Transportation Research Part F 69 (2020) 235–250
Contents lists available at ScienceDirect
Transportation Research Part F journal homepage: www.elsevier.com/locate/trf
Assessing the driving distraction effect of vehicle HMI displays using data mining techniques Jun Ma a, Zaiyan Gong a,⇑, Jianjie Tan a, Qianwen Zhang b, Yuanyang Zuo b a b
School of Automotive Studies, Tongji University, No. 4800, Caoan Road, Shanghai 201804, China College of Design and Innovation, Tongji University, No. 281, Fuxin Road, Shanghai 200092, China
a r t i c l e
i n f o
Article history: Received 17 May 2019 Received in revised form 24 January 2020 Accepted 28 January 2020
Keywords: Vehicle HMI Secondary task Driving distraction Eye tracking Data mining Random forest
a b s t r a c t With the rapid development of human–machine interface (HMI) systems in vehicles, driving distraction caused by HMI displays affects road safety. This study presents a data mining technique to model the four driving distraction indicators: speed deviation, lane departure standard deviation, dwell time, and mean glance time. Driving distraction data was collected on a real-car driving simulator. 3 secondary tasks in 13 mass produced cars were tested by 24 drivers. The random forest algorithm outperformed linear regression, extreme gradient boosting, and multi-layer perceptron as the best model, demonstrating good regression performance as well as good interpretability. The result of random forest showed that the importance of target speed is large for all driving distraction indicators. Among the variables of interaction and user interface design, less step and less onscreen distance of finger movement are efficient for lowering lane departure standard deviation and dwell time. The position of right point is another important variable, and should be between 37 and 47 degrees on a typical sample in this study. A larger angle leads to bigger lane departure, while a smaller angle leads to bigger mean glance time. Most variables of HMI display positioning themselves are not important. This study provides one driving distraction assessment method with a variable impact trend analysis for HMI secondary tasks in an early phase of product development. Ó 2020 Elsevier Ltd. All rights reserved.
1. Introduction Safety is one of the most important issues of road traffic, and driving distraction has been identified as a major contributor to road crashes and incidents (Née et al., 2019). Secondary tasks are among the most common distraction activities (68.7%) reported by drivers, being nearly as common as lack of concentration (71.8%), and much higher than viewing outside people, objects, or events (57.8%); talking to passengers (39.8%); drinking (11.3%); eating (6.0%); or smoking (10.6%) (McEvoy, Stevenson, & Woodward, 2006). There is an obvious trend that secondary task engagement has increased, according to studies across multiple countries from 1999 to 2015 (Huemer, Schumacher, Mennecke, & Vollrath, 2018). A naturalistic driving study in Germany reported that drivers are engaged in secondary tasks for approximately 40% of their driving time (Metz, Landau, & Just, 2014), whereas American drivers spend 34% of their driving time engaged in secondary tasks (Sayer, Devonshire, & Flannagan, 2005).
⇑ Corresponding author. E-mail address:
[email protected] (Z. Gong). https://doi.org/10.1016/j.trf.2020.01.016 1369-8478/Ó 2020 Elsevier Ltd. All rights reserved.
236
J. Ma et al. / Transportation Research Part F 69 (2020) 235–250
Recently, many in-vehicle secondary tasks are being accomplished through central displays of human–machine interface (HMI), which are becoming important for driving and commonly equipped in new vehicles. The HMI central displays can show a large amount of useful, high complexity information, and accommodate a greater number of in-vehicle controls in a smaller space (Crundall, Large, & Burnett, 2016), but they have also become a significant source of driving distraction. 1.1. Distraction categorization and assessment Drivers’ attention resources are limited, and if a driver attempts to perform any secondary tasks, the redistribution of attention may lead to deterioration in driving performance (Choudhary, 2017). Distraction caused by secondary tasks is a combination of three categories: visual distraction, when drivers take their eyes off the road to interact with a device, leading to observation errors; cognitive distraction, when attention is withdrawn from driving and assigned to secondary tasks, leading to information encoding and retrieval errors; and manual distraction, when drivers take their hands off the steering wheel to manipulate a device, leading to action errors (Strayer, Watson, & Drews, 2011; Young & Salmon, 2012). Objective assessment indicators of driving distraction refer to driving performance and visual demand. Driving performance is indicated by longitudinal speed control, headway distance control, brake reaction time, lateral steering error, lane departure, etc. Meanwhile, visual demand is indicated by number of glances, mean glance duration, single glance duration, percent dwell time, total dwell time, etc. (Hofmann, Tobisch, Ehrlich, & Berton, 2015; Liang & Lee, 2010; Gaspar et al., 2016; Li et al., 2018; Rosenthal, 1999). Visual distraction deteriorates perception, increases glance duration, and is accompanied by larger lateral lane departure. Additionally, visual distraction also raises drivers’ vigilance, as they reduce speed and increase their headway distance to compensate for their impaired response to potential sudden events (Muhrer & Vollrath, 2011; Kountouriotis, Spyridakos, Carsten, & Merat, 2016; Liang & Lee, 2010). Cognitive distraction influences anticipation of the possible future actions of other car drivers (Muhrer & Vollrath, 2011). In a simple road condition without a lead car or sudden events, many studies found that cognitive distraction reduces gaze dispersion as well as standard deviation of lane departure (Engström, Markkula, Victor, & Merat, 2017; Kountouriotis et al., 2016; Liang & Lee, 2010; Metz, Schoch, Just, & Kuhn, 2014). However, the finding of improved lane keeping by cognitive distraction is still controversial. Li et al. (2018) presented timeto-line crossing as a more significative index than lane departure standard deviation, which showed worse lane keeping safety during cognitive distraction. Visual tasks while driving always have a greater impact on driving safety than cognitive tasks, and the impairment of perception due to visual distraction leads to a slower reaction than cognitive distraction (Jin, Xian, Niu, & Bie, 2015, Muhrer & Vollrath, 2011). Manual distraction mainly deteriorates lateral control, and is intensified by complex physical manipulation (Libby, Chaparro, & He, 2013). 1.2. Distraction impact factors of HMI displays To study the driving distraction caused by secondary tasks in HMI displays, 3 categories of independent variables are analyzed by previous researches: HMI display size and position, interaction design of secondary tasks, and UI (User Interface) design. HMI display position is closely linked to visual distraction, and is recommended to be mounted in a position where the downward viewing angle is no larger than 30 degrees at the geometric center of the display, according to AAM (2002). The reason is that on the vertical meridian, human visual detection thresholds between 10 degrees and ca. 25 degrees show a plateau of constant sensitivity before detection ability again decreases (Strasburger & Rentschler, 1996). Thus, a lower positioned HMI display results in worse lane departure and lower driving speed (Olaverri-Monreal, Hasan, Bulut, Körber, & Bengler, 2014; Wittmann et al., 2006). A larger display size does not always lead to good visibility. Gong and Ma (2018) reported that finding an icon on a 17-in. display takes longer time than on a 12.3-in. display. HMI display position also impacts manual distraction, but it is not extensively discussed by previous studies. One reason could be that most HMI displays and buttons in the driving simulators used in previous studies are positioned to be easily approached by drivers, while in a real car the positions are not always ideal because of compromises in interior design. Kim and Song (2014) found that the display elevation angle should be slightly bigger or adjustable for touch control such that HMI display gestures alleviate fatigue on a user’s wrist, but other dimensions of HMI display position were not discussed in this research. Interaction design of HMI display includes task procedure, touch gesture, menu architecture, etc. Many studies described task procedures in terms of task complexity and completion time, which affect the length of visual and cognitive distractions. More complex secondary tasks usually result in greater interference with lane keeping (Tijerina, Parmer, & Goodman, 1998). Longer task completion times are associated with higher mean glance frequencies to the infotainment device, leading to more aggressive lateral control actions of the drivers, and the rate of lane departure becomes increasingly large over time (Gellatly & Kleiss, 2000; Ma, Shi, Fu, & Guo, 2015). Reading longer sentences on central displays also results in greater lane departures (Peng, Boyle, & Lee, 2014). The International Organization for Standardization (2002) mentioned that glances should be shorter than 1.5 s. Simons-Morton, Guo, Klauer, Ehsani, and Pradhan (2014) reported that crash risk increased with the duration of the single longest glance during all secondary tasks (OR = 3.8 for >2 s). Glances away from the roadway that exceed 1.6–2.0 s are known to increase crash risk (Caird, Johnston, Willness, Asbridge, & Steel, 2014). Regarding touch screens, tapping the touch buttons is discrete, but flicking a list requires continuous finger manipulation, thus the eye dwell time required for flicking is longer than that of tapping, and is likely to exceed 1.5 s (Kim & Song, 2014). In addition, menu
