Driver Performance in a Cooperative Adaptive Cruise Control String

Transcription

1 Proceedings of the Human Factors and Ergonomics Society 2016 Annual Meeting 1184 Driver Performance in a Cooperative Adaptive Cruise Control String Vaughan W. Inman 1, Steven Jackson 1, Brian H. Philips 2 1 Leidos, 2 Federal Highway Administration Cooperative Adaptive Cruise Control (CACC) has been proposed as a method to increase highway capacity and possibly enhance safety. Two experiments were conducted in a driving simulator to verify that drivers with CACC would effectively monitor the system s longitudinal control and override the system in the event that greater braking authority was needed than the system was designed to provide. In the first experiment, the emergency response of drivers with the CACC was compared with that of drivers who manually controlled following distance within a string of vehicles. The CACC group experienced markedly fewer crashes and had longer mean time-to-collision. The second experiment examined whether the CACC safety benefit was the result of the CACC system s limited automatic braking authority, an auditory alarm, or both. The results suggest that both auto-braking and an auditory alarm are necessary to achieve a crash reduction benefit, although the alarm alone may promote less severe collisions. Not subject to U.S. copyright restrictions. DOI / BACKGROUND Cooperative Adaptive Cruise Control (CACC) combines three driver assist systems: (1) conventional cruise control, which automatically maintains the speed a driver has set, (2) adaptive cruise control (ACC), which uses radar or lidar sensors to automatically maintain a driver-specified gap between the driver s vehicle and a vehicle ahead, and (3) dedicated short-range communications (DSRC) to transmit and receive data with surrounding vehicles (i.e., vehicle-tovehicle (V2V) communications). The system can more quickly respond to changes in the speed of other CACC equipped vehicles than is possible with ACC, and can even respond to changes that the driver cannot see (Jones, 2013). One goal of CACC is to enable strings of equipped vehicles to travel at higher speeds with smaller gaps between vehicles while affording greater safety than is possible without CACC technology. Such strings could increase highway capacity (Milanés & Shladover, 2014). Like conventional cruise control, CACC relieves the driver of the need to manually maintain longitudinal control. Yet the driver is still responsible for monitoring separation between their vehicle and others, and may need to resume control in some situations. OVERVIEW OF GENERAL APPROACH Two experiments are presented that examine the performance of CACC drivers within a string of CACC vehicles. Experiment 1 compares driver performance using CACC to performance using manual longitudinal control. Experiment 2 compares performance of drivers using CACC with the performance of drivers using adaptive cruise control. In both experiments, the primary interest was in how the participant drivers would react to an emergency braking event after an extended period of uneventful driving. Dependent measures in both experiments included collision incidence, brake reaction time, and minimum time-to-collision. Experiment 1 also included a workload measure and eye tracking. Both experiments were conducted in a driving simulator. Method EXPERIMENT 1 Participants. Participants were 25 licensed drivers recruited from the Washington, DC metropolitan area. There were 12 drivers in the manual control group and 13 in the CACC group. Each group had an approximately equal number of males and females and a median split in age of 46 years. The mean age of the younger group was 30 years and the mean age of the older group was 60 years. The CACC group had one additional younger male. Participants were recruited from a database of previous participants in Federal Highway Administration (FHWA) driving simulator studies, where those studies did not include surprise emergency events. Participants were paid 60 dollars for their participation. Materials and Apparatus. The experiment was conducted in the FHWA Highway Driving Simulator. The simulator consisted of a compact sedan mounted on a six-degree of freedom motion base placed within a cylindrical projection screen with a radius of 2.7 m. Three projectors provided a 200-degree (horizontal) by 40-degree (vertical) field-of-view. Each projector provided a nominal resolution of 4,096 by 2,400 pixels. Highway sounds were also provided. Participants drove in a dedicated center lane on a simulated 8-lane interstate highway (4 lanes in each direction). The center lanes were separated from the other lanes by barriers. Figure 2 shows a typical section of roadway. Entrance to the dedicated-lane was via ramps on the left side of the roadway. The simulation began with the participant s vehicle in the third position within a string of four CACCenabled vehicles. When a ramp meter turned green, the string accelerated and merged into the CACC lane and cruised at 113 km/hr. Vehicles in CACC mode were set to maintain a 1.1 s gap. Participants in the manual control group had an auxiliary display on the center stack that showed their distance from the vehicle ahead on a ribbon type indicator. Calibration of the Perceived Following Distance. Given the interest in an emergency event with a relatively close following, it was important that the perceived gap between vehicles be realistic. In previous experiments in the FHWA simulator, participants had shown a reluctance to follow other

