Learning from errors in a driving simulation: effects on driving skill and self-confidence

Two experiments investigated the effect of making errors during training (error training) on a driving simulator versus learning from examples of errors (guided error training) on driving skill and confidence. Experiment 1 indicated that compared with errorless learning (where participants drove through a training run not designed to elicit errors), error training led to significantly better transfer to driving tests that were analogous to those situations encountered in training and more effective use of strategies for coping with a novel driving situation. Error training also reduced self-confidence in driving skill at the end of training relative to errorless learning. Experiment 2 provided weak evidence of the superiority of guided error training over errorless learning (where the driver in the video did not make any errors) on analogous tests, and no evidence of transfer to a novel test. Furthermore, guided error training did not influence self-confidence in driving skill. The potential value of driving simulators in providing active processing during driver training is discussed, along with the effects of passive and active exposure to errors on driver confidence.     

Relationships between driving simulator performance and driving test results

This article is considered relevant because: 1) car driving is an everyday and safety-critical task; 2) simulators are used to an increasing extent for driver training (related topics: training, virtual reality, human – machine interaction); 3) the article addresses relationships between performance in the simulator and driving test results–a relevant topic for those involved in driver training and the virtual reality industries; 4) this article provides new insights about individual differences in young drivers' behaviour. Simulators are being used to an increasing extent for driver training, allowing for the possibility of collecting objective data on driver proficiency under standardised conditions. However, relatively little is known about how learner drivers' simulator measures relate to on-road driving. This study proposes a theoretical framework that quantifies driver proficiency in terms of speed of task execution, violations and errors. This study investigated the relationships between these three measures of learner drivers' (n = 804) proficiency during initial simulation-based training and the result of the driving test on the road, occurring an average of 6 months later. A higher chance of passing the driving test the first time was associated with making fewer steering errors on the simulator and could be predicted in regression analysis with a correlation of 0.18. Additionally, in accordance with the theoretical framework, a shorter duration of on-road training corresponded with faster task execution, fewer violations and fewer steering errors (predictive correlation 0.45). It is recommended that researchers conduct more large-scale studies into the reliability and validity of simulator measures and on-road driving tests.     

Relationships between driving simulator performance and driving test results

This article is considered relevant because: 1) car driving is an everyday and safety-critical task; 2) simulators are used to an increasing extent for driver training (related topics: training, virtual reality, human-machine interaction); 3) the article addresses relationships between performance in the simulator and driving test results--a relevant topic for those involved in driver training and the virtual reality industries; 4) this article provides new insights about individual differences in young drivers' behaviour. Simulators are being used to an increasing extent for driver training, allowing for the possibility of collecting objective data on driver proficiency under standardised conditions. However, relatively little is known about how learner drivers' simulator measures relate to on-road driving. This study proposes a theoretical framework that quantifies driver proficiency in terms of speed of task execution, violations and errors. This study investigated the relationships between these three measures of learner drivers' (n=804) proficiency during initial simulation-based training and the result of the driving test on the road, occurring an average of 6 months later. A higher chance of passing the driving test the first time was associated with making fewer steering errors on the simulator and could be predicted in regression analysis with a correlation of 0.18. Additionally, in accordance with the theoretical framework, a shorter duration of on-road training corresponded with faster task execution, fewer violations and fewer steering errors (predictive correlation 0.45). It is recommended that researchers conduct more large-scale studies into the reliability and validity of simulator measures and on-road driving tests.     

Comparative evaluation of performance measures for human driving skills

In this paper, we evaluate the adequacy of several performance measures for the evaluation of driving skills between different drivers. This work was motivated by the need for a training system that captures the driving skills of an expert driver and transfers the skills to novice drivers using a haptic-enabled driving simulator. The performance measures examined include traditional task performance measures, e.g., the mean position error, and a stochastic distance between a pair of hidden Markov models (HMMs), each of which is trained for an individual driver. The emphasis of the latter is on the differences between the stochastic somatosensory processes of human driving skills. For the evaluation, we developed a driving simulator and carried out an experiment that collected the driving data of an expert driver whose data were used as a reference for comparison and of many other subjects. The performance measures were computed from the experimental data, and they were compared to each other. We also collected the subjective judgement scores of the driver’s skills made by a highly-experienced external evaluator, and these subjective scores were compared with the objective performance measures. Analysis results showed that the HMM-based distance metric had a moderately high correlation between the subjective scores and it was also consistent with the other task performance measures, indicating the adequacy of the HMM-based metric as an objective performance measure for driving skill learning. The findings of this work can contribute to developing a driving simulator for training with an objective assessment function of driving skills.     

