Concept for Safety-Related Development of Deep Neural Networks in the Automotive

Concept for Safety-Related Development of Deep Neural Networks in the Automotive

연구 분야: Verification

학회: 2020 Fourth International Conference on Multimedia Computing, Networking and Applications (MCNA)

The performance of Deep Neural Networks in field of computer vision as most attractive segment of current trends in automotive is not just remarkable, but also quite beneficial in comparison to the classic SW approaches, since many hundred thousand lines of codes can be easily replaced. On the other side, a completely different development process makes them not fully conform to the requirements of current standards for functional safety in automotive, ISO 26262 and SOTIF. Analyzing Deep Learning Networks from the perspective of classic SW approach was the first try to address the drawbacks and deviations from the current standards. Although worthwhile, such studies didn't provide concepts, how to develop DNNs in a systematic, transparent and assessable way.This paper introduces a novel concept for the development of deep neural networks considering their characteristics, overcoming the drawback with tool evaluation process of wide established frameworks (such as TensorFlow, Keras etc.) and providing a systematic, traceable and assessable solution.