Understanding and modeling driver’s lateral control behavior accurately are essential for developing effective human–machine shared takeover (HMST) system during emergency collision avoidance under dynamic traffic environment. However, little attention has been devoted to modeling driver behavior during HMST. Moreover, existing methods mostly focus on the interaction between a single vehicle and the driver, making it difficult to effectively characterize the nonlinear and uncertain behavior of drivers in dynamic traffic environment. Therefore, this study focuses on modeling of driver lateral control behavior during HMST under dynamic traffic environment, and a Transformer-based modeling method with multi-head attention is proposed. To validate the performance of the purposed approach, comparative experiments against traditional modeling approaches are carried out. And twelve drivers are recruited to perform HMST tasks in different emergency collision avoidance scenarios based on a driving simulator. Results of comparative experiments of different driving scenarios demonstrate the superior accuracy of the proposed method. In addition, the shapley (SHAP) additive explanations is employed to enhance the interpretability of the proposed model, thereby providing a reliable foundation for the development of efficient HMST systems.

In the rapidly evolving field of Intelligent Transportation Systems, users’ low acceptance caused by psychological barriers has emerged as a significant obstacle to realizing the vision of a safe, efficient, and smooth mobility experience. Building on the foundation of technology acceptance research, this study introduces and validates the concept of situational acceptance within the context of automated driving, and establishes a two-dimensional model of acceptance based on a driving simulator study. This model comprises two dimensions: positivity (the tendency to accept or reject) and firmness (the strength of that tendency). Each dimension showed correlations with objective indicators, including decision-making timing, gaze fixation, pupil diameter, and eyelid opening, and these relationships were further explained by psychological theories. Furthermore, multiple machine learning models demonstrated strong classification performance on the feature set (four-class accuracy up to 0.6379 and binary accuracy up to 0.8608), further confirming the utility of the selected features. The incorporation of the two-dimensional model maintained or even enhanced classification accuracy while substantially reducing computational cost, with FLOPs decreased by up to 96%, laying the groundwork for developing a real-time, objective acceptance estimation method. Additionally, the study provides implications to enhance acceptance in interface transparency, motivational mechanisms, and adaptive personalization, offering actionable insights for developing real-time, adaptive in-vehicle systems.

This study addresses the gap in understanding drivers’ eye movement behaviors during automated lane changes and explores the application of driver monitoring in personalizing highly automated driving systems. Through a driving simulator experiment, this study thoroughly investigated the relationship between participants’ eye movement behaviors and subjective perceptions. 37 participants each experienced 48 different automated lane changes in the experiment. Their subjective feedback on situational trust and system aggressiveness was collected through questionnaires, and eye movements related to gaze, blink, saccade, fixation, pupil diameter, and eyelid opening were monitored. The analysis identified 25 significant eye movement features. The experimental results show that lower situational trust during automated lane changes correlated with increased attention to side mirrors, decreased attention to the front vehicle in the original lane, reduced blink frequency and suppressed saccade amplitude which increased the duration of effective visual information acquisition, smaller pupil diameter and greater fluctuations indicating increased cognitive load. The extracted features proved effective in building predictive models for drivers’ subjective evaluations. This study provides practical recommendations for selecting eye movement features and implementing personalized automated lane-change systems through eye movement monitoring. Future research will focus on expanding the dataset, improving predictive model performance, and exploring the application of these models.

To practically apply level 2 automated driving in complicated conditions including intersections where events that require manual interventions occur frequently, it is necessary to consider the influences of provided traffic information on driver behaviors. This study performed driving simulator experiments to evaluate the effects of two kinds of information on driver behaviors while interacting with oncoming vehicles at intersections, of which static information informed drivers of the approaching intersections, and sensor information offered the real-time object detection results of the system to drivers. It was observed that the distances to oncoming vehicles at takeover decreased when only the static or sensor information was displayed, and it could be improved when both kinds of information were provided, compared to the condition when no information was offered. Meanwhile, drivers’ feeling of safety significantly increased with the presentation of both kinds of information. The results indicated that the combination of the static and sensor information might improve drivers’ feeling of safety during level 2 automated driving, without delaying drivers’ intervention.

