Probability of detection of electric vehicles with and without added warning sounds Open Access

The National Highway Traffic Safety Administration's (NHTSA) Traffic Safety Facts (National Center for Statistics and Analysis, 2020) revealed that “in 2018 there were 6,283 pedestrians killed in traffic crashes, a 3.4-percent increase from the 6,075 pedestrian fatalities in 2017.” Although this increase can be attributed to a variety of causes, steadily increasing sales of quiet vehicles are altering how vehicles are perceived within traditional roadway environments due to their quiet operating noise relative to internal combustion engine (ICE) counterparts. According to the Edison Electric Institute, “Global EV sales totaled about 2.1 × 10⁶ for 2018, an increase in 64% compared to the total sold in 2017. 2018 EV sales increased 79% in the United States, 78% in China, and 34% in Europe compared to 2017. United States EV sales represented approximately 17% of global EV sales in 2018” (Edison Electric Institute, 2019). This trend shows no signs of slowing down, with automakers, such as General Motors, suggesting shifts toward electric vehicles (EVs) within the next decade (General Motors, 2019). In response to concerns regarding decreased detectability, auto-manufacturers are producing non-ICE vehicles with an additive noise component aimed at signaling vehicle presence in the same way road users have learned to identify approaching vehicles through engine noise. The Virginia Tech Transportation Institute (VTTI) conducted an evaluation of quiet car detectability as part of a General Motors-funded project in 2015–2016 (Neurauter et al., 2017; Roan et al., 2017). This evaluation conducted vehicle warning signal testing based on methods outlined by the United Nations Economic Commission of Europe (2015) and the developing NHTSA regulations available at the time. The primary focus involved a pedestrian detectability component, where vision-impaired participants were positioned on the side of the road as they evaluated four vehicle types approaching at specific speeds in two noise conditions. These vehicles included a 2011 Chevrolet Volt (EV, no additive sound), a 2014 Cadillac ELR (EV, production additive sound), a 2013 Toyota Prius [hybrid vehicle (HV), production additive sound under EV mode], and a 2013 Cadillac SRX (ICE benchmark). Despite each vehicle exceeding the United Nations Economic Commission of Europe (UNECE) minimum, this initial evaluation revealed that none of the vehicles, including the ICE benchmark, were immune to missed or late detections. Furthermore, the ICE benchmark significantly outperformed the other three vehicles under the 10 km/h (kph) steady approach, but these differences largely disappeared at 20 kph due to increased tire and road noise. Trends of improved detectability offered by the additive noise signals were observed but did not demonstrate a significant advantage over the EV with no additional noise component. Since that original project, NHTSA has released their final version of the Federal Motor Vehicle Safety Standard (FMVSS) 141, outlining “Minimum Sound Requirements for Hybrid and Electric Vehicles” (National Highway Traffic Safety Administration, 2018). A new project aimed to demonstrate differences in detectability by replicating the previous study but with newer, FMVSS 141 compliant sounds. This project produced all of the data used to generate results presented in this paper. Further details of the project can be found in Neurauter et al. (2020).

There has been other work in the detectability of approaching EVs by vision-impaired pedestrians. (Altinsoy, 2013) used binaural recordings to show that pedestrians detect the sound of electric vehicles at an average of 14 m, much closer than the sound of vehicles with ICEs at a distance of roughly 36 m. Parizet et al. (2014) provide results comparing sighted and non-sighted participants who listened to binaural dummy-head recordings of vehicle pass-bys. This work showed that there was no statistically significant difference in detection performance between the two groups. Wall Emerson et al. (2013) also studied detection of EVs by vision-impaired pedestrians, showing that detection performance dependencies included “average wind speed, amplitude modulation of the signal, hearing loss in the 500 Hz range, vehicle velocity, minimum ambient sound level, and overall vehicle sound level in units of A-weighted decibels.” Parizet et al. (2013) evaluated the influence of various timbre parameters on sound detectability, showing that an electric vehicle equipped with a particular low level sound was as detectable as a diesel vehicle. Yamauchi (2016) provides an in-depth perspective on the detectability of the EV additive warning sounds compared to the background noise levels and concludes that the additive sound approach could “solve problems only in limited scenarios.” The environmental noise impact of additive sounds for EVs (Weinandy et al., 2019) is also a topic of considerable discussion. Steinbach et al. (2017) used artificial neural network (ANN) detectors to model listener responses to show that the change of the speed-scaling of additive sounds greatly influences the detection time. The authors also conclude that the use of ANN provides a model that can predict the annoyance or the warning effectiveness of future electric vehicle sounds. Lee et al. (2017) have developed a useful metric called the whine index for evaluating the annoyance and detectability of EVs and also show that an amplitude modulated signal was the best sound in regard to detectability and annoyance (a result also supported by the current work below).

