17-exploring of field experiment radio tasks with naturalistic driving data- a comparison with shrp2 nest_h

1、Adjunct Proceedings of the 8th International Conference on Automotive User Interfaces and Interactive Vehicular Applications (AutomotiveUI 16), October 2426, 2016, Ann Arbor, MI, US.Exploring Generalizability of Field Experiment Radio Tasks with Naturalistic Driving Data: A Comparison with SHRP2 NES

2、TAbstractIn this study we compare glance patterns observed in field experiment driving studies with glance patterns observed in the naturalistic SHRP 2 NEST database. We describe the methodology used to identify appropriate naturalistic epochs and to prepare glances for comparison to field experimen

3、t data, and graphically show points of similarity and points of contrast between the two sets of data.Overall, glance patterns observed in field experiments appear to hold in naturalistic data, with a few caveats. Using naturalistic glance data to validate experimentally-acquired glance data appears

4、 to show promise and provides confidence for conclusions drawn from behaviors observed in controlled on-road driving scenarios.Sean SeamanTouchstone Evaluations, Inc. 440 BurroughsDetroit, MI Bruce MehlerMIT AgeLab & New England University Transportation Center 77 Mas

5、sachusetts Ave, E40-215 Cambridge, MA Joonbum LeeMIT AgeLab & New England University Transportation Center 77 Massachusetts Ave, E40-215 Cambridge, MA Bobbie Seppelt Touchstone Evaluations, Inc. 440 BurroughsDetroit, MI Linda AngellTouchst

6、one Evaluations, Inc. 440 BurroughsDetroit, MI Bryan ReimerMIT AgeLab & New England University Transportation Center 77 Massachusetts Ave, E40-209 Cambridge, MAAuthor KeywordsDriver distraction; naturalistic driving study; driver glance behavior.Permissio

7、n to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page Copyrights for third- party c

8、omponents of this work must be honored For all other uses, contact the Owner/AuthorCopyright is held by the owner/author(s)AutomotiveUI 16 Adjunct, October 24-26, 2016, Ann Arbor, MI, USA ACM 978-1-4503-4654-2/16/10http:/dx doi org/10 1145/3004323 3004350ACM Classification KeywordsH.1.2. User-machin

9、e systems: human factorsWork-in-Progress and InteractiveDemosAutomotiveUI Adjunct Proceedings 16, Ann Arbor, MI, USAIntroductionField experimentsdriving experiments conducted on open roadsare the closest studies to real world driving that can be conducted in an experimental fashion; track, simulator

10、, and bench studies are each progressively farther removed from significant aspects of the natural driving environment, while naturalistic and field operational studies are observational, where very little should be manipulated. However, because field experiments are still largely contrived, thereMe

11、thodThis study analyzed data collected across five field experimental driving studies conducted by the MIT AgeLab and the SHRP2 NEST naturalistic driving database.Field Experiment DataComplete methodological details of the fieldexperiments can be foundeveral technical reports5,6,7,8,9. All of the st

12、udies included radio interactions that were adapted from the visual-manual preset selection and the radio tuning tasks employed in the Crash Avoidance Metrics Partnership (CAMP) Driver Workload Metrics project 2 The implementation of multi-step radio tuning was compatible with the NHTSA radio refere

13、nce task 10. Baseline driving (“justdriving”) was included as a reference period for comparative analysis with secondary task periods within each study.behaviors 1. The goal of this study is to demonstrate the feasibility of comparing behaviors from naturalistic driving to a set of field experiments

14、, specifically focused on secondary task performance and glances.Numerous field experiments have examined secondary task behavior during driving 2; a number of these have also specifically looked at glances during drivingwhile engagingecondary tasks e.g., 3. In thisNEST DataThe NEST database contast

15、udy, we analyzed glance behavior from thedetailed trip and glanceNaturalistic Engagementecondary Tasks (NEST)information about 1,180 trips (944 Baseline epochs and 236 Safety Critical Event SCE epochs: 85 Near-Crash epochs, and 151 Crash epochs). There are 198,159 samples of glance coding in the NES

