This paper aims to intuitively display the details of drivers' visual perception and related driving behavior in the lane-changing process by developing a multi-view collaborative visualization-based lane-changing graph. Specifically, driving behavior data related to lane-changing process are extracted from a simulated expressway, which is carried out by a driving simulator. The lane-changing graph is developed by coordinating parallel coordinates, count diagram, and bar chart with lane-changing trajectory. Following the analysis of 40 data sets of lane-changing behavior using the multi-view technique and the criteria for qualified lane-changing area, the lane-changing behavior is then classified into"Qualified""Barely Qualified", and"Unqualified". Meanwhile, the reasons of the unqualified lane-changing processes are also studied. The results show that the proportions of"Qualified""Barely Qualified", and"Unqualified"processes are 10.00%, 12.50%, and 77.50% respectively. The average standard deviations of the turning speed of the steering wheel, acceleration, and lateral acceleration observed over the unqualified processes (6.57°; 0.91 m/s²;0.41 m/s²) are larger than those observed over the qualified processes (4.55°; 0.34 m/s²;0.17 m/s²). The reasons for showing unqualified processes are mainly twofold: excessive lateral acceleration due to a large turning angle of the steering wheel and excessive change of longitudinal acceleration due to inappropriate operation of the gas panel. In general, the lane-changing graph can analyze and diagnose the lane-changing process accurately, which can provide supports for optimizing driver behavior in the lane-changing process.

Based on current data collection techniques from connected vehicles, this paper aims to classify driving styles(i.e., driving habits or behavior)by analyzing driving modes which is defined as time headway in 3 s. Specifically, driving modes are quantitatively classified by time headway and typical driving modes reflecting each driving style are identified according to the calibrated driving styles. Evaluation score is assigned to each typical driving mode using a fuzzy classification method and the thresholds of each driving style is proposed based on the evaluation scores and the calibrated driving styles. The thresholds are applied to a test data set, which includes driving behavior data of 44 drivers, to verify the accuracy of the proposed method. In summary, three types of driving styles are identified: the evaluation score S < 64.67 is seen as the conservative driving style(CDS), the score 64.67 ≤ S < 181.20 is classified as the"regular"(that is, neither-conservation-nor-aggressive(NCNA))driving style; and the score S ≥ 181.20 is grouped into the aggressive driving style(ADS). Study results show that the accuracy of the proposed method against the testing data set is 85.7%. The proposed method uses simplified driving parameters (headway)for driving-style classification, which provides a new way for driving-style classification.

This paper aims to investigate the effects ofthe existence and content of information from connected vehicles and infrastructure (CVI) ondriving workload and behavior of young drivers at signalized and non-signalized intersections. Driving simulationsfor such intersectionsin urban areas are developed, in which 26 young drivers aged between 22 and 30 are involved. The results show that: such information can significantly reduce the workload of young driversand the increase in heart rate reduced by 1.95 beats/min for signalized intersections and 2.96 beats/min or 3.29 beats/min for non-signalized intersections, respectively. In addition, such information can significantly reduce the response time for braking actions of young drivers with 2.35 s at signalized intersections and 2.71 s or 2.09 s at non-signalized intersections respectively. It is also found that it can improve the stability of vehicles in reducing the standard deviations of vehicle speed by 31.33% for signalized intersections and 47.40% or 60.23% for non-signalized intersections, respectively. In addition, when thered phase of the vehicle moving direction at signalized intersections is about to end, the command information from CVI can significantly reduce the response time of young drivers by 3.47s, and the standard deviation of vehicle speed by 39.10%, compared to the effectiveness of regular instruction information.