This work presents approaching-vehicle detection results from 16 listeners, plotted as probability of detection (Pd) vs distance (rather than average detection distance) for four regulatory-compliant additive sounds and the no-sound condition. This was done using a high-resolution differential GPS system placed on the test vehicle while subjects performed the task of vehicle detection. Detections were made in two background noise level conditions of 55 and 60 dBA and two approach speeds of 10 and 20 kph. Plots of Pd vs distance show that although the sounds exceed the average detection regulatory criteria, at the 10 kph speed condition, Pd does not reach 100% at the minimum detection distance. For higher vehicle speeds, increased tire noise made the vehicle much more detectable at longer ranges consistent with previous work (Neurauter et al., 2017). The results also include insights into false alarm rates for the test subjects for different background noise conditions.

The paper is organized as follows. Section II provides information on the test methods including the test set up, details on the transmitted additive warning sounds, participant demographics, measurement hardware, and data analysis. Section III provides the detection results given as Pd vs distance. Section IV provides points of discussion regarding the results and testing methodology as well as suggestions for future research studies.

This section provides an overview of all tasks performed up to and including formal data collection during the listener evaluations.

For each session, four subjects were seated on a flat, isolated section of the VTTI Smart Road (Fig. 1). The Smart Road is a closed test bed adjacent to VTTI in Blacksburg, VA, and provides a safe (controlled) environment, low ambient noise levels, a level roadway, a road surface representative of typical roadways, and a site with length appropriate for dynamic maneuvers. Four subjects on each day of trials were seated adjacent to the Smart Road test lane. The seating arrangement is shown in Fig. 2. An electric vehicle (2019 Chevy Bolt) drove along a trajectory perpendicular to the forward facing orientation of the listeners (Fig. 3). The EV was equipped with a pair of loudspeakers (see Fig. 5) that broadcast one of the four test sounds and also made passes with no additive sound. Surrounding the subjects were six JBL LSR308 loudspeakers and a subwoofer directly adjacent to the participants. These speakers broadcast a noise signal with the spectrum shown in Fig. 4. This spectrum is identical to that in National Highway Traffic Safety Administration (2011) and represents an “average urban background noise spectrum.” This noise signal was broadcast at either 55 or 60 dBA as measured by microphones placed above the heads of the four participants. Between the vehicle speeds, test sound number, and background noise level, the listeners were presented with one of 20 constant-velocity scenarios. The scenarios used in this study are provided in Table I. Four trials of each scenario were done.

Subjects wore a sleep mask to block vision and were instructed to press a hand-held button upon detecting an approaching vehicle and to release the button when they felt that the vehicle had passed. The demographics of the subjects are summarized below.

Sixteen individuals volunteered for participation in this study. Consistent with the previous study, specific age groups were not targeted (Neurauter et al., 2017). Table II provides further detail pertaining to participants. Vision-impairment was not required for eligibility, as it was not expected to impact vehicle detectability (Parizet et al., 2014). All participants were recruited in accordance with the Virginia Tech Institutional Review Board (IRB) approval #15-729 “Electric Vehicle Detectability: Impact of Artificial Noise on Ability of Pedestrians to Safely Detect Approaching Electric Vehicles.”