Role of formal instruction and informal practice in learning to drive

Abstract

Learning to drive a motor vehicle is important for the mobility of the majority of people in industrialized countries. Although a great deal is known about the acquisition of psychomotor skills in the laboratory and in some other practical settings, there has been little scientific study of learning to drive. One major practical question concerns the value of formal tuition with a qualified instructor and informal practice with friends or relatives. This paper reports a cross-sectional study of 805 learner drivers in the UK who had undergone at least 5 h of formal tuition to examine associations between their history of formal tuition and practice and current levels of skill and confidence as assessed by both pupils and their instructors. The results indicated an increase in instructor ratings of pupil skill with both increasing practice and tuition. However, the relationship between tuition and instructor-rated skill was only observed in pupils who had had no practice. Pupils' self-confidence did not increase with either tuition or practice; instructors' feelings of confidence in and safety with the pupil increased with pupils' prior hours of practice but not tuition. Instructors' ratings of the likelihood that pupils would pass the driving test first lime were positively associated with prior hours of practice and negatively associated with prior formal tuition. The results suggest that informal practice constitutes an important element of the process of learning to drive. Longitudinal and experimental studies are now required to confirm this finding and to establish whether recommendations should be made for a component of driving tuition involving supervised practice.     

On-line driver workload estimation. Effects of road situation and age on secondary task measures

In order to develop a driver-car interface that adapts the presentation of messages generated by in-vehicle information systems to driver workload, two experiments investigated potential determinants of driver visual and mental workload as indicated by performance on two secondary tasks. Experiment 1 suggested that road situation is a major determinant of visual and mental workload of the driver and that the processing resources of older drivers are somewhat more limited than those of younger and middle-aged drivers. Familiarity with the area of driving (when guided) and time of day (associated with traffic density) showed no secondary task effects. Experiment 2 showed that the categorization of road situations, proposed in Experiment 1, could underlie adaptation of visually loading messages to the workload incurred by driving. This was not found with respect to mentally loading messages.     

Components of driving skill: experience does not mean expertise

Using an instrumented car driven in normal traffic, we assessed the driving skills of trained experts, of normal, experienced drivers, and of novices. Previous research suggests a fairly simple picture of improvement in driving skills with experience, and for aspects of car control (e.g., steering path, speed of manoeuvres) our results confirm this: normals largely resembled experts, while novices performed more poorly. There are reasons to suspect, however, that experience may not always be so beneficial. For example, one important variable may be feedback, or information indicating to the driver that a particular action is or is not important in terms of overall goals (progress, safety, etc.). Where feedback is good, simple experience may bring expertise, but where feedback is poor, skills may fail to improve or even deteriorate once explicit tuition is removed. Correspondingly, our findings showed that for scanning patterns (e.g., mirror checking), anticipation (e.g., braking into an intersection), and safety margin (e.g., close following on the motorway), it was often the normal, experienced drivers who performed worst, novices sometimes even resembling experts. The data make it clear that for many aspects of driving skill experience is no guarantee of expertise.     