2 Proceedings of the Human Factors and Ergonomics Society 2016 Annual Meeting 1185 vehicles with gaps that are commonly observed in the real world. To explore this phenomenon, five individuals were recruited from laboratory staff to follow the same simulated vehicle, a 2007 sport-utility vehicle, on a real freeway and in the simulator (Balk, Inman, & Perez, 2015). Each individual was asked to follow at two distances: (1) a comfortable distance, and (2) the minimum distance that they thought was safe. Between comfortable and minimum safe trials, participants were asked to back off and drive at a considerably longer distance to foster independence between trials. Each following distance was repeated at least two times. The results confirmed that the perceptions of following distance were indeed different between the simulator and the real world, even though the simulated lead vehicle was scaled to subtend the same visual angle as the actual SUV. The observed following distance in the simulator was about 1.3 times the following distance in the real world. To achieve the same perceived following distance in the simulator, the simulated vehicles were scaled to be 75 percent of the size for a 1:1 visual angle. Five weeks after the first following distance tests, the same five individuals were again asked to follow the reduced size simulated lead vehicle at comfortable and minimum safe distances. Given the results shown in Figure 1, the current experiments were conducted with the reduced size simulated vehicles. Figure 1. Following distance in field and in simulator as a function of the scaling of the lead vehicle. the vehicle ahead, which briefly reduced the time gap to 0.5 s. With the addition of this vehicle, the participant s vehicle became the fourth in a five car string. A NASA-TLX (National Aeronautics and Space Administration, 2009) workload assessment was verbally administered after this merge event and again 10 minutes later. At the 56 km mark, a vehicle rapidly came down an onramp on the participant s left and disappeared from the participant s view as it passed in front of the five car string and overturned. This critical event caused the lead car of the string to begin decelerating at a rate of 9.8 m/s 2. This in turn caused all of the CACC vehicles in the string to automatically decelerate at a rate of 3.9 m/s 2 after a 0.1 s delay. As the autobraking commenced, an auditory alarm sounded (a 1000 Hz warning tone of four beeps of 140 ms separated by 22 ms of near silence). The brake lamps of the third vehicle, the vehicle visible to the participant, illuminated when the driver of that simulated vehicle began decelerating a 9.8 m/s 2, i.e., overrode the CACC system and applied maximum braking force. This happened 1.9 s after the lead vehicle began decelerating. This chain of events meant that the CACC participants decelerated at a rate of 3.9 m/s 2 beginning 0.1 s after the lead vehicle began decelerating, whereas the manual control participants first cue that a response was required was the looming of the vehicle ahead as it decelerated at 3.9 m/s 2. Thus the CACC group had two early warnings that the manual control group did not: their own vehicle decelerating at 3.9 m/s 2, and an auditory alarm. Procedure. Upon arrival, participants were asked to review and sign an informed consent. This was followed by the health screen administered to ensure that the participants were not at increased risk of simulator sickness as a result of illness or lack of sleep. Participants were asked to show a valid driver s license. An eye chart was used to verify visual acuity equal to or better than 6/12, with correction if necessary. A PowerPoint presentation with embedded videos was shown to explain the CACC concept. Participants assigned to the CACC conditions were introduced to the warning tone that is triggered when more braking is needed than the CACC system provides. The CACC related instructions were as follows: Set the gap to Near [defined in the PowerPoint as 1.1s]. Set the speed to 70 MPH. You will control steering follow the car in front. The system will accelerate and brake up to a limit. You need to monitor the situation at all times the system can fail. You can take over control by pressing the accelerator or brake. Pressing the brake disengages the CACC system if you need to take control. Press ENGAGE as soon as possible when the situation allows. Figure 2. Typical section of simulated roadway. Simulation Scenario. Participants in both manual and CACC groups drove 56 km. At 9.8 km into the drive a vehicle merged into the string between the participant s vehicle and Except for the above instructions, the control group instructions were the same. The instructions unique to the control group were as follows:

3 Proceedings of the Human Factors and Ergonomics Society 2016 Annual Meeting 1186 Aside from maintaining 1.1 s gap, you should drive as you normally would. Stay alert for unexpected events. The PowerPoint presentation concluded with an explanation of the NASA-TLX, which was to be verbally administered while the participants were driving. Following the PowerPoint briefing, participants were seated in the simulator cab where the controls and displays were reviewed, and the instructions repeated. While seated in the cab, participants were asked to complete a simulator sickness questionnaire (SSQ: Kennedy, Lane, Berbaum, & Lilienthal, 1993) to provide a symptoms baseline. Finally, the eye-tracking system was calibrated; a procedure that generally took 5 to 10 min. With the preliminaries completed, participants were asked to perform a brief (less than 10 min) practice drive. On the practice drive, participants were asked to enter the dedicated CACC lane, which was free of traffic, and accelerate to 70 mi/h (112.7 km/h). They were asked to gently brake then accelerate followed by a request to brake hard then accelerate. To enable adaptation to the lateral control, participants were asked to gently move from the travel lane into the breakdown lane and then move back into the travel lane. This was followed by a request to quickly change into and out of the breakdown lane. In the CACC condition, participants were then asked to engage the CACC system, which was accomplished by pressing a touch screen control on the center stack. The system accelerated to 121 km/h until it closed on a string of CACC vehicles traveling at 88.5 km/h. The string traveled at that speed for 2 minutes and then accelerated to 113 km/h. In the control condition, participants were asked to accelerate to 70 mi/h (113 km/h) and maintain that speed until they closed on a string of CACC vehicles. They were then asked to follow with a 1.1 s time gap and refer to a ribbon gap display, located on the touch screen display on the center stack, as necessary. The other vehicles in the string behaved in the same manner as for the CACC condition. After traveling in the string at 113 km/h for 2 min, the NASA-TLX was administered to further familiarize participants with the workload assessment tool. At the conclusion of the workload assessment, participants were asked to take the next available off-ramp and come to a complete stop. After completing the practice drive, participants were asked to exit the vehicle and complete an SSQ. The experimental session began with the participant seated in the third of a string of four vehicles. The string was stopped on a ramp at a ramp meter showing a red indication. When the ramp meter turned green, the vehicles ahead began to accelerate down the ramp towards the CACC travel lane. At this time, participants in the CACC conditions were asked to release the brake and engage CACC. Next, the participant s vehicle followed the two preceding vehicles in the string with a 1.1 s time gap. Participants in the control condition were asked to follow the preceding vehicles and try to keep the gap close to the 1.1 s target. About 5 min into the drive, a CACC vehicle came down a ramp on the left and merged into the gap directly in front of the participant s vehicle. This momentarily cut the gap to half of what it had been. The CACC-equipped vehicles behind the merged vehicle responded by decelerating with engine braking until the gap was again 1.1 s. If necessary, a researcher would remind control participants to return to the 1.1 s following distance. As soon as the now five car string s stability was reestablished, which generally took about 30 s, the NASA-TLX was administered to assess workload during the merge event (described as during the preceding minute or so ). From the conclusion of the NASA-TLX about 10 min elapsed before another NASA-TLX was administered. This administration was intended to assess workload during uneventful cruising. The cruise was again uneventful for the next 31 min until the critical event (the deceleration in response to the overturning vehicle). At the conclusion of the critical event response, a final NASA-TLX was administered, after which the participant was asked to take the next exit ramp and come to a stop. After exiting the simulator, participants were asked to complete a final SSQ, were debriefed, and paid for their participation. Results The time gap ribbon display was successful in enabling the manual control group to maintain the same 1.1 s time gap as the CACC group. The mean time gaps of both groups measured just before the critical event are shown in Table 1. The difference between means, assuming unequal variances, was not statistically significant, t (10.5) = 1.44, p > Table 1. Mean Following Gap. Group Mean Headway Std. Error Manual Control CACC The crash results for the critical event are shown in Table 2. The difference in number of crashes between groups was significant by Fisher s Exact Test (p < 0.02). Table 2.Number of crashes as a function of experimental group. Group Crashed Avoided Manual Control 5 6 CACC 1 12 The adjusted time-to-collision (ATTC), as described by Brown (2005), also showed a CACC safety benefit. Positive values of ATTC can be considered as the amount of extra time the driver had to react in the case of an avoided collision, and negative values the amount of additional time the driver needed to react to avoid a collision. Table 3 shows the mean ATTC in Experiment 1, where the 95 percent confidence limits were computed using a generalized linear model (GLM) with gamma response distribution and inverse link function. Because the gamma distribution cannot include negative values, the ATTC values were transformed before model