16、T Baseline database; 25,299 in the Near-Crash database; and 44,816 samples in the Crash database. Glance data is provided at a sampling rate of 10 Hz. The average number of samples across Baseline epochs is 209.91 samples per epoch (20.991 s), ranging from a minimum of 97 (9.7 s) to a maximum of 211

17、 (21.1 s), with a median of 210 (21.0 s).database 4, which was collected as a part of the Strategic Highway Research Program (SHRP 2), to compare to glance behavior collected across multiple field studies conducted by MIT AgeLab.In this paper, we lay out the methodology used to prepare naturalistic

18、glance data for comparison to field data, describe some behavioral differences between the two sets of data, and graphically show points of similarity and difference between naturalistic and field glances during radio interactions and baseline driving.In order to match the glance coding scheme of th

19、e studies conducted by MIT AgeLab, many of the NEST112Field Experiment Method:VehiclesThe five on-road studies utilized six standard production vehicles: (a) 2010 Lincoln MKS (with Ford SYNC system), (b) 2013 Chevrolet Equinox (with MyLink system), (c) 2013 Volvo XC60 (with Sensus system),(d) 2014 C

20、hevrolet Impala (with MyLink system), (e) 2014 Mercedes CLA (with COMMAND infotainment system), and (f) 2015 Toyota Corolla (with Entune Premium Audio with navigation).ParticipantsRecruitment drew from the greater Boston area using online and newspaper advertisements. Four age groups (20-24, 25-39,

21、40-54,and 55 over) were formed to generally conform to National Highway Traffic SafetyAdministrations recommendations 9; 18-19 year-olds were not recruited. A total of 364 drivers data were analyzed in the main analysis.Work-in-Progress and InteractiveDemosAutomotiveUI Adjunct Proceedings 16, Ann Ar

22、bor, MI, USAAOIs were recoded and the dataset filtered accordingly. Frames coded as “Transition” were recoded as the AOI of the next frame, following ISO 15007 IS 11 when the next frame was known. Furthermore, AOIs were reassigned new codes, and sometimes grouped with other AOIs to match the glance

23、coding scheme used in the AgeLab studies. Our analyses comparing field experiments to naturalistic driving studies focus on theFrom this filtered set of epochs we identified sub- epochs in which the driver interacted with the center stack. These periods are coded as beginning when a driver first beg

24、an to move a hand toward the center stack, or initiated a glance to the center stack immediately before reaching; they are coded as ending when the driver ended contact with the center stack. If a second interaction with the center stack was initiated within five seconds of breaking contact, these t

25、wo periods of interaction were grouped together. Based on these criteria, 68 Baseline, 6 Near-Crash, and 17 Crash epochs contained radio adjustment periods. These periods ranged in duration from 2 (0.2 s) to 200 samples (20.0 s), with a mean duration of 38.75 samples (3.875 s). These epochs were fur

26、ther filtered to remove those that contained any fully or partially overlapping additional visual-manual tasks, including eating or drinking, holding or manipulating an object, talking on a handheld cell phone, or reaching for something that was not the center stack, producing 57 Baseline, 5 Near-Cr

27、ash, and 16 Crash epochs.AgeLab AOIs: Center Stack,trument Cluster, Left(Window), Right (Window), Rearview Mirror, and Road. Coded samples that were excluded from analyses when the glance location was unknown.Secondary task activity in the NEST database was evaluated using a combination of trip-leve

28、l task summary data and sample-by-sample detailed secondary task analysis. For the first ten and second ten seconds of each Baseline, Near-Crash, and Crash epoch, six potential secondary tasks were coded in NEST. For the task of “Adjusting/monitoring radio”, there were 75 Baseline epochs, 7 Near-Cra

29、sh epochs, and 17 Crash epochs. Of the 75 Baseline epochs, four also contained climate control adjustment during the first twenty seconds; these were removed from the set. “No Task” epochs represent those that contained no secondary task activity across the first twenty seconds of each epoch a total

