Scalable Data Set Distillation for the Development of Automated Driving Functions

Since highly automated driving functions are obligated to work safely without human interference, careful and comprehensive testing is mandatory. Due to the open world context, automated driving functions must be able to cope with a huge variety of driving situations. To achieve this goal, huge data sets are being recorded. These data sets are used in the development and in open loop tests where the data is replayed to examine driving functions under diverse conditions. The driving functions design is based on machine learning to cope with the variant environment. For the generation of meaningful testing evidences, well-tailored data sets are necessary. To avoid costly processing of voluminous sensor data, the focus lies on a method based on metadata. Therefore, we present a methodology based only on the position and temporal information for the distillation of automotive data sets. The method is intended to support the driving function experts and can be easily parameterized by them according to their requirements. This enables broad scalability and allows to efficiently find specific and potentially challenging situations. The process is presented exemplarily for sun positions and geographical data.