Overall A-weighted sound pressure level (SPL) and 1/3 octave band SPLs (also A-weighted) were recorded at the four microphones above the heads of the participants. The major components of the noise measurement data acquisition system (DAS) were as follows:

• Four G.R.A.S. (Holte, Denmark) 46AQ TEDS microphones

• One National Instruments (Austin, TX) cDAQ USB data acquisition rack

• One National Instruments NI 9234 analog to digital converter module

• A Dell (Round Rock, TX) Inspiron desktop PC (8 GB RAM 1 TB hard drive) running matlab

• National Instruments labview software

• The National Instruments labview Acoustics and Vibrations Measurement Suite

This hardware and software combination provides a class 1 system that meets or exceeds the following standards: IEC 61 260: 1995, class 1, IEC 61 672: 2002, class 1, JIS C 1509-1: 2005, JIS C 1514: 2002, ANSI S1: 11-2004, class 1, ANSI S1.4: 1983, ANSI S1.42: 1986, ISO 8041: 2005(E), ISO 532: 1975, DIN 45 631: 1991, DIN 45631/A: 2008. This acquisition package was configured as shown in Fig. 5. Microphone output voltage was sampled at 50 kHz passed through an A-weighting filter and passed to a sound level meter to calculate overall SPL (0.125 s exponential averaging) and 1/3 octave band SPLs. Overall SPL and 1/3 octave band calculations were logged on the PC and transmitted over ethernet to the VTTI DAS at a rate of 10 Hz.

Before and during all testing, each microphone was calibrated using a G.R.A.S. 42AA pistonphone calibrator. The 42AA (114 dB at 250 Hz) complies with all the requirements of International Electrotechnical Commission (IEC) Standard 942 (1988) Sound Calibrators Class 1 and was corrected with a G.R.A.S. ZC0002K barometer. The data acquisition and analysis system described above was used for both the vehicle noise tests and for the listener testing. Ambient background measurements were made periodically throughout all testing. Table III provides the daily averages of the ambient levels. The very low level on December 13, 2018 is due to the roughly ten inches of snowfall at the site.

A modified VTTI NextGen DAS shown in Fig. 6 logged distance and speed simultaneously with the acoustical measurements. The DAS was configured to communicate with a differential global positioning system (DGPS) unit installed in the test vehicle. This DGPS unit consisted of a NovAtel (Calgary, Canada) antenna on the vehicle roof and an AvaLAN transmitter placed on the passenger-side dashboard.

This instrumentation approach provided continuous recording of base-to-vehicle distance and speed. An experimenter calibrated the transmitter and receiver at the beginning of each test session, ensuring accuracy of the recorded output. Based on known positions of each participant's seated location with respect to vehicle path and location of antenna relative to the front bumper, accurate distances were calculated.

Subjects were presented with one of four additive alert sounds and a no-sound condition. The four additive sounds were initially created to produce certain aesthetics that satisfied the sound designer. The sounds were then modified using a graphic equalizer to boost selected octave bands to satisfy the proposed NHTSA regulations pertaining to FMVSS 141 (National Highway Traffic Safety Administration, 2018). These are, in short, “Vehicles complying with the 2-band option must meet minimum SPLs in two non-adjacent one-third octave bands between 315 Hz and 3150 Hz, with one band below 1000 Hz and the other band at or above 1000 Hz. The two bands used to meet the 2-band option also must meet a minimum band sum level.” These modifications were intended to make the sounds FMVSS 141 compliant but minimally impact the aesthetic of the sound. Both stationary and drive-by measurements were made. These are provided in Secs. II D 1 and II D 2.

Using the geometry in Fig. 7, 20-s-long measurements were taken and recorded by the DAS. These measurements were completed in accordance with Section 13 of the FMVSS 141 standard (National Highway Traffic Safety Administration, 2018). The reason for this choice of measurement geometry was to use regulatory-compliant sounds and measurement approach to ultimately generate Pd curves that are relevant to FMVSS 141. The loudspeaker arrangement at the front of the car shown in Fig. 7 is unconventional when compared to production additive sound alerting systems. In production acoustic vehicle alerting systems (AVASs), the loudspeakers are inside the boundary of the vehicle, e.g., in the engine compartment. The location of the speakers at the front of the vehicle was chosen for these experiments as a matter of convenience so that their positions and volumes could be adjusted and so that overall levels on the driver and passenger sides of the vehicle would be the same. This goal was achieved, as shown in the left-hand plots of Figs. 8–11. Figures 8–11 provide the results for both overall A-weighted SPL (left) and 1/3 octave bands (right). The overall SPL plots on the left provide a short time history of the overall A-weighted SPL at each of the microphones (driver, passenger, and front). At the bottom of the overall SPL plots, the peak level is given for each microphone position. The 1/3 octave band plots on the right of each figure are presented for the time that matches the peak of the overall SPL for the microphone position with the lowest overall SPL. For example, in Fig. 8, the microphone with the lowest overall level is the passenger position. The peak overall level (50.6 dBA) for this microphone occurs at 11.6 s (marked by a circle marker). The plot on the right of Fig. 8 gives the 1/3 octave band levels at the time of the max overall SPL (i.e., 11.6 s). Additionally, the 1/3 octave plot provides the sum of the upper and lower range non-adjacent bands that exceed the NHTSA-prescribed octave band levels (in this case 48.2 dBA). Sound #1's lowest overall level was from the passenger side. The 1/3 octave band results in Fig. 8 show that several bands in both the upper and lower ranges exceeded the NHTSA minima and that the two-band sum met the 48 dB requirement. The passenger-side overall peak level was 1 dBA below the 52 dBA requirement.

