One of the main challenge of AutoMate project is to develop models that predict the driver status, behavior and intentions in a way that allows to validly assess if the driver is able to perform the tasks that are or will be assigned to her/him. Currently, research is focused on mostly isolated aspects within the different driver modelling categories, using a variety of different and sometimes incompatible modelling approaches. By now, no mature attempts have been made for incorporating the intra- and inter-dependencies between many different aspects of driver states, typologies, and/or control behavior into a profound human driver model. In this context, AUTOMATE advanced the state of the art by developing a probabilistic structure for modelling driver states, typologies, and control behavior in an integrated way, by unifying modelling approach based on Dynamic Bayesian Networks and including the information of driver's states (distraction and drowsiness) coming for the Driver Monitoring System (DMS) installed on the demonstrators. Furthermore, online learning algorithms have been added, in order to incorporate new observed data during runtime to continuously recalibrate the driver model and thus to adapt the TeamMate Car to the characteristics of individual drivers and to new situations.
Bayesian probabilistic formalism is also used to gather information on individual objects into one coherent model, by placing objects into relation with each other and to infer additional plausible information about objects based on the recognized scene. We have innovated object-tracking algorithms to be able to handle Multi-Object-Tracking, taking into account state uncertainties, existence uncertainties and contradicting information in a single algorithm.
In the past years much work was done in the field of driving maneuver planning and execution for automated vehicles. We have reused such automation functions, but upgraded them with new driver adaptive functionality. In particular, on one hand, the planning algorithms can incorporate and involve the driver in the driving task to generate a dynamically type of responsibility assignment. On the other hand, the system is able to learn from the driving style of the human driver in a wide range of situations. Thus, the system anticipates and incorporates inter- and intra-individual differences in driving style and driver decisions.
Another crucial challenge is to assess a huge number of possible evolutions of the traffic scene in real time. Current approaches to risk assessment aim at checking whether any given trajectory is feasible or will likely lead to a collision. However, these approaches only consider the safety of individual actions and only address the fully automated case. In order to consider dynamic task distributions between driver and automation in AUTOMATE, we have implemented situation-dependent corridors of safe actions.
The last innovation is about the Human-Machine Interaction (HMI) aspect, whose main goal is to keep the driver sufficiently in the loop or to get her/him back in the control-loop, according to her/his actual state and driving tasks. Currently, there has not been any research yet on finding the most comprehensive way to convey the rationale for autonomous actions to drivers (some studies exist on applying the Ecological Interface Design (EID) approach for communicating automation behavior, but this has only been achieved for isolated automation functions). AUTOMATE used EID in a completely new way to integrate all relevant information on the traffic, driver and automation, by showing safe driving corridors and constraints on these corridors using graphical means. Therefore, we are creating a Navigation-Centred Driving Cluster (NCDC), that significantly improves the initial concept by integrating all TeamMate relevant information (e.g. driver and automation state to intuitively show imminent risks, as well as distant hot spots where the vehicle may request the support of the driver). Moreover, we will research Personalised Multi-model Communication Preferences, in form of “concurred abbreviations”. The intention is to communicate “Why” information in a personalized way, since this can improve trust and acceptance in the automation, supporting the driver in the decision-making process, via different HMI channels.
