Design and implementation of a torque-based predictive steering assistance for human-centered and safe automated driving

Road traffic deaths continue to be a major global public-safety and health problem. According to the report ofWorld Health Organization (WHO), globally more than 1.24 million people die every year due to road traffic crashes on the other hand between 20 to 50 million people suffer non-fatal injuries and disabilities. The automotive industry and researchers have focused on developing the vehicle safety systems in order to mitigate the severity of injuries and even to avoid the collisions either by issuing warning signals or intervening the vehicle dynamics. With the advances in sensor, actuator and computing technologies, the automation is introduced in vehicle control through systems such as Anti-lock Braking System (ABS) and Electronic Stability Control (ESC) which are proven to increase vehicle safety and efficiency. Although the human errors are the ultimate cause of the collisions, the humans are reluctant to give a complete authority to an automated driving technology because of the reliability issues. In order to support the drivers in such demanding tasks, the advanced driver assistance systems (ADAS) are designed to provide various functionalities while keeping the driver in the center of authority. One particular commercialized implementation of ADAS is the lateral vehicular safety systems such as lane departure prevention (LDP), lane keeping assistance system (LKAS) and blind spot intervention (BSI). The human-in-the-loop nature of these systems makes the design problem non-trivial and challenging since the influence of automation on the driver should be considered in terms of human factors thoroughly. There is a vast of human factor studies investigating the conflicts arising in human-machine systems and one approach has received considerable attention in academic research to fulfill the desired design guidelines. Haptic shared control concept is build on the idea of two systems (e.g., human operator and automation) acting together on the manual control interface and performing the same task cooperatively. This framework aims to provide a bi-directional information flow between both systems while increasing the efficiency and reducing the workload of human operator. One of the most popular implementation in automotive research is the haptic guidance system which shares the steering control in order to support driver in lane keeping/changing and obstacle avoidance maneuvers. These systems communicate with the driver through feedback forces and so the demanding visual workload of the driver is relieved by utilizing the under-used haptic sensory channel. The various studies show that these systems are effective under routine driving tasks when the goals of both systems meet but there is still some topics which should be investigated in order to analyze the effectiveness of these systems during emergency maneuvers and/or conflict situations. This thesis investigates the open issues in the literature of shared steering control problem by proposing an assistance system design based on the model predictive control (MPC) approach. MPC is a model-based optimal control routine which employs receding horizon technique. In this approach, an optimal input sequence is calculated by minimizing a cost function which depends on the future prediction of state trajectories that are calculated by using a mathematical model of the plant. This constrained optimal control problem is solved at every time step for a finite time horizon. Only the first input of the optimal sequence is applied to the plant and the optimization routine is solved at next time step. The design of the predictive controller could be separated into four elements: A prediction model, a user-defined objective function, the constraints and tuning of the associated weights and prediction horizon. We emphasis the trade-off between the control performance and the complexity of the optimization problem with respect to the prediction model that is used. Using a relatively simple but yet accurate model could decrease the complexity while not deteriorating the control performance. In this thesis, we design and implement a driver steering assistance system (DSAS) which utilizes the torque-based steering actuator (i.e., a torque overlay is superimposed instead of angular position) of the vehicle in order to generate a torque guidance for safe and assisted automated driving. Most of the previous studies design the guidance forces through the combination of a preview controller, which calculates the optimal steering angle, and a proportional feedback loop. The tuning of these forces are heuristic and it may lead to conflicts and increase in physical workload even in routine driving tasks. The studies show the affects of neuromuscular system (NMS) properties of the driver's arms while interacting with the guidance forces. The adaptability of these properties is a challenging issue in designing such systems. In order to over come these issues, we consider the NMS properties of the driver's arms coupled with the steering system and we propose a novel driver-in-the-loop steering model whose parameters are identified online by using a recursive least squares (RLS) scheme. This allows the controller to adapt the dynamic changes in the human-machine system and update the parameters of the prediction model in order to optimize a more suitable guidance force. We also design the objective function such that there is a trade-off between amount of utilized guidance forces and amount of deviation from the optimal trajectory. If the driver's and system's goals are matched (i.e., no conflict), the controller utilizes a suitable amount of guidance with respect to NMS properties. However when the goals do not match (i.e., conflict) as the controller increases the guidance torque and the driver opposes it by increasing the impedance, the controller decreases the guidance torque gradually since the input cost dominates the tracking cost in the objective function. On the other hand, if the driver's actions are most likely to cause a safety issue the system intervenes at all cost whether the driver accepts or opposes. We first test the proposed assistance system through closed-loop simulations in lane departure and blind spot intervention scenarios. The results show that when the NMS properties are known correctly in the prediction model, the control performance increases with minimum intrusion and control effort. Next, the system is validated in a test vehicle which is instrumented with a production level electric power assisted steering (EPAS) system which is called motor driven power steering (MDPS) system. We performed the experiments with human drivers in lane departure scenarios. The results show the effectiveness of the proposed system in adapting with respect to driver's response. The haptic guidance is generated according to how the driver interacts with the system's activity. A better control performance is achieved by updating the prediction model in the controller. We also conducted an experimental study with five participants to validate the performance. The results show that the assistance system adapts the guidance torque systematically with respect to the interaction level. This approach could resolve conflicts between the driver and the controller during routine driving tasks when there exists disparity in control objectives.