28
1 Quantifying and recommending seat belt reminder timing using naturalistic driving video data *Daniel V. McGehee 1,2 , Cheryl A. Roe 1 , Pranaykumar Kasarla 1,2 , Chao Wang 1,2 University of Iowa 1 National Advanced Driving Simulator 2 Department of Industrial and Systems Engineering *Corresponding Author: Daniel V. McGehee, National Advanced Driving Simulator, University of Iowa, 2401 Oakdale Blvd, Iowa City, Iowa 52245; [email protected]; Tel +1-319-430-3052

Quantifying and recommending seat belt reminder timing

  • Upload
    others

  • View
    3

  • Download
    0

Embed Size (px)

Citation preview

1

Quantifying and recommending seat belt reminder timing using naturalistic driving video data

*Daniel V. McGehee1,2, Cheryl A. Roe1, Pranaykumar Kasarla1,2, Chao Wang1,2

University of Iowa

1National Advanced Driving Simulator

2Department of Industrial and Systems Engineering

*Corresponding Author: Daniel V. McGehee, National Advanced Driving Simulator, University of Iowa,

2401 Oakdale Blvd, Iowa City, Iowa 52245; [email protected]; Tel +1-319-430-3052

2

Abstract

Introduction To better understand the timing of when people buckle their seat belt, an analysis of a

naturalistic driving study was used. The study provided a unique perspective inside of the vehicle where

the entire seat belt was visible from the time the driver entered the vehicle to one minute of driving

forward or 32 kph. Method: Thirteen of these drivers were identified for a seat belt sequencing which

identified the points when the vehicle was put into ignition, shifted, when vehicle movement began, and

when the seat belt was buckled. The speed at belt closure was also identified. The timing from ignition to

buckle and to shifting into forward gear were examined in order to identify the speed and appropriate

timing for seat belt reminders. Results: The data shows that drivers were buckled in over 92% of the

3,102 drives. In 70% of those total drives, the drivers were buckled before the vehicle began movement.

Of greater interest for seat belt reminders/interlocks are those drives when drivers buckle after movement.

When considering time from ignition to seat belt closure, the mean was 27.5 seconds. Because higher

speeds are typically reached when travelling forward rather than reverse, it was important to know the

time duration from shifting into drive to buckling. Conclusions: Since higher speeds lead to an increased

risk of injury and/or death, a recommendation of a 30 second time from forward shift and a 25 kph

threshold for reminder systems should be implemented. The regression analysis also validates that most

of the predicted seat belt buckling times are within 30s. Practical Applications: This would reduce

perception of nuisance alerts and protect the driver from higher speed unbuckled crashes. The seat belt

buckling time prediction model also demonstrates good potential for developing tailored buckling

warning system for different drivers.

Keywords: seat belt; nuisance alarm; safety interlock; buckle timing

3

Quantifying and recommending seat belt reminder timing using naturalistic driving video data

1. Introduction

A continued increase in seat belt use is fundamental to creating safer roads and reducing the

societal burden of crash injuries and fatalities. It is estimated that in 2020, seat belt use for front seat

occupants reached 90.3% in the United States (National Center for Statistics and Analysis, 2021), a 9.1%

gain from 2006. Furthermore, seat belts saved 14,955 passenger vehicle occupants aged five and older in

the U.S. during 2017 (NHTSA, 2020). While seat belt use continues to slowly rise, further increasing belt

use is still the most effective way to reduce fatalities and minimize injuries in motor vehicle crashes

(Campbell, 1987; Evans, 1986; Sato, 1987), as millions still do not buckle up. In 2018, nearly 23% of all

passenger vehicle occupants who died in crashes were unrestrained at the time of the crash (NHTSA,

2020).

Over the years, seat belt reminder systems have played a vital role in increasing seat belt use.

While auditory or visual warnings have been shown to be effective in getting occupants to buckle their

seat belt (Fildes, Fitzharris, Koppel, Vulcan, & Brooks, 2003; Krafft, Kullgren, Lie, & Tingvall, 2006;

Williams, Wells, & Farmer, 2002), they haven’t always proven successful. At issue is the timing. If the

reminder alerts the drivers too soon after engaging forward motion, it could be viewed as a nuisance

(Regan et al., 2006). This could result in manually disconnecting the reminder alert. Reminding a driver

too late may lead to increased risk of injury as they may already be traveling at a higher speed without

being restrained. The change of velocity is captured by delta-v, which is directly related to the severity of

injury and increases the risk of fatality (Richards, 2010). Most seriously, a single car crashing into a tree

or infrastructure would decelerate from some initial velocity to zero in a short time, and even relatively

low speed crashes can be fatal. According to 2018 Fatal Analysis Reporting System (FARS) data, 34.4%

of single car crashes occurred at or below 64 kph, and 52.8% of single car crashes occurred below 80 kph

(NHTSA, 2018). Averaging speeds for the top 17 crash types for vehicle-to-vehicle (V2V) applications,

48% of them occur at or below 56 kph, while 45% occur between 56 and 88 kph (Najm et al., 2013).

4

Three of the 17 crash types (Running stop sign, Straight crossing path, Non-signal, and Turn at Non

signal) occurred at less than 56 kph more than 60% of the time.

