AI protection – Safety for autonomous driving
QualityMinds tracks down “corner cases” and supports the quality assurance of artificial intelligence
Follow us
Safety for autonomous driving in Berlin
The AI protection project is embedded in the VDA flagship initiative for autonomous and connected driving and is funded by the Federal Ministry for Economic Affairs and Climate Protection.
In addition to QualityMinds, a further 24 partners from industry and academia were involved, investigating key issues relating to the validation of AI-based functions for highly automated driving. The project is funded with 41 million euros.
Deep neural networks play a decisive role in the realization of highly automated and autonomous driving, as they can perceive and distinguish various sensor data from the environment, such as pedestrians or obstacles. This is the only way to enable highly automated driving, in which the vehicle can and must react appropriately to different situations.
The Challenge:
Creation of a synthetic data set & analysis of “corner cases”
The central aim of the project was to derive an industry consensus for a systematic methodology for the acceptance of AI functions in the field of autonomous driving.
The quality of AI systems is of crucial importance, particularly in the complex field of autonomous driving.
Here, algorithms, traffic law and sensor technology must interact perfectly, even in the most unusual situations, in order to make the right decisions.
Because where human lives are at stake, it is undoubtedly important to keep the error rate as low as possible.
The requirements for QualityMinds were to create a synthetic data set and to identify and analyze“corner cases” (exceptional situations).
This was to provide indications of the areas of the AI functions in which the corresponding neural network was not working as anticipated and where further action was required.
The solution:
QualityMinds’ contribution to autonomous driving
Through the investigation of the corner cases and the data set jointly created in the project, important challenges with regard to the realization of autonomous driving technology were identified.
QualityMinds is leading a work package dedicated to the development of a methodology for identifying corner cases (situations in which the neural network reacts to its environment in unexpected ways) and thus taking them into account for the quality of the training and test data.
In addition, QualityMinds acts as a Gatekeeper (“Quality Gate”) by critically reviewing requirements and ensuring that quality criteria are met before a project moves on to the next phase. The project also produced scientific papers, e.g. on the automatic creation of a corner case data set using a tooling pipeline, which were subsequently presented by Niels Heller and Namrata Gurung at DATA 2022 in Lisbon.
Facts at a glance
Support in building a synthetic data set, e.g. through requirements engineering
Function as a “quality gate” (verification of quality criteria)
Investigation of “corner cases” to identify areas where the neural network reacts unexpectedly and further action is required
Application of hedging methods in national development & international standardization
Creation and presentation of scientific papers, e.g. via the automatic creation of a corner case data set using a tooling pipeline
Recognizing the challenges of autonomous driving, e.g. identifying risk zones behind corners
Participating in the KI-Absicherung project was an exceptional collaborative research experience in the field of autonomous driving.
QualityMinds played a pivotal role in the ‘Synthetic Data’ subproject, focusing on requirements engineering and establishing a ‘quality gate’ for synthetic datasets. We meticulously examined corner cases—scenarios where neural networks react unexpectedly—to identify areas needing improvement.
This collaborative effort led to significant advancements in autonomous driving safety, including the development of the Synthetic Pedestrian Dataset (SynPeDS), which supports performance evaluation of AI-based perception functions. We also further explored a novel method for automated corner cases extraction as published in our paper, ‘Highly Automated Corner Cases Extraction: Using Gradient Boost Quantile Regression for AI Quality Assurance,’ presented at the DATA 2022 conference.
Additionally, we contributed to ‘Inspect, Understand, Overcome: A Survey of Practical Methods for AI Safety,’ providing a comprehensive overview of AI safety mechanisms.
The synergy among all partners fostered innovation and excellence, contributing to the future of autonomous technology.
Dr. Namrata Gurung QualityMinds
Senior Data Scientist