Road safety is still a challenging issue. In 2020, 1.35 million people have died as a result of traffic accidents, where the number one cause of death for young adults between the age of 5 and 29 is car accidents. In an attempt to improve road safety, the automotive industry has developed numerous types of Advanced Driver Assistance Systems (ADAS). These systems are in general effective in improving safety. However, these systems will only be used if and only if drivers perceive the assistance as intuitive and cooperative. It is recently found that 61% of drivers sometimes switch off the assistance, 23% feel that current assistance are annoying and bothersome, whereas only 21% find them helpful. A safe system that is not used has no safety benefits. A promising way to improve driver acceptance and to increase safety is to employ haptic shared control (HSC), which is an effective way of keeping drivers in the active control loop. Support in the form of HSC benefits situation awareness and ensures effective monitoring of the environment and automation. However, torque conflict resulting from opposing intentions of driver and automation is reported to be a bottleneck for drivers' acceptance of HSC. Particularly, such conflicts are found to be most debilitating in curves. With each driver having an individual driving style, with different preferences and skill levels, the current standard 'one-size-fits-all' assistance approach to HSC, and driver support in general, is not satisfactory for every individual. An effective approach to increase acceptance in ADAS, and a reliable way to align the automation to the driver's preferences, is through personalisation. Here, personalisation is generally defined as 'making something suitable for the needs and preferences of a particular person'. For HSC, personalisation can be effectively realised by adapting the system's adopted trajectory to that of the driver. Therefore, the personalisation of HSC requires a driver modelling approach that predicts an individual driver's behaviour. Before this thesis, the personalisation of HSC was attempted by adjusting the gains of a corrective feedback HSC, as though it were a driver steering model itself. What was missing was 1) a HSC that allows for personalisation, i.e., a framework where a personalisable reference trajectory is independent of the haptic controller and, 2) a computational driver model or a data-driven driver classification approach that is able to describe individual drivers. When this thesis was started, a theoretical HSC concept, the 'Four-Design-Choice-Architecture' (FDCA) was introduced within our group. This promising concept was, however, not realised or implemented yet. As for modelling individual drivers, it was not known what type of driver steering and trajectory model(s) are suitable to generate personalised trajectories, if any, due to the lack of a standardised way to compare and evaluate the output performance of driver behaviour models with different structures and complexities. It was not known exactly how to achieve successful personalisation in curves, nor was the needed level of personalisation understood, i.e., adapting to the intricacies of each individual or adapting to a more general style. Moreover, whether personalisation in itself improves the acceptance of HSC systems, was still to be verified. These challenges are addressed in the four parts of this thesis: 1) Driver model assessment: The development of an assessment method and application on prominent control-theoretic driver models in the literature. %This was done to gain in-depth understanding of what is needed to model and describe individual drivers. 2) Driver trajectory classification: Understanding and categorising the types of individual driver trajectories present in the driving population. 3) Driver prepositioning: Understanding and modelling driver prepositioning behaviour, a behaviour found to be an essential, yet mostly overlooked aspect of curve-driving behaviour. 4) Application to Haptic Shared Control: Apply and evaluate personalised haptic shared control. This thesis has achieved it's highest level goal, which is to improve the acceptance of the haptic shared control driver support. This thesis provides an improved understanding and new insights into 1) how the novel FDC HSC has solved much of the acceptance issue put forward, and 2) an understanding of how to personalise with the FDC HSC. In terms of modelling tools and methods, this thesis has contributed with: 1) a model assessment procedure that can highlight the strengths and weaknesses of any control theoretic model, 2) a trajectory classifier, which can categorise different types of drivers, 3) a prepositioning path model, which, when combined with the Van Paassen control-theoretic driver model results in the first individual control-theoretic driver model, i.e., a model that can capture all main styles of individual driver behaviour and 4) the first personalisable HSC, where the developed modelling methods are applied to evaluate personalised haptic shared control. The findings and insights from this thesis have contributed to design guidelines and, can accelerate future research. Some examples include 1) using the individualised driver steering model, personalisation of ADAS can now be done in real-time, 2) using the developed trajectory classifier, explicit personalisation can be achieved, i.e., the driver can select the type of trajectory guidance he may want, and, 3) the driver trajectory modelling methods developed in this thesis can be used for the personalisation of path-planning in fully autonomous-vehicles. @en