Sound #2 had the highest overall SPL in front of the vehicle. The driver-side microphone had the lowest overall level. The 1/3 octave band results in Fig. 9 show that several bands in the lower range and one band in the upper range exceeded the NHTSA minima and that the two-band sum met the 48 dB requirement. The passenger-side overall peak level exceeded the 52 dBA requirement by 3 dBA, a very high measurement due to the two-band requirement. For the single upper-range octave band (1000 Hz) to exceed the NHTSA requirement (41 dBA), the overall level had to be increased. This created a very loud sound at the front of the vehicle.

Sound #3 had a tempo that was almost impulsive, as evidenced by the time series plot of the overall SPL. The passenger-side microphone had the lowest overall level. The 1/3 octave band results in Fig. 10 show that one band in both the upper and lower ranges exceeded the NHTSA minimum and that the two-band sum met the 48 dB requirement. The passenger-side overall peak level was 1 dBA below the 52 dBA requirement.

Sound #4 spectrum contained much more lower-frequency energy than the other sounds. The passenger-side microphone had the lowest overall level. The 1/3 octave band results in Fig. 11 show that several bands in both the upper and lower ranges exceeded the NHTSA minimum and that the two-band sum met the 48 dB requirement. The passenger-side overall peak level exceeded the 52 dBA requirement by 1 dBA.

In all of the cases above, the overall A-weighted SPL in front of the vehicle was 3–6 dBA above the lowest overall SPL for that sound. Sound #2 was the extreme at 6 dB above the level on the sides of the vehicle. These differences can be attributed to the loudspeaker arrangement used to transmit the test sounds. First, the speakers are not omni-directional. Using two speakers increased coverage to the sides of the vehicle but also increased the overall level in front of the vehicle. To fully understand the directionality issues of the loudspeaker arrangement, measurements in an anechoic chamber on a turntable would be required.

The dynamic testing procedure consisted of measuring the overall A-weighted SPL as the vehicle moved through the test area at either 10 or 20 kph. These tests were designed to match the measurement criteria provided in the FMVSS 141 test measurement procedure document (National Highway Traffic Safety Administration, 2018). The test area and microphone locations are illustrated in Fig. 12. Background noise-level measurements were also made throughout the testing procedure (see Table III). Figures 13–17 provide the drive-by results with the vehicle moving from left to right. The right to left measurements are consistent with the left to right and have been omitted for brevity. All of the drive-by measurements illustrate the directionality of the loudspeaker source where the sound had a higher SPL in front of the vehicle. This is evidenced by the peak occurring roughly 10 m before the car reaches the microphones. The higher speed of 20 kph causes an increase in overall SPL of roughly 10 dB seen in all of the plots. This is also seen in the no-sound condition plot in Fig. 15, particularly as the car's noise emerges from the background noise at a distance of roughly 20 m. The no-sound condition peaks at 9 m from the two microphones at 52 dBA for the 10 kph case and 62 dB for the 20 kph case. Sound #2 had the highest SPL in front of the vehicle as shown in the stationary measurements above Fig. 7. The drive-by plot in Fig. 12 shows that the effect of this is a peak SPL occurring ever farther away from the 0 m mark at roughly 15 m.