This study aims to study the effects of high, medium, and low levels of immersion of non-driving-related (NDR)tasks on the takeover behavior under automated driving. Test scenarios are developed using a driving simulator. There are four takeover scenarios developed based on the task of non-driving-related activity(entertainment and work task)and time to collision(5 s and 10 s). To collect data on takeover behavior, volunteer drivers are recruited for a driving simulation. In order to measure takeover behavior, four metrics are used: speed, lateral deviation, responding time, and takeover correct time(TCT). The results show that: ①the vehicle speed decreases as the immersion level of NDR tasks decreases, and the deceleration rate increases after taking over the vehicle. When the time to collision is 5 s, the level of immersion of NDR tasks significantly affects lateral deviation. ②When the time to collision is 10s, the level of immersion of NDR tasks has a weaker significance on the takeover response time (p =0.056 < 0.1), and it decreases as the immersion level increases(low immersion level=3.94 s; medium immersion level=3.45 s; high immersion level=3.21 s). A significant difference between the level of immersion of NDR task and the time of takeover has been found(time to collision 5s: p =0.031 < 0.05;time to collision 10 s: p =0.019 < 0.05). That is, the time for takeover decreases as the level of immersionofNDR tasks increases. ③ With the same NDR task, the takeover responding time decreases as the immersion level increases. According to the results of statistical analysis, the interaction between NDR tasks and their immersion level does not affect the response time for takeover significantly, however, it does affect the TCT.

There are issues for the decision support method for automated driving based on reinforcement learning, such as low learning efficiency and non-continuous actions. Therefore, an end-to-end decision-making method for autonomous driving is developed based on the Twin Delayed Deep Deterministic Policy Gradient with Discrete (TD3WD)algorithm, which can be used to fuse the information from different action spaces over a network. In the network of traditional Twin Delayed Deep Deterministic Policy Gradient(TD3)algorithm, an additional Q network that outputs discrete actions is added to assist exploration training. Weighted fusion of the output actions of TD3 network and additional Q network is performed. The fused actions interact with the environment, in order to fully explore the environment and enhance the efficiency of the environment exploration. When the Critic network is updated, the output of the attached network is merged into the target actions as noise to encourage the agent to explore the environment and obtain better action estimates. Instead of the original images, image feature obtained from the pre-trained network is used as the state input to reduce the computational cost in the training process. The proposed model is tested under a set of simulated autonomous driving scenarios generated by Carla simulation platform. The results show that the convergence speed of the proposed method is about 30% higher than that of traditional reinforcement learning algorithms like TD3 and Deep Deterministic Policy Gradient(DDPG)under the training scenarios. Under the testing scenarios, the proposed method shows better convergent performances and the average rate of lane-crossing and the change rate of steering angle are reduced by 74.4% and 56.4% respectively.

Advanced driver assistance systems have been widely installed on various types of vehicles. To study the effectiveness of the collision warning systems, a field test was conducted on real-world expressways. A group of 15 vehicles underwent a comparative test equipped with or without the system. During the testing, a method for evaluating the effectiveness of the early warning system has been proposed from the following three aspects: interactions between vehicles, safety risk, and acceptance of the driver. Study results show that, at the microscopic level, under the scenarios that the vehicles are equipped with the system, when accidents take place during car-following and lane-changing for overtaking action, the average time headway (THW) observed increases by 0.37 s and 0.34 s, respectively. It is also found that using the system has a significant impact on the THW (p < 0.05). In contrast, at the mesoscopic level, the frequency of the above two types of accidents drops by 16.0% and 23.7%, respectively, as a result of early warning provided by the system. The dispersion degree of the vehicle speed in the group is also found to decrease significantly. The questionnaire survey shows that 86.7% of drivers will take safety measures after receiving the warningfrom the system, and 73.3% of drivers agree that the early warning system improves road safety.

Traffic accidents are strongly correlated with driving style, and driving style can be intuitively represented by driving behavior. In order to further advance understanding of the relationship between driving behavior and driving style, this paper explores thedifferences between driving styles and identifies factors that affect the classification. A driving-style recognition model is then proposed and evaluated. Based on the experimental data from connected vehicles, a K-means++ algorithm is proposed and used to classify data of driving behavior under different driving styles and a support vector machine-recursive feature elimination(SVC-RFE)and a random forest-recursive feature elimination(RF-RFE)algorithm are used to rank the importance of features of driving behavior. A classification model for driving styles based on neural network and the above selected features is developed. The results show that: ①when the number of selected features is set as n = 6, the correct ranking rate of both feature ranking algorithms is above 85% and the correct rate of the RF-RFEalgorithm is up to 90%.②The indicator with the highest importance in feature ranking is the maximum speed, and its difference among the three driving style groups is up to 10 m/s. ③When only the maximum speed is used as input, the accuracy of the driving-style recognition model is 86.1% and therefore, it can be concluded that maximum speed can effectively distinguish driving styles.