Another type of reminder related to reminder timing is the seat belt safety interlocks. The timing

of this type of system is crucial to prevent it from becoming a nuisance. Interlocks are designed to limit

some of the systems operating in the vehicle, such as the radio or other entertainment systems, accelerator

pedal (providing resistance), ignition, or shifting capabilities (Mizruchi, 1996). In the 1970s, the National

Highway Traffic Safety Administration (NHTSA) mandated seat belt ignition interlocks, prohibiting

vehicles from being started unless the seat belt was buckled. These interlocks were widely unaccepted by

the general public and were quickly removed from vehicles. Negative perceptions of these interlocks still

linger today (Eby, Molnar, Kostyniuk, Shope, & Miller, 2004; Kidd, McCartt, & Oesch, 2014).

Furthermore, it is widely understood that wearing a seat belt correctly saves lives and mitigates

injuries during crashes. However, a considerable percentage of the population still does not to wear seat

belts. Understanding why people choose not to wear their seat belt is only part of the issue. Reminders

have been effective, but safety interlocks have not shown to have the same success. Understanding how to

effectively design safety interlocks so that they will be accepted could play a role in getting more people

to buckle up.

To remedy these issues, belting behavior is often studied through crash data, surveys,

observation, or with interviews. More recently, naturalistic studies have been used to investigate seat belt

behavior as they provide a unique perspective, capturing drivers in more everyday environments.

Although researchers in previous studies have identified seat belt users, there is no standard for how this

is done. A seat belt user has been identified by frequency (Kidd et al., 2014; Reagan, McClafferty, Berlin,

& Hankey, 2013) or by the time of buckling. The timing is often associated to different actions taking

place in the vehicle, such as seat belt reminders (Bao, Xiong, Buonarosa, & Sayer, 2015) or shifting into

gear (Malenfant & Van Houten, 2008). Although the research includes some of the actions going on in

the vehicle, there is no research that provides a detailed analysis of the seat belt timing sequence from

5

before ignition through the start of driving (McGehee & Roe, 2017). Our study fills this gap and makes

recommendations on seatbelt timing.

2. Methods

In order to investigate seat belt buckling behavior, data from a previous naturalistic driving study

(NDS) with 30 drivers were examined in the seat belt buckling context. Observations during that study,

which was conducted in suburban Iowa, focused on the participants’ entry into their vehicle until

acceleration to a driving speed as well as the parking maneuver at the end of each drive. Each participant

had a four-camera system (interior cabin view, forward view, and two foot-well cameras) along with an

accelerometer, two small infrared lights (one to light the interior cabin and one in the footwell), and GPS

installed into their vehicle for about one month. These data were collected as part of another study

(McGehee et al., 2016). Video and audio data were recorded at the start (door opening to 32 kph or 1

minute) and end of each drive (last minute), and at accelerometer threshold settings of 0.5g or greater.

These settings were chosen based on the guidance and experience of the manufacturer, as well as on those

used in other naturalistic driving studies. Dingus et al. (2006); McGehee, Raby, Carney, Lee, & Reyes

(2007); and Carney, McGehee, Lee, Reyes, & Raby (2010) used -0.5g as the threshold for defining hard

braking and ±0.4g and ±0.55g as the threshold for defining rapid steering maneuvers (McGehee & Roe,

2017). This analysis used the audio, video, and GPS data collected during the NDS to identify seat belt

behaviors.

2.1 Participants

A total of 30 participants from two age groups completed the NDS in Iowa City, Iowa: ten

between 18 to 35 years and twenty over 65 years of age. Although seat belt use was determined for all 30

participants, only 13 of them (4 younger (2 males and 2 females) and 9 older (6 males and 3 females)

were selected for the seat belt timing sequence that will be used to determine the timing of seat belt

reminder system recommendations (see selection criteria in procedure section). Participants were licensed

drivers with normal color vision and a visual acuity of at least 20/40. They had to make at least one round

trip per day (e.g., to the store and back) and drive a 1996 or newer vehicle model.

6

2.2 Procedure

The video data from the NDS provides a camera view of the interior of the cabin that allows for

the identification of drivers’ seat belt buckling behavior as well as the sequences that drivers take when

buckling (Figure 1).

Figure 1. Interior cabin view

Seat belt use was classified into four groups: driver puts seat belt on before car is moving, driver puts

seat belt on during start-up maneuver, driver puts seat belt on during driving, and no seat belt

throughout the drive. To better understand the details regarding buckling after movement, the drivers

that buckled after vehicle movement at least 25% of the time and had at least ten drives in this

category were identified for the seat belt sequencing. This sequencing identified various time points

from entering the vehicle through seat belt closure. Using this criteria, 13 drivers were identified for

which ten drives were randomly selected for the time series for buckling. Time points at ignition, gear

shifts, grabbing the seat belt, and seat belt closure were collected. Gear selection and speed at closure

were also recorded. Seat belt closure was defined as the point at which the male end of the seat belt

connected to the female end. Coding was completed by two trained coders with discrepancies being

reviewed and mediated between them. Time points were identified from the video and/or audio

available from the interior camera, and speed was identified from the GPS device installed in the

vehicle.

7

Descriptive statistics were calculated for seat belt use, timing to closure from ignition, shifting

into initial gear, and shifting into forward gear as well as speed at buckle. Furthermore, data cleaning and

model selection criteria techniques were applied on the data and regression analyses were performed to

build models for determining the drivers’ seat belt bucking behaviors and buckling time.

3. Results

Out of 3,102 drives, drivers were buckled 92% of the time. The individual participant breakdown

showed buckling rates between 74% and 100%, with 21 out of the 30 drivers buckling more than 90% of

the time. Results show that buckling prior to movement occurred in 70% of the drives. Of the drives when

buckling took place prior to movement, there were:

• 11% who buckled prior to ignition,

• 56% who buckled after ignition and prior to shifting, and

• 6% who buckled after both ignition and shifting but prior to vehicle movement.

