Analyzing and Predicting Highway Congestion through Camera Systems

Abstract

This thesis investigates the short-term prediction of highway congestion using traffic data collected from surveillance cameras installed along the Belgian highway network. These cameras provide aggregated vehicle speeds and flow rates, which were enriched with derived measures such as traffic density, weather conditions, and temporal features. The goal is to forecast congestion levels, categorized into free flow, low congestion, and high congestion, at specific highway locations over horizons of 5, 15, and 60 minutes. The primary aim of this work is to support proactive traffic management by enabling timely interventions, such as variable speed limit system, before congestion escalates. The thesis begins by outlining the relevance and potential benefits of congestion forecasting in improving road safety and traffic efficiency. It then presents a section on related work that traces the evolution of traffic prediction systems. A significant part of the study is devoted to data preparation and exploratory analysis, with particular attention to the challenge of class imbalance in congestion events. To address this, the study introduces tailored evaluation metrics that emphasize correct detection of rare but critical congested states. These metrics guide the comparison of a wide range of models, from rule based baselines to advanced machine learning algorithms, in order to identify the most effective models. Ultimately, boosting ensembles prove to offer the most robust and accurate results across all prediction horizons. The thesis concludes by highlighting the practical implications of these findings for deployment in real-time traffic monitoring systems and suggests several directions for future work, including the integration of additional data sources and further generalization to other highway segments.