Trade-off analysis using synthetic training data for neural networks in the automotive development process

Applications in the field of Deep Learning are constantly increasing the need for extensive, annotated data sets. Simulation software offers the possibility to create data sets of almost any size and thus the potential to cover this need. In the context of autonomous driving the test effort in the development processes increases to an extent that the use of virtual driving environments comes into focus. Deep Learning offers the opportunity to learn advanced driving strategies, starting from image recognition up to trajectory implementation.In this paper different driving simulation environments are used as data sources for the training of a neural network and examined for applicability on the basis of the working example of vehicle detection in image data.The visualizations include CarMaker, Carla and Grand Theft Auto V with different procedures for exporting ground truth information. For object detection, a pre-trained Convolutional Neural Network (CNN) is used and the effects on the detector quality of the data source and data size are investigated.An important part of the task of generating suitable data is the annotation process. The quality of the generated data is a decisive factor in achieving high performance training results. The test executions show that virtual data sources in training achieve lower detection rates in contrast to real data, but are more cost-effective in the overall context by scales.