J. Ma et al. / Transportation Research Part F 69 (2020) 235–250
237
architecture impacts task time and mean glance time, according to comparisons between hierarchical layout and checkerboard layout in different driving scenarios (Li, Chen, Sha, & Lu, 2017). UI design mainly refers to text, icon and color blending for vehicle HMI display. Good readability reduces cognitive load and shortens visual distraction time. Both the driving safety and the usability of in-vehicle HMIs increases as its touch-key size increased, up to approximately 17.5 mm (Kim, Kwon, Heo, Lee, & Chung, 2014). However, such a large size is not always feasible for HMI display content. Crundall et al. (2016) recommended that designers of interfaces for HMI displays should avoid using text sizes of 6.5 mm or smaller, because the number of glances and glance duration increase significantly. For Chinese characters, based on the result of Huang, Patrick Rau, and Liu (2009), the optimum font size for providing good readability is 5.25 mm (with a viewing distance of 70 cm, resolution 200 dpi). In addition, properly designed icons reduce system complexity and mental workload, providing less cognitive difficulty than textual user interfaces (García, Badre, & Stasko, 1994). The primed product comparisons method can efficiently find an optimized set of icons for time-critical applications and is beneficial to driving safety (Silvennoinen, Kujala, & Jokinen, 2017). 1.3. Experiment design and multifactor analysis Two methods are commonly applied in tests of driving distraction for HMI systems. The first method is the naturalistic driving study (Dingus et al., 2006), which allows researchers to collect data simultaneously with actual driving on the road. However, this method only enables the collection of CAN-bus data and in-vehicle video (Metz, Landau, et al., 2014, Nevile & Haddington, 2010, Klauer, Dingus, Neale, Sudweeks, & Ramsey, 2006), which is not sufficient for a deep analysis of driving safety. The visual detection task is another way to collect drivers’ reaction speeds (Young, Salmon, & Cornelissen, 2013), but real daily HMI tasks are not included. The second method is a driving simulator. Both the simple open-cab (Choudhary, 2017) and real-car modified-cockpit (Hofmann et al., 2015; Jin et al., 2015) simulators are popular. Driving simulators facilitate precise detection of secondary tasks without being significantly different from actual driving (Risto & Martens, 2014), and the resulting patterns between the simulator and real driving are very similar (Strayer et al., 2015). However, the HMI display positioning as well as interaction and UI design inside the display are always fixed, or are provided with very limited selections in driving simulators, which does not create an experience identical to real production car cockpits. Many previous studies of driving distraction involved a limited amount of impact factors, rather than combining the numerous variables from HMI display size and position, interaction design, and UI design. Recently, data mining techniques have been applied to explore the main drivers of driving distraction in a complex system. Zhang, Owechko, and Zhang (2004) performed one of the earliest studies for detecting a driver’s cognitive distraction, using a decision tree classifier to identify distraction based on driving performance and eye-gaze related features. Liang, Reyes, and Lee (2007) investigated real-time cognitive distraction detection using SVM (Support Vector Machines). Atiquzzaman, Qi, and Fries (2018) and Englund, Nilsson, and Voronov (2016) compared the performance of several algorithms for the regression of driver and bicyclist behaviors, respectively. The aim of this study is to assess the driving distraction effect of HMI displays with extensive independent variables, including HMI display size and position, interaction design of secondary tasks, and UI design. However, other individual differences among drivers and secondary task purposes are considered to be insignificant factors, and are therefore not selected as independent variables in this study. To involve up to 18 independent variables, a driving simulator was established to collect driving distraction data from HMI secondary tasks in 13 real mass produced cars. To analyze the importance and impact trend of each independent variable, four algorithms were selected and compared to model driving distraction indicators. The algorithm with the best performance can be used to assess driving distraction led by HMI display secondary tasks, and can be used to create improvement suggestions for driving safety at an early stage of car product development. 2. 2. Materials and methods 2.1. Driving simulator Conventional driving simulators use an irreplaceable dashboard, and are difficult to use to simulate the differences of HMI secondary tasks among various real production cars. Thus, a real-car driving simulator was established, allowing mass produced cars to be connected to virtual driving scenes. The real-car driving simulation system is composed of 4 parts: (a) A screen with a 240° curve with a diameter of 8 m, allowing the driver to become immersed in driving. Five projectors are used (Fig. 1, top). (b) A 53-km virtual driving scenario based on Unity 3D, a platform for visualization and simulation, widely used in games and driving simulation (Ojados Gonzalez et al., 2017; Dols et al., 2016). A 13.8 km route is defined for the test, including urban road (red line, 3.5 + 3 km), elevated road (blue and purple lines, 3.5 km), and suburban road (yellow line, 3.8 km). All roads are defined and designed based on real Chinese road scenarios (Fig. 1, bottom left).
238
J. Ma et al. / Transportation Research Part F 69 (2020) 235–250
Fig. 1. The real-car driving simulation system: curved screen for immersion (top), test route in virtual driving scenario (bottom left), and red iron turn plate (bottom right). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
(c) A fast connection system which connects any production car with the virtual driving scenario in approximately 60 min. When a car is driven into the lab, the two front wheels are positioned in the middle of two iron turn plates (Fig. 1, bottom right). The two plates are mounted in a lateral groove and can move left and right to match the wheel track of each car. In the running state, when the front wheels rotate, the plates rotate synchronously. The angle sensor below the left plate produces angle data of the virtual driving scenario. Another 2 angle sensors are installed next to the original brake and accelerator pedals to obtain pedal data. The driving behavior is transmitted and used to control the virtual driving scene through the turn plates and pedal sensors. (d) A data-acquisition system which can acquire vehicle data (i.e., accelerator, brake, steering wheel angle), road data (i.e., speed, lane departure), and ergonomic data (i.e., eye tracking). The real-car driving simulation system has the following four advantages: (1) a real HMI can be tested to avoid deviations from the center display position as well as the interaction and UI design of a normal driving simulator; (2) the immersion and realism of the 8-m curved screen are better than those of the smaller open-cab driving simulator; (3) standard scenarios and events can be performed in the simulator; (4) various data are collected in real time for further analysis. 2.2. Test models and participants To evaluate real HMIs in the market, 13 mass produced cars were selected and tested on the real-car driving simulator. The criteria were as follows: both Chinese brands and global brands were included, both premium and non-premium brands were included, car model years were included from between 2016 and 2018, HMI software had been officially updated within the last 2 years, and both vertical and transverse HMI display layouts were included. All center displays allowed touch control (Fig. 2). The HMI displays had sizes from 8 in. to 17 in., which covers the mainstream size range in the market. All car models were left-hand-drive with an HMI display on the right side of the driver.