Previous results for the data collected in this work were expressed as mean detection distances (Neurauter et al., 2020; Roan et al., 2017). Human detection performance in this work is presented as Pd vs distance and is based on passive sonar signal processing as discussed in Van Trees (2004). The process is illustrated in Fig. 18. The data from a 4 m window that corresponds to the length of the vehicle are used to calculate Pd and probability of missed detection directly. While a vehicle is in a given range cell, there are a fixed number of samples of the listener's button presses (1 for detection, i.e., button pushed, and 0 for no detection, i.e., no vehicle present). The total number of button-push data points in the range cell gives the number of opportunities to make a decision. As the vehicle was always present in the range cell, the Pd is the number of 1's in the range cell divided by the number of opportunities. The number of missed detections is 1-Pd (i.e., the vehicle is present, but listeners did not detect it). Sliding the range cell along the data set and averaging across data sets for a given scenario gives Pd vs distance for that scenario.

Each plot in Figs. 19–23 consists of two subplots. The plot on the left, labeled as (a), is for the 55 dBA background noise condition for vehicle speeds of 10 and 20 kph. The plot on the right, labeled as (b) is for the 60 dBA background noise condition for vehicle speeds of 10 and 20 kph. For both (a) and (b) plots, the left-hand axis is Pd. The right-hand axis is A-weighted SPL from the microphones above the listeners' heads. Error bars indicate standard error.

For sound #2, the results in Fig. 20 show increased detection distance due to the higher overall SPL of the additive sound in front of the vehicle. Additionally, the slope of the detection curve is less negative than the other sounds, indicating a higher overall Pd at farther distances. However, the Pd of this sound is actually lower than sound #1 in the close range (less than 10 m). The Pd for this sound never reaches 1 for the 10 kph case in either background condition, and the Pd for 20 kph reaches 1 at a shorter distance.

Sound #3 was different from the other sounds in that it was more of a periodic impulsive sound, as shown in Fig. 8. This sound has the poorest overall performance of the four sounds. This is evidenced by the steepest negative slope of the Pd curve. However, at closer distances for the 10 kph case, the performance was better than sounds 1 or 2 with a Pd of roughly 1 inside of 11 m. This may indicate that a more pulsing sound is more detectable at closer ranges but has a lower impact on the soundscape at farther distances where detecting the vehicle is unnecessary.

Sound #4 had similar detection performance to sounds 1 and 2 for the 10 kph cases but had superior detection performance at 20 kph. In the 55 dBA background condition for 20 kph approaches, sound #4 had a Pd of 1 out to a distance of 30 m and to 20 m for the 60 dBA background. The 10 kph curves never reach a Pd of 1 for either 55 or 60 dBA background. Sound #4 was the only sound with significant energy in the 1600 Hz octave band (followed by sound #3). The sound also had the highest number of bands that exceeded the NHTSA minimums, giving the sound the highest level of frequency diversity.

The no-sound condition provided the worst performance in terms of detectability. The Pd curves have the steepest negative slopes and never reach a Pd of 1 for any of the test conditions. Tables IV and V provide several points of interest from the Pd vs distance curves. The second column in the tables gives the Pd at the NHTSA minimum detection distance of 5 m (National Highway Traffic Safety Administration, 2011). In the 55 dB background noise for 10 kph vehicle speed, all of the sounds exceeded a Pd of 90%, but only sound #1 reached 100% detection at the minimum prescribed distance. At 20 kph, in 55 dBA background, all sounds achieved 100% detection. At the mean detection distances, all sounds hovered in the 40–60% Pd range. As expected, the no-sound case performed poorly. In the 60 dB background noise for 10 kph vehicle speed, performance was reduced with no sounds achieving 100% detection. At 20 kph and 60 dBA background, all sounds achieved 100% detection. At the mean detection distances (all greatly reduced from the 55 dBA background cases), all sounds again had 40–60% Pd ranges.

Figures 19–23 provide Pd vs distance for all 16 subjects with four trials per speed and background noise condition. In terms of general trends, the higher speed case of 20 kph, as expected, produced higher overall detection probabilities due to increased tire noise. These curves had the lowest negative slope, indicating higher Pd at farther distances. The NHTSA minimum distances for detection for 10 and 20 kph approaches are 5 m and 11 m, respectively (National Highway Traffic Safety Administration, 2011). The average detection distances (Tables IV and V) far exceed the minimum required distances. Observing the Pd at these distances provides a clearer picture of actual performance (i.e., Pd in the 0.4–0.6 range). The average detection distances were consistent with the authors' previous work (Roan et al., 2017) and that in Altinsoy (2013).

