONFERENCE 2020

© AET 2020 and contributors 5

Table 1: The elements composing the different LODs of road scenarios. road, vertical signs, road markings buildings roofs materials, fences, walls, streetlamps LOD 0 LOD 1 LOD 2 LOD 3

(6)

EUROPEAN TRANSPORT CONFERENCE 2020

© AET 2020 and contributors 6

3. DESCRIPTION OF THE EXPERIMENT

To assess the impact of the proposed scenarios on the driver’s behaviour, an experimental campaign was organised. The experiment involved 26 subjects, see Table 2 for more details.

Table 2: Descriptive statistic of samples involved in the experiment.

Gender 16 males (62%), 10 females (38%)

Age [years] Min 21, Max 25, Mean 23, SD 1.19

Driving experience [years] 62% between 2 and 5, 38% more than 5

To become more familiar with the simulator all subjects firstly drove for 3 minutes through an adaptation scenario where they could check visibility, “gas” pedal, brake pedal, and steering wheel reactions. After the adaptation procedure the subjects drove through the four testing scenarios with different LODs; the sequence of the scenarios was varied for each participant to avoid bias among four different sequences of LOD, called A,B,C,and D. Sequence A identifies the series of LOD (0, 1 ,2 and 3), sequence B that of LOD (1, 2, 3, 0), sequence C that of LOD (2, 3, 0, 1) and, finally, sequence D that of LOD (3, 0, 1, 2). In practice each sequence is obtained by a left shift of the preceding one.

Subjects run the path twice for each LOD. After each second lap, the participant answered a questionnaire on the driving experience. After the test on all the scenarios a short discussion about the driving experience was done with the participants.

Figure 2: The driving simulator used during the experimental test.

A fixed-based driving simulator was used to simulate the implemented driving scenarios. Driving simulator comprised a set of vehicle control such as steering wheel with force feedback, gear shifter with automatic transmission, and brake and gas pedals (Figure 2). The driving scene was displayed in three screens providing 36 degrees and 175 degrees

(7)

EUROPEAN TRANSPORT CONFERENCE 2020

© AET 2020 and contributors 7

on vertical and horizontal field of view, respectively. Virtual driving simulation was managed IPG CarMaker (IPG, 2020), which is a high-end car simulator software This software was also used to interface input and output devices of the simulator system. The driving simulator was finally equipped with the commercial physiological sensor Biograph Infinity system from Thought Technology (TT, 2020) to record SC and HRV, for identifying drivers’ state. Physiological and vehicle data have been collected by using an external network application which is able to synchronise both signals at 10 Hz to simplify the subsequent data processing.

4. DATA PROCESSING

In order to compare collected data, it is necessary to process them. In fact, during driving simulation data are collected at regular time intervals (100ms). Since speed (also average speed) changes driver by driver and according to road characteristics, collected measures (both physiological for drivers and physical for vehicles) may have a different number of records: namely, the higher the number the slower the speed. This implies that comparison on time dimension is quite cumbersome and could not lead to profitable results.

For the sake of data comparison, we transformed data from time to space, so that data can be sampled at regular space intervals. By doing this, it was possible to use the road segment length for the comparison; indeed, all data series have the same length and events (e.g. when cornering) are more clearly comparable and attributable to road geometry.

Since the relationship time-space is strictly increasing monotonic (vehicles can go only in one direction and cannot stop) and then bijective, it could be easily built by using a classical fitting function. Then, by using an array of equally spaced distance intervals (e.g. we use a 0.1m step to increase resolution), we could calculate the new corresponding values of time.

a) b)

Figure 3: An example of curve transformation, a) SC signal by Time (before transformation) b) SC signal by Distance (after transformation).

Until there is only one relationship between time and space (there is one driver with many signals) the operation is simple. When more relationships must be managed, due for

(8)

EUROPEAN TRANSPORT CONFERENCE 2020

© AET 2020 and contributors 8

example to different drivers, there are no particular issues to face thanks to the bijective property of those relationships. Obviously, it is crucial to check that total length is the same for all curves. If not, this entails that the number of data may be different and comparison between transformed curves could be affected.