239
J. Ma et al. / Transportation Research Part F 69 (2020) 235–250
Fig. 2. Interior of 13 tested cars: the first row from left to right: RR-model, BF-model, FM-model, LO-model, GB-model; the second row from left to right: RV-model, BR-model, TS-model, NE-model; the third row from left to right: RM-model, PC-model, PP-model, GG-model.
24 participants (14 male, 10 female) aged 20–45 years (mean = 28.6; SD = 5.3) were included in the study. The participant heights were in the range of 1.65–1.81 m (mean = 1.72; SD = 0.053). Eight participants drove an annual mileage considered short to medium (3000–9000 km/year), ten participants had a medium annual mileage (9000–20,000 km/year), and six participants had a high annual mileage (>20,000 km/year). Participants did not have any specific experience with the HMI test of the vehicle driving simulator, and motion sickness was not apparent on this simulator. 2.3. Experiment design 2.3.1. Secondary task selection Driving itself is the main task for drivers, and all other tasks that require interaction with the HMI are secondary tasks. There are lots of secondary tasks in every test car model. For example, 139 tasks can be operated on the HMI display of the TS-model. Because we hypothesized that the difference among secondary task purposes is not significant in this study, it is unnecessary to select very many tasks. For example, consider the tasks of selecting the most recent call in the contact list and selecting the most recent destination in the navigation list. If these 2 tasks take same number of steps and have similar icon positionings and sizes on the display, the driving distraction of these 2 tasks should be equivalent. The first selection requirement for selected secondary tasks on all test models is to be controlled by touch on the display. Temperature adjustment cannot be selected because many test models only use physical knobs or buttons for this task. The second requirement is for the touch gesture to be performed with only taps on all test models, rather than slide, swipe, or press and hold, which could be difficult to compare. The third requirement is for the task to be finished within 5 steps on all test models. Tasks that require more than 5 steps are suggested for operation while the car is stopped. Based on the above requirements, 3 secondary tasks were selected: skipping to the next track of music, calling a contact, and answering a call (Table 1). All tasks were performed starting from the homepage of HMI system. 2.3.2. Testing process Before each test, the car was driven into the lab and connected to the virtual driving scenario. Drivers sat in the test car and drove it from the starting point to the end. They were required to maintain a certain speed and remain in the center of the left lane. The target driving speed was 45 km/h on urban roads and 60 km/h on suburban and elevated roads. The real-time driving speed was projected on the curved screen. When the driving speed was close to the target speed, the projected number was green. If the deviation between the driving and target speeds exceeded 3 km/h, the projected speed number turned yellow, and if the deviation exceeded 8 km/h, the number turned red to alert the driver. According
Table 1 Secondary task definition in the experiment. Task
Code
Description
Min step among 13 car models
Max step among 13 car models
Next track of music Calling contact
Task-NT Task-CC
1 2
3 5
Answering a call
Task-AC
When one track is playing, skip to the next track Select the first contact in the contact list and call (smartphone connected the HMI via Bluetooth in advance) Accept an incoming call (smartphone connected the HMI via Bluetooth in advance)
1
1
240
J. Ma et al. / Transportation Research Part F 69 (2020) 235–250
to China Light-vehicle Test Cycle (2018), the maximum speeds in city and suburban test scenarios in China are 48.1 km/h and 71.2 km/h, respectively. In this study, because the speed target is continuous for drivers, rather than being a transitory peak, the 5–15% lower deviation from this standard is reasonable. The completion of all tests lasted for more than 3 months. Each week, one car model was rented and tested for four days in the simulator. Every day, 6 drivers participated in the one-hour test in sequence. Each driver was required to drive all of the 13 car models. However, because of the long duration of this test, their attendance rate was 91% in total. Within the testing of one model, the driver drove on the route 3 times. Before driving, the driver statically attempted all secondary tasks until they were familiar with all tasks. The first drive was a warm-up. The drivers drove the car without any secondary tasks to familiarize themselves with the route and vehicle handling. During the second and third drives, drivers performed each secondary task 4 times using touch control on the HMI display. The tasks were arranged randomly. Each task was performed 2 times at 45 km/h on urban roads, and was performed once each at 60 km/h on both suburban and elevated roads. The drivers cannot test the cars in a random order, because the test cars are rented one by one. However, learning effect did not strongly impact on the result. The algorithm and conclusion of this study were based on abstractly simplified variables, rather than specific car models nor secondary tasks. 2.4. Data collection All data was collected from 3 source types: cockpit and HMI display measurement by laser finder and ruler, driving performance from the driving simulator, and gaze monitoring by an eye tracker. Drivers’ subjective opinions are excluded in this research. 2.4.1. Driving distraction indicators Driving performance and visual demand are the two key categories for evaluating driving safety during HMI secondary tasks. The driving simulator collects driving performance data in real time, including driving speed and lane departure, with a sampling frequency of 10 Hz. Speed accuracy is 0.1 km/h; lane departure is 0.1 m. Speed deviation (SpDev) and lane departure standard deviation (LDSD) are two separate indicators. SpDev is the mean deviation between the target and driving speeds during the process of one secondary task. LDSD is the standard deviation of the lateral distance between the car center and the lane center. As it is difficult to drive a car in the exact center of a lane, the standard deviation is more reasonable than the average deviation of the lane departure from the lane center. Smaller SpDev and LDSD values mean less distraction from secondary tasks and safer driving. A head-mounted SMI ETG II eye tracker collected visual demand data. The time accuracy is 1 ms. Dwell time (Dwl) and mean glance time (MnGlc) are indicators based on eye tracking data. In a given secondary task, Dwl is the total time that the driver moved their eyes from the virtual road on the external curved screen to the interior of the car, including individual fixation, saccade, and blink time (Jin et al., 2015). MnGlc is the dwell time divided by the visit count. The visit count refers to the number of times the driver moves their eyes from the virtual road to the interior during a single task. 2.4.2. HMI display size and position The display positioning is used to determine whether a driver’s gaze moves to it quickly and whether driver’s hand reaches it comfortably. To measure the position of the HMI display, a spherical coordinate was established as in Fig. 3. When the seat is adjusted to the lowest position, the origin is from the center of the driver’s eyes (1.72 m, the average height of the 24 drivers). Every point on the display can be represented as (r, h, u). r is the distance from driver’s eyes to the point; h is the side horizontal angle; u is the vertical angle. The horizontal right ahead direction is (r, 0, 0) in this spherical coordinate. The distance and angles are measured by a laser finder installed in the position of driver’s eyes. The HMI display size can also be represented in this spherical coordinate with a unit of square degree. This measurement of display size avoids the impact of display distance from the driver’s eyes, which describes the visual demand more precisely. Otherwise, an HMI display with 10-in. absolute area, for example, could result in different visual distraction when positioned at 650 mm and 750 mm away from the driver’s eyes. The measurement results of the 13 car models are shown in Fig. 4. The positioning of the HMI displays concentrated in a common area, excepting the BF-model and the RM-model. 7 indicators as listed in Table 2 were selected to describe the HMI display size and position. 2.4.3. Interaction and UI design The interaction and UI design determine the complexity of a secondary task and impacts the driver’s gaze movement and the convenience of the touch controls. The interaction design of HMI touch-controlled tasks refers to the number of steps and the touchpoint position on the HMI display. The 10 indicators are shown in Table 3. One step represents one tap, one slide, one swipe, or one press and hold on the HMI display. For the 3 selected secondary tasks, only the tapping gesture is common among the 13 car models, which makes the step factor more comparable among the car models. A touchpoint is an icon area to be touched on the HMI display. The number of touchpoints is always equal to the number of steps for tapping gestures. Considering that tasks shared among different car models can require different numbers of touchpoints, it is difficult to compare the positions of every touchpoint across the car models. Thus, the touchpoint on the top and the touchpoint on the right end are defined as interaction design indicators, because these two touchpoints are the farthest from drivers’ right hand and
J. Ma et al. / Transportation Research Part F 69 (2020) 235–250
Fig 3. Drivers’ spherical coordinate of HMI display measurement.