All of the sounds except sound #2 in 60 dBA background noise reached 100% Pd at distances exceeding the NHTSA minimum of 11 m. Sound #2 reached 95% at 11 m for the 60 dBA background for the 20 kph approach speed. Sound #2 had the highest overall SPL, as shown by Figs. 7 and 12, but also had the lowest energy in the 1250 and 1600 Hz 1/3 octave bands. Sound #3 presents an interesting case in that its overall energy was quite low. This was the only sound that had significant dynamic amplitude shifts as shown in Fig. 8(a). This sound provided a Pd that reached 100% for all of the scenarios except for the 10 kph approach in 60 dBA background, where it reached 0.87 at 5 m. Sound #3 provided high Pd at close ranges but had the steepest negative slope for all cases at distances farther than 10 m. This result is consistent with the work by Wall Emerson et al. (2013) regarding amplitude modulation of the signal. Additionally, this type of detection curve is what would be needed to begin to satisfy the balance between safety at close ranges and annoyance at farther ranges. Sound #4 also provided interesting Pd results. In the 10 kph case, it had performance similar to the other sounds, but it had superior performance in the 20 kph cases. As noted above, Sound #4 had the highest levels of low frequency content (Fig. 9) but also had a peak SPL in the drive-by tests that was the latest to arrive at the listeners (Fig. 14). Last, the no-sound case did not in any case reach 100% detection.

The no-sound condition also provides insight into the false alarm rate of the human detectors (Van Trees, 2004). Figure 23(a) shows a leveling of the Pd curve at distances greater than 50 m for the 55 dBA background condition at a Pd level of 2.4%. Figure 23(b) shows a leveling of the Pd curve at distances greater than 70 m for the 60 dBA background condition at a higher Pd level of 5.0%. In classical detection and estimation theory a false alarm rate is selected and a threshold is applied to the detection statistic. Values of the detection statistic above the threshold are declared detections, and those below it are non-detections (Poor, 2013). The number of correct detections given a true hypothesis divided by the number of opportunities is the Pd. These types of detectors are referred to as constant false alarm rate (CFAR) detectors. In analogous sonar problems, as the signal to noise ratio (SNR) goes down into the noise, the Pd asymptotically approaches the false alarm rate as the source moves away from the detector and Pd levels off as we see in Figs. 23(a) and 23(b). These false alarm rates are relatively low and are consistent in that when the background noise level is almost doubled, the listeners become more likely to guess when the vehicle sound is extremely low.

There are several future research topics of interest. Efforts should be made to produce FMVSS-141 compliant sounds that are uniform in SPL around the front of the vehicle. Given the tonal nature of the requirements, this may prove difficult, as many factors will impact directionality. Speaker placement will play an important role in producing sound outside the perimeter of the vehicle. A second area for future research is to perform a regression analysis relating the octave band levels to detection performance. Once a relationship is established relating required octave band content to detection performance, various approaches to sound design should be considered. Sound aesthetics plays a very important role in the perception of vehicle quality (Genuit, 2004). Therefore, sounds are not likely to be designed with only regulatory requirements in mind, and methods of incorporating aesthetics will be required. Last, immersive audio techniques could be employed to provide both test uniformity and diversity in terms of road surfaces and other environmental considerations at lower cost and higher repeatability. The tests performed in this study were done on a closed section of roadway that was isolated from many environmental acoustic sources, such as traffic and industrial noise, but still required stopping testing several times a day due to aircraft fly-overs, changes in wind, and locomotive pass-bys. Last, in regard to background noise, it would be very beneficial to understand detection performance in realistic backgrounds rather than the broadband, uniform continuous noise tested to date. This background should include the presence of several slow moving EVs in the area around the pedestrian to determine the impact of many alert sounds possibly confusing the listener.

The authors of this report would like to acknowledge the support of the stakeholders of the National Surface Transportation Safety Center for Excellence (NSTSCE): Tom Dingus from the Virginia Tech Transportation Institute; John Capp from General Motors Corporation; Chris Hayes from Travelers Insurance; Terri Hallquist and Nicole Michel from the Federal Motor Carrier Safety Administration; Cathy McGhee from the Virginia Department of Transportation and the Virginia Transportation Research Council; and Jane Terry from the National Safety Council. Additional project funding was received from the General Motors Corporation. The NSTSCE stakeholders have jointly funded this research for the purpose of developing and disseminating advanced transportation safety techniques and innovations.