Thesis Topics

Devise small scale cyberattacks (e.g., GPS spoofing, manipulating traffic lights, injecting phantom vehicles to a navigation app, etc.) that creates the most chaos to the traffic flow on the network. Model their impact on a calibrated traffic simulator. 

The implementation of Edge AI and TinyML on resource-constrained embedded systems is a critical step toward sustainable computing. This project focuses on deploying machine learning models on low-power hardware to reduce cloud dependency, latency, and overall energy consumption. The student will select, hardware-optimize (e.g., via quantization or pruning), and deploy neural networks directly onto a microcontroller platform. A core requirement of the research is the critical evaluation of the necessary trade-offs between model accuracy, execution time, and power efficiency to achieve a viable and sustainable local AI solution.

Manual modeling of rare but dangerous traffic situations (“edge cases”) is a difficult and time-consuming task in traffic simulations. Large language models (LLMs), on the other hand, may be capable of creating diverse, extreme scenarios (sudden lane changes, incidents, etc.). The task is therefore to use LLM to generate critical situations in the SUMO environment, then run them automatically and analyze the reactions of vehicles and changes in traffic dynamics.

Building a natural language interface (e.g., Python-based framework Gemini or OpenAI engine) for generating SUMO-based traffic scenarios. The student’s task is to develop a framework that generates the necessary files based on English text descriptions (e.g., “Create a 2-lane 1 km highway section, etc.”). The research question is to what extent the LLM is able to interpret traffic engineering terminology and convert it into syntactically correct SUMO xml files. The output is essentially a “Prompt-to-Simulation” software tool.

Implementation of a real-time traffic digital twin on a selected motorway section. The SUMO or VISSIM simulation tool must be able to track reality. Reality is provided by real (M1-M7) data and/or virtually real data generated by SUMO. The question is how SUMO digital twins can reproduce real traffic dynamics more accurately and with as little delay as possible. Another question is how digital twins can be used to effectively manage traffic disruptions (e.g., proactive control by “running ahead” with digital twins).

Developing a methodology for optimal traffic light placement based on multiple criteria. Developing a method that helps us determine traffic light-controlled intersections on an urban road network with a given topology and dynamic traffic demand, taking into account the dynamic, adaptive behavior of individual travelers.

PLCs (Programmable Logic Controllers) are embedded devices that are widely used in automation and manufacturing technology. These devices are renowned for their reliability in safety-critical application environments and their ease of programming compared to other embedded systems. The student’s task is to design an intelligent traffic light or VJT control system based on a PLC system.

There are models of varying levels and accuracy in the literature for describing road vehicle traffic. In this work, a new approach is taken to create (identify) traffic models using control theory system identification solutions based on SUMO traffic simulation “measurements.” Standard solutions (e.g., MATLAB System Identification Toolbox or Python SIPPY) and/or AI-based approaches should be used for identification.

Assess the trade-off of using the co-simulation of vehicle dynamics (e.g., CARLA, IPG CarMaker) and microscopic road traffic simulation (e.g., SUMO). Assuming individual simulators are well-calibrated, evaluate how the accuracy of the co-simulation is reduced compared to a full-scale vehicle dynamics simulation. Different co-simulation ideas and use-cases are welcome.