Information fusion for robust ‘context and driver aware’ active vehicle safety systems

Although there is currently significant development in active vehicle safety (AVS) systems, the number of accidents, injury severity levels and fatalities has not reduced. In fact, human error, low performance, drowsiness and distraction may account for a majority in all the accident causation. Active safety systems are unaware of the context and driver status, so these systems cannot improve these figures. Therefore, this study proposes a ‘context and driver aware’ (CDA) AVS system structure as a first step in realizing robust, human-centric and intelligent active safety systems. This work develops, evaluates and combines three sub-modules all employing a Gaussian Mixture Model (GMM)/Universal Background Model (UBM) and likelihood maximization learning scheme: biometric driver identification, maneuver recognition, and distraction detection. The resultant combined system contributes in three areas: (1) robust identification: a speaker recognition system is developed in an audio modality to identify the driver in-vehicle conditions requiring robust operation; (2) narrow the available information space for fusion: maneuver recognition system uses estimated driver identification to prune the selection of models and further restrict search space in a novel distraction detection system; (3) response time and performance: the system quickly produces a prediction of driver’s distracted behaviour for possible use in accident prevention/avoidance. Overall system performance of the combined system is evaluated on the UTDrive Corpus, confirming the suitability of the proposed system for critical imminent accident cases with narrow time windows.     

Analysis and modeling of human driving behaviors using adaptive cruise control

We describe a driver model based on the feedback-error learning scheme using neural network (NN) for adaptive cruise control (ACC) use in driving and show the applicability of the feedback-error learning scheme as a behavior model of adaptability of human. The focus of the study is on the adaptation process of driving behaviors using ACC. The driver model for computer simulations was implemented by using a NN. The developed simulation model is used for predicting control performance of skilled driver using ACC. In the experiments, we used a fixed-base driving simulator (DS) installed ACC system for collecting driver’s data. Headway time when lane changing in a row and lateral deviation from road center were investigated as driver behavior characteristics during ACC use and manual driving, respectively. The simulation results of the lateral deviation were compared with the experimental results and showed that control performance with ACC use would be better than that of manual driving. Furthermore, it was found that human error occurred during ACC use in the DS experiments and provided an explanation of the cause in terms of Neisser’s perceptual cycle model.     

The relationships between organizational and individual variables to on-the-job driver accidents and accident-free kilometres

Highway fatalities are the leading cause of fatal work injuries in the US, accounting for approximately 1 in 4 of the 5900 job-related deaths during 2001. The present study focused on the contribution of organizational factors and driver behaviours to on-the-job driving accidents in a large Western Canadian corporation. A structural equation modelling (SEM) approach was used which allows researchers to test a complex set of relationships within a global theoretical framework. A number of scales were used to assess organizational support, driver errors, and driver behaviours. The sample of professional drivers that participated allowed the recording of on-the-job accidents and accident-free kilometres from their personnel files. The pattern of relationships in the fitted model, after controlling for exposure and social desirability, provides insight into the role of organizational support, planning, environment adaptations, fatigue, speed, errors and moving citations to on-the-job accidents and accident-free kilometres. For example, organizational support affected the capacity to plan. Time to plan work-related driving was found to predict accidents, fatigue and adaptations to the environment. Other interesting model paths, SEM limitations, future research and recommendations are elaborated.     

司机分心驾驶检测研究进展

随着车辆工业和世界经济的快速发展,私家汽车数量不断增加,导致交通事故越来越多,且交通安全问题已经成为全球关注的焦点问题。司机分心驾驶检测的研究主要分为传统计算机视觉（CV）算法和深度学习算法两种。基于传统CV算法的司机分心检测通过尺度不变特征转换（SIFT）、方向梯度直方图（HOG）等特征算子提取图像特征,然后结合支持向量机（SVM）建立模型并对图像进行分类。然而传统CV算法具有对环境的要求高、运用范围较窄、参数多、计算量大的缺点。近年来深度学习在提取数据特征方面表现出速度快、精度高等优异的性能,因此研究人员开始将深度学习引入到司机分心驾驶检测中。基于深度学习的方法可以实现端到端的司机分心驾驶检测网络,而且取得了很高的准确度。介绍了传统CV算法和深度学习算法在司机分心驾驶检测的研究现状,首先,阐释了传统CV算法用于图像领域和司机分心驾驶检测研究的情况;接着,介绍了基于深度学习的司机分心驾驶研究;而后,从准确度、模型参数量等方面对不同司机分心驾驶检测方法进行比较分析;最后,对现有的研究进行了总结并提出了未来司机分心驾驶检测需要解决的三个问题：驾驶过程中司机分心状态以及分心程度划分规范需进一步完善,需要综合考虑人－车－路三者以及如何才能更有效地减少神经网络参数。     