Fig 4. Spherical coordinates measurement of 13 car models.
241
242
J. Ma et al. / Transportation Research Part F 69 (2020) 235–250
Table 2 Indicators of HMI display size and position. Indicator
Unit
Description
Average of 13 car models
SD of 13 car models
Ball area Aspect ratio (width/tall) Center height
Deg2 – Deg
227.10 1.68 30.35
130.72 0.86 6.87
Center side horizontal angle
Deg
42.04
3.99
Center distance Depression angle Horizontal angle
mm Deg Deg
The area of HMI display in the spherical coordinate The length ratio of HMI display width (mm) and height (mm) The downward viewing angle of HMI display geometric center from horizontal plane The side horizontal angle of HMI display geometric center from right ahead vertical plane The distance from driver’s eye to HMI display geometric center The depression angle of HMI display surface from vertical plane The horizontal deflection angle of HMI display perpendicular from the car longitudinal axis
733.25 26.65 9.06
55.87 8.17 4.87
Table 3 Indicators of interaction and UI design. Indicator
Unit
Description
Step Right point Top point On-screen distance Avg touchpoint area Touchpoint text Touchpoint pic Touchpoint text pic
Step Deg Deg mm mm2 % % %
Touchpoint pic text
%
Touchpoint box
%
The number of steps in one touch-controlled task The side horizontal angle of the touchpoint on the right end The vertical angle of the touchpoint on the top The distance from the first touchpoint to the last touchpoint in a sequence The average area of all touchpoints in a touch-controlled task The percentage of ‘‘text type” touchpoints among all touchpoints in a touch-controlled task The percentage of ‘‘picture type” touchpoints among all touchpoints in a touch-controlled task The percentage of ‘‘text with small picture type” touchpoints among all touchpoints in a touch-controlled task (picture height is similar with text height) The percentage of ‘‘big picture with text type” touchpoints among all touchpoints in a touch-controlled task (picture height is obviously larger than text height) The percentage of ‘‘box type” touchpoints among all touchpoints in a touch-controlled task
are expected to exert stronger impact on driving distraction. Additionally, on-screen distance from the first touchpoint to the last touchpoint in a sequence indicates the workload of drivers’ arms. The UI design of HMI touch-controlled tasks refers to the icon size, color, pattern, etc. Most UI features, such as color blending and icon readability, are difficult to describe as quantified indicators. Thus, only two indicators are defined for UI design: average touchpoint area and touchpoint types. Touchpoint types are shown in Table 3 and examples are shown in Fig. 5.
Fig. 5. Samples of touchpoint types: the first row from left to right: text type, picture type, text with small picture type; the second row from left to right: big picture with text type, box type.
243
J. Ma et al. / Transportation Research Part F 69 (2020) 235–250
2.5. Model construction This study uses data mining techniques to model driving distraction indicators respectively to determine what the important impact factors are and to measure their impact. The performance of four algorithms are compared, including linear regression, random forest (RF), eXtreme Gradient Boosting (XGB), and Multi-Layer Perceptron (MLP). Among all algorithms, the four driving distraction indicators are dependent variables respectively, and are affected by target speed and the 17 other independent variables of HMI display size and position, as well as interaction and UI design indicators. This study used Python 2.7.0 to construct models for the four dependent variables for each of the four algorithms. There were 2310 effective samples processed by the algorithms. The samples involved 13 car models, 3 secondary tasks with 4 times of operation, and 24 drivers with 91% attendance rate. The invalid sample occurred due to obvious driving error that could be influenced by other factors rather than secondary tasks, and data collection error or missing, especially for the eye tracker. Because this study hypothesized individual differences among drivers, car models and secondary tasks are not significant, and effective samples were not divided into groups. 2.5.1. Linear regression The ordinary least square (OLS) is a common method of linear regression. It minimizes the sum of squared error (SSE) to find the best regression line for the data set. This is derived using matrix calculus, and it is computational efficient and easy to interpret. However, the OLS underperforms on the factors expected as nonlinear, which could be a limitation in this study. 2.5.2. RF model Tree-based models have gained popularity as a result of their ability to capture highly nonlinear and complex relationships between dependent and independent variables. The RF method, as proposed by Breiman (2001), is an ensemble learning approach based on the predictions of several decision trees. Instead of using a single decision tree, RF uses a group of decision trees from which a majority vote makes the predictions. In a similar method, called bagging, each tree uses all of the available variables. Each tree in the RF method only uses a few variables to minimize the correlations among different trees (Jahangiri, Rakha, & Dingus, 2016). 200 trees were constructed in this study. 2.5.3. XGB model Similar to RF, XGB is another tree-based algorithm. It integrates many tree models to form a strong classifier. CART (classification and regression tree) regression trees are used in XGB. The algorithm constantly splits features to grow new trees. Adding a tree is done by learning a new function to regress the residual of the last prediction. When k trees are trained, based on features of a given sample, there are scores given in every tree for each leaf node. The prediction result is sum of the scores in all trees. In this research, the maximum depth was 3, the learning rate was 0.1, and the quantity of trees was 200. 2.5.4. MLP model The MLP model is one of the typical ANN (artificial neural networks) algorithms and has shorter training time and accurate prediction (Ortiz-Catalan, Hkansson, & Brnemark, 2014). This feedforward model maps sets of input data onto a set of appropriate outputs. An MLP consists of multiple layers of nodes in a directed graph, with each layer fully connected to the next one. Except for the input nodes, each node is a neuron with a nonlinear activation function. The learning rate was 0.001 and the hidden layer size was 100 in this study. 3. Results for driving distraction indicators 3.1. Algorithms comparison The mean squared error was a common indicator used to evaluate the results of both the linear regression and the nonlinear algorithms (RF, XGB, MLP). The mean squared error for the four dependent variables and the four algorithms is shown in Table 4. Data was normalized before calculation. Because the weights among the four dependent variables were not
Table 4 Normalized mean Squared Error comparison among 4 algorithms. Algorithm
Linear regression RF XGB MLP
Dependent variables SpDev
LDSD
Dwl
MnGlc
1.542 1.076 1.245 1.783
1.056 1.096 1.258 1.165
1.592 0.860 0.915 0.934
2.950 0.971 1.066 1.444
Average
SD
1.785 1.001 1.121 1.331
0.704 0.094 0.141 0.317
244
J. Ma et al. / Transportation Research Part F 69 (2020) 235–250
defined in this study, their average values and standard deviation with identical weights were calculated as a reference for comparison. Although linear regression has the best interpretability among the four algorithms, its regression performance is the worst. In particular, mean squared error for MnGlc is up to 2.95. Among the nonlinear algorithms, RF performed better than MLP and XGB for all four independent variables. The smallest SD of RF indicated a stable performance for various dependent variables. Moreover, the interpretability of RF is stronger than XGB. For example, it is evident that more steps of secondary task results in a longer operation time and a longer Dwl. Thus, step should be one of the most important indicators for Dwl. The importance of step for Dwl is 0.232 in RF, but only 0.080 in XGB. Thus, RF was chosen as the most suitable algorithm for this study because of its smaller and stable mean squared error, as well as better interpretability. 3.2. Variable importance of RF The RF model is a nonlinear model combined with decision trees. Importance is a reference indicator to evaluate the contributions of each variable in the decision trees. The importance of variables is evaluated by data from a bag. The bag is a set of samples taken each time for training, and the data not taken is considered to be outside the bag. The importance evaluation method is as follows. First, use the out-of-bag data to calculate the error Ej of the decision tree j. Then randomly dis0
order the variable i out of the bag and calculate the error Eij of the decision tree j. The squared sum of the error differences is P 0 2 f i ¼ nj ðEij Ej Þ , and n is the number of decision trees and f i is the importance of variable i. The result of variable importance of the RF is shown in Table 5, and the importance figures larger than 0.07 are marked as bold. Among the 18 variables, target speed is an important impact factor for all distraction indicators. Other important variables with average importance larger than 0.07 are horizontal angle, step, right point, on-screen distance and avg touchpoint area. Based on the importance comparison, the variables of interaction and UI design play more significant roles than HMI display size and position. 3.3. Variable impact trend regression for a typical sample Among different car models, the impact trends from one variable could be different, because the other 17 variables changed across car models. To interpret the difference, four car models (PP-model, RM-model, RR-model, LO-model) were selected as an example, and had very different measured values for all 7 indicators of HMI display size and position. In the figures showing average touchpoint area impact trends for MnGlc (Fig. 6) and right point impact trends for LDSD (Fig. 7), the curves of car models had similar shapes, but their ordinate positions and detailed trends were different. This variability and similarity of impact trends also appeared for all other independent variables and distraction indicators. Due to the variability of the impact trends, the curve of a given car model cannot be used to represent other car models. Thus, one abstract dummy car model with a dummy secondary task was constructed as AVG-Model&Task (Table 6). All variables of HMI display size and position on the dummy car were set as the average values among the 13 tested car models, while all interaction and UI design variables of the dummy task were set as the average values of Task-NT, Task-CC and Task-AC among the 13 tested car models. Driving speed was the average speed of city and suburban driving. Due to the sim-
Table 5 Importance of driving distraction impact indicators of the RF Bold figures: importance larger than 0.07. Dependent variables
Scenario HMI display size and position
Interaction and UI design
Target speed Ball area Aspect ratio (width/tall) Center height Center side horizontal angle Center distance Depression angle Horizontal angle Step Right point Top point On-screen distance Avg touchpoint area Touchpoint text Touchpoint pic Touchpoint text pic Touchpoint pic text Touchpoint box
Average
SpDev
LDSD
Dwl
MnGlc
0.170 0.061 0.016 0.030 0.029 0.127 0.042 0.046 0.021 0.092 0.089 0.057 0.078 0.026 0.028 0.042 0.041 0.003
0.153 0.028 0.009 0.013 0.017 0.036 0.045 0.051 0.161 0.081 0.069 0.125 0.046 0.031 0.028 0.007 0.013 0.089
0.121 0.027 0.012 0.018 0.025 0.034 0.067 0.042 0.232 0.033 0.049 0.174 0.049 0.010 0.019 0.038 0.043 0.006
0.148 0.122 0.010 0.019 0.034 0.051 0.029 0.154 0.017 0.078 0.042 0.052 0.146 0.016 0.027 0.017 0.035 0.002
0.148 0.060 0.012 0.020 0.026 0.062 0.046 0.073 0.108 0.071 0.062 0.102 0.080 0.021 0.026 0.026 0.033 0.025
J. Ma et al. / Transportation Research Part F 69 (2020) 235–250
Fig. 6. Avg_touchpoint_area impact trends for MnGlc on 4 selected car models.
Fig. 7. Right_point impact trends for LDSD on 4 selected car models.
Table 6 AVG-Model&Task variable definition. Scenario
Target speed
52.5 km/h
HMI display size and position
Ball area Aspect ratio (width/tall) Center height Center side horizontal angle Center distance Depression angle Horizontal angle Step Right point Top point On-screen distance Avg touchpoint area Touchpoint text Touchpoint pic Touchpoint text pic Touchpoint pic text Touchpoint box
207.6 mm2 1.69 30.6 deg 42.1 deg 732.0 mm 26.2 deg 8.5 deg 2.07 42.6 deg 29.1 deg 445.6 mm 879.4 mm2 19.9% 39.8% 11.1% 23.9% 5.3%
Interaction and UI design
245
246
J. Ma et al. / Transportation Research Part F 69 (2020) 235–250
ilarity of the impact trends among car models, the AVG-Model&Task can be a typical sample to represent all tasks among the 13 car models. When one variable on AVG-Model&Task was analyzed, the other 17 variables were all fixed. The analysis range for the variable was from the minimum value to the maximum value of all tasks among the 13 tested car models. For SpDev and LDSD, the two driving performance indicators and variable impact trends (importance larger than 0.07 in Table 5) are shown in Fig. 8 and Fig. 9. Higher target speed resulted in higher distraction, because the driving attention requirement increased, and secondary tasks became more difficult. Very small center distance and very low top point led to a larger SpDev, possibly resulting from increased hand movement from the steering wheel and increased gaze movement from the road. An average touchpoint area smaller than 600 mm2 led to a larger SpDev, but a much bigger area did not have a continuously positive impact. A larger right point angle did not have a strong impact on SpDev, but LDSD increased hugely after 47 degrees, as many drivers are uncomfortable approaching the very far touchpoints. Increasing steps beyond 4 and having on-screen distance beyond 75 mm had significant impact on LDSD. As for Dwl and MnGlc, the two visual demand indicators, variable impact trends (importance larger than 0.07 in Table 5) are shown in Fig. 10 and Fig. 11. In contrast to driving performance, higher target speed resulted in lower Dwl and MnGlc, because drivers paid more attention to driving and tried to return their gaze to the road quicker. Similar to LDSD, Dwl also increased significantly when step and on-screen distance increased. An average touchpoint area smaller than 700 mm2 led to larger MnGlc, which is also similar to LDSD. Horizontal angle of display strongly impacted MnGlc, demonstrating that the angle should be larger than 7 degrees. Surprisingly, a larger display size with a larger ball area did not impact MnGlc significantly when it was larger than 120 deg2, which is about 8 in. for the HMI display size. The ideal right point for MnGlc was about 38–43 deg. This is very different from the ideal right point for LDSD. For any other secondary task on one specific car model, the RF algorithm can output another variable impact trend plot based on corresponding variable value input, and make suggestions to optimize HMI design. Because the RF is a nonlinear regression model, the variable impact trend outputs could be slightly different among various tasks and various car models. 4. Discussion In this study, the driving distraction effect of HMI displays was accessed by extensive independent variables. Data mining techniques were demonstrated as effective methods to regress and assess the driving distraction caused by HMI display. Among linear regression, RF, XGB, and MLP, the RF algorithm outperformed the other models in assessing driving distraction indicators due to its smaller and stable normalized mean squared errors. The RF and other tree-based algorithms were proven as an efficient method to detect and assess driving performance by previous studies. Atiquzzaman et al. (2018) used RF to detect texting and eating distractions based on driving behavior and eye tracking, with result of accuracy higher than 81% and having better robustness relative to than SVM and linear algorithms. Englund et al. (2016) selected RF to model bicyclists’ self-assessed visual distraction level with 71 variables from a simulator and a questionnaire to predict the distraction level. Because many extensive independent variables are involved in the RF algorithm, this study can discuss some complicated issues of practical HMI display design that many previous studies using fewer variables could not address. Four driving distraction indicators were defined in this study, including SpDev, LDSD, Dwl, and MnGlc. The SpDev is mainly affected by visual distraction and cognitive distraction. Visual distraction from seeing the real-time speed indicator on the curved screen prevented the driver from adjusting the accelerator. Cognitive distraction was created by drivers comparing real-time speeds and target speeds frequently, disrupting accurate control of the accelerator. Lane departure is not only affected by visual and cognitive distraction, but also by manual distraction (Owens, McLaughlin, & Sudweeks, 2011). When the driver takes one hand off the steering wheel for interactions, the slight change in body balance will cause the other hand to be unstable on the steering wheel. Dwl is the total time of visual distraction. MnGlc refers to the visual distraction time of each glance. MnGlc is always below 700 ms based on RF algorithm result (Fig. 11), which is significantly less than the longest single glance duration suggestion from the ISO standard of 1.5 s (2012) and from previous studies at 2 s
Fig. 8. Variable impact trends on AVG-Model&Task for SpDev.
J. Ma et al. / Transportation Research Part F 69 (2020) 235–250
247
Fig. 9. Variable impact trends on AVG-Model&Task for LDSD.
Fig. 10. Variable impact trends on AVG-Model&Task for Dwl.
Fig. 11. Variable impact trends on AVG-Model&Task for MnGlc.
(Simons-Morton et al., 2014, Caird et al., 2014). However, considering a car runs about 12 m every 700 ms at 60 km/h, shorter glances are still necessary for the improvement of safety. The choice of transverse or vertical HMI display layout is a widely discussed question. However, based on this research, it is not an important issue for driving distraction, because the importance of the differences in aspect ratio is no larger than 0.016 for any driving distraction indicator. Ball area is a more important variable, especially for MnGlc. A bigger display size is expected to result in longer MnGlc, because drivers find it difficult to search for a specific icon on a such large area. Although this trend is not significant to AVG-Model&Task in this study, it should be noted for other car models and tasks, because the RF result is nonlinear and the importance of ball area reached 0.122 for MnGlc. Gong and Ma (2018) reported
248
J. Ma et al. / Transportation Research Part F 69 (2020) 235–250
that finding a specific icon on a 17-in. display takes 20% more time than on a 12.3-in. display. The variables of HMI display positioning are not important for most driving distraction indicators, if the positioning does not bring limitations of interaction design. For example, to optimize SpDev, the importance of top point (interaction design) is much larger than center height (HMI display positioning), which are 0.089 and 0.030, respectively. Compared with HMI display size and position, interaction and UI design played more important roles in the impact of driving distraction in this study. The sum of variable importance is up to 0.65 for LDSD and Dwl. However, according to previous research, three kinds of UI designs for radio-channel shifting have no obvious difference in impact on maintaining speed and amount of lane departure (Mitsopoulos-Rubens, Trotter, & Lenné, 2011). One important reason is that secondary tasks were performed by a rotary knob and buttons, rather than a touch screen. If the UI menu structure changed, the operation area of the driver’s hand also changed accordingly when using a touch screen, which could affect a driver’s body balance and result in lane departure. However, using a knob and button controller, the operation area of the driver’s hand did not change. When assessing HMI secondary tasks with a touch screen, Li et al. (2017) reported that there was a different impact on task time and mean glance time between hierarchical UI layout and checkerboard UI layout. Thus, interaction and UI design are more important for touch screens than some other interaction modalities. Among the variables of interaction and UI design, step, on-screen distance, right point, and average touchpoint area are identified as very important variables. Previous studies did not deeply analyze this ranking because secondary tasks were always compared directly, and tasks were not decomposed into those abstract variables (Young & Salmon, 2012, Xie, Zhu, Guo, & Zhang, 2013, Kim & Song, 2014). Step is noticeably the most important impact factor, especially for LDSD and Dwl, because more steps mean a task becomes more complicated and produces a stronger distraction. The on-screen distance of finger movement is always positively correlated with step. However, even for a secondary task with more than 3 steps, the on-screen distance can still be minimized by arranging the sequential touch icons in similar positions, so LDSD and Dwl would decrease. The right point is an important variable in terms of both manual and visual distraction which many previous studies did consider. A right point larger than 47 degrees results in a huge LDSD increase on AVG-Model&Task, as many drivers have to move their entire body rightward to approach the very far touchpoint. However, a right point smaller than 37 degrees also led to larger MnGlc, because this area is close to the right edge of the steering wheel and is not easy to view clearly. Ma, Li, Gong, and Yu (2017) tested 3 mass produced car models with an eye tracker and drew the similar conclusion that the icons on the left parts of HMI displays always take more time for drivers to identify. A touchpoint area smaller than 700 mm2 leads to greater MnGlc, but it is unnecessary to make the touchpoint much bigger. The impact of touchpoint types is not obvious in this research. A possible reason is that touchpoint types are difficult to classify, which should be further explored in future researches. Target speed is an important impact factor on driving distraction. Higher speed results in obvious increases of SpDev and LDSD, but slight decreases of Dwl and MnGlc, because drivers paid more attention to driving and tried to return their gaze to the road quicker. The car speed varied in different driving scenarios. 51 km/h and 66 km/h were reported as the maximum comfort and safe speeds in daytime on city and suburban roads in the US, respectively (Mikoski, Zlupko, & Owens, 2019). Thus, for the secondary tasks that were frequently used in higher speeds, other important impact factors should be designed carefully to minimize the SpDev and LDSD increases, such as requiring fewer steps, less on-screen distance, and an appropriate right point. There are certain limitations to this study, which should be thoroughly investigated further. First, any secondary task in any car can be abstractly simplified as 18 variables for modeling, including target speed, 7 variables of HMI display size and position, and 10 variables of interaction and UI design. More variables should be evaluated and added to the algorithm to strengthen the model performance. This is especially true of UI design, as touchpoint area and touchpoint type classification were only partially considered in this study, while color blending, font size, icon identifiability, and some other necessary factors were not considered. Second, complex driving tasks should be added, including car following and use of the emergency brake. These driving tasks are effective in analyzing cognitive distraction. Third, other interaction modalities, including physical buttons and voice control, can be analyzed and compared against the use of a touch screen. Fourth, more mass produced cars should be tested to collect richer data to improve the algorithm performance. 5. Conclusion The secondary tasks on vehicle HMI displays result in driving distraction and affect road safety. This study presents a data mining technique to model four driving distraction indicators: speed deviation (SpDev), lane departure standard deviation (LDSD), dwell time (Dwl), and mean glance time (MnGlc). A real-car driving simulator was developed to test 13 mass produced cars and collect driving distraction data from 3 kinds of secondary tasks. The RF algorithm was verified as the best model, having good and stable regression performance as well as good interpretability. This research provides one assessment method for an HMI secondary task to reduce driving distraction in an early phase of product development. Even before the first prototype is produced, this method can assess the driving distraction resulting from an HMI central display and show the impact trend as a reference for optimization. With more relevant variables added and an increased test sample size, the performance of the RF algorithm will be continuously enhanced in future studies. According to the result of the RF algorithm, 6 variables have high average importance larger than 0.07: target speed, horizontal angle, step, right point, on-screen distance, and average touchpoint area. The importance of target speed is large for