In figure 3 two applications to SC curves (LOD 0,1,2,3 in red, blue, green, and black respectively, for the 1st loop) are shown: a) before transformation, and b) after transformation.

4.1 Segmentation of the path

The circuit has been subdivided into 112 elementary segments, 10 meters long each (Figure 4a). To each segment characteristics are joined: road width, number of lanes, curvature radius, pedestrian crossing, presence on left and right side separately of buildings, sidewalk, trees, light poles, traffic lights, etc.

Then, aggregation is made considering separately rectilinear (1/r=0) and curvilinear stretches (r < 100m), resulting in 8 rectilinear segments (total 730m) and 8 curvilinear segments (total 390m) (Figure 4b).

a) b)

Figure 4: Segmented path and aggregation by type (rectilinear and curve stretches) for the test bed circuit.

4.2 Tools used for analysis

We used both empirical approaches (findspots function) and Statistical analysis (correlation and ANOVA).

Findspot programme

Findspot looks for minima or maxima separately occurring in a group of waveforms in the same spot along the path (first and/or second loop). In the present version, the group of

(9)

EUROPEAN TRANSPORT CONFERENCE 2020

© AET 2020 and contributors 9

waveforms is always made up by 4 waveforms, one for each LOD, and they refer to the same signal, same sample, same loop. Two parameters control the computation:

- The first controls the percentage of variation of signal (used 20%) that is if the current maximum (or minimum) peak is far from previous minimum (or maximum) peak at least the 20% of overall range;

- The second defines the breadth of the spot (used 20m, two segments) inside which peaks can be considered stacked.

Peaks occurring within the same spot are then extracted and shown on the map (see figure 5 for an example).

The underlying idea is that the more the spots the less the effect of LOD on signal. In fact, if changes in waveforms (of different LOD) of the same signal occur at the same place it is likely that LOD is not a control variable.

Correlation

We use canonical correlation, rxy, to analyse similarity between two waveforms, x and y, of the same signals (for the same or different samples, loops and LOD sequence).

𝑟

𝑥𝑦

=

∑ (𝑥𝑖−𝑥̅)(𝑦𝑖−𝑦̅)

𝑛 𝑖=1

√∑𝑛𝑖=1(𝑥𝑖−𝑥̅)2∑𝑛𝑖=1(𝑦𝑖−𝑦̅)2

(1)

It requires the same number of data for arrays, but it is not sensitive to change in averages. rxy value, the coefficient of correlation, is in the range [-1, +1]. The higher the correlation the more similar the waveforms in signals. It is worth noting that high correlation does not mean a causal effect between waveforms.

To make the outcomes easy to read, we adopted the following scheme to classify correlation and therefore wave forms interactions:

• 0 – 0.3 none (dark blue) • 0.3 – 0.5 weak (cyan) • 0.5 – 0.7 moderate (green) • 0.7 – 1.0 strong (red)

Average of correlation values, needed when comparing groups of waveforms, is calculates by applying the Fisher’s Z transformation (Fisher, 1915) to improve normality of data. It is calculated by transforming correlation values firstly by the arc-tanh function, averaging, and then back transforming by tanh function. Averaging (and comparison) of correlation values is allowed only for correlations between waveforms referred to the same segments or their clusters (also not adjacent).

(10)

EUROPEAN TRANSPORT CONFERENCE 2020

© AET 2020 and contributors 10

ANOVA

ANOVA is a procedure for determining whether variation in the response variable arises within or among different groups (levels) of some factors (e.g. LOD, sample, loop, etc.). We use N-way ANOVA that considers more than one factor a time and allows to assess also linear interaction between factors. Graphical interpretation of the result (box plot) is given for one-way ANOVA.

5. RESULTS

We considered two types of signals: 1. Physiological

a) Skin Conductance (SC), varying in the range [8.9, 42.5] [mSiemens]; 2. Mechanical

a) Speed, the speed of vehicle, varying in the range [0, 26.5] [m/s]; b) DMGas, the ‘gas’ pedal use, varying in the range [0, 0.76];

c) RoadDist, the distance of vehicle from road median line, varying in the range [-3.45 , 3.28][m]; it can be considered the trajectory of vehicle.

