Prediction of traffic accident severity

Abstract

Road traffic accidents pose a significant global challenge, affecting safety, congestion, and productivity. In Malta, the increasing vehicle population and dense urban infrastructure necessitate predictive safety solutions. This dissertation examines the use of supervised machine learning algorithms to classify traffic accident severity based on historical data, categorising incidents into four classes: “Slight,” “Grievous,” “Fatal,” and “Insignificant.” The research has four main objectives. First, the data preprocessing phase aimed to ensure quality for the dataset from 2004 to 2021. The second objective was to develop three classification algorithms: Naive Bayes (NB), Support Vector Machine (SVM) and Random Forest (RF), using R language. Techniques such as Laplace smoothing for NB, kernel transformations for SVM and bootstrapped tree aggregation for RF are to be used to enhance performance. The third objective was to assess the classifiers using metrics like accuracy, recall, precision and F1 score. Monte Carlo Cross-Validation (MCCV) was employed for model validation. Results showed RF performed best, with accuracy at 98.71%, recall of 95.25%, a precision of 93.13% and an F1-score of 93.77%. Finally, the study aimed to analyse features influencing accident severity, revealing hotspots and risk zones around critical infrastructure. Notably, accidents involving children clustered near schools, and there was an increased incidence of ