The article presents an experimental study on the efects of driving disengagement on takeover performance in a simulated driving environment. Takeover performance was measured from participants (N=28, 14 females, age M=30.46, SD =10.67) as the response time (RT) required to complete the transition from automated to manual driving. Several other potential factors for takeover performance were also examined, including driver age, gender, simulator experience, driving-related data, and automotive user interface (UI) complexity (baseline vs. head-up display). A significant efect on RT was found for the type of disengagement (task vs. rest), as well as for the interaction efect of gender and disengagement. Males had significantly longer RT than females (diference in RT: M=2353.14 ms) when engaged in a secondary task. Machine learning was performed to examine the predictive performance of several regression models and the significance of the features (gender, age, driving disengagement, simulator experiance, average speed) on RT. The LightGBM regressor performed well (training accuracy: 0.89, test accuracy: .73, mean absolute error (MAE): .14). In addition to average speed and age, the disengagement features task, rest, and eyes-of-road ratio were the most important predictors of RT.
As automation progresses, drivers become less focused and their situational awareness tends to
decrease, especially in highly automated vehicles (HAVs) [
This study examines driver performance in a conditionally automated vehicle and in mixed situations (urban and suburban) with changing trafic conditions and obstacles. It focuses on driving disengagement, which is defined as an event where the driver is not involved in driving or is not focused on the driving situation.
The aim of the presented study is to investigate the efects of driving disengagement on driver takeover performance (RT) when taking control of the automated vehicle.
The efects of diferent types of driving disengagement (resting vs. completing a non-driving related task (NDRT)) on driver response time to the takeover request are investigated. Specifically, rest, NDRT engagement, and eyes of the road are analyzed as diferent types of driving disengagement and as possible predictors of driver response time in takeover situations when taking control of the automated vehicle. The efects of two automotive user interfaces are also examined in terms of their complexity (basic instrument cluster (baseline) vs. advanced head-up display (HUD)) and their efects on takeover performance.
In what follows, we briefly present the related work. We then present the experimental setup of the simulated driving environment. The statistical analysis and machine learning are presented in the Results section. The article concludes with a discussion of the results of the presented research and the potential for future work.
Experimental studies in driving simulators show that user behavior in automated driving poses
several risks and safety implications related to takeover performance during the transition
to manual driving [
To this end, most previous research has focused on takeover requests (TOR) and driver
performance (in terms of safety, comfort) when transitioning to manual mode in critical situations
where response times (RT) are relatively short, even when the driver is engaged in NDRT [
The ability to regain control of the vehicle may depend on many factors, both situational (driving
conditions, engagement in secondary tasks) and human (age, gender, driving experience). For
example, Li et al. investigated the efect of age and driving disengagement on the takeover
performance in a driving simulation study (SAE level 3 [
An interesting insight into driver behavior and disengagement in HAVs (SAE level 3 [
The study was conducted in a simulated driving environment consisting of a motion-based
driving simulator [
The HUD was assessed for driving a conditionally automated vehicle (SAE level 3 [
Each trial featured four requests to turn on the ADS and four requests to take over control of
the vehicle. The requests to take control of the vehicle occurred due to both critical (e.g., a busy
pedestrian crossing or complicated intersection) and non-critical events (this was to simulate
the vehicle simply losing communication with the infrastructure or the vehicle sensor system
failure). The trial always started and ended in the manual driving mode, also referred to by [
The HUD evaluated in this study was developed based on the results of an exploratory study [9, 10] that provided insight into what and how information should be presented in a HUD for L3 vehicles. The HUD featured elements presented in two dimensions (2D) and using augmented reality (AR). Throughout the journey, the following elements were displayed on the HUD at automation level L0 and L3: • vehicle speed, • speed limit, • speeding, • available ADAS, • time to collision < 2 seconds, • level of automation the vehicle is in (L0 or L3), • display of important trafic signs 150 m before their location in the environment (stop, yield, pedestrian crossing, etc.), • GPS directional information directly on the road lane (via AR), and • short messages.
When automated driving was no longer available, the driver received a visual and auditory takeover request. The visual takeover request was displayed on the HUD 15 seconds before automated driving was turned of with a ”Takeover” sign and a numeric countdown indicating the time remaining until takeover. The auditory takeover request was played 5 seconds before the automation shutdown as a secondary reminder and as an additional notification to draw the driver’s attention to the request. The auditory notification was a pure 4000 Hz tone [ 11] played at a volume of 65 dB from the start of the takeover notification until the driver took control of the vehicle.
During the takeover request, the HUD displayed the following elements: • vehicle speed, • speed limit, • active ADAS, • ADL level (L3), • highlighting important road participants that may afect the takeover using read bounding boxes (via AR), • visual takeover notification and countdown (15 seconds lead time), and • auditory takeover notification (5 seconds lead time).
Twenty-eight drivers (14 women and 14 men) between the ages of 21 and 57 years (M=30.46, SD=10.67) and with a valid driver’s license (M=11.98, SD=10.24) participated in the study. The participants’ only task was to drive the vehicle safely and follow the GPS to reach the destination. Participation in the study was on a voluntary basis, and participants could end their participation at any time. As a thank you for their participation in the study, participants received a €10 gift voucher. The experimental design was prepared according to the rules and guidelines for experiments with human participants issued by University of Ljubljana [12]. 3.4. Data In the presented study, we focus only on the automated driving intervals. The following data are used for statistical analysis.
Dependent variable: • Response Time (RT). RT is measured (in ms) as the time interval from the request to take over the vehicle (triggered by the visual TOR in UI) to the participant taking over the vehicle in one of the three predefined ways (braking, steering the steering wheel or pressing the ADS button).
Independent variables: • Automotive UI. Two user interfaces with diferent levels of complexity were compared: Baseline (a typical instrument cluster - a head down physical dashboard, featuring speedometer, tachometer, fuel level indicator and indicator of the level of automation of the vehicle) vs. HUD, as described above; • Disengagement. Disengagement represents events in which the participant was not engaged in driving or focused on the driving situation. Two types of disengagement are compared: Rest vs. Task. Rest vs. Task is binary information. In rest, the participant is not attentive to driving and is not engaged in secondary tasks. If the participant was engaged in both at the time of the individual automated driving interval, the type of disengagement with the longer duration is chosen. • Age. Age is divided based on the age distribution of the participants into two classes: younger drivers (participant age < = 25 years) and older drivers (participant age > 25 years). This selection was first tested to obtain an even distribution of classes. • Gender: male and female.
The Shapiro-Wilk test showed that the dependent variable was not normally distributed (W=.879, p <.001). Several nonparametric statistical tests were used: Mann-Whitney U test for comparison between two groups, Kruskal-Wallis test for multiple groups, and pairwise Mann-Whitney test for post-hoc analysis (with Bonferroni correction for multiple interactions). Summary statistics of response time (RT) grouped by sex, age, and automotive UI type is provided in Table 1.
Several statistical analyzes were performed to examine the relationship between the independent variables and RT, as shown in Figure 1.
Mann-Whitney U tests were performed to examine possible efects of the independent variables on RT. No significant efects were found on RT for age, gender, and the type of automotive UI.
A significant efect on RT was found for the type of disengagement (Task vs. Rest), with the participants engaged in a task having longer RT (U=6294.50, p <.001, and efect size CLES=.70).
The analysis also showed a significant efect of the interaction between gender and driving disengagement on RT (H=29.08, p <.001). A pairwise Mann-Whitney test was used to analyze the interactions (Bonferroni correction was used for the interactions). A significant diference in RT was found between males (M=10921 ms, SD =4227.13) and females (M=8568.24 ms, SD =4192.73) when engaged in a NDRT (p<.001, Hedges’ g=-.55). Males had a significantly longer response time than females (diference in RT: M=2353.14 ms).