It is becoming a common scene to observe the usages of level 2 automated driving in our daily life. To ensure driving safety while using the level 2 automated driving systems in various traffic conditions, it is an essential issue to clarify the influences of the level 2 automated driving on driver behaviors. Previous studies normally focused on drivers’ reactions to emergency events while using level 2 automated driving. However, it is still unclear that how will drivers interact with the level 2 automated driving systems during the periods when no emergency event occurs. Therefore, a driving simulator experiment was performed, and the differences in driver behaviors, especially eye-gaze behaviors, under level 2 automated driving and manual driving were analyzed. Meanwhile, to simulate non-driving related tasks that may occur in real driving environment, a visual task, Surrogate Reference Task (SuRT), was applied in the experiment. It was observed that the percentage of gaze that fell within the road center area and speedometer significantly decreased, and the gaze time to the left and right mirrors, and the SuRT display significantly increased during level 2 automated driving. Meanwhile, the eyelid closure time were significant longer and the subjective evaluation scores of attention to front and surroundings were significant lower while applying level 2 automated driving. The results indicated that drivers’ attention levels, especially for the front areas, might be significantly reduced during level 2 automated driving, compared to that of manual driving.

The lane-change decision-making module of automated and connected vehicles (ACVs) is one of the most crucial and challenging issues to be addressed. Motivated by human beings' underlying driving paradigm and the convolutional neural network's (CNN) dramatic capability of extracting features and learning strategies, this article proposes a CNN-based lane-change decision-making method via the dynamic motion image representation. Human drivers take proper driving maneuvers after they subconsciously construct the dynamic traffic scene representation in their brains, so this study first proposes the dynamic motion image representation method to reveal informative traffic situations in the motion-sensitive area (MSA), which provides a full view of surrounding cars. Then, this article develops a CNN model to extract the underlying features and learn driving policies from labeled datasets of MSA motion images. Besides, a safety-constrained layer is added to avoid vehicle collisions. We build a simulation platform based on the simulation of urban mobility (SUMO) to collect traffic datasets and test our proposed method. In addition, real-world traffic datasets are also involved to further investigate the proposed method's performance. The rule-based strategy and reinforcement learning (RL)-based method are used to compare with our approach. All results demonstrate that the proposed method performs lane-change decision-making much better than prevailing methods, which suggests our scheme has huge potential to accelerate the deployment of ACVs and is worth further study.

Automated driving is transforming the nature of passenger interaction and user experience. In highly automated vehicles, establishing passengers' trust is essential for interactions and key to achieving a positive Quality of Experience (QoE). Improving system transparency is widely employed among the various methods for calibrating trust. However, increasing transparency can also harm QoE by increasing passenger workload. This research examines this trade-off through an experiment conducted in a driving simulation environment. Our findings reveal a positive correlation between system transparency and trust but a non-monotonic correlation with the workload. This suggests that there may be better options than maximizing system transparency for QoE design, and workload management should be considered to optimize system transparency. Based on our design practice and verification, we provide suggestions for improving system transparency, including information integration, synergy, feedback, and auxiliary prediction.

Thorough performance evaluation of automated vehicles (AVs) is an essential prerequisite for AVs' release and deployment. The challenges posed by dynamics performance appraisal of AVs are centered around the complexity of chassis dynamics, performance diversity, and lack of unified quantitative metrics. Therefore, this article proposes a novel quantitative evaluation metric for AVs' chassis-domain performance. We reveal mathematically explicit chassis steady boundaries of various vehicle maneuvers based on the modeling of chassis-domain dynamics and vehicle spatiotemporal signal analysis for safety-critical AVs. By defining and analyzing the multiperformance appraisal problem, this article gives mathematically prerequisites for evaluation metrics. Then, a rigorous metric is developed to quantify AVs' safety and comfort performance comprehensively. Wherein, the steady boundaries are leveraged to the metric normalization. We demonstrate the effectiveness of the proposed quantitative evaluation methodology in various scenarios. Test results illustrate that the proposed method provides a quantitative way to test AVs' integrated dynamics performance.

Haptic-shared steering (HSS) systems have been reported to enhance vehicle safety and reduce the workload for drivers. However, few studies have focused on modeling the lateral vehicle control behaviors of distracted drivers for HSS. The current study models this type of behavior in a series of high-fidelity driving simulator experiments with 18 participants. Two experimental conditions for a double lane change task are tested: HGT-Constant (haptic guidance torque with a constant gain) and HGT-Adaptive (haptic guidance torque with an adaptive gain). A gated recurrent unit (GRU) network is used to model lateral control behavior during driving. The effectiveness of the GRU network-based lateral control model of distracted drivers is benchmarked using a state-of-the-art long short-term memory network, a back propagation network, an extreme learning machine, and a traditional two-point visual model with neuromuscular dynamics. Experimental results indicate that the GRU network has the highest accuracy in terms of root mean square error, mean absolute error, mean absolute percentage error, and determination coefficient. In addition, a simulation model of the driver-vehicle-road closed-loop system demonstrates that the proposed model predicts driver behavior with acceptable lateral position error.

It is an essential issue for autonomous vehicles to keep driving safety while sharing spaces with other road users, especially moving pedestrians. An ideal trajectory planning of autonomous vehicles should be collision-free and feasible for the vehicles to execute. Previous efforts on trajectory planning mainly focused on traditional roads, and the research into pedestrian-vehicle interaction in shared spaces were still insufficient. This study proposed a prediction-planning collaboration method for autonomous vehicles to avoid collisions with moving pedestrians by predicting their motions, and generate feasible trajectories based on Frenet coordinate in shared spaces. Pre-crash scenarios were designed based on reported crashes, and simulations were performed to evaluate the collision avoidance performances of the proposed method. Meanwhile, validations with real world dataset were conducted to verify the practicability of the proposed framework in real driving environment. The results indicated that the proposed prediction-planning collaboration method could effectively predict the motions of moving pedestrians and ensure the safe interactions between autonomous vehicles and pedestrians in shared spaces.

Driving maneuver decision-making is critical to the development and mass deployment of automated vehicles (AVs). The prevailing approaches are stuck with either specific optimization objectives or separate maneuver planning. Inspired by the inherent driving mechanism of human beings, we develop a comprehensive deep learning scheme that extracts and establishes informative image-based representation of dynamic traffic flows and then learns driving strategies from collected driving datasets based on convolutional neural networks (CNNs). Our scheme can effectively capture comprehensive and dynamic features of traffic flows surrounding the ego car and construct dynamic motion images through processing the spatio-temporal signals of all neighbor cars. These constructed virtual images cover all informative dynamic states of the neighbor cars, located in the maneuver-critical area, defined as the motion-sensitive area (MSA). Then, sequential spatio-temporal images together with labeled driving behaviors are fed into our proposed CNN model. The network can extract underlying motion patterns and learn proper driving behaviors, including both the lateral maneuver and longitudinal speed. We demonstrate the effectiveness of the proposed scheme in a typical highway scenario. Results suggest that the proposed scheme merits further investigation to promote the launch of AVs.

Fast and accurate long-term trajectory prediction of surrounding vehicles (SVs) is critical to autonomous driving systems. In high-density traffic flows, strongly correlated vehicle behaviors require considering the interactions among multiple SVs when predicting their future trajectories. However, existing interactive prediction methods, most based on Long Short-Term Memory (LSTM), are suffering from slow prediction because they analyze SVs one by one and analyze trajectory sequence node by node. This paper presents a fast interactive trajectory prediction method called Structural Transformer which learns both spatial and temporal dependencies among multiple SVs in parallel. Specifically, our model first removes the internal states and loops of LSTM and replaces with a weighted self-reference mapping to realize parallel computation. Then, it embeds the relative spatial information of multiple SVs into trajectory states and reorganizes the self-reference mapping with neighbor-only interaction masks to achieve interactive prediction. Results on the NGSIM dataset show satisfyingly speed and accuracy performance on long-term trajectory prediction of multiple SVs. The longitudinal and lateral errors are reduced to 2.67m and 0.25m over 5s time horizon. The computational time of each step is only 12ms on a 2080ti GPU, which is over 4 times faster than the Structural LSTM.

In this article, the authors present a novel method to learn the personalized tactic of discretionary lane-change initiation for fully autonomous vehicles through human-computer interactions. Instead of learning from human-driving demonstrations, a reinforcement learning technique is employed to learn how to initiate lane changes from traffic context, the action of a self-driving vehicle, and in-vehicle user's feedback. The proposed offline algorithm rewards the action-selection strategy when the user gives positive feedback and penalizes it when negative feedback. Also, a multi-dimensional driving scenario is considered to represent a more realistic lane-change trade-off. The results show that the lane-change initiation model obtained by this method can reproduce the personal lane-change tactic, and the performance of the customized models (average accuracy 86.1%) is much better than that of the non-customized models (average accuracy 75.7%). This method allows continuous improvement of customization for users during fully autonomous driving even without human-driving experience, which will significantly enhance the user acceptance of high-level autonomy of self-driving vehicles.

