KemptenCity - Semantic Segmentation of Urban Areas for Simulation

Abstract

Autonomous driving and traffic flow simulation requires a realistic and accurate representation of the environment. Therefore, this research focuses on the semantic segmentation of aerial images for simulation purposes. Initially, a dataset was created based on true orthophotos from 2019 and Kempten’s street cadaster, with true orthophotos being fully rectified aerial images. The chosen classes were oriented towards the subsequent conversion and usage in simulation. The proposed labeling workflow used cadaster data and demonstrated significant time efficiency compared to state-of-the-art datasets. Subsequently, a neural network was implemented that was trained and tested on the dataset. In addition, the network was also trained only on the lane markings to compare the network’s performance. Both cases demonstrated excellent segmentation results. The generalizability was then tested on true orthophotos from 2021. The results indicated a solid generalizability, but still needs to be improved. Finally, the aerial information was converted into a 3D environment, that can be used in simulations. Our results confirm the usage of aerial imagery and street cadaster data as a basis for the simulations.