Simulation-based testing of classical driver assistance functions (ABS, ESP) involves mainly the EGO vehicle and some dummy objects. Specifying the scenario in simulation is simple. However, testing complex, cooperative, or perception-based driver assistance functions (e.g., a traffic jam pilot) requires many vehicles inside the simulation. Therefore, microscopic traffic simulation is often used together with vehicle simulation to generate swarm traffic. However, computing the dynamics, car following behavior, and rendering the cars in 3D is computationally burdensome, limiting the size of scenarios.
On the other hand, simulating an entire district or city with thousands of vehicles present is superfluous since vehicles hundreds of meters away from the EGO vehicle have minimal impact in EGO-centric scenarios. Thus, simulation accuracy far from the EGO vehicle can be traded for simulation speed.

Adopting the idea of the level of detail from computer graphics, modeling swarm vehicles is carried out in less detail farther away from the EGO vehicle. These levels are defined using the classical (macroscopic, mesoscopic, and microscopic) categorization of traffic modeling. The macroscopic traffic model is responsible for traffic demand and traffic assignment. The mesoscopic model (Network Shockwave Profile Model) is capable of capturing the fluctuating nature of traffic on lane level. Since the mesoscopic model is continuum-based, it has constant time complexity with respect to vehicle number, as opposed to the polynomial time complexity of the microscopic car-following logic. Closer to the EGO vehicle, microscopic traffic simulation (SUMO) is employed while the EGO vehicle is modeled in full detail, including vehicle dynamics too (using CARLA). Thus, the number of vehicles that are simulated in high detail is bounded, enabling EGO-centric scenarios on arbitrary large road networks. The challenge in the proposed methodology is transitioning between the mesoscopic and the microscopic models, i.e., selecting the boundary and spawning/destroying vehicle agents.
To this end, we have developed a middleware between the mesoscopic model and the SUMO-CARLA co-simulation framework to realize the above trade-off efficiently.

Our proposed solution can significantly increase the simulation’s real-time factor while retaining the accuracy around the EGO vehicle. Testing our solution on CARLA’s example network Town04, we could achieve a simulation speed gain of 3-5 times (depending on traffic volume).

Related publication