4 Proceedings of the Human Factors and Ergonomics Society 2016 Annual Meeting 1187 estimation. On average, in the manual control condition drivers required an additional 0.6 seconds to react sufficiently to avoid a collision. Table 3. Estimated Mean ATTC and Confidence Limits as a Function of Group Group Mean ATTC Lower Cl Upper Cl Manual Control CACC Brake reaction time was defined as the time between when the car immediately ahead began braking and when the participant began to depress the brake pedal. One manual control participant and two CACC participants either (1) never hit the brake, or (2) swerved out of the travel lane before braking, were excluded from brake reaction time analyses. There was no significant difference in brake reaction time between groups; a GLM with a gamma response distribution and inverse link function yielded a Wald χ 2 (1) = 0.42, p = The manual control group tended to brake more vigorously than the CACC group. The time between the onset of braking and full brake pedal depression was shorter for the manual control group, although not significantly so, Wald χ 2 (1) = 3.07, p = Because brake reaction times and brake intervention deceleration rates were similar, the main source of difference in crash rates between the groups appears to be the result of the CACC system automatically decelerating at 3.9 m/s 2 (0.4 g) prior to participant intervention. The eye tracking results showed that the manual control group devoted a significantly greater amount of time looking at the instrument cluster (speedometer) than did the CACC group (6.4 percent versus 1.2 percent, t(12.6) = 3.55, p < 0.01). Glances to the center stack did not differ between groups. Discussion The Experiment 1 results suggest that drivers can reliably monitor and take over control from a CACC system in the event of an extreme (1 g) braking event in which they cannot see what initiated the emergency situation. Drivers without CACC were in much greater jeopardy in the same situation. These results apply to a 1.1 s time gap between vehicles. It is not clear from these results what gave the CACC group a safety advantage: the onset of partial braking (0.4 g), the auditory alarm, or both. Furthermore, the brake lamps of the preceding vehicle did not illuminate with the onset of partial braking, which disadvantaged the manual control group. The finding that the manual control group devoted more time glancing to the instrument cluster, presumably to monitor their speed (i.e., it appears that maintaining the same speed as the other vehicles, 113 km/h, was a more efficient means to maintaining the 1.1 s following distance than closely monitoring the gap distance ribbon), also provides an alternative explanation for the crash and ATTC differences. Experiment 2 was conducted to disambiguate the sources of the CACC safety benefit. EXPERIMENT 2 In Experiment 1, the CACC group was less likely to experience a crash, and had longer ATTCs than the manual control group. The safety benefit that accrued to the CACC group may have resulted from the CACC system s partial braking. However, it might also have been the result of the auditory alarm that initiated simultaneously with the automatic braking. Additionally, because the brake lamps of the vehicles ahead did not illuminate until the (simulated) drivers began manually braking, the manual control group was at an unintended disadvantage, as the brake lamps should have come on when the automatic deceleration commenced. In Experiment 2, the brake lamps were set to illuminate on all CACC vehicles when the 3.9 m/s 2 automatic deceleration began. To make the CACC and control groups more comparable in terms of visual demands or distraction, the control group was equipped with ACC, so that the driving demands were the same for all groups until the onset of the critical event. The ACC system responded to the automatic deceleration of the CACC vehicles by decelerating, with a 0.3 s delay, at 2 m/s 2. To distinguish between the effects of the automatic deceleration and the auditory alarm, these variables were factorially crossed in Experiment 2, as shown in Table 4. Method Table 4. Factorial Design of CACC Groups in Experiment g Automatic Braking No Yes No ACC Control CACC B Auditory Alarm Yes CACC A CACC AB Participants. One-hundred twelve participants completed the study, 28 in each of four groups. Participants in Experiment 1 were excluded from participation. To roughly balance the groups on participant age, half the recruits in each experimental group were under the age of 47. Each group was balanced as well with respect to gender. Materials and Apparatus. The simulator and scenario were the same as in Experiment 1 except for the following: (1) the critical event occurred after 31 km of driving, (2) the brake lights of CACC vehicles illuminated when the automatic deceleration commenced, (3) the NASA-TLX was not administered, and (4) eye tracking was not used. Results Crashes. Table 5 shows the number of crashes, and crashes avoided by each group, the maximum likelihood estimates of crash probability, and the 95 percent confidence limit (CL) for those estimates. A GLM with binomial response distribution and logit link function indicated a significant between group difference in the probability of a crash, χ 2 (3) = 10.6, p = Post hoc testing showed that only the CACC-AB significantly differed from the ACC control (p = 0.003).