We collected also physiological signals related to the HRV but because of their nature requiring aggregation on whole path they are not used in the following analysis.

Findspot analysis was applied for each sample, loop, and signal, and min/max spots are calculated as reported in table 3. Average correlation between LOD curves is also reported. Figure 5 shows an example of this analysis.

(11)

EUROPEAN TRANSPORT CONFERENCE 2020

© AET 2020 and contributors 11

The four signals have different outcomes both between signals and samples. SC has the smallest number of common min/max while RoadDist has the largest one. This is a clue, though not yet an evidence, that the needed observance of road geometrical characteristics can vanish the effects of graphical details for what concerns mechanical signal. Note that outcome for DMGas is counterintuitive because it has a relative low average correlation but a relative high number of common spots. It may be due to many zero values that characterize this signal.

Table 3: Correlation between LODs waveforms, number of common maxima and minima by sample and signal.

Correlation is calculated for:

1. different LOD (LOD0/LOD1, …, LOD2/LOD3 2. different LOD Sequence (A/B, A/C, …, B/D, C/D* 3. different loop (1 - 2 )

Factors (and levels in brackets) used in ANOVA are: 1. Samples (26)

2. LOD (4) 3. Loop (2)

4. Sequence (A/B, A/C, …, B/D, C/D) (6) 5. Code (LOD0/LOD1, …, LOD2/LOD3) (6)

6. Homogeneous segments (HS) (8 curvilinear, 8 rectilinear)

The first analysis deals with data of the whole path (really the first 200m of the path, ten segments, are not considered because in loop 1 wave forms have a very different shape from those in loop 2). In table 4 the synthesis of outcome for the whole path is reported. It can be observed, by analysing only LOD column (correlations between waveforms of the same signal, sample, and loop) that RoadDist and Speed are the most correlated, SC and DMGas the least. This means that only SC and DMGas can be considered affected by LOD.

In Figure 6 two examples of ANOVA outcomes are reported. In Figure 6a) the role of samples is shown; this is a general outcome since differences between samples always result significant. In Figure 6b) the effect of LOD on samples correlation is reported.

(12)

EUROPEAN TRANSPORT CONFERENCE 2020

© AET 2020 and contributors 12

Table 4: Synthesis of outcomes by signal, type of correlation for whole path (in small letters significant factors are reported with related correlation value range and significant interactions between factors).

a) b)

Figure 6: Box plots for whole path outcomes: a) Speed LOD correlation by samples, b) Road-dist sample correlation by LOD (red crosses represent outliers).

• *[]=Minimum and maximum correlation values for factors

Whole path

LOD

(0,1,2,3) LOD Sequence (A,B,C,D) Loop (1,2)

Speed

0.72 0.63 0.75 Samples Loop Samples*Code Samples*Loop [0.36 , 0.91] [0.68 0.71] Samples Lod Samples*Lod [0.25 , 0.76] [0.50 , 0.75]

DMGas

0.43 0.30 0.49 Code Samples Samples*Code Samples*Loop [0.39 , 0.49] [0.17 , 0.74] Lod Lod*Loop [0.73 , 0.80] Samples Lod Samples*Lod [0.06 , 0.46 [0.18 , 0.48]

Road

Dist

0.79 0.61 0.78 Samples Code Samples*Code [0.44 , 0.93] [0.75 , 0.82] Lod Sequence Loop Lod*Sequence Lod*Loop Sequence*Loop [0.88. 0.94] [0.88 , 0.93] [0.91 0.92] Lod Samples Samples*Lod [0.13 , 0.80] [-0.07, 0.76]]

Skin

Conductance

0.44 0.19 0.40 Samples Samples*Loop [-0.01, 0.67] Lod Sequence Loop [0.38 , 0.60] [0.33 , 0.66] [0.41 , 0.63] Samples Lod Samples*Lod [-0.75, 0.77]] [0.30 , 0.40]

(13)

EUROPEAN TRANSPORT CONFERENCE 2020

© AET 2020 and contributors 13

Table 5: Synthesis of outcomes by signal, type of correlation and road type (rectilinear and curvilinear stretches).

