Rigid barrier or not? : Machine Learning for classifying Traffic Control Plans using geographical data

Sammanfattning: In this thesis, four different Machine Learning models and algorithms have been evaluated in the work of classifying Traffic Control Plans in the City of Helsingborg. Before a roadwork can start, a Traffic Control Plan must be created and submitted to the Traffic unit in the city. The plan consists of information regarding the roadwork and how the work can be performed in a safe manner, concerning both road workers and car drivers, pedestrians and cyclists that pass by. In order to know what safety barriers are needed both the Swedish Association of Local Authorities and Regions (SALAR) and the Swedish Transport Administration (STA) have made a classification of roads to guide contractors and traffic technicians what safety barriers are suitable to provide a safe workplace. The road classifications are built upon two rules; the amount of traffic and the speed limit of the road. Thus real-world problems have shown that these classifications are not applicable to every single case. Therefore, each roadwork must be judged and evaluated from its specific attributes. By creating and training a Machine Learning model that is able to determine if a rigid safety barrier is needed or not a classification can be made based on historical data. In this thesis, the performance of several Machine Learning models and datasets are presented when Traffic Control Plans are classified. The algorithms used for the classification task were Random Forest, AdaBoost, K-Nearest Neighbour and Artificial Neural Network. In order to know what attributes to include in the dataset, participant observations in combination with interviews were held with a traffic technician at the City of Helsingborg. The datasets used for training the algorithms were primarily based on geographical data but information regarding the roadwork and period of time were also included in the dataset. The results of this study indicated that it was preferred to include road attribute information in the dataset. It was also discovered that the classification accuracy was higher if the attribute values of the geographical data were continuous instead of categorical. In the results it was revealed that the AdaBoost algorithm had the highest performance, even though the difference in performance was not that big compared to the other algorithms.