An additional 107 drives (3%) were confirmed as buckling prior to vehicle movement, however due to

video sampling errors the timing of seat belt closure was unknown. Figure 2 shows the breakdown of the

buckling sequence of the drives.

304(11%)

1603(56%)

167(6%)

673(24%)

0

200

400

600

800

1000

1200

1400

1600

1800

Buckled prior to ignition Buckled after ignition,prior to shift

Buckled after ignition andshift

Buckled after ignition andshfit

Nu

mb

er

of

dri

ves

Ignition Gear Shift Vehicle Movement

8

Figure 2. Summary of buckle sequence

In order to recommend the appropriate requirements for seat belt reminder warnings, drives when

drivers buckled after movement began were identified. The time duration from ignition to seat belt

closure showed a mean of 27.5 seconds with a range from 4.6 to 86.6 seconds. This duration could lead to

nuisance alarms because this duration would include any actions the driver performs before the act of

driving begins. If considering the act of driving to be when the vehicle was put into gear, the amount of

time that elapsed from initial shift to seat belt closure was 22 seconds with a range of 2.8 to 92.4 seconds.

The majority of drives in this category began with the initial shift position into reverse, however 61% of

the drives had belt closure during forward movement. The time duration from “shift to drive to buckle”

showed a mean of 16.2 seconds with a range of 3.7 to 92.4 seconds. Table 1 shows the details of the

various time durations to buckle. Ninety-one percent of buckling after forward movement took place

under 30 seconds, which can be seen in the cumulative distribution in Figure 3.

The mean speed at buckling when travelling forward was 15.3 kph (4.2 m/s) with a range of 7.9

to 50.7 kph (2.2 to 14.1 m/s). Twelve of the thirteen drivers mean speeds were below 21 kph (5.8 m/s).

The cumulative distribution shown in Figure 3 shows that in 90% of the drives, drivers were buckled at

speeds lower than 28.7 kph (8.0 m/s). Note that the horizontal axis represents both the duration (in unit

second) and the speed (in unit kph), where the two units are aligned to share the same scale.

Table 1. Time duration to buckle

Number of

drives with

duration

Mean

Duration

(s)

Median

Duration

(s)

Standard

Deviation

(s)

Min

(s)

Max

(s)

Ignition to buckle 103 27.5 23.7 18.0 4.6 86.6

Initial shift to buckle 110 22.0 19.7 15.4 2.8 92.4

Shift to drive to

buckle 85 16.2 13.3 12.2 3.7 92.4

9

Figure 3. Distribution and comparison of “time for forward shift to buckle” and “speed at buckle”

Identifying speed relative to how much time has elapsed from “forward shift to buckle” provides

an opportunity to see how speed changes as time increases. Error! Reference source not found. 4

identifies each drive with valid speed and time duration (forward shift to buckle). There were only 63

points identified. Speeds of 0 kph are included in the figure. A gray box was added to the figure to

illustrate the drives that would fall inside the recommendations (25 kph and 30 seconds from forward

shift) discussed in the discussion section. Out of our analysis, 14 drives (22%) fell outside of one of these

recommendations: of these, six fell outside due to buckle taking place at speeds greater than 25 kph,

seven after a time duration (forward shift to buckle) greater than 30 seconds, and one drive fell outside the

range for both. For the drives when drivers buckled more than 30 seconds after forward shift, three were

buckled within 31 seconds.

0

10

20

30

40

50

60

70

80

90

100

0 10 20 30 40 50 60

Per

cen

tage

KPH at buckle

Duration from forward shiftto buckle (s)

90

%

Time (s)/Speed (kph)

10

Figure 4. Time from ‘forward shift to buckle’ and speed

3.1 Regression Analysis

To better facilitate the analysis of the factors that influence drivers’ seat belt buckling behavior

and timing, regression was used. Table 2 highlights the four categories used in the analysis, which are:

• Driver puts seat belt on before car is moving: If participant buckled seat belt prior to any

movement of the vehicle, in reverse or forward direction.

• Driver puts seat belt on during start-up maneuver: If participant put the seat belt on during the

start-up sequence. The start of the drive sequence was defined as the moment the car was put into

gear to the point when the vehicle was completely in the roadway and traveling forward.

• Driver puts seat belt on during driving: If participant put on seat belt after the start-up sequence

had been completed.

• No seat belt throughout the drive: If participant did not have the seat belt on at any time during

either the start-up or parking sequence. As previously discussed, the video system did not

continuously record the entire drive. If the video available showed no belt worn at start-up and

0

5

10

15

20

25

30

35

40

45

50

55

0 10 20 30 40 50 60 70 80 90 100

Spee

d a

t b

uck

le (

kph

)

Time from forward shift to buckle (s)

11

parking, these drives were “assumed no seat belt.” However, it is possible that these drivers could

have buckled and unbuckled during the section of the video that wasn’t available.

Table 2. Seat belt usage – participant drive breakdown

In this analysis, the following research questions were set to investigate the seat belt buckling

behavior and buckling time of the drivers by analyzing the data provided:

1. What is an appropriate model that can represent the driver’s seat belt buckling behaviors?

2. What is an appropriate model that can represent the driver’s seat belt buckling time during

driving?

3. What are the significant variables (contributors) which have influence on the driver’s seat belt

buckling behavior and buckling time during driving?