In table 5 outcomes obtained by grouping together all rectilinear and curvilinear segments (RS and CS respectively) are reported. It is worth noting that these results require specific calculations and cannot be achieved by, for instance, outcomes achieved for single segments; it is also evident that overall outcome for whole path (Total) is not the (weighted) average of rectilinear and curvilinear outcome. Their interpretation is more complex because there are no pronounced trends though Speed and RoadDist signals have the highest values both for RS and CS. By considering correlations by LOD we can observe a little higher values for CS, but this result is overturned when considering the other types of correlation (LOD sequence and Loop). SC does not change much in curves between the two loops, instead it changes much when considering the LOD sequence. Table 6 shows a further detailed analysis performed for each segment. Only LOD correlation is reported for brevity. It is easy to see that RoadDist is the signal with the largest number of strong correlation (in four segments their values are over 0.94), followed by Speed and then, SC and DMGas (the latter has five segments with none correlation, and less than 0.15). The very low values for Speed and DMGas in the last two segments

By road type

Synthesis

LOD (0,1,2,3) LOD Sequence (A,B,C,D) Loop (1,2) Spee d RS 0.48 0.33 0.40 CS 0.75 0.69 0.35 Total 0.72 0.63 0.38 DMG as RS 0.37 0.25 0.78 CS 0.35 0.26 0.56 Total 0.43 0.41 0.53 RoadDis t RS 0.50 0.61 0.42 CS 0.57 0.22 0.39 Total 0.79 0.30 0.49 SC RS 0.41 0.21 0.52 CS 0.40 0.15 0.78 Total 0.44 0.19 0.75

(14)

EUROPEAN TRANSPORT CONFERENCE 2020

© AET 2020 and contributors 14

of the loop probably mean that driving style (but not trajectory) is very different from one LOD to another one.

Table 6: Synthesis of outcomes obtained for LOD correlation and each segment (R=rectilinear, C=curvilinear).

Figure 7 shows the box plots of ANOVA for all rectilinear segments and Speed correlation by LOD (a) and by Loop (b). A part the conspicuous number of outliers in some segments which point out the great variability in the process, results in a) confirm those in table 6, but they also show the non uniform ranges in which they vary. Besides, figure 13b underlines the low correlation between loops for the first segment AB.

Figure 8 resumes LOD correlation per each signal and confirms previous presented results: RoadDist has the highest correlation and smallest range (a part a few outliers), on the opposite DMGas and SC have the lowest correlations.

Figure 9 shows the average correlation between loop 1 and 2 considering all signals. First two segments have the lowest values (as expected) but last two have low values probably witnessing a drop in attention at the end of second loop and path.

By road

segments

Speed

DMGAS

RoadDist

SC

LOD

(0 ,1 ,2 ,3 ) AB (R)

0.8383

0.5677

0.6931

0.4668

BC (C)

0.5822

0.5086

0.7968

0.4209

CD (R)

0.7450

0.5992

0.7271

0.4466

DE (C)

0.5206

0.6995

0.9451

0.1146

EF (R)

0.6869

0.4842

0.7079

0.4859

FG (C)

0.4568

0.1300

0.8994

0.7218

GH (R)

0.8682

0.4340

0.7506

0.4171

HI (C)

0.5581

0.4316

0.8162

0.4501

IJ (C)

0.9296

0.3234

0.7043

0.2902

JK (R)

0.2741

0.1117

0.9490

0.6703

KL (C)

0.7941

0.3058

0.8014

0.4446

LM (R)

0.7969

0.1490

0.9501

0.2942

MN (C)

0.8481

0.6401

0.9539

0.2121

NO (R)

0.4538

0.3109

0.7985

0.5892

OP (C)

0.1758

0.0392

0.8629

0.5820

PA (R)

0.0155

0.1125

0.4830

0.5983

(15)

EUROPEAN TRANSPORT CONFERENCE 2020

© AET 2020 and contributors 15

a)

b)

(16)

EUROPEAN TRANSPORT CONFERENCE 2020

© AET 2020 and contributors 16

Figure 8: Different LOD correlations by signal.

Figure 9: Correlation between wave forms in loop 1 and 2 along the path (blue points), average value (red point).