A lot of research has been done to model and predict a driver's behaviors to improve driving safety. Inferring lane change maneuver can be a critical one among them. However, the lane change prediction problem is generally treated as a classification task in which the labels represent the probability of whether the driver will make a lane change in the upcoming few seconds. In our work, we formulate this problem as a regression task. The process of lane change behavior is analyzed to build a Gated Recurrent Units (GRU) network for predicting two time points during lane change behavior: a) When the driver will shift lane. b) When the lane change will be completed. We make a comparison of Long Short Term Memory (LSTM) network and Support Vector Machine (SVM) regression performance to show that our method can give a more precise prediction time. This work can be used to improve the safety performance of driver assistance systems and help other traffic participants having a safer environment.

One challenge of haptic guidance steering system is to be both efficient on reducing lane departure risk and as comfort as possible for driver acceptance. This paper focuses on design and evaluation of a comfort-oriented haptic guidance steering system via deep reinforcement learning for lane keeping purpose by numerical simulation. The design is based on a Deep Q Network to learn an end-to-end mapping from environmental observation and driver input to agent action values, with task reward as the only form of supervision. The evaluation is based on lane keeping accuracy measured by mean absolute lateral error from the centerline of lane and steering control activity measured by mean absolute steering wheel angle. By comparison, other two types of haptic guidance steering, namely continuous haptic guidance with constant gain and threshold-based haptic guidance are addressed. The results indicate that haptic guidance via reinforcement learning has the best performance on reducing steering control activity while remaining a high lane keeping accuracy. Furthermore, the proposed system shows its capability of being adaptive to individualized drivers. This paper suggests the potential of using deep reinforcement learning to design haptic guidance steering system for improving individualized driver safety and comfort.

This paper presents a driver model in shared control that focuses on driver interaction with haptic guidance steering for a lane-following task. A weight parameter in the driver model was developed to describe driver interaction and reliance on haptic guidance steering. A driving simulator experiment with 5 participants was conducted to identify the model parameters. One experimental condition was driving with the haptic guidance steering system, and another condition of manual driving was conducted as a comparison. The identification results showed the relationship between driver model parameters and characteristics of driver behavior, including the driver interaction with haptic guidance steering and the steering effort. Taking the degree of driver reliance on haptic guidance steering into account improved the fitness of modeling driver steering behavior. The proposed driver model in shared control driving was shown to be appropriate, as the model matched the driver's input torque with a fitness of 68% on average.

A hidden Markov model (HMM) has been applied to describe the dynamic process of driver fatigue over time. However, during the HMM training, the initial value selection of the confusion matrix has a great influence on the accuracy ofthe HHM, which brings a lot of inconveniences to the practical application. Therefore, this study applied a particle swarm optimization (PSO) algorithm to simplify the training process without affecting the accuracy of the HMM. In the beginning, an improved HMM was built based on the PSO algorithm, and then was trained by the adoption of the collected data of percentage of eye closing time over a certain period (PERCLOS), which was measured from twenty participants in a driving simulator experiment. Finally, the improved model was compared with the original model, and it didn’t require the initial value selection process based on the prior condition to achieve the global optimum and improve accuracy. It indicates that the proposed method can provide an effective way for driver fatigue identification.

The collision avoidance warning system, as an intelligent assistance driving technology, provides the driver a safety and comfortable driving experience, by transferring early warning signals to avoid collision while detecting potential danger. In view of the shortcomings of traditional safety distance models, this paper focuses on the driver style and environmental factors to establish novel models of warning distance, dangerous distance and expectation safety distance. The fuzzy control was used to analyze the relationship among driver style, environmental factors and vehicle velocity variable with the safety distance. Based on the relative distance and speed, a sliding mode variable structure control method was employed to get desired acceleration for the upper controller, and and a PID method was applied to design the lower controller. The tracking effectiveness of the controller was verified by numerical analyses. The simulating results show that the proposed system, which can enable the rear vehicle to follow the speed change of the front vehicle to remain the expectation safety distance, and to keep the running state safety.

This paper focuses on the effect of haptic guidance steering on driver lane following performance by simulation study and, in particular, different degrees of driver reliance on the assistance system are addressed. Driver interaction and reliance on the assistance system is represented by two parameters in a driver model: target steering angle to torque gain and neuromuscular reaction gain for haptic feedback. The haptic guidance system is designed based on a two-point visual model and applies a proportional-derivative control theory. The lane following performance with different degrees of driver reliance on the assistance system is compared and evaluated using three different sets of driver parameters, and a manual (unassisted) driving condition is conducted as comparison. Moreover, lane following performance is also investigated in the case of a system failure. The results indicate that the haptic guidance system is effective on improving lane following performance, and the performance is sensitive to the degree of driver reliance on the assistance system, especially in the case of a system failure. The cost-benefit analysis suggests that the driver determines the degree of reliance on the assistance system by balancing the improvement of lane following performance and maintenance of own driving strategy.

Visual feedback from the road ahead is required for steering a car. When visual feedback is limited or only partial road is visible, driver steering performance declines. To solve this problem, haptic feedback is expected to assist drivers by providing guidance torque on the steering wheel. This paper focuses on the effect of haptic guidance on driver steering performance during curve negotiation when visual feedback is limited. Twelve subjects participated in the experiment conducted in a high-fidelity driving simulator. Levels of haptic guidance were none, weak, and strong, and levels of visual feedback were whole, near, medium and far. The steering performance was assessed by drivers' turning maneuver when approaching and leaving curves, and time-to-lane crossing during curves. Results indicate that mean value of time-to-lane crossing decreased due to the implementation of strong haptic guidance when visual feedback was limited. The start point of turning maneuver was earlier resulting from the implementation of strong haptic guidance under the condition of visual feedback from near segment. In conclusion, the drivers tended to rely on haptic guidance to achieve better steering performance when visual feedback was limited.

Highly automated driving tremendously reduces workload of drivers. However, the drivers may lose their visual attention to road ahead due to the out-of-the-loop problem. Haptic guidance has been developed to reduce drivers' workload while keeping the drivers in the control loop. The haptic guidance system continuously provides assistant torques on the steering wheel so that both the driver and the system contribute to the steering input. This paper focuses on analyzing drivers' visual attention to road ahead and response to critical events when driving with different levels of haptic guidance. A simulator experiment with five participants was conducted. There were five driving conditions: hands free, hands ready, strong guidance, weak guidance, and manual driving. Autopilot driving was realized by the highest level of haptic guidance in the conditions of hands free and hands ready. Results indicate that percent road center was significantly higher in the condition of weak guidance compared to hands ready. Root mean square of steering torque was significantly lower in the condition of weak guidance compared to manual driving. Moreover, the drivers' response to the critical event indicates that lane change maneuver was smooth in the conditions of weak guidance and strong guidance.

Prolonged driving on monotonous roads often leads to a reduction in task load that causes drivers passive fatigue. Passive fatigue results in loss of driver alertness and is detrimental to driver safety. This paper focuses on the effect of a haptic guidance steering system on improving behaviors of passively fatigued drivers. By continuously exerting active torque on a steering wheel, the haptic system guides drivers to follow the centerline of a lane; meanwhile, the drivers sense the torque and interact with it while operating the steering wheel. An experiment was conducted with 12 healthy participants in a high-fidelity driving simulator. A monotonous driving course was designed, and vehicle speed was fixed in order to induce drivers' passive fatigue. A treatment session was arranged with the haptic guidance steering system, and a control session was conducted as a comparison. Driving performance, assessed by standard deviation of lane position, mean absolute lateral error, and time-to-lane crossing, was significantly improved when haptic guidance was activated. Results of physiological measures, including heart rate variability and percentage of eye closure, revealed that passively fatigued drivers were aroused when they were aware of the active torque on the steering wheel. In conclusion, the activation of haptic guidance can be regarded as an effective countermeasure for the passively fatigued drivers who have performed a prolonged monotonous driving task.

Haptic steering technology has been developed to support drivers for more accurate and reliable vehicular control. This paper investigates the evaluation of steering performance when a haptic guidance steering system is implemented for passive fatigued drivers. The haptic system continuously produces torque on the steering wheel to inform drivers about the lateral lane deviations; consequently, drivers are aware of the active torque and then contribute to the steering task by interacting with the system. An experiment with 12 participants was conducted in a high-fidelity driving simulator. In a within-subject counterbalanced design, one session included implement of haptic system and the other excluded haptic system. A monotonous long-period driving course was designed to induce drivers' passive fatigue. Electromyography signals of the brachioradialis muscles were measured to examine grip strength of a driver holding a steering wheel. Results show that the risk of lane departure was significantly reduced when the haptic system was activated. In addition, grip strength of the drivers decreased when cooperating with the haptic system.