3.2 Regression Modeling

Both qualitative and quantitative regression analysis was performed on the data, and significant

input variables were identified to build models for determining the seat belt bucking behaviors and

buckling time. Note that in the case of buckling time, we only focus on determining the driver’s seat belt

12

buckling time when the driver falls in the Puts seat belt during driving category. This is because the

“during driving” period is when the seat belt functions most significantly. The Puts seat belt during

driving corresponds to the “vehicle movement” portion in Figure 2.

Selecting the regression model depends on the type of the data to be analyzed. Multinomial

Logistic Regression (Hilbe, 2009), denoted as Model (a), was used for the research question 1 and 3, as

the response variable – “buckling behavior” has four categories. Whereas, Linear Regression, denoted as

Model (b), was used for the research question 2 and 3, as the response variable – “buckle time” is a

continuous variable.

Model - a) Input variables:

For all of the input variables, their range, categories, and how they are represented in the

multinomial logistic regression model are shown in Table 3 below:

Table 3. Input variables for buckling behavior

Model - a) Response variable:

The multinomial logistic regression model gives the log odds of a new category to a reference

category of the response variable. In this case, it is 𝑙𝑜𝑔 (𝑃𝑖

𝑃4) for all i = 1,2,3 where P1, P2, P3, P4 are the

probabilities of each categorical membership of buckling behavior.

P1: Probability of being in category No seat belt throughout drive

P2: Probability of being in category Puts seat belt during driving

P3: Probability of being in category Puts seat belt during startup maneuver

P4: Probability of being in category Puts seat belt on before car is moving

Model - a) Multinomial logistic regression model equation:

13

As we have four categories in the response variable, multinomial logistic regression estimates

parameter values for 𝛽0, 𝛽1, ⋯ , 𝛽7 via maximum likelihood method of all other three categories with

respective to the reference category. Applying multinomial logistic regression to our data, by taking Puts

seat belt on before car is moving category as a reference level for response variable, results in the

following equation for all 𝑖 = 1, 2, 3.

𝑙𝑜𝑔 (𝑃𝑖

𝑃4) = 𝛽0

𝑖 + 𝛽1𝑖 𝑋1 + 𝛽2

𝑖 𝑋2 + 𝛽3𝑖 𝑋3 + 𝛽4

𝑖 𝑋4 + 𝛽51𝑖 𝟏(𝑋5 = 𝑋51) +

𝛽52𝑖 𝟏(𝑋5 = 𝑋52) + 𝛽61

𝑖 𝟏(𝑋6 = 𝑋61) + 𝛽62𝑖 𝟏(𝑋6 = 𝑋62) +

𝛽63𝑖 𝟏(𝑋6 = 𝑋63) + 𝛽7

𝑖 𝑋7

(1)

Note: 𝟏(𝒙) is an indicator function, where 𝟏(𝒙) = {1 𝑖𝑓 𝑥 𝑖𝑠 𝑡𝑟𝑢𝑒0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

Model - b) Input variables:

For all the input variables, their range, categories, and how they are represented in the linear

regression model are shown in Table 4 below. Weather was not involved in the regression model related

to buckling time due to the model diagnostic that removes outliers and missing values.

Table 4. Input variables for buckling time

Model - b) Response variable:

The response variable “buckle time” is the time taken by the driver to buckle the seat belt from

the moment the car started moving forward after the startup maneuver (e.g., reversing).

14

Model - b) Linear regression model equation:

Multiple models were generated with all the linear terms and two-way interactions terms of the

input variables. The generalization of these models is as follows, where ‘Y’ is the buckle time.

𝐹𝑢𝑙𝑙 𝑀𝑜𝑑𝑒𝑙 𝑌 ~ 𝑋1 + 𝑋2 + 𝑋3 + 𝑋4 + 𝑋5 + 𝑋6 + ∑ 𝑋𝑖𝑋𝑗

1≤𝑖<𝑗≤6

(2)

3.3. Regression analysis results and discussion

Model - a) Multinomial logistic regression model results:

Model selection criteria: Model selection is an important part of any statistical analysis. It is used

to balance the model complexity and prediction accuracy. The Akaike Information Criterion (AIC)

(Sakamoto, Ishiguro, & Kitagawa, 1986) was applied on our models to select the most representative

input variables.

Model results from AIC:

By applying AIC model selection criteria, preferred model with the minimum AIC value of

2882.37 has included all the input variables. The multinomial logistic regression is then applied to

provide detailed analysis results.

To make it easier to understand the results generated by the model, we provide two examples for

explaining the coefficients: continuous input variable (Height) and categorical input variable (Rain/Mist

category of Weather). All other coefficients of continuous input variables can be explained like the

coefficient of the Height. Similarly, all other coefficients of categorical input variables can be explained

like the coefficient of Rain/Mist. Note: An asterisk (*) indicates the significant variables (p-value < 0.05)

in the model equations.

Scenario 1:

𝑙𝑜𝑔 (𝑃1

𝑃4) = −32.079∗ − 0.010𝑋1 − 0.085𝑋2 + 0.111𝑋3

∗ + 0.011𝑋4 +

13.165 𝟏(𝑋5 = 𝑋51)∗ + 11.588 𝟏(𝑋5 = 𝑋52)∗ + 11.297 𝟏(𝑋6 = 𝑋61)∗

+10.523 𝟏(𝑋6 = 𝑋62)∗ + 11.334 𝟏(𝑋6 = 𝑋63)∗ − 0.008𝑋7∗

(3)

15

𝛽31 = 0.111: If a subject were to increase his Height by one inch, the multinomial log-odds of

preferring category No seat belt throughout drive to category Puts seat belt on before car is moving

would be expected to increase by 0.111 units while holding all other variables in the model constant.

