Road Collision Analysis and Prediction Using Machine Learning Approaches

Road travel accounts for most traffic accidents worldwide. Improvements in road safety, education, recent technology advancements, and other environmental factors have decreased the number of collisions in developed nations. Many countries, provincial, and local governments envision the possibility of zero fatalities or serious injuries in the near future. Thus, it is essential to develop road traffic accident prediction models to support such a vision. On the one hand, classical statistical models have been applied to develop prediction models throughout the literature. These models provide interpretable parameters at the expense of poor generalization when faced with complex and nonlinear relationships. On the other hand, data-driven methods utilizing Machine Learning (ML) approaches have been used recently to deal with the drawbacks of classical models, which showed promising results. Road accidents result from many factors, including spatial, temporal and external factors. Those factors may influence the occurrence of accidents differently, according to the location and time of accidents. Thus, it is essential to consider the area-specific influential factors while analyzing and developing prediction models. Canada is the second coldest country globally, and its extreme weather has a higher effect on accidents than the other countries, which must be addressed. This thesis seeks to explore determinants of road collisions, emphasizing Canadian weather. It then compares classical and ML models for collision prediction. Furthermore, it introduces the most influential factors in crashes with respect to Calgary's weather. All study parts are performed on the collisions data in Calgary, Alberta, Canada, between 2017 to 2020. It is shown that all the weather attributes are correlated to collisions. It shows the importance of considering the weather attributes in accident analysis and prediction. Based on the nature of the collisions dataset, which is tabular and heterogeneous, Neural Networks showed higher performances than the other investigated model, with 92% accuracy. The proposed models can be used for policy-making and individual usage in Canadian cities since the effect of all the weather features is already embedded in the models. In order to demonstrate the thesis's applicability, a new speed limit is recommended utilizing the developed models for Deerfoot TR SE. Results showed, for instance, if the speed limit is decreased from 100 to 90 km/h on Deerfoot TR SE, a 5% accident reduction is predicted.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.