Enhancing Realistic Floating Car Observers in Microscopic Traffic Simulation

We present a system that enables a realistic detection of traffic participants using Floating Car Observers (FCOs) directly within microscopic simulations. Point clouds are utilized to transform the two-dimensional simulation into a three-dimensional environment. In this environment, vehicles can be equipped with up to four camera sensors. Employing computer vision strategies, we identify which traffic participants would be detected by real-world sensor systems. We utilize the resulting system to generate datasets. In this process, various vehicles move within a simulation, recording which traffic participants would be recognized by the approach. Additionally, for each step of the simulation, an image from the current simulation and the position vectors of the traffic participants are documented. We employ the dataset to train neural networks, enabling them to replicate the results achieved using the CV method. The trained Vision Transformer and ResNet architectures achieve accuracies of up to 90%. Compared to the CV approach, the neural networks facilitate as much as 18-fold speed enhancement. We have made the source code, datasets, and trained models openly accessible at: github.com/urbanAIthi/SUMO-FCO.