J. Ma et al. / Transportation Research Part F 69 (2020) 235–250
249
all driving distraction indicators. For tasks frequently used at higher speed, it is advised to pay more attention to prevent a decrease in driving performance. The horizontal angle of the HMI display positioning should be more than 7 degrees toward the driver to decrease MnGlc. Reducing step and on-screen distance of finger movement is efficient for continuously lowering LDSD and Dwl. The position of the right point is another important variable, and should be between 37 and 47 degrees. A larger angle leads to bigger LDSD, while a smaller angle leads to bigger MnGlc. Average touchpoint area should be larger than 700 mm2 to maintain a lower SpDev and MnGlc. CRediT authorship contribution statement Jun Ma: Conceptualization, Funding acquisition, Supervision. Zaiyan Gong: Conceptualization, Methodology, Writing original draft, Writing - review & editing. Jianjie Tan: Data curation, Software. Qianwen Zhang: Investigation, Visualization. Yuanyang Zuo: Investigation. Acknowledgement This study was partially supported by Banma Technologies Co., Ltd. We thank the participants of the study for participating. References Alliance of Automobile Manufacturers (AAM), Driver Focus-Telematics Working Group, 2002. Statement of Principles, Criteria and Verification Procedures on Driver Interactions with Advanced In-Vehicle Information and Communication Systems—Version 2.0. Atiquzzaman, M., Qi, Y., & Fries, R. (2018). Real-time detection of drivers’ texting and eating behavior based on vehicle dynamics. Transportation Research Part F: Traffic Psychology and Behaviour, 58, 594–604. Breiman, L. (2001). Random forests?. Machine Learning, 45(1), 5–32. Caird, J. K., Johnston, K. A., Willness, C. R., Asbridge, M., & Steel, P. (2014). A meta-analysis of the effects of texting on driving. Accident Analysis and Prevention, 71, 311–318. Crundall, E., Large, D. R., & Burnett, G. (2016). A driving simulator study to explore the effects of text size on the visual demand of in-vehicle displays. Displays, 43, 23–29. Choudhary, P. (2017). Mobile phone use during driving: Effects on speed reduction and effectiveness of compensatory behaviour. Accident Analysis & Prevention, 106(June), 370–378. Dols, J. F., Molina, J., Camacho, F. J., Marín-Morales, J., Pérez-Zuriaga, A. M., & Garcia, A. (2016). Design and development of driving simulator scenarios for road validation studies. Transportation Research Procedia, 18(June), 289–296. Dingus, T. A., Klauer, S. G., Neale, V. L., Petersen, A., Lee, S. E., Sudweeks, J., ... Bucher, C. (2006). The 100-car naturalistic driving study phase II–Results of the 100-car field experiment (Report No. DOT HS 810 593). Washington: National Highway Traffic Safety Admin. (NHTSA). Englund, C., Nilsson, M., & Voronov, A. (2016). The application of data mining techniques to model visual distraction of bicyclists. Expert Systems with Applications, 52, 99–107. Engström, J., Markkula, G., Victor, T., & Merat, N. (2017). Effects of cognitive load on driving performance: The cognitive control hypothesis. Human Factors, 59(5), 734–764. García, M., Badre, A. N., & Stasko, J. T. (1994). Development and validation of icons varying in their abstractness. Interacting with Computers, 6(2), 191–211. Gaspar, J. G., Ward, N., Neider, M. B., Crowell, J., Carbonari, R., Kaczmarski, H., ... Loschky, L. C. (2016). Measuring the useful field of view during simulated driving with gaze-contingent displays. Human Factors, 58(4), 630–641. Gellatly, A. W., & Kleiss, J. A. (2000). Visual attention demand evaluation of conventional and multifunction in-vehicle information systems. Proceedings of the Human Factors and Ergonomics Society Annual Meeting, 44(20), 282–285. Gong, Z., & Ma, J. (2018). Vehicle central display layout evaluation based on driver distraction simulator test. 2018 JSAE Annual Congress (Spring) Proceedings, 23-25 May 2018, Yokohama, Japan, No. 20185265. Hofmann, H., Tobisch, V., Ehrlich, U., & Berton, A. (2015). Evaluation of speech-based HMI concepts for information exchange tasks: A driving simulator study. Computer Speech & Language, 33(1), 109–135. Huang, D. L., Patrick Rau, P. L., & Liu, Y. (2009). Effects of font size, display resolution and task type on reading Chinese fonts from mobile devices. International Journal of Industrial Ergonomics, 39(1), 81–89. Huemer, A. K., Schumacher, M., Mennecke, M., & Vollrath, M. (2018). Systematic review of observational studies on secondary task engagement while driving. Accident Analysis and Prevention, 119(May), 225–236. International Organization for Standardization (2002). ISO 15005: 2002 Road vehicles - ergonomic aspects of transport information and control systems dialogue management principles and compliance procedures. Jahangiri, A., Rakha, H., & Dingus, T. A. (2016). Red-light running violation prediction using observational and simulator data. Accident Analysis and Prevention, 96, 316–328. Jin, L., Xian, H., Niu, Q., & Bie, J. (2015). Research on safety evaluation model for in-vehicle secondary task driving. Accident Analysis and Prevention, 81, 243–250. Kountouriotis, G. K., Spyridakos, P., Carsten, O. M. J., & Merat, N. (2016). Identifying cognitive distraction using steering wheel reversal rates. Accident Analysis and Prevention, 96, 39–45. Klauer, S. G., Dingus, T. A., Neale, V. L., Sudweeks, J. D., & Ramsey, D. J. (2006). The impact of driver inattention on near-crash/crash risk: An analysis using the 100-car naturalistic driving study data. Report No. DOT HS 810 594. Washington: U.S. Department of Transportation National Highway Traffic Safety Administration. Kim, H., & Song, H. (2014). Evaluation of the safety and usability of touch gestures in operating in-vehicle information systems with visual occlusion. Applied Ergonomics, 45(3), 789–798. Kim, H., Kwon, S., Heo, J., Lee, H., & Chung, M. K. (2014). The effect of touch-key size on the usability of In-Vehicle Information Systems and driving safety during simulated driving. Applied Ergonomics, 45(3), 379–388. Li, P., Merat, N., Zheng, Z., Markkula, G., Li, Y., & Wang, Y. (2018). Does cognitive distraction improve or degrade lane keeping performance? Analysis of timeto-line crossing safety margins. Transportation Research Part F: Traffic Psychology and Behaviour, 57, 48–58. Li, R., Chen, Y. V., Sha, C., & Lu, Z. (2017). Effects of interface layout on the usability of In-Vehicle Information Systems and driving safety. Displays, 49, 124–132. Liang, Y., Reyes, M. L., & Lee, J. D. (2007). Real-time detection of driver cognitive distraction using support vector machines. IEEE Transactions on Intelligent Transportation Systems, 8(2), 340–350.
