Real-World Traffic Scenarios for ADAS and AD Development

Testing and validation is currently the limiting factor in the deployment of automated driving. This bottleneck can only be overcome with simulation that allows to significantly decrease the required test miles in the real world. However, results from simulators can only be trusted if they reflect the full complexity and diversity of reality. It is therefore indispensable to bring real- world scenarios into simulation to bridge the gap between development and deployment much faster. For this purpose, this paper leverages the unique advantages of aerial data to reconstruct traffic scenarios in a highly flexible and cost-effective way. A novel deep learning-based 3D computer vision approach is introduced that overcomes the severe limitations of existing 2D methods in terms of accuracy and quality of the reconstruction. The robustness and generalizability of this 3D approach is demonstrated in four traffic datasets that were collected in highly interactive urban and suburban environments. All four datasets are several hours long and contain a map of the road network. Real-world scenarios are extracted from these datasets for testing and validation of automated vehicles. Test results provide insights about which environments are most challenging and where coverage needs to be increased. This scenariobased validation process allows to significantly reduce the number of test miles in the real world enabling the deployment of automated driving at scale.