(17)

EUROPEAN TRANSPORT CONFERENCE 2020

© AET 2020 and contributors 17

6. DISCUSSION

Correlation varies much by signal, sample, LOD, LOD sequence and Loop. Outcomes show that also interaction between these variables is significant, making data processing more complex.

Path segmentation shows that rectilinear road stretches have different figures than curvilinear ones and this depends very much on the considered signal. Particularly, RoadDist (that is, vehicle trajectory) is the most correlated signal and it turns out it depends more on geometrical features than on LOD. Correlation values generally range in a wide interval but all averaged values for signal and whatever long segment are positive (in fact, trends of wave forms are generally similar). In some sense, we can say that driving process is oriented, and driving behaviour generates wave forms with same slope for the same segments. But by dividing path into clusters of segments (rectilinear or curvilinear aggregation) or into basic homogenous segments, outcomes may be counterintuitive (in fact, it is worth noting that correlation is not a linear operator). This means that the correlation of a certain path stretch cannot be obtained by summing the correlation of its basic segments.

The H0 hypothesis that LOD affects behaviour can be accepted for some signals only, like DMGAS and SC. This can be explained by considering that RoadDist and, at a lesser extent, Speed are much more constrained by road geometrical features. May be that these two variables could be affected more by interacting effects of circulation, like pedestrians, other circulating vehicles, which could be part of a furthermore complex scenario, a kind of LOD4. Instead, the secondary hypothesis on the effect of LOD sequence cannot be accepted for any signal.

Moreover, we stress that interactions between factors (Sample, Loop, Sequence, …) have such an important role that they could deserve a specific analysis for each different path.

7. CONCLUSIONS AND FUTURE WORKS

Four scenarios of the same road path were modelled with different LODs for running driving simulation tests in order to compare drivers’ behaviours. In particular, LOD 0 was a simple carriageway with road signs, in LOD 1 we added 2.5D volumes without texturing, in LOD 2 roofs were added, and in LOD 3 the model became photorealistic by texturing existing buildings facades and by adding urban details, such as lampposts.

As a matter of fact, driving scenario with missing or poor environmental details (low LODs) negatively affect the fidelity of the driving experience.

Specific programmes (Findspot) and classical statistical analyses (correlation and ANOVA) were applied to analyse collected data for getting both qualitative and quantitative outcomes. For this phase they have been revealed sufficient to achieve the stated goal. They have shown the different features related to collected signals. RoadDist (i.e vehicle trajectory) strongly depends on road characteristic and nearly nothing on LOD. On the

(18)

EUROPEAN TRANSPORT CONFERENCE 2020

© AET 2020 and contributors 18

other hand, SC show a quite strong dependence on LOD and the two other signals, Speed and DMGas (acceleration pedal) are in an intermediate position.

The planned next steps of the research are: in the short time, to consider further factors of road scenario (e.g. buildings, trees, lampposts) and check if wave forms change when the vehicle approaches them; to make a joined analysis of these data with those collected by eye-tracker device.

ACKNOWLEDGMENT

This research was supported by the i.Drive Lab (http://www.idrive.polimi.it/). The authors are thankful to all the survey participants for investing their time in this study.

BIBLIOGRAPHY

Ellison, A. B., Greaves, S.P., & Bliemer, M.C.J., Driver behaviour profiles for road safety analysis. Accident Analysis & Prevention, 76, 118–132.https://doi.org/10.1016/J.AAP.2015.01.009, 2015 Fisher, R.A., Frequency distribution of the values of the correlation coefficient in samples of an indefinitely large population. Biometrika. 10 (4): 507–521. https://doi.org/10.2307/2331838, 1915 Godley, S.T., Triggs, T.J., and Fildes, B.N., Driving simulator validation for speed research. Accident Analysis & Prevention., 34(5), 589-600. https://doi.org/10.1016/S0001-4575(01)00056-2, 2002

Groeger, J.A., Understanding driving: Applying cognitive psychology to a complex everyday task. Psychology Press, 2000

Johnson, M.J., Chahal, T., Stinchcombe, A., Mullen, N., Weaver, B. and Bedard, M., Physiological responses to simulated and on-road driving. International journal of Psychophysiology, 81(3), 203-208. https://doi.org/10.1016/j.ijpsycho.2011.06.012, 2011

Kaptein, N.A., Theeuwes, J., van der Horst, R., Driving simulator validity: Some considerations. Transportation Research Record, 1550(1), 30–36, 1996

http://doi.org/10.1177/0361198196155000105

Kolbe, T.H., Representing and Exchanging 3D City Models with CityGML. In J. Lee & S. Zlatanova (Eds.), 3D Geoinformation Sciences. Berlin, Springer, 2009.