𝛽521 = 11.588: If the weather were to change from Snow/sleet/hail to Rain/Mist, the multinomial

log-odds of preferring category No seat belt throughout drive to category Puts seat belt on before car is

moving would be expected to increase by 11.588 units while holding all other variables in the model

constant.

All other coefficients from the model can be explained like the above explained coefficients.

Similarly, the results from scenario 2 and 3 can be explained except that scenario 2 gives the log-odds of

preferring category Puts seat belt during driving to category Puts seat belt on before car is moving and

scenario 3 gives the log-odds of preferring category Puts seat belt during startup maneuver to category

Puts seat belt on before car is moving.

Scenario 2:

𝑙𝑜𝑔 (𝑃2

𝑃4) = −10.017∗ + 0.022𝑋1

∗ + 1.024𝑋2∗ + 0.020𝑋3 + 0.021𝑋4

∗ +

1.590 𝟏(𝑋5 = 𝑋51)∗ + 1.015 𝟏(𝑋5 = 𝑋52)∗ − 0.072 𝟏(𝑋6 = 𝑋61) + +0.325 𝟏(𝑋6 = 𝑋62) + 0.433 𝟏(𝑋6 = 𝑋63) − 0.0002𝑋7

(4)

Scenario 3:

𝑙𝑜𝑔 (𝑃3

𝑃4) = −0.171∗ − 0.010𝑋1 + 1.062𝑋2

∗ + 0.003𝑋3 − 0.009𝑋4∗ +

0.839 𝟏(𝑋5 = 𝑋51)∗ + 0.649 𝟏(𝑋5 = 𝑋52)∗ + 0.224 𝟏(𝑋6 = 𝑋61) − 0.160 𝟏(𝑋6 = 𝑋62) − 0.401 𝟏(𝑋6 = 𝑋63) − 0.0002𝑋7

(5)

The analysis results in this model provide answers for the research questions 1 and 3: The

multinomial logistic regression model can well represent and identify the significant variables for driver’s

seat belt buckling behaviors. The model interpretations in Equations 3–5 provide quantitative impacts on

the buckling behaviors when changing each variable, and the asterisk (*) in Equations 3–5 identify the

significant variables that have impacts on the buckling behaviors. The significance is determined by

16

choosing variable with the p-value smaller than 0.05 in the analysis of variance in multinomial logistic

regression.

Model - b) Linear regression model results:

The performance of all the models from Equation 2, were evaluated by calculating the Root Mean

Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE)

values. To reduce variability, ten-fold cross-validation was performed using ten different partitions

(seeds) and then an average of the results was taken for each model. In addition, outliers with a Cook’s

distance greater than 0.5 were identified in every round and were removed from the data. The lower the

error values the better the performance of the model. Hence, the model with minimum mean RMSE,

MAE, and MAPE values was selected.

In addition, Box Cox transformation (Sakia, 1992) was applied on this model and Box Cox

transformed model was generated. Box Cox transformation is a way to relax the “normality” assumption

in response variables. The core of the Box Cox transformation is an exponent, lambda (𝜆), which varies

from -5 to 5. All values of λ are considered and the optimal value for the data is selected. The “optimal

value” is the one which results in the best approximation of a normal distribution curve.

AIC model selection criteria was applied on both these models. Resulted Normal model with a

minimum AIC value of 189.95 and Box Cox model with a minimum AIC value of -253.86 have included

only the Age and Sex input variables. These models will be referred as Normal final model and Box Cox

final model in this report.

In order to check the performance of these two models, twenty-fold cross-validation was

performed using twenty different partitions (seeds) and then average of RMSE, MAE, and MAPE values

were calculated for each model. Normal final model has the average error values of RMSE = 4.87, MAE

= 3.86, and MAPE = 27.86, and Box Cox final model has the average error values of RMSE = 4.89, MAE

= 3.65, and MAPE = 24.45. Minimum RMSE, MAE, and MAPE values were used in determining the best

partition (seed) for both Normal final model and Box Cox final model. These best seed models’ results are

interpreted and explained in the following sections.

17

Normal final model results:

The best partition (seed) of the Normal final model, with minimum error values of RMSE = 3.40,

MAE = 2.60, and MAPE = 17.33 gave the following results.

𝑌 = 14.163∗ + 0.044𝑋1 − 4.072𝑋2 (6)

𝛽1 = 0.044: If a subject were to increase his Age by one year while holding all other variables in

the model constant, the (buckle time) time taken by the driver to buckle the seat belt from the moment the

car started moving forward after the startup maneuver would be expected to increase by 0.044 seconds.

Box Cox final model results:

The best partition (seed) of the Box Cox final model, with minimum error values of RMSE =

3.72, MAE = 2.77, and MAPE = 16.90 gave the following results.

𝑌𝑏𝑜𝑥 𝑐𝑜𝑥 = 1.451∗ + 0.002𝑋1 − 0.091𝑋2∗ (7)

𝛽2 = −0.091: If the Sex were to change from female to male, while holding all other variables in

the model constant, the 𝑌𝑏𝑜𝑥 𝑐𝑜𝑥 would be expected to decrease by 0.091 units.

Also, for the better visualization of the model evaluation metrics, RMSE, MAE, and MAPE

values from the twenty-seeds of both “Normal” to “Box Cox” models were plotted. The box plot in

Figure 5, contains the twenty seeds error values from both these models.

18

Figure 5. Normal and Box Cox models twenty-fold evaluation metrics boxplots

Figure 6 is for the visualization of the Normal and Box Cox final models’ performance on the best

partition of the data, after eliminating outliers. The red circles represent the actual buckle time from the

training dataset on which the models were trained, the blue circles represent the actual buckle time from

the validation dataset on which the model performance was tested. The black asterisks represent the

predicted buckle time by the Normal model, and the purple square represent the predicted buckle time by

the Box Cox model. Every blue circle will have a corresponding black asterisk and purple square, as the

performance of models was tested only on the validation dataset.

Figure 6 shows most of the predictions approximate the real buckling time well, which

demonstrates the capability of the model for predicting the buckling time. This also reveals the potential

for developing a tailored seat belt buckling warning system for drivers with different characteristics, e.g.,

Sex and Age.

19

Figure. 6 Buckle time predictions of Normal and Box Cox models

The analysis results in this model provide answers for the research questions 2 and 3: The linear

regression model can quantify the driver’s seat belt buckling time, and the AIC helps select two important

variables, i.e., Age and Sex, that contribute to the seat belt buckling time.

4. Discussion

As these results show, this study provided new insights into driver buckling behavior from

vehicle entry, ignition, general shifting, forward shifting, and unbuckling at the end of the drive. Of

greatest interest was the timing from the forward shift point and the speed at seat belt closure. These data

can help better inform seat belt reminder designers of the most appropriate time and speed to alert a

driver.

Among 3,102 drives, drivers in our sample were buckled at a very high rate: 92.0%. This number

is nearly identical to 2017 driver buckling rates cited by the Iowa DOT, when 91.4% of drivers were

observed to be buckled (Larson, Fox, & Berg, 2017). In 70% of the drives for the current study, the

drivers were buckled prior to vehicle movement. For this study, buckling took place after ignition in 89%

of the drives, 70% of drivers buckled prior to vehicle movement, and about 31% were after gear selection.

20

This is in contrast to Malenfant and Van Houten (2008), who found that only 68.7% of drivers were

buckled after ignition and about 20% buckled after placing the vehicle in gear. The primary difference

between these studies is that Malenfant and Van Houten (2008) looked at between-subject observations,

and this study looked at within-subjects across repeated drives. This University of Iowa study was able to

capture more sustainable behavioral trends within drivers.

In terms of belt timing, initial shift to buckling duration showed a mean of 22.0 seconds with a

range from 2.8 to 92.4 seconds. Bao, Funkhouser, Sayer, and Toyoda (2016) found that their “late

bucklers” had a mean duration of 35.1 seconds. The difference between Bao et al. (2016) and the current

analysis could be explained in that Bao et al. (2016) had a larger sample size covering a wider age range.

It is also possible that because Bao et al. eliminated shorter drives with slower speeds, that could impact

the time duration observed. In our analysis, we observed that those that buckled after movement often put

the vehicle into reverse initially, however, did not buckle until traveling forward. As buckling in reverse

was rare in our sample and buckling occurred at lower speeds, we concentrated our focus on shifting into

forward gear to buckle rather than just the initial shift. The duration from shifting into forward gear to

buckling showed a mean of 16.2 seconds with a range from 3.7 to 92.4 seconds. When considering the

shift into forward gear, a majority of the drives (81%) were buckled under 20 seconds and, if increasing to

30 seconds, 91% of the sample would be buckled.

Speed at buckling while traveling forward showed a mean of 15 kph; 76% of the drives were

below 25 kph, with a range was from 2 kph to 51 kph. It must be kept in mind that there were seven out

of 130 drives (5%) when speed was unknown due to unavailable video data, and the impact these drives

had on the speed cannot be determined. It is possible the drivers were traveling at lower or higher speeds

than 15 kph, or even stopped when buckle occurred. In this study, drivers were buckled by the time they

reached 28 kph in 90% of the drives. Given that the age of the drivers for this sample was skewed toward

older drivers (nine aged > 65 and four 18 – 35), this may explain slower speeds. One would expect that,

among a larger sample of young and middle-aged drivers, the speeds might be higher.

21

We use the multinomial logistic regression to model the various contributors’ impact on the

drivers’ seat belt buckling behavior, and linear regression to model the various contributors’ impact on the

drivers’ seat belt buckling time. AIC model selection criteria was applied to select the most appropriate

contributors to construct the regression models. The key results from the analysis in Section 3.3 are

summarized below.

In the seat belt behavior analysis, the results in Equations 3-5 demonstrate that:

1) In a comparison between No seat belt throughout drive and Puts seat belt on before car is

moving, the Age, Height, Weight, Duration, Environment, and Weather are significant

contributors;

2) In a comparison between Puts seat belt during driving and Puts seatbelt on before car is moving,

the Age, Sex, Weight, Environment, and Weather are significant contributors;

3) In a comparison between Puts seat belt during startup maneuver and Puts seat belt on before car

is moving, the Age, Sex, Weight, Environment, and Weather are significant contributors.

For predicting the buckle time, the results in Equations 6-7 demonstrate that Age and Sex input

variables are selected as contributors after applying AIC model selection criteria. Meanwhile, the

regression analysis also shows that Sex is a significant input variable in the buckle time

prediction model.