30、 of 198 Baseline epochs, 7 Near-Crash epochs and 3 Crash epochs. The relatively lownumber of “No Task” SCE epochs in the NEST database is to be expected given that it was designed specifically to study secondary task activities. The 10 “No Task” SCE epochs are those that contain post-precipitating-

31、event secondary task activity, and were therefore removed.Table 1 shows average task duration for epoch type for radio interactions using both sets.Table 1. Radio task length by type of epoch.113Mean Task Dur. (s)NBaseline3.82957Near-Crash1.5765Crash7.68116Precipitating EventsFor SCE epochs, we trim

32、med the number of samples at the point at which the precipitating event was coded as beginning: this is the point, coded for each SCE, at which the events surrounding the crash or near-crash began. In order to exclude glance behavior potentially elicited by the presence of contextual factors of an S

33、CE, we limited our glance analysis to coded frames that occurred from the onset of each SCE epoch to the last frame immediately before the initiation of the precipitating event. This yielded 16,782 Near-Crash and 29,514 Crash samples. The average number of samples across Near-Crash epochs was197.43

34、(19.743 s), ranging from a minimum of 57 (5.7 s) to a maximum of 238 (23.8 s), with a median of 200. The average across Crash epochs was 195.56 (19.556 s), ranging from a minimum of 37 (3.7 s) to a maximum of 201 (20.1 s), with a median of 200 (20.0 s).Work-in-Progress and InteractiveDemosAutomotive

35、UI Adjunct Proceedings 16, Ann Arbor, MI, USAResults and DiscussionWe chose to use only the Baseline data for comparing the demand characteristics of naturalistic driving with and without radio task engagement with field experiment data for three reasons: First, because it appears that Crash epochs

36、contain potentially different radio tuning tasks than Baseline epochs (given the substantially different task durations); second, because the Near-Crashes have so few available epochs and appear to be substantially different from Crashes; third, because the goal of this (in progress) effort is to va

37、lidate experiment data obtained in the field by using naturalistic data, not to establish a relationship between experimentally obtained glance data and crash risk.AgeLab AOIs, including percentage of glance time and percentage of glance frequency (Figure 1). Glances to AOIs look remarkably similar,

38、 with some contrastsappearing in thetrument Cluster location (moreseen in AgeLab) and Left Window (more seen in NEST). This distinction could be due to driving environments, with AgeLab studies primarily conducted on highways and many NEST epochs occurring elsewhere.Glances were also tallied to all

39、of the AOIs not labeled “Road” combined (“off-road glances”), following the conventions of 13. We applied two NHTSA off-road glance metrics: (a) mean single off-road glance duration, and (b) percentage of long duration off-road glances (note that “total off-road glance time” was not applied given th

40、at task length was not comparable), plotted in Figure 2. Overall, naturalistic glances appear to be shorter than field experiment glances, with aUsing the Baseline radio and no-task data, we computed glance statistics two ways: first, to theFigure 1. Glance time (%) and frequency (%) by location for

41、 field experiments and NEST. Note that CS indicates CenterStack; IC indicatesMirror.trument Cluster; L indicates Left Window; RM indicates Rearview Mirror; and R indicates Right114Task BoundariesBecause this is naturalistic data, task beginnings and ends sometimes occurred at the boundary of an epoc

42、h, as was the case for 35 of the 78 available epochs (44.9%),with 19 epochs (24.4%) starting (potentially) in media res and 17 epochs (21.8%) ending during a radio task; one radio task subtended the entire epoch. This is important for the purpose of reporting task time because artificially truncated

43、 task times may lower task length statistics.However, by excluding epochs with radio tasks that extended to epoch boundaries, we actually reduced average task duration. It appears that eliminating tasks that extend to epoch boundaries unfairly penalizes longer tasks, which has a much greater effect