Lee, H.C., Cameron, D. and Lee, A.H., Assessing the driving performance of older adult drivers: on-road versus simulated driving. Accident Analysis & Prevention. 35(5), 797-803.

https://doi.org/10.1016/S0001-4575(02)00083-0, 2003

Lenneman, J.K. and Backs, R.W., Cardiac autonomic control during simulated driving with a concurrent verbal working memory task. Human Factors, 51(3), 404–418.

https://doi.org/10.1177/0018720809337716 , 2009

Piga, B., Caruso, G., Ferraioli, A., Mussone, L. . Modeling Virtual Road Scenarios for Driving Simulators: a Comparison of 3D Models with Different Level of Details. Proceedings of the 42° International Conference UID (Unione Italiana per il Disegno), Reggio Calabria, Sept. 2020

Robbins, C.J., Allen, H.A., Chapman, P., Comparing drivers’ visual attention at junctions in real and simulated environments. Applied Ergonomics, 80, 89–101. http://doi.org/10.1016/j.apergo.

(19)

EUROPEAN TRANSPORT CONFERENCE 2020

© AET 2020 and contributors 19

Ruscio, D.; Bascetta, L.; Gabrielli, A.; Matteucci, M.; Ariansyah, D.; Bordegoni, M.; Caruso, G.; Mussone, L., Collection and comparison of driver/passenger physiologic and behavioural data in simulation and on-road driving. In 5th IEEE International Conference on Models and Technologies for Intelligent Transportation Systems (MT-ITS). IEEE, http://doi.org/10.1109/mtits.2017.8005705, 2017

van der Horst, R., de Ridder, S., Influence of Roadside Infrastructure on Driving Behavior: Driving Simulator Study. Transportation Research Record, 2018(1), 36–44. https://doi.org/10.3141%2F2018-06 , 2007

Wang, Y., Mehler, B., Reimer, B., Lammers, V., D'Ambrosio, L.A.; Coughlin, J.F., The validity of driving simulation for assessing differences between in-vehicle informational interfaces: A comparison with field testing. Ergonomics, 53(3), 404–420.

http://doi.org/10.1080/00140130903464358, 2010

Websites

IPG, https://ipg-automotive.com/ (last visit August 2020) SketchUp, https://www.sketchup.com/ (last visit August 2020) TT, http://thoughttechnology.com/ (last visit August 2020)

Riferimenti

Documenti correlati

We will relate the transmission matrix introduced earlier for conductance and shot noise to a new concept: the Green’s function of the system.. The Green’s functions represent the

In this setting, we characterise compliance of quantified ABoxes, and use this characterisation to show how to compute the set of all optimal compliant anonymisations of a

The investigation of attachment experiences in the subsample of the most severe psychopaths, who showed very high average scores on the items denoting devaluation of attachment bonds,

One way of preventing macula detachment during a fluid–air exchange procedure is to drain via a relatively posterior iatrogenic drainage retinotomy (or with a flute needle with a

Finally, is it sensible to assume that the elasticity of substitution across products of the same …rm does not di¤er from that across products of di¤erent …rms, thus limiting the

In Section 4, the characterization provided by Theorem 3.2 leads to combina- torial results: we investigate the structure of the partial order of likeness among dendrites, showing

Title: Effect of the change of social environment on the behavior of a captive brown bear (Ursus arctos).. Article Type: Original

Proposed by Olivier Bernardi (USA), Thaynara Arielly de Lima (Brazil), and Richard Stanley (USA).. Let S 2n denote the symmetric group of all permutations