We offer a software library for human-like automated driving. We support development of highly automated Advanced Driver Assistance Systems (ADAS) that are trusted and accepted by human drivers.
The car industry strives to realize the dream of automated driving. An important challenge is to not forget the driver. Will the driver accept to be driven automatically? Will the driver be scared when the car drives very different from own personal style? Will the driver trust the automated car? Will the driver understand when he has to take over?
Trust and understanding can only be developed if the automated car communicates and interacts with the driver in an intuitive and transparent way. Thus, the HMI is of utmost importance.
Our solution: We offer a driver model which learns and predicts driver decisions and driver behaviour which can be used by automation systems to drive in a way that is understandable and predictable for the driver. Such automated behaviour is likely to be accepted by the driver. It will also be perceived as natural by other drivers and is a means to integrate highly automated cars swiftly into mixed traffic situations.
Internationally recognized: Our driver model is based on a scientific approach that has been published in more than 50 publications at conferences like IEEE Intelligent Vehicles Symposium and journals like Transportation Research.
Real-time performance: Our driver model is able to learn and perform individual driving styles in real-time
High Technology Readiness Level (TRL): Our driver model is currently tested in real cars. We are currently at TRL 5 and heading further towards 6. This work is done in cooperation with Centro Ricerche Fiat and Peugeot Citroen (PSA) in the European Project AutoMate.
We offer a software library implemented in C++ with functions to automatically learn individual driving preferences from driving data. In total, the software consist of approx. 118 classes with approx. 33000 lines of code. The software learns the structure and probability contributions of a Bayesian Network that is able to predict driving decisions and driving styles of human drivers. We offer services to integrate the software into your application for automated driving:
Step 1: In the first step we analyse the functional requirements for the driver model in your application. Our driver model can compute different types of output: driving decisions (e.g. when or if to overtake), speed profiles and steering profiles.
Step 2: In the second step we analyse the technical requirements for the driver model in your application, this includes e.g. the data input format, data output format, computation platform.
Step 3: In the third step we will produce a tailored configuration of the driver model according to your requirements.
Step 4: In the fourth step we will test the tailored driver model with driving data provided by our customer.
Step 5: In the fifth step we will support the technical integration of the tailored driver model in the customer's application and we will support any test activities in this environment.
Probabilistic Modelling: We use Dynamic Bayesian Networks to model probabilistic dependencies between driver's actions and information on the current driving and traffic situation. The model is defined as a (conditional) Dynamic Bayesian Network assuming first-order Markov and time invariance. The main assumption of our modeling approach is that complex human behaviour can be represented by a sequence and mixture of simpler sensor-motor schemata according to skill hierarchies. This hierarchical approach allows to compose a complex Bayesian Network from smaller simpler ones. Computation is done on the simpler models and then integrated according to the skill hierarchy. In this way, real-time computation is possible.
Machine Learning: The parameters and structure of component-models is learnt from multivariate time-series of human behaviour traces.
Our driver model is a software module that mimics driver behaviour (e.g. characteristic overtaking manoeuvres). It can be used to infer and predict characteristics of individual driving styles. The driver model acts as a virtual “clon” of the individual driver.