From the lab to the road: How TU Graz is making driver assistance systems safer

 (Image: Pixabay CC0)
(Image: Pixabay CC0)

In the Christian Doppler Laboratory under his leadership, Franz Wotawa has developed test and monitoring procedures with company partner AVL that make common driver assistance systems safer.

Intelligent Speed Assist, Emergency Brake Assist, Emergency Lane Keeping Assist, Driver Drowsiness Assist, Reversing Assist, Low Concentration Warning System or Emergency Brake Lights - from July 2024, a whole range of safety and driver assistance systems will be mandatory for all new cars in the European Union. Since October 2017, the Christian Doppler Laboratory for Methods of Quality Assurance of Autonomous Cyber-Physical Systems at Graz University of Technology has been working with corporate partner AVL List GmbH to ensure that these systems work as intended and really do ensure greater safety. The laboratory team, led by Franz Wotawa from the Institute of Software Technology at Graz University of Technology, has used basic research to develop new methods to rule out sources of error in driver assistance systems in advance and analyze them during operation. Based on this, AVL was able to add new methods and procedures to its portfolio in the field of Advanced Driver Assistance Systems (ADAS).

Small deviations with a big impact

Specifically, Franz Wotawa and his team had to face the challenge that even minor deviations in a certain traffic scenario can significantly influence the reaction of driver assistance systems. As the systems should not have to learn these deviations during operation, a process was developed for the automated generation of test cases based on ontologies. Ontologies are descriptions of the environment in which the vehicle is located in the respective test case. These descriptions contain information such as the existing road network, traffic lights, road signs, other vehicles or pedestrians.

For test case generation, the team has adapted a search-based and a combinatorial test procedure and, building on this, has created an algorithm-based link between the ontologies and an input model. This makes it possible to automatically derive and run through even better and more comprehensive test scenarios - regardless of the assistance or security system being tested. For example, it was possible to find some undetected errors in an emergency brake assistant during the tests, which could then be analyzed in more detail.

View of real conditions

Despite the sophisticated test procedures, it is essential to keep an eye on ongoing operations, as unforeseen situations can always occur. Here, the team compares collected car sensor data with the expected behaviour of the vehicles and tries to combine this with formalized knowledge about object movements. The focus here is on object recognition in order to formalize the object movement from a sequence of images by means of logical deduction. By tracking the objects over several image frames, they can be classified as potentially dangerous or harmless and the appropriate measures can be derived - for example, whether a tree is approached directly and must be avoided or the journey passes it after all. These findings are subsequently incorporated into updates to the assistance systems. In addition, the data from the experience gained in real-life operation can also be used to generate further test cases.

With frame-by-frame analysis in combination with a logic model for spatial perception, object recognition can also be improved and object movements that are not consistently recognized can be deduced. This is useful, for example, if an object is visible for a few frames but is not recognized for one frame due to reflections or a sensor error. An assistance system could then think that there is no longer any danger in this area. Thanks to the logic model, however, the software deduces that the object must still be there because it cannot simply disappear.

Rapid transfer of knowledge to the industry

For Franz Wotawa, the results achieved so far by the CD Laboratory, which will run until the end of September 2024, are proof that the combination of basic research with concrete applications through the corporate partner offers many advantages. "We have a very direct exchange with AVL, each doctoral student also works five to ten hours a week at the company. As a result, we are very familiar with the problems from industry and can conduct basic research on this basis. On the other hand, knowledge is transferred to industry very quickly because the employees have direct access to AVL’s infrastructure. This has enabled us to jointly advance our research into the security of autonomous cyber-physical systems," explains Franz Wotawa.

Mihai Nica, Global Head ADAS, Automated Driving and Connectivity AVL, adds: "In the rapidly evolving world of autonomous driving, AVL relies on innovative testing methods. The application of AI gamification and ontology-based testing provides the ability to generate critical scenarios and test autonomous driving under extreme and complex conditions that are difficult to replicate in the real world. This is crucial to ensure the reliability and safety of the technology and helps to build public confidence. This trust is crucial for the successful integration of autonomous vehicles into our transportation networks of the future."

This research project is anchored in the Field of Expertise "Information, Communication & Computing", one of five strategic focus areas at TU Graz.

From Falko Schoklitsch