As described, the logistic regression analysis performed to understand the driver’s seat belt

buckling behaviors provided some interesting results (Equations 3 to 5). The input variables Weather,

Height, and Weight have similar impact on the driver’s seat belt buckling behaviors. For example, if the

Weather changes from snow/sleet/hail to clear/cloudy or rain/mist, the log odds of drivers’ changing their

seat belt behavior from Puts seat belt on before car is moving to other three categories (Puts seat belt

during startup maneuver, Puts seat belt during driving, or No seat belt throughout drive) increase. This

implies that drivers tend to wear the seat belt before the car is moving when the weather condition has

snow/sleet/hail. Similarly, if the driver’s height increases the log odds of drivers changing their seat belt

22

behavior from Puts seat belt on before car is moving to one of the other three categories, this implies that

as their Height increases the drivers tend to not wear the seat belt before the car moves. These analysis

results provide solutions for the research questions 1 and 2.

In the linear regression analysis, just the two input variables Age and Sex were selected to predict

the buckle time. From the results obtained, we can infer that the buckle time increases with the increase in

Age, and the buckle time decreases when the driver’s Sex changes from female to male. The regression

analysis resulted a mean buckle time of 14 seconds, which is similar to Malenfant and Van Houten (2008)

who found a mean buckle time of 12 seconds. Moreover, mean buckle time of 14 seconds is less than the

mean buckle time of 22 seconds obtained from the descriptive statistics. This is because some buckling

data are treated as outliers in the regression analysis, which also reveals most of the outliers are for those

taking a long time to buckle. Thus, the regression analysis provides some complement to the observations

in descriptive statistics.

Finally, the results from the regression analysis demonstrate that the models developed can be

used to predict the driver’s seat belt buckling behaviors and buckling time from the driver’s demographics

and drive conditions. This prediction helps to provide customized reminders to individual drivers by

predicting their seat belt behaviors and buckle times. This can reduce the perception of nuisance alerts

and protects the drivers from higher speed unbuckled crashes.

4.1 Practical applications

The data presented in this paper and the primary takeaway suggest that a 30 second threshold

from forward shift and 25 kph could be used in seat belt reminder systems. This recommendation would

reduce perception of nuisance alerts and protect the driver from higher-speed unbuckled crashes. Using

such timing would ensure that drivers are buckled before they reach hire speeds where injuries become

more severe.

5. Limitations

Naturalistic driving studies have a number of advantages and disadvantages. While we have real world

observations, NDS data have no experimental control. What you see is what you get. NDS studies suffer

23

from this limitation in general and as a result, in order to compare like conditions, some drivers must be

excluded. Sample size, age of participants can therefore be an issue, too. However, this is balanced by the

sheer number of events captured. This study was also not specifically designed to examine seat belt

behavior, so more detailed controls were not possible. Although the video data provided the interior view

of the vehicle to identify the key points associated with buckling the seat belt, the data doesn’t allow for

studying unbuckling during the middle of the drive. We only sampled approximately the first and last

minute of driving in this NDS study, so we were not able to capture the speed or time duration at buckle

or how the placement of the shoulder belt changes after the first minute. We had seven instances out of

130 (5%) when drivers buckled after the video stopped recording. The impact of these seven drives is

unknown as we have no way of knowing what speed drivers were traveling at or how soon after that one

minute they buckled. When considering seat belt placement, this study did not examine how the belt

shifted when buckling prior to movement. It is possible that the shifting of the belt may take place when

drivers are turning their torso and traveling in reverse which is missing from our analysis. Additionally,

our sample size was small, with a majority of the population being 65 years of age and older. An

additional examination of a sample including a wider age range of drivers, as well as additional data

throughout the drive, would be useful in helping to quantify the timing and speeds necessary for the seat

belt safety interlocks, as well as identifying seat belt shifting.

24

Funding

This work was supported by the Toyota Collaborative Safety Research Center (CSRC) and Safety

Research using Simulation University Transportation Center (SAFER-SIM) [U.S. Department of

Transportation’s University Transportation Centers Program, grant number 69A3551747131]. However,

the U.S. government assumes no liability for the contents or use thereof.

25

References

Bao, S., Funkhouser, D.S., Sayer, S.R., & Toyoda, H. (2016). An examination and categorization of

unbelted driving behavior: Insights from naturalistic driving data analysis. Submitted by UMTRI to

TRB 2017 Annual Meeting.

Bao, S., Xiong, H., Buonarosa, M.L., & Sayer, J.R. (2015). Using naturalistic driving data to examine

drivers’ seatbelt use behavior: Comparison between teens and adults. Journal of Safety Research,

54, 69–73. https://doi.org/10.1016/j.jsr.2015.06.006

Campbell, B.J. (1987). Safety Belt Injury Reduction Related to Crash Severity and Front Seated Position.

The Journal of Trauma, 27(7), 733–739. https://doi.org/10.1097/00005373-198707000-00007

Carney, C., McGehee, D.V., Lee, J.D., Reyes, M.L., & Raby, M. (2010). Using an event-triggered video

intervention system to expand the supervised learning of newly licensed adolescent drivers.

American Journal of Public Health, 100(6), 1101–1106. https://doi.org/10.2105/AJPH.2009.165829

Dingus, T.A., Klauer, S.G., Neale, V.L., Petersen, A., Lee, S.E., Sudweeks, J., Perez, M.A., Hankey, J.,

Ramsey, D., & Gupta, S. (2006). The 100-car naturalistic driving study, Phase II-results of the 100-

car field experiment. No. DOT-HS-810-593. United States Department of Transportation. National

Highway Traffic Safety Administration.