Speed-of-processing and driving simulator training result in improved driving performance

Useful field of view, a measure of processing speed and spatial attention, can be improved with training. We evaluated the effects of this improvement on older adults' driving performance. Elderly adults participated in a speed-of-processing training program (N = 48), a traditional driver training program performed in a driving simulator (N = 22), or a low-risk reference group (N = 25). Before training, immediately after training or an equivalent time delay, and after an 18-month delay each participant was evaluated in a driving simulator and completed a 14-mile (22.5-km) open-road driving evaluation. Speed-of-processing training, but not simulator training, improved a specific measure of useful field of view (UFOV), transferred to some simulator measures, and resulted in fewer dangerous maneuvers during the driving evaluation. The simulator-trained group improved on two driving performance measures: turning into the correct lane and proper signal use. Similar effects were not observed in the speed-of-processing training or low-risk reference groups. The persistence of these effects over an 18-month test interval was also evaluated. Actual or potential applications of this research include driver assessment and/or training programs and cognitive intervention programs for older adults.     

The limits of path error-neglecting in straight lane driving

Abstract

It is argued that driving cannot simply be considered as a permanent, closed-loop task. Time-to-line-crossing (TLC) is used as a measure to quantify the potential role of visual open-loop and path-error-neglecting strategies. Basically, TLC represents the time available for a driver to neglect path errors until the moment at which any part of the vehicle reaches one of the lane boundaries. The strategy adopted by drivers during error-neglecting should be represented in terms of decision rules, describing how drivers switch from error-neglecting to error-correcting when approaching the edge of a lane. The experiment to be presented in this paper was designed to provide these rules for a straight lane-keeping task. Drivers were instructed to neglect the vehicle path error and to switch to error-correcting only at that moment when the vehicle heading could still comfortably be corrected to prevent a crossing of the lane boundary. The results show that the lateral distance from the lane boundary at which drivers switch to error-correction increases about linearly with the lateral approach speed. This mechanism results in an approximately constant TLC (time) distance at the moment of decision: this result being consistent over a broad range of speeds.     

Learning headway estimation in driving

OBJECTIVE: The main purpose of the present study was to examine to what extent the ability to attain a required headway of 1 or 2 s can be improved through practical driving instruction under real traffic conditions and whether the learning is sustained after a period during which there has been no controlled training. BACKGROUND: The failure of drivers to estimate headways correctly has been demonstrated in previous studies. METHODS: Two methods of training were used: time based (in seconds) and distance based (in a combination of meters and car lengths). For each method, learning curves were examined for 18 participants at speeds of 50, 80, and 100 km/hr. RESULTS: The results indicated that drivers were weak in estimating headway prior to training using both methods. The learning process was rapid for both methods and similar for all speeds; thus, after one trial with feedback, there was already a significant improvement. The learning was retained over time, for at least the 1 month examined in this study. CONCLUSION: Both the time and distance training of headway improved drivers' ability to attain required headways, with the learning being maintained over a retention interval. The learning process was based on perceptual cues from the driving scene and feedback from the experimenter, regardless of the formal training method. APPLICATION: The implications of these results are that all drivers should be trained in headway estimation using an objective distance measuring device, which can be installed on driver instruction vehicles.     

Effects of an in-vehicle collision avoidance warning system on short- and long-term driving performance

Many new in-vehicle systems focus on accident prevention by facilitating the driving task. One such driving aid is an in-vehicle collision avoidance warning system (IVCAWS), used to alert the driver to an impending collision. Our study evaluated the effects of an imperfect IVCAWS both on driver headway maintenance and on driver behavior in response to warning system errors. Our results showed that drivers tend to overestimate their headway and consequently drive with short and potentially dangerous headways, and that IVCAWSs are a useful tool for educating drivers to estimate headway more accurately. Moreover, our study showed that after a relatively short exposure to the system, drivers were able to maintain longer and safer headways for at least six months. The practical implications of these results are that the use of an IVCAWS should be considered for inclusion in driver education and training programs.     

Learning to make errors: evidence from a driving task simulation

Human errors represent a mismatch between the demands of an operational system and what the operator does. If they cannot be reversed, their consequences may be severe. Errors are frequently classified as design-or operator-induced. A third class of errors may also be identified, namely process-induced errors. Such errors arise out of on-going processes which typically extend over time. One such process is that of learning. In relation to the acquisition of skills, for example, learning frequently involves a trial-and-error component. Accidents by inexperienced drivers may represent a severe consequence of such errors. Errors may also arise out of particular learning experiences which provide a distorted underestimate of objective risk and/or motivate high risk behaviour. These phenomena are investigated in a computer simulation of the driving task. The relationship is discussed between various kinds of learning experience and the development of situations in which the possibility of error recovery declines. Some suggestions for reducing the frequency of irreversible errors and for increasing the data base for human error in vehicle driving are made.     

Drivers' errors in,and out of,context

Error depends for its definition, commission, and the seriousness of its consequences on the circumstances in which it occurs. As such, it is argued, in this overview of a large number of contemporary papers on (driver) error, that an erroneous act is only a useful index of behaviour where the background to that act is properly understood. The role of error in the development of skill, and its relationship to accident causation and risk-taking is discussed from this point of view.     

Errors and driver support systems

Recent technological developments seem to pave the way to sophisticated electronic co-driver systems that may help automobile drivers to cope with an ever increasing information load, to avoid certain errors, and to recover from others. GIDS—which stands for Generic Intelligent Driver Support—is a research project (under the EEC DRIVE Programme) to study the feasibility of an adaptive co-driver system. The conceptualization of a GIDS system requires close attention to performance errors as they may occur in certain subtasks of the driving task. One important issue that should be considered in some detail is that GIDS may eliminate errors as well as introduce them. Should various types of errors be represented formally and, if so, how they should be represented in order that GIDS can detect and cope with behavioural errors that drivers are likely to make under certain conditions? The requirements imposed by the project's goal to actually implement various driver support functions into a GIDS system is imposing tight constraints on error definition and identification. Some of the requirements will be discussed in terms of Soar. Soar is an intelligent computer architecture which is the embodiment of the theory of human problem solving formulated by Newell and Simon (1972). To the extent that the driving task is representable in Soar, the error theory that is required for any type of GIDS system to function must also be representable in Soar.     

Aberrant driving behaviour: errors and violations

The present study was a replication of the research of Reason et al. (1990). Its aim was to confirm the distinction between driving errors and violations in a Western Australian driving population. Sixty-one male drivers and 74 female drivers completed a questionnaire containing items on driver demographics, driving penalties incurred, driving convictions and accident history and driver behavioural aberrations drawn from the Driver Behaviour Questionnaire (DBQ). In agreement with Reason et al. factor analysis revealed three factors; in the present study these were general errors, dangerous errors, and dangerous violations. Young drivers committed more dangerous errors and dangerous violations than older drivers. Females reported more dangerous errors than males. Males reported more dangerous violations than females. Drivers who reported a high level of road exposure and those who reported having been convicted for speeding reported more dangerous violations. Differences in the results of the two studies can largely by accounted for by differences in the representation of age and gender in the two populations studied.     

Impacts of anger on driving performance: A comparison to texting and conversation while driving

Traditionally, driver distraction has been categorized into four types: visual, biomechanical, auditory, and cognitive. However, the place of emotion in driving research is largely undefined. The present study investigates the specific influences of anger – representative emotion arisen while driving, on driving performance, compared to those of traditional distraction tasks. In total, seventy-eight participants were recruited and placed into one of four driving conditions: physical (visual-biomechanical) distraction, cognitive (cognitive-auditory) distraction, emotional (anger), and control conditions. The results demonstrated that anger degrades driving performance as much as or more than other distraction types, specifically, in a yellow traffic signal situation. The causes for these results, underlying mechanisms, and other considerations are discussed with implications for future research.