Performance evaluation of a machine learning environment for intelligent transportation systems

While automotive manufacturers are already implementing Autonomous Driving (AD) features in their latest commercial vehicles, fully automated vehicles are still not a reality. In addition to AD, recent developments in mobile networks enables the possibility of Vehicle-to-Infrastructure (V2I) and Vehicle-to-Vehicle (V2V) communication. Vehicle-to-Everything (V2X) communication, or vehicular Internet of Things (IoT), can provide solutions that improve the safety and efficiency of traffic. Both AD and vehicular IoT need improvements to the surrounding infrastructure and vehicular hardware and software. The upcoming 5G network not only reduces latency, but improves availability and massively increases the amount of supported simultaneous connections, making vehicular IoT a possibility.


Developing software for AD and vehicular IoT is difficult, especially because testing the software with real vehicles can be hazardous and expensive. The use of virtual environments makes it possible to safely test the behavior of autonomous vehicles. These virtual 3D environments include physics simulation and photorealistic graphics. Real vehicular hardware can be combined with these simulators. The vehicle driving software can control the virtual vehicle and observe the environment through virtual sensors, such as cameras and radars.


In this thesis we investigate the performance of such simulators. The issue with existing open-source simulators is their insufficient performance for real-time simulation of multiple vehicles. When the simulation is combined with real vehicular hardware and edge computing services, it is important that the simulated environment resembles reality as closely as possible. As driving in traffic is very latency sensitive, the simulator should always be running in real-time. We select the most suitable traffic simulator for testing these multi-vehicle driving scenarios. We plan and implement a system for distributing the computational load over multiple computers, in order to improve the performance and scalability.


Our results show that our implementation allows scaling the simulation by increasing the amount of computing nodes, and therefore increasing the number of simultaneously simulated autonomous vehicles. For future work, we suggest researching how the distributed computing solution affects latency in comparison to a real-world testing environment. We also suggest the implementation of an automated load-balancing system for automatically scaling the simulation to multiple computation nodes based on demand.