This paper introduces a systematic assessment method which quantitatively assesses computational driver steering models with respect to their suitability for online identification of individual driver steering behaviour. This methodology is based on three criteria: (1) descriptiveness, the model's ability to capture different types of steering behaviour, (2) identifiability, the ability of the model for unique mapping between a steering behaviour and a parameter combination, and (3) realism, the parameter span resulting in realistic steering behaviour. The utility of the introduced assessment method is shown by analysing and comparing two driver models from literature which are based on the same high-level concept. Both models assume proportional control on a predicted lateral position, however one uses a linear prediction for lateral position and the other uses a nonlinear prediction. The proposed assessment method distinguishes between the performance of the models by showing that the nonlinear model outperforms the linear model in terms of descriptiveness (66% compared to 33% of the linear model), better inherent identifiability for steering angle (3.8 compared to 7.5), better inherent identifiability for lateral position (0.01 compared to 0.5), better curve-cutting experimental identifiability and a 2.72 times larger realistic parameter span allowing for more flexibility for parameter selection. This quantitative assessment method has successfully reflected the effect of merely altering the way the lateral position is predicted in two driver models. Thereby, this method can be used to give a fair assessment by giving a model an absolute classification that also allows for quantitative comparison with many more driver models.

@en

When taking a curve, drivers follow their own unique trajectory. Most driver style classifiers in literature are based on inertial inputs, denoting whether a given driver is aggressive or calm. However, this does not give any indication of a drivers trajectory style, i.e. whether a driver is curve cutting. To fill this void, this paper introduces a novel rule based classifier that categorises seven different trajectory styles. The classifier is applied to data from a fixed-base driving simulator study in which 45 subjects drove on three roads, comprising three different velocities: 25, 50 and 80 km/h, with three corresponding radii: 20, 80 and 204 m. The results show that some classes are more prevalent than others, with biased outer curve negotiation performed by a majority of the subjects and with no drivers classified as centerline drivers. The proposed trajectory classifier is shown to exhibit high levels of consistency, with 93% of drivers exhibiting consistent trajectory classes for at least 66% of the right curves driven and 84% exhibits consistent trajectory classes for atleast 66% of the left curves driven. Where this consistency indicates a potential for generalising the classification results to other curves. Additionally, this classifier can be used to adapt trajectory-driven advanced driver assistance systems, thereby serving as an alternative to driver modelling.

@en

Did you know that most drivers swing left before taking a right curve? In fact, this is a given for all race car drivers and a rule for efficient curve negotiation. This distinct way of approaching a curve is called prepositioning. In a recent study it is found that incorporating knowledge of this prepositioning phase is crucial for the reliability and acceptance of some trajectory-guiding advanced-driver-assistance-systems. Unfortunately, our understanding of prepositioning behaviour is still limited, with most driver models unable to account for this phenomenon. In an attempt to improve our understanding, the effects of changing velocity and road radius on prepositioning behaviour are studied experimentally in a fixed-base driving simulator. Twenty-four participants drove four conditions comprising two different fixed-speed velocities (50 and 80 km/h) and two different curve radii (204 and 350 m). The results show that the drivers' maximum prepositioning position significantly increases with increasing velocity and significantly decreases with increasing radius. With 88% of the runs exhibiting a significant displacement, i.e. larger than 0.05 m relative to a constant road bias. The findings suggest that drivers adjust their prepositioning behaviour to the Time-to-Line-Crossing (TLC) of the road environment, in an attempt to maximise TLC. Incorporating these findings in future driver modelling will bridge the gap between straight road and in-curve driving behaviour, thereby bolstering the descriptive capacity of these models.

@en

Vehicle control tasks require simultaneous control of multiple degrees-of-freedom. Most multi-axis human-control modeling is limited to the modeling of multiple fully independent single axes. This paper contributes to the understanding of multi-axis control behavior and draws a more realistic and complete picture of dual-axis manual control. A human-in-the-loop experiment was performed to study four distinctive phenomena that can occur in multi-axis control: performance degradation, axis asymmetry, crossfeed, and intermittency. In a simulator, three conditions were tested in the presence and absence of physical motion: the full dual-axis control task, single-axis roll task, and single-axis pitch task. Controlled element dynamics, stick dynamics, and forcing functions were equal in all cases. Results show that performance is worse in dual-axis tasks. Performance in roll axis is consistently worse than pitch, thereby proving axis asymmetry. Physical motion improves the performance and stability of the system. The application of independent forcing function signals in both controlled axes resulted in the detection of crossfeed in dual-axis tasks from spectral analysis. Using a novel extended Fourier coefficient method, the identified crossfeed dynamics can explain up to 20% of the measured control inputs and improves modeling accuracy by up to 5%. Dual-axis control behavior is less accurately modeled with linear time-invariant models and is more intermittent.