5 Proceedings of the Human Factors and Ergonomics Society 2016 Annual Meeting 1188 Table 5. Crash results by experimental group. Avoided Crash Lower 95 Upper 95 Crashed Condition Crash Probability Percent CL Percent CL ACC CACC-AB CACC-A CACC-B Total Adjusted Time-to-Collision. The ATTC means and 95 percent confidence limits are displayed in Table 6. These data are based on a sample size of 92 participants. The remaining 20 participants had uninterpretable ATTC estimates. Three of those 20 never applied the brakes. The remaining 17 were decelerating at a rate less than that of the lead vehicle (also decelerating) at the time of impact, thereby generating ATTC values of negative infinity. None of the CACC-AB participants had a negative infinity ATTC, and only one of the CACC-B participants did. Thus it appears that the automatic braking contributed to mitigating the probability of inadequate braking responses. A GLM models with normal response distribution and identity link function showed the effect of condition was significant, χ 2 (3) = 8.54, p = The CACC- AB group had a substantial positive ATTC, i.e., on average members of this group had almost 0.6 s spare time to react to the collision event. The ACC and CACC-B groups had significantly lower mean ATTC values than the CACC-AB group. The CACC-A group mean was not significantly different from any of the other three group means. Table 6. ATTC as a Function of Condition. Condition Mean L-CL U-CL ACC CACC-A CACC-AB CACC-B Reaction time. The brake onset reaction time means and confidence limits for the crash event are shown in Table 7. Three participants in the CACC-B group never reacted and therefore were not included in the reaction time analysis. A GLM with normal response distribution and identity link function showed the condition effect significant, χ 2 (3) = 59.2, p < Post hoc testing showed that the ACC group mean reaction time did not differ significantly from the CACC-AB group mean, but that all the other group mean comparisons yielded significant differences. Table 7. Mean Brake Onset Reaction Time. Condition Mean L-CL U-CL ACC CACC-AB CACC-A CACC-B Discussion and Conclusions The safety benefit that accrued to the CACC group in Experiment 1 can be attributed to the combination of partial (0.4 g) automatic deceleration and an auditory alarm. Crash rates were indistinguishable from ACC when either the autobraking or alarm were missing. The alarm did produce somewhat (but not significantly) lower ATTC than CACC-B and ACC, which may suggest that the presence of an alarm without auto-braking might result in less severe crashes. Reaction times were significantly faster with an alarm alone than in any of the other three conditions. The faster reaction time was probably the source of the slightly reduced ATTC in this condition. Faster reaction time with the alarm alone suggests that drivers in the other CACC conditions delayed their braking override response until it was clearer that the CACC auto-braking would be insufficient. The greatest automation benefit was seen in this study when the driver and automation cooperated effectively (Flemisch, Bengler, Bubb, Winner, & Bruder, 2014); the vehicle provided an early brake response and notified the driver, via alarm, that additional braking assistance was needed. Whether the combination of alarm and automated braking will be effective with other CACC implementations, e.g., implementations with shorter gaps or different levels of automated braking, remains to be explored. ACKNOWLEDGEMENTS The research described in this paper was performed under FHWA Contract Number DTFH61-13-D-00024, Human Factors On-Site Support Services. REFERENCES Balk, S., Inman, V., & Perez, W. (2015). Visual Perception and Illusions in a Driving Simulator - Little Cars, Big Signs. Journal of Vision, 15(12), doi: / Brown, T. L. (2005). Adjusted Minimum Time-to-Collision (TTC): A Robust Approach to Evaluating Crash Scenarios. Paper presented at the Driving Simulation Conference 2005 North America, Orlando, FL. Flemisch, F. O., Bengler, K., Bubb, H., Winner, H., & Bruder, R. (2014). Towards Cooperative Guidance and Control of Highly Automated Vehicles: H-Mode and Conduct-by-Wire. Ergonomics, 57(3), doi: / Jones, S. (2013). Cooperative Adaptive Cruise Control: Human Factors Analysis. McLean, VA: Federal Highway Administration. Kennedy, R. S., Lane, N. E., Berbaum, K. S., & Lilienthal, M. G. (1993). Simulator Sickness Questionnaire: An Enhanced Method for Quantifying Simulator Sickness. The International Journal of Aviation Psychology, 3(3), Milanés, V., & Shladover, S. E. (2014). Modeling Cooperative and Autonomous Adaptive Cruise Control Dynamic Responses Using Experimental Data. Transportation Research Part C: Emerging Technologies, 48, pp National Aeronautics and Space Administration. (2009). Task Load Index:( NASA-TLX). Moffett Field, CA: Retrieved from pdf.