Eby, D.W., Molnar, L.J., Kostyniuk, L.P., Shope, J.T., & Miller, L.L. (2004). Developing an Optimal In-

Vehicle Safety Belt Promotion System. Report No. UMTRI-2004-29. In University of Michigan,

Ann Arbor, Transportation Research Institute.

Evans, L. (1986). The effectiveness of safety belts in preventing fatalities. Accident Analysis &

Prevention, 18(3), 229–241. https://doi.org/10.1016/0001-4575(86)90007-2

Fildes, B., Fitzharris, M., Koppel, S., Vulcan, P., & Brooks, C. (2003). Benefits of seat belt reminder

systems. Annual Proceedings/Association for the Advancement of Automotive Medicine, 47, 253.

Hilbe, J.M. (2009). Logistic regression models. New York: CRC press.

26

Kidd, D.G., McCartt, A.T., & Oesch, N.J. (2014). Attitudes toward seat belt use and in-vehicle

technologies for encouraging belt use. Traffic Injury Prevention, 15(1), 10–17.

https://doi.org/10.1080/15389588.2013.792111

Krafft, M., Kullgren, A., Lie, A., & Tingvall, C. (2006). The use of seat belts in cars with smart seat belt

reminders—results of an observational study. Traffic Injury Prevention, 7(2), 125–129.

https://doi.org/10.1080/15389580500509278

Larson, J.M., Fox, J.R., & Berg, E. (2017). Iowa Seat Belt Use Survey 2017 Data Collection Methodology

report. http://www.iowadot.gov/mvd/stats/survey17.pdf

Malenfant, J.E.L., & Van Houten, R. (2008). Observations of how drivers fasten their seatbelts in relation

to various startup tasks. Accident Analysis & Prevention, 40(1), 309–314.

https://doi.org/10.1016/j.aap.2007.06.009

McGehee, D.V., & Roe, C. (2017). An Investigational Analysis of Driver Seat Belt Use, Timing, and

Position in Naturalistic Driving. Toyota Technical Report, CSRC. Ann Arobor, MI.

McGehee, D.V., Raby, M., Carney, C., Lee, J.D., & Reyes, M.L. (2007). Extending parental mentoring

using an event-triggered video intervention in rural teen drivers. Journal of Safety Research, 38(2),

215–227. https://doi.org/10.1016/j.jsr.2007.02.009

McGehee, D.V., Roe, C.A., Boyle, L.N., Wu, Y., Ebe, K., Foley, J., & Angell, L. (2016). The wagging

foot of uncertainty: data collection and reduction methods for examining foot pedal behavior in

naturalistic driving. SAE International Journal of Transportation Safety, 4(2), 289–294.

https://doi.org/10.4271/2016-01-1526

Mizruchi, M.S. (1996). What do interlocks do? An analysis, critique, and assessment of research on

interlocking directorates. Annual Review of Sociology, 22(1), 271–298.

https://doi.org/10.1146/annurev.soc.22.1.271

27

Najm, W.G., Ranganathan, R., Srinivasan, G., Smith, J.D., Toma, S., Swanson, E., & Burgett, A. (2013).

Description of light-vehicle pre-crash scenarios for safety applications based on vehicle-to-vehicle

communications (Final Report No. DOT HS 811 731). United States. National Highway Traffic

Safety Administration.

National Center for Statistics and Analysis. (2021). Seat belt use in 2020 – Overall results. (Traffic Safety

Facts Research Note. Report No. DOT HS 813 072). National Highway Traffic Safety

Administration.

NHTSA. (2018). Vehicles Involved by Speed Limit and Crash Type - State : USA, Year : 2018.

https://www-fars.nhtsa.dot.gov/Vehicles/VehiclesLocation.aspx

NHTSA. (2020). 2018 Traffic Safety Facts A Compilation of Motor Vehicle Crash Data. US Department

of Transportation, National Highway Traffic Safety Administration, Washington, D.C. Report No.

DOT HS 812 981.

Reagan, I.J., McClafferty, J.A., Berlin, S.P., & Hankey, J.M. (2013). Using naturalistic driving data to

identify variables associated with infrequent, occasional, and consistent seat belt use. Accident

Analysis & Prevention, 50, 600–607. https://doi.org/10.1016/j.aap.2012.06.008

Regan, M.A., Triggs, T.J., Young, K.L., Tomasevic, N., Mitsopoulos, E., Stephan, K., & Tingvall, C.

(2006). On-road evaluation of intelligent speed adaptation, following distance warning and seatbelt

reminder systems: Final results of the TAG SafeCar project. Monash University Accident Research

Centre Reports, 253, 270.

Richards, D.C. (2010). Relationship between speed and risk of fatal injury: pedestrians and car

occupants. London, UK: Department of Transport, Transport Research Laboratory.

Sakamoto, Y., Ishiguro, M., & Kitagawa, G. (1986). Akaike information criterion statistics. Dordrecht,

The Netherlands: D. Reidel. 1986; 81. Taylor & Francis.

Sakia, R.M. (1992). The Box‐Cox transformation technique: a review. Journal of the Royal Statistical

Society: Series D (The Statistician), 41(2), 169–178. https://doi.org/10.2307/2348250

28

Sato, T.B. (1987). Effects of seat belts and injuries resulting from improper use. The Journal of Trauma,

27(7), 754–758.

Williams, A.F., Wells, J.K., & Farmer, C.M. (2002). Effectiveness of Ford’s belt reminder system in

increasing seat belt use. Injury Prevention, 8(4), 293–296. http://dx.doi.org/10.1136/ip.8.4.293