250
J. Ma et al. / Transportation Research Part F 69 (2020) 235–250
Liang, Y., & Lee, J. D. (2010). Combining cognitive and visual distraction: Less than the sum of its parts. Accident Analysis and Prevention, 42(3), 881–890. Libby, D., Chaparro, A., & He, J. (2013). Distracted while driving: A comparison of the effects of texting and talking on a cell phone. In Proceedings of the human factors and ergonomics society (pp. 1874–1878). Ma, J., Li, J., Gong, Z., & Yu, J. (2017). Impact of in-vehicle touchscreen size on visual demand and usability. SAE Technical Papers, September, 1298. Ma, Y., Shi, Y., Fu, R., & Guo, Y. (2015). Impact of driver’s distracted driving time on vehicle lane departure. Journal of Jilin University (Engineering and Technology Edition), 45(4), 1095–1101. McEvoy, S. P., Stevenson, M. R., & Woodward, M. (2006). The impact of driver distraction on road safety: Results from a representative survey in two Australian states. Injury Prevention, 12(4), 242–247. Metz, B., Landau, A., & Just, M. (2014). Frequency of secondary tasks in driving - Results from naturalistic driving data. Safety Science, 68, 195–203. Metz, B., Schoch, S., Just, M., & Kuhn, F. (2014). How do drivers interact with navigation systems in real life conditions? Transportation Research Part F: Traffic Psychology and Behaviour, 24, 146–157. Mikoski, P., Zlupko, G., & Owens, D. A. (2019). Drivers’ assessments of the risks of distraction, poor visibility at night, and safety-related behaviors of themselves and other drivers. Transportation Research Part F: Traffic Psychology and Behaviour, 62, 416–434. Mitsopoulos-Rubens, E., Trotter, M. J., & Lenné, M. G. (2011). Effects on driving performance of interacting with an in-vehicle music player: A comparison of three interface layout concepts for information presentation. Applied Ergonomics, 42(4), 583–591. Muhrer, E., & Vollrath, M. (2011). The effect of visual and cognitive distraction on driver’s anticipation in a simulated car following scenario. Transportation Research Part F: Traffic Psychology and Behaviour, 14, 555–566. Nevile, M., & Haddington, P. (2010). In-car distractions and their impact on driving activities. Department of Infrastructure and Transport, Road Safety Grant, Report RSGR-2010-001. National Technique Committee of Auto Standardization (2018). China Automotive Driving Cycle Part I: Light-duty vehicles (draft). Beijing, China. Née, M., Contrand, B., Orriols, L., Gil-Jardiné, C., Galéra, C., & Lagarde, E. (2019). Road safety and distraction, results from a responsibility case-control study among a sample of road users interviewed at the emergency room. Accident Analysis and Prevention, 122(May 2018), 19–24. Ojados Gonzalez, D., Martin-Gorriz, B., Ibarra Berrocal, I., Macian Morales, A., Adolfo Salcedo, G., & Miguel Hernandez, B. (2017). Development and assessment of a tractor driving simulator with immersive virtual reality for training to avoid occupational hazards. Computers and Electronics in Agriculture, 143(July), 111–118. Olaverri-Monreal, C., Hasan, A. E., Bulut, J., Körber, M., & Bengler, K. (2014). Impact of in-vehicle displays location preferences on drivers’ performance and gaze. IEEE Transactions on Intelligent Transportation Systems, 15(4), 1770–1780. Ortiz-Catalan, M., Hkansson, B., & Brnemark, R. (2014). Real-time and simultaneous control of artificial limbs based on pattern recognition algorithms. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 22(4), 756–764. Owens, J. M., McLaughlin, S. B., & Sudweeks, J. (2011). Driver performance while text messaging using handheld and in-vehicle systems. Accident Analysis and Prevention, 43, 939–947. Peng, Y., Boyle, L., & Lee, J. (2014). Reading, typing, and driving: How interactions with in-vehicle systems degrade driving performance. Transportation Research Part F: Traffic Psychology and Behaviour, 27, 182–191. Risto, M., & Martens, M. H. (2014). Driver headway choice: A comparison between driving simulator and real-road driving. Transportation Research Part F: Traffic Psychology and Behaviour, 25(PART A), 1–9. Rosenthal, T. J. (1999). STISIM Drive User’s Manual. Hawthorne, CA: Systems Technology Inc.. Sayer, J. R., Devonshire, J. M., & Flannagan, C. A. (2005). The effects of secondary tasks on naturalistic driving performance (Report No. UMTRI-2005-29). Ann Arbor, Michigan: The University of Michigan. Simons-Morton, B. G., Guo, F., Klauer, S. G., Ehsani, J. P., & Pradhan, A. K. (2014). Keep your eyes on the road: Young driver crash risk increases according to duration of distraction. Journal of Adolescent Health, 54(5 SUPPL.), S61–S67. Silvennoinen, J. M., Kujala, T., & Jokinen, J. P. P. (2017). Semantic distance as a critical factor in icon design for in-car infotainment systems. Applied Ergonomics, 65, 369–381. Strasburger, H., & Rentschler, I. (1996). Contrast-dependent dissociation of visual recognition and detection fields. European Journal of Neuroscience, 8, 1787–1791. Strayer, D. L., Watson, J. M., & Drews, F. A. (2011). Cognitive distraction while multitasking in the automobile. In B. Ross (Ed.), The psychology of learning and motivation: Advances in research and theory, vol. 54 (pp. 29–58). San Diego, CA: Elsevier Academic Press. Strayer, D. L., Turrill, J., Cooper, J. M., Coleman, J. R., Medeiros-Ward, N., & Biondi, F. (2015). Assessing cognitive distraction in the automobile. Human Factors, 57(8), 1300–1324. Tijerina, L., Parmer, E., & Goodman, M. J. (1998). Driver workload assessment of route guidance system destination entry while driving: A test track study. In Proc. of the 5th ITS World Congr. ITS. Seoul, CD-ROM. Wittmann, M., Kiss, M., Gugg, P., Steffen, A., Fink, M., Pöppel, E., et al (2006). Effects of display position of a visual in-vehicle task on simulated driving. Applied Ergonomics, 37(2), 187–199. Xie, C., Zhu, T., Guo, C., & Zhang, Y. (2013). Measuring IVIS impact to driver by on-road test and simulator experiment. Procedia - Social and Behavioral Sciences, 96(Cictp), 1566–1577. Young, K. L., & Salmon, P. M. (2012). Examining the relationship between driver distraction and driving errors: A discussion of theory, studies and methods. Safety Science, 50(2), 165–174. Young, K. L., Salmon, P. M., & Cornelissen, M. (2013). Missing links? The effects of distraction on driver situation awareness. Safety Science, 56, 36–43. Zhang, Y., Owechko, Y., & Zhang, J. (2004). Driver cognitive workload estimation: A data-driven perspective. In Proceedings of the 7th international IEEE conference on intelligent transportation systems, 3–6 October 2004 (pp. 642–647).