Evaluating traffic signal optimisation systems

Matilya, Norbert

Evaluating traffic signal optimisation systems

Thesis / Dissertation

2026

Publisher

University of Cape Town

Department

Department of Civil Engineering

Faculty

Faculty of Engineering and the Built Environment

Abstract

Intersections are important components of a road network. They facilitates the crossing of two or more road alignments. Traffic signal systems are one of the most effective methods to manage traffic flow, at intersections in high-volume areas. These systems are governed by Traffic Signal Control (TSC). They are specifically designed to regulate the timing and sequencing of the signals at an intersection. Various TSCs can be installed at an intersection, however, there is a need for a standardized way to compare the efficiency of different TSC based on multiple factors. The primary objective of this research project was to design a way to assess different TSC systems, using Multi Criteria Decision Analysis (MCDA) as the evaluation method. The TechnoPark intersection located in Stellenbosch, Western Cape, South Africa was used as the case study site. Three MCDAs techniques were used to compare each TSC based on various traffic-related criteria. The criteria used for each MCDA method were collected during multiple Simulation of Urban MObility (SUMO) simulations. SUMO is an open-source traffic simulation software package that handles large road networks. Its ability to model vehicular movement, road layouts, traffic signals and collect various Key Performance Indicators (KPI) during simulations made it a good fit for the project. Static, delay-based, vehicle-actuated, and a Reinforcement Learning (RL) based TSC systems were compared. The static system has a fixed traffic sequence and duration for each phase. Vehicle Actuated Signal (VAS) rely on input from vehicle detection devices on one or more approaches. The delay-based control system is designed to extend traffic signal phases by considering the accumulated time lost from vehicles queuing. An innovative approach, using a RL based signal system was also used. It is a form of Machine Learning (ML) that can observe an environment and learn to take expected actions through a trial and error approach. The dynamic nature of RL allows the system to learn and optimise the signal timing and sequencing continually, thus adapting to observed vehicle movement from vehicle detectors found throughout the intersection. The research project has a systematic methodology divided into distinct phases: Data analysis, vehicle generation, simulations, and MCDA development. Initial data analysis involves processing observed data from a real-world intersection. Vehicle count data was obtained from ByteFuse, which were obtained by video image processing. To enhance the accuracy of simulations on SUMO, the statistical significance of the observed counts was assessed to best replicate these counts. The Kolmogorov Smirnov (KS) test was used to find the statistical significance of vehicles during the morning am (6h00 to 10h00) and evening pm (16h00 to 20h00) peak intervals of the observed data. After analysis, the gamma distribution was found to most constantly fit the observed data. The distribution was used to model the vehicle counts using best fit parameters obtained from the vehicle counts. This was done to accurately simulate the traffic flow in a virtual environment, allowing SUMO to collect various KPIs, which were then used for analysis. A workflow was developed to facilitate the production of multiple KPIs. It was a series of Python scripts that leveraged its ability to manage files and initialise SUMO. The TSC systems were assessed, resulting in a ranking as follows Delay-based, Actuated, RL, and Static. The results suggest that for this particular case study, the Delay-based system would be the most favourable alternative among the considered TSC options. However, future research should focus on generalising the intersection that can be analysed. This will broaden the range of intersections that could be evaluated.