As already explained in our earlier highlight topic, handling of AI based functionalities is hardly imaginable without the use of virtual environments. Virtual training offers the advantage that ground truth comes for free due to the scene description, thus eliminating the need for complex labeling of real data.
Boundary conditions such as lighting conditions, weather conditions, road surfaces and traffic conditions can be altered by scene description and thus a high variety of training data can be generated. This way more training data can be generated specifically for weak points of the algorithms.
This is illustrated by the example of an initially poorly functioning lane detection algorithm. After training the algorithm using virtually generated scenarios, good lane detection is achieved. The featured lane detection algorithm is part of the software applied in FKFS research and development vehicle, thus spanning the whole range of development from complete virtual environments to real world applications.