Autonomous vehicles must be able to perceive its environment and react adequately to it. This environment recognition must be able to interpret the movements of other road users, such as pedestrians, and derive intentions for their further behaviour. In highly automated vehicles, these tasks are increasingly performed by artificial intelligence (AI). Such AI functional modules based on machine learning are thus developing into a key technology.
One of the biggest challenges in integrating these technologies into highly automated cars is to ensure the usual functional safety of previous systems. Existing and established safety processes cannot simply be transferred to machine learning methods.
“To me, enabling artificial intelligence to train algorithms on the basis of virtual worlds is the top class of digital visualization. While many of our previous use cases only required perfecting visual quality, there is another component, the physical correctness of a virtual world. In the course of the further development of autonomous driving, there are still many hurdles that we have to overcome together. On the one hand, the focus is on the sheer mass of physical test kilometres, which we are successively replacing with virtual simulations. On the other hand, we have the possibility to implement scenarios and situations based on our virtual world that are very difficult or impossible to test in the real world. These two aspects exemplify the advantages of this technology. To contribute to this is a great opportunity for us to show how versatile data-based visualization can be used.”
Kian Saemian, Vice President Future Technologies, Mackevision – Part of Accenture Interactive
In order to solve this challenge, the research project AI protection was launched on July 1, 2019, and consists of 25 partner companies and institutions under the leadership of Volkswagen AG. The consortium is pursuing the goal of establishing a stringent and verifiable chain of arguments for the validation and release of AI functional modules in the context of highly automated driving.
For this purpose, the project will create a process chain with open standardized interfaces for the generation of high-quality and reproducible synthetic training and test data sets. Furthermore, AI algorithms for pedestrian recognition will be programmed and mainly trained and tested with these generated synthetic data. Thus, exemplary methods and measures can be developed, which are suitable to substantiate the chain of argumentation for the principle of protection with measurable performance and safety measures for AI functional modules. With the help of the knowledge gained in the project, the basis for an industry consensus on the safeguarding of such AI function modules is to be laid in dialogue with standardization committees and certification bodies.
In the project AI Securing, leading experts from industry and science from the fields of AI algorithms, 3D visualization and animation as well as functional safety, which up to now have operated largely independently of each other, are working together for the first time. Over the next three years, they will work together to develop solutions for more targeted observation, evaluation and testing of AI function modules and correspondingly reliable and transparent safeguarding.
The research project is part of the German government’s AI strategy, which aims to establish Germany as a location for the new key technologies in the long term and, among other things, to secure the market leadership of the German automotive industry with regard to automated driving 2 in the long term. The project is funded by the Federal Ministry for Economic Affairs and Energy with 19.2 million euros.
Duration: 1. July 2019 – 30. June 2022, 36 months
Total budget: 41 mio. EUR
Volkswagen AG (Konsortialführer), AUDI AG, BMW Group, Opel Automobile GmbH
Continental Automotive GmbH, Robert Bosch GmbH, Valeo Schalter und Sensoren GmbH, ZF Friedrichshafen AG, EFS
Automotive Safety Technologies GmbH, Intel Deutschland GmbH, Luxoft GmbH, Mackevision Medien Design GmbH, Merantix AG, QualityMinds GmbH, umlaut systems GmbH
Fraunhofer IAIS (Stellv. Konsortialführer und Wissenschaftlicher Koordinator), Deutsches Forschungszentrum für Künstliche Intelligenz, Deutsches Zentrum für Luft- und Raumfahrt, FZI Forschungszentrum Informatik
Universities: Bergische Universität Wuppertal, TU München, Universität Heidelberg
External technology partners: BIT Technology Solutions GmbH, neurocat GmbH, understand ai GmbH
European Center for Information and Communication Technologies –EICT GmbH