44、on task duration than truncating task durations.Work-in-Progress and InteractiveDemosAutomotiveUI Adjunct Proceedings 16, Ann Arbor, MI,USAcombined when evaluating radio interaction demand during real world driving, as these sets appear to consist of different blends of radio tasks, supporting the c

45、onclusions of 12: SCEs are heterogeneous. Radio interaction durations may be a boundary representing that heterogeneity.AcknowledgmentSupport for this work was provided by the US DOTs Region I New England University Transportation Center at MIT and the Toyota Class Action Settlement Safety Research

46、and Education Program. The views and conclusions being expressed are those of the authors, and have not been sponsored, approved, or endorsed by Toyota or plaintiffs class counsel.References1Angell, L.S. (2014). An opportunity for convergence? Understanding the prevalence and risk of distracted driv

47、ing through the use of crash databases, crash investigations, and other approaches. AAAM, 58, 40-59. /pmc/articles/PMC4 001669/Angell, L., Auflick, J., Autria, P.A., et al. (2006). Driver Workload Metrics Project Task 2 Final Report. NHTSA, Washington, D.C. http:/www.nhtsa.

48、gov/DOT/NHTSA/NRD/Multim edia/PDFs/Crash%20Avoidance/Driver%20Distr action/Driver%20Workload%20Metrics%20Final%20Report.pdfVictor, T. W., Harbluk, J. L., & Engstrm, J. A. (2005). Sensitivity of eye-movement measures to in-vehicle task difficulty. Transportation Research Part F: Traffic Psychology an

49、d Behaviour, 8, 167-190. /10.1016/j.trf.2005.04.0142Figure 2. Mean single off-road glance duration (s) and proportion of long glances (%) for field and NEST data.lower proportion of long glances except in comparisonto AgeLabs baseline data.3The difference in task times for Near-Crash

50、 and Crash epochs (1.1 s vs. 6.2 s, respectively; see sidebar)suggests that these categories of SCEs shouldnt be115Naturalistic Task DurationsThere were considerable differences in task length by epoch type in the NEST data (see Table 1), with Crash radio tasks subtending over twice as many seconds

51、as Baseline, and around five times as many seconds as Near-Crash radio tasks.These different types of epochs may be reflecting different types of radio tasks. Perez et al. 11 found a median infotainment interaction duration of 2.2 seconds in naturalistic infotainment use; in NEST, the median is 1.7

52、seconds for Baseline epochs, suggesting that these interactions are likely primarily the most common types of radio interactions: power and volume operations. The median Crash epoch task was6.0 seconds, which suggests that this set of tasks probably includes other task types, such as searching for s

53、omething new to which to listen.Work-in-Progress and InteractiveDemosAutomotiveUI Adjunct Proceedings 16, Ann Arbor, MI, USA4Owens, J. M., Angell, L., Hankey, J. M., Foley, J., & Ebe, K. (2015). Creation of the Naturalistic Engagementecondary Tasks (NEST) distracted driving dataset. Journal of Safet

54、y Research, 54, e29-36. /10.1016/j.jsr.2015.07.001Mehler, B., Kidd, D., Reimer, B., Reagan, I., Dobres, J., & McCartt, A. (2016). Multi-modal assessment of on-road demand of voice and manual phone calling and voice navigation entry across two embedded vehicle systems.Ergonomics, 59,

55、344367. /10.1080/00140139.2015.1081 412Mehler, B., Reimer, B., Dobres, J., & Coughlin,J.F. (2015). Assessing the Demands of Voice Based In-Vehicle Interfaces - Phase II Experiment 3 - 2015 Toyota Corolla. Massachusettstitute of Technology, 2015. /files/Publication

56、s/Mehler_ etal_MIT_AgeLab_Techical%20Report_2015- 6A_Impala.pdfMehler, B., Reimer, B., Dobres, J., et al. (2014). Further Evaluation of the Effects of a Production Level “Voice-Command” Interface on Driver Behavior: Replication and a Consideration of the Significance of Training Method. MIT AgeLab, Cambridge, MA. /files/MIT_AgeLab_Technic al_Report_2014-