Comprehensive data kit provides new impetus
A well-founded and comprehensive database is essential for the successful use of AI-based functions in vehicles. In addition to as many scenarios as possible, this must also include various sensor modalities. Until now, however, there have only been isolated solutions for the data of individual sensor types, very limited scenarios or parts of the training and evaluation chain. In the KI Data Tooling research project, tools, and methods for the provision of data from different sensor modalities for AI-based functions were developed and investigated for the first time in their entirety.
The project focused primarily on the efficient development of the database and pursued a complete data kit for the training and validation of AI-based automated driving functions from the outset. Real, synthetic and augmented data from various sensor modalities were generated for this purpose. Methods were also developed for the analysis, quality assessment and efficient, resource-saving storage and transmission of this data. Finally, an optimized AI training strategy was developed on this basis.
Getting autonomous vehicles safely on the road with machine learning
Over the past three years, the researchers from the Machine Learning competence field worked as associated partner together with BMW Group on several challenges in the project. In these efforts, fortiss colleagues Tianming Qiu, Carlos Valle, and Stephan Rappensperger focused on data-driven methods, such as deep learning techniques, which are used for comprehensive end-to-end training of AI functions in autonomous driving.
In this context, the quality of the data is a particularly important concern. Only with such a database is it possible to successfully train and validate AI functions. However, the provision of data is associated with numerous challenges. They include various facets, which the fortiss team investigated. These include the efficient labeling of data, the integration of synthetic data and the handling of unknown or rare scenarios (corner case detection) and the generation of contextual information (long-tail scenarios).
These research contributions, which have been developed over the past three years in the team of competence field leader Dr. Hao Shen, have been integrated into the project consortium's comprehensive data tooling pipeline. There, they will significantly benefit future research and development activities in the German automotive industry.
At the high-profile final event, the fortiss researchers presented their results in detail. The event also offered the opportunity for specialist deep dives and in-depth discussions among experts to discuss future fields of research.
