Discover more about the topics and technologies to be discussed at this year's conference, via a series of exclusive interviews with a selection of our expert speakers
Mark Eilers, principal scientist at Humatects GmbH, on a modeling suite for the development of probabilistic driver models and their utilization in a variety of use cases including driver intention recognition, traffic prediction and autonomous control.
What is your presentation about?
Advanced driver assistance systems require models for the monitoring, understanding, assessment and anticipation of human drivers. We present a modeling suite for the development of probabilistic driver models and their utilization in a variety of use cases including driver intention recognition, traffic prediction and autonomous control. We realize driver models as dynamic Bayesian networks that capture temporal evolutions and statistical relationships between driver’s state and behavior, vehicle and environment. Inputs are selected objectively from a variety of features proposed by psychological literature. The structure and parameters of the models are estimated offline from data or online using Bayesian machine learning approaches.
How does your modeling suite differ from others?
Understanding the human driver is essential for the development and personalization of driver assistance and automated driving systems. Our Driver Modelling Suite (DMS) is a software library for the development, utilization and machine learning of driver models. The driver models learned using our library can be used in driver assistance and automated driving systems for various purposes including driver intention recognition, human-like vehicle control and the personalization of existing assistance functionality.
Despite its importance, there are only very few modeling suites available to explicitly support driver modeling. General-purpose suites like R, MATLAB, Simulink and SCADE can be used for aspects of driver modeling, but lack the overarching toolchain to support a rapid development prototyping from raw data toward executable models and tangible results.
DMS has been developed using the feedback of automotive partners concerning the unique needs, requirements and use cases of the industry. DMS focuses on probabilistic driver modeling that can deal with the inescapable uncertainty of real-world applications. It allows the user to combine driver and environment in a coherent, dynamic, probabilistic model. The causal and statistical relations between a driver’s perception, state, and observable behavior and effects are described using probability distributions with clear semantics and interpretation. Using probability calculus, observable evidence can then be used to automatically reason about hidden information of interest, such as intentions and likely future behavior.
Since building a driver model heavily depends on the application for which the model will be used, we offer not only the DMS but also customer-specific modeling services. This includes personal support, assistance and adaptation. Every customer gets a guarantee that their requirements will be implemented immediately by our teams of modeling experts.
What are its advantages?
DMS has a lot to offer to support the task of driver modeling. It provides a variety of predefined and psychologically motivated input, output and driver state parameters, distribution types, and ready-made machine-learning algorithms for parameter and structure learning. It also provides tools and applications for data pre-processing annotation, runtime utilization and visualization, and the evaluation and diagnostics of driver models and parameters. Step-by-step example workflows and configurations enable easy access for beginners, while the powerful API provides complete freedom to experts.
DMS can learn driver models incrementally and even during runtime. In many cases, a full data set on driver behavior is not available right from the start; for example, when driver modeling is applied to predict overtaking intentions. In such cases, DMS allows starting with an intention model of an average driver. With every new overtaking maneuver driven by the driver, the DMS software updates the model and learns overtaking characteristics of the individual driver.
DMS allows smooth integration to third-party middleware (such as RTMaps) and driving simulators (such as SiLAB, Scanner). For this purpose, specific interfaces and plug-ins have been developed and are available.
How do you use psychology to select inputs?
DMS offers a large and growing set of psychologically motivated input, output and driver-state parameters. To collect these parameters and features, we constantly evaluate the psychological research literature and translate promising theories into machine-usable parameters and features. The DMS experts additionally investigate which of these parameters describe driver behavior in various cases and test their utility for driver modeling in empirical studies in driving simulators and real vehicles. This knowledge is then integrated into the DMS software library.
What do you mean by ‘probabilistic driver models’?
Human driving behavior is very dynamic and variable, and the underlying cognitive processes are inherently complex. It remains to be seen whether future research will be able to fully understand the details of the inner workings of the human mind. We therefore require models that can deal with the uncertainty arising from our lack of knowledge, imperfect observations, and inherent non-determinism of the real world and human behavior. In contrast to other approaches like control-theoretical models, cognitive architectures and neural networks, probabilistic models use the tools of probability theory as a formal framework to quantify and reason under uncertainty. By observing and recording data on drivers, we get a powerful key to learn the characteristics and regularities of driver behavior.
Don’t miss Mark Eilers’ presentation at the Future of Automotive Interiors Conference Europe. More information about the conference and its most current line-up can be found here.