@en

For automated vehicles (SAE Level 2-3) part of the challenge lies in communicating to the driver what control actions the automation is taking and will take, and what its capabilities are. A promising approach is haptic shared control (HSC), which uses continuous torques on the steering wheel to communicate the automation’s current control actions. However, torques on the steering wheel cannot communicate future spatiotemporal constraints, that might be required to judge appropriate overtaking or obstacle avoidance. A visualisation of predicted vehicle trajectory, along with velocity-dependent constraints with respect to achievable trajectories is proposed. The goal of this paper is to experimentally compare obstacle avoidance behaviour while driving with the designed visualisation against driving with a previously designed HSC, as well as the two support systems combined. It is expected that adding visual feedback improves obstacle avoidance and user acceptance, and reduces control effort with respect to HSC only. In a driving simulator experiment, 26 participants drove three trials with each feedback condition (visual, HSC, and combination) and had to avoid obstacles that appeared with a Time to collision of either 1.85 s (critical) or 4.7 s (non-criticall). Results showed that, compared to HSC only, the HSC and visual combination yielded slightly smaller safety margins to the obstacle, a significant reduction of control activity on straights, and increased subjective acceptance rating. Visual and HSC offered a beneficial synergy, as it seemed the visual feedback allowed drivers to anticipate the effect of their steering actions on the car’s trajectory more accurately, and the HSC reduced the intra-subject variability. Future research should investigate the effects of added visual feedback in more detail, specifically in terms of the effectiveness to communicate automation capabilities and driver gaze behavior.@en

When drivers have opposing intentions to a haptic shared controller which, like the driver, can continuously control the vehicle through torques on the steering wheel, the driver has to fight against the controller torque to reach their goal. This phenomenon is called haptic shared control (steering) conflicts and are a reason for drivers to reject such automation. This study is the first to realise an implementation of the novel ”Four-Design-Choice-Architecture” design philosophy for shared control, hypothesized to reduce conflicts through its inherent control structure. The implemented haptic shared controller decouples reference trajectory from independent feedback and feedforward haptic control. The implemented Four-Design-Choice haptic shared controller is compared to the baseline (predecessor) Meshed haptic shared controller through a simulator experiment. The results show that the new shared controller significantly reduces occurrence of conflicts by a factor 2.3 and significantly reduces driver torque by a factor of 3.2. Analysis shows that the novel feed-forward haptic torque and a reference trajectory supporting the drivers future (curve-entry) intentions are the dominant players in conflict reduction. The findings show that the Four-Design-Choice-Architecture is proven very effective and has large potential to further reduce conflicts with different design settings.@en

In many practical control tasks, human controllers (HC) can preview the trajectory they must follow in the near future. This paper investigates the effects of the length of previewed target trajectory, or preview time, on HC behavior in rate tracking tasks. To do so, a human-in-the-loop experiment was performed, consisting of a combined target-tracking and disturbance-rejection task. Between conditions the preview time was varied between 0, 0.1, 0.25 0.5 0.75 or 1 s, capturing the complete human control-behavioral adaptation from zero- to full-preview tasks, where the performance remains constant. The measurements were analyzed by fitting a HC model for preview tracking tasks to the data. Results show that optimal performance is attained when the displayed preview time is higher than 0.5 s. When the preview time increases, subjects exhibit more phase lead in their target response dynamics. They respond to a single point on the target ahead when the preview time is below 0.5 s and generally to two different points when more preview is displayed. As the model tightly fits to the measurement data, its validity is extended to different preview times@en

While many realistic manual control tasks require human operators to control multiple degrees-of-freedomsimultaneously, our understanding of such multi-axis manual control has not moved far beyond considering it simply as the control of multiple fully-independent axes. This investigation aims to further our understanding of multi-axis control by focusing on one phenomenon that is known to occur in such tasks: crossfeed. Crossfeed occurs when operators’ inputs in one controlled axis feed into another controlled degree-of-freedom, thereby affecting overall control performance. A human-in-the-loop experiment, in which operators performed a dual-axis aircraft roll and pitch tracking task with physical motion feedback, was conducted in the SIMONA Research Simulator at TU Delft. Three conditions were tested: the full dual-axis control task, supplemented with reference single-axis roll and pitch tasks. Through the use of independent target and disturbance forcing function signals in both controlled axes, we were able to detect the presence of crossfeed in this dual-axis task from spectral analysis. Furthermore, these signals facilitated the objective identification of the dynamics of the crossfeed contribution, in parallel with estimating operators visual and motion responses. The crossfeed dynamics were found to resemble the well-known dynamics of human operators’ visual responses. The crossfeed contribution was found to explain up to 20% of the measured control inputs, thereby indicating that crossfeed can be a factor of significance in multi-axis manual control.@en