For machine learning, all continuous features (Task, Rest, Average Speed) were normalized and
transformed into a range [
LightGBM regression was chosen as an advanced machine learning algorithm that makes no assumptions about the normality distribution of the data. The target variable was RT with the following predictors: • the independent variables used in the statistical analysis (age, gender, type of automotive
UI), and additional variables related to disengagement and driving: • Task (NDRT; duration in ms): a disengagement variable. • Rest (duration in ms): a disengagement variable. • Eyes-Of-Road Ratio: a disengagement variable categorized into three levels: low, medium, high.
driving. • Simulator driving experience: 0=Never 1=Once 2=Few times 3=Multiple times. Simulator driving experience might have an efect on RT. • Average speed in automated mode (ms). • Timely Transition Count: a count of timely transitions (transition within the 15 second takeover request time before the automation was turned of by the vehicle) to manual • Reaction: reaction to transition request: 0=none, 1=brake, 2=steer, 3=accelerate, 4=other (e.g. pressing a button on the steering wheel).
The LightGBM model performed relatively well: Accuracy on the training set=.89, Accuracy on the test set=.73, Mean Squared Logarithmic Error (MSLE)=.01, and Mean Absolute Error (MAE)=.14. Figure 2 shows the importance of each feature for predicting RT. The feature importance is calculated with a ’split’ method used for tree-based models: the method counts how many times the tree nodes split on each feature, assuming higher importance for the features with more splits. In addition to average speed and age, the disengagement features Task, Rest, and Eyes-Of-Road Ratio were the most important predictors of RT. Interestingly, similar to the statistical analysis above, the type of automotive UI with two diferent levels of complexity (baseline vs. HUD) did not have a strong influence on RT.
Several other regression models were trained and evaluated. Figure 3 shows their performance
based on 2. 2 is a goodness-of-fit measure that measures the strength of the relationship the
model and the dependent variable on a scale [
There are several important findings from the presented research. The results show that
engaging in secondary tasks leads to significantly longer RT, which is consistent with several
existing studies [
Task
Age sEyesOffRR e r tua Rest eF SimExp
Gender Reaction 0.8 0.6 ^20.4 R 0.2 0.0
UI 0 16 20 107 40 60 80 Feature importance 100 120
The efects of disengagement on RT were found to be significant and men had significantly
longer RT than women while engaged in a task. However, no efects were found for age (unlike
in[
A comparison of the machine learning models shows that several of them perform well: The Random Forest Regressor, the Gradient Boosting Regressor, and the LightGBM Regressor had similar performance on the testing set. An interesting observation is the predictive performance of age in relation to RT, which was not significant in statistical tests. This might also be due to random efects within and between participants and should be further evaluated in future work.
The work presented in this paper was financially supported by the Slovenian Research Agency within the project Modelling driver’s situational awareness, grant no. Z2-3204 and program ICT4QL, grant no. P2-0246, and by the European Union’s Horizon 2020 research and innovation program for the project HADRIAN, grant agreement no. 875597. This document reflects only the authors view, the Innovation and Networks Executive Agency (INEA) is not responsible for any use that may be made of the information it contains. [8] Avsimulation. scaner studio, 2022. https://www.avsimulation.com/scanerstudio/. [9] K. Stojmenova, G. Jakus, S. Tomažič, S. Jaka, Is less really more? a user study on visual in-vehicle information systems in automated vehicles from a user experience and usability perspective, in: Proceedings of the 13th AHFE International Conference on Usability and User Experience, New York, USA, July 24-28, 2022. New York: AHFE Open Access, 2022. [10] K. Stojmenova, G. Strle, S. Jaka, Uporabnik ima vedno prav: uporabniška izkušnja, zaznana uporabnost in voznikove želje o zaslonih v pogojno avtomatiziranih vozilih. / the user is always right: user experience, perceived usability and driver’s preferences on userinterfaces in conditionally automated vehicles., in: Proceedings of the 31st International Electrotechnical and Computer Science Conference ERK 2022, Portorož, Slovenija, 19. - 20.
September 2022., 2022. [11] K. Stojmenova, F. Policardi, J. Sodnik, On the selection of stimulus for the auditory variant of the detection response task method for driving experiments, Trafic injury prevention 19 (2018) 23–27. [12] Guidelines for ethical behavior in research work with people, 2022. https://www.uni-lj.si/ raziskovalno_in_razvojno_delo/etika_in_integriteta_v_raziskovanju/.