The development of Advanced Driver Assistance Systems (ADASs) has reached a stage where, in addition to the traditional challenges of path planning and control, there is an increasing focus on the behavior of these systems. Assistance functions shall be personalized to deliver a full user experience. Therefore, driver modeling is a key area of research for next-generation ADASs. One of the most common tasks in everyday driving is lane keeping. Drivers are assisted by lane-keeping systems to keep their vehicle in the center of the lane. However, human drivers often deviate from the center line. It has been shown that the driver’s choice to deviate from the center line can be modeled by a linear combination of preview curvature information. This model is called the Linear Driver Model. In this paper, we fit the LDM parameters to real driving data. The drivers are then clustered based on the individual parameters. It is shown that clusters are not only formed by the numerical similarity of the driver parameters, but the drivers in a cluster actually have similar behavior in terms of path selection. Finally, an Extended Kalman Filter (EKF) is proposed to learn the model parameters at run-time. Any new driver can be classified into one of the driver type groups. This information can be used to modify the behavior of the lane-keeping system to mimic human driving, resulting in a more personalized driving experience.

Automated driving systems are often used for lane keeping tasks. By these systems, a local path is planned ahead of the vehicle. However, these paths are often found unnatural by human drivers. In response to this, this paper proposes a linear driver model, which can calculate node points reflective of human driver preferences and based on these node points a human driver preferred motion path can be designed for autonomous driving. The model input is the road curvature, effectively harnessed through a self-developed Euler-curve-based curve fitting algorithm. A comprehensive case study is undertaken to empirically validate the efficacy of the proposed model, demonstrating its capacity to emulate the average behavioral patterns observed in human curve path selection. Statistical analyses further underscore the model's robustness, affirming the authenticity of the established relationships. This paradigm shift in trajectory planning holds promising implications for the seamless integration of autonomous driving systems with human driving preferences.

Modern lane centering assistance systems often use the mid-lane as a reference line for tracking. However, human drivers prefer to keep an offset to the mid-lane. Therefore, we have proposed in our previous work a driver model which calculates human-like offset values and enable the path planner to plan human-like paths. The driving data of 15 drivers have been analyzed. It has been revealed that drivers behave differently in left or right curves and choose to drive on either the left or the right side in straight road sections. Therefore, we extend our model to handle these traits. We have shown that the extended model outperforms the previous symmetric model.

Vehicles with driver assistance and even autonomous driving capabilities have become widely spread on the roads in the last decades. Replacing human drivers even partially is a complex issue, as these driving systems must be safe and reliable under various conditions. The complexity is increased further by the fact that driving systems must interact with human participants of the traffic: the driver and passengers of the given vehicle as well as drivers in other vehicles or pedestrians. The market trends show that even though more and more driving assistance solutions are available, the vehicle users often refuse to use them as the behavior of these systems feels unnatural. Therefore, manufacturers have initiated the development of personalized driving assistance which motivates research in driver modeling and driving style classification. One of the most controversial assistance function is lane keeping. In our paper, we propose various different model structures that are able to capture the lane offset selection behavior of human drivers. It is shown that a static linear regression model provides reasonable accuracy and robustness in modeling the lane offset. While Gaussian Process models offer more accuracy, their training time is more demanding. It is shown, that a combination of linear and polynomial third-order basis functions offers a good tradeoff to efficiently describe the lane offset selection of human drivers. Our observations are based on studying real driving data from 41 drivers, whose behavior is modeled with the considered model structures. Then, using the corresponding model parameters, drivers are clustered into two driving style groups. Using the collective data of these groups the models are fitted to all data records in the two driver groups. Any new driver can be classified into one of these groups, and the aggregated models can be further personalized to to increase user satisfaction with the lane keeping systems.

The evolution of Advanced Driver Assistance Systems (ADAS) has reached a stage of maturity where the social acceptance of these systems is becoming as important as coping with technical challenges. Most users complain that ADAS functions behave unnaturally, and they feel anxious because the driving system behaves in a way that is far from what the user prefers. One of the most controversial ADAS functions is Lane Keeping Assistance (LKA). This system detects the edge of the lane and drives the car actively by intervening on the steering wheel. Most available lane keeping follows the center of the lane, while the vehicle controller oscillates and often deviates from this line. Most people find this operation disturbing and unreliable. In our work, we propose a comprehensive Model Predictive Planner and Control that preserves the basic concept of how human drivers drive: they instinctively create the combined representation of the environment and the vehicle, and predictevely plan their action trajectory (i.e., steering angle). This method does not divide the driving task into planning and control, thereby improving tunability. On the other hand, the method provides direct relation to vehicle dynamics and high-level policies, and therefore it can be validated easier than end-to-end trained functions, such as neural networks. The initial algorithm is implemented in MATLAB and validated via simulation. The test script and the data can be accessed upon request¹¹https://github.com/jkk-research/jkk_controllers/tree/prototype/mppc