Effects of Simulator Fidelity on Automated Vehicle Object Perception Accuracy

The development of intelligent transportation systems, and intelligent vehicles in particular, requires significant time and effort, with one of the primary phases being the training of the automated driving system. This training can be conducted using real-world annotated data or within a simulation environment. Simulation tools can provide a wide array of traffic scenarios unavailable in the real world. However, simulator development is complex and time-consuming, necessitating numerous compromises throughout the process. Among these challenges is the effects of graphical fidelity, which encompasses the level of detail and image resolution, on object detection accuracy in automated vehicles. This article aims to determine the importance of simulator graphical fidelity for automated vehicle development and define guidelines for simulator fidelity requirements. CARLA simulator was used to collect images for the AV control and object detection system. Experiments show that increasing image resolution while maintaining a constant level of detail leads to a decreasing improvement in model accuracy, and reducing image quality does not lead to a significant loss in object detection accuracy. It is also shown that graphical fidelity should be balanced with budget and computational constraints. It is concluded that in AV simulation, the priority should not be the pursuit of maximal graphical fidelity but rather the focus on the accuracy of traffic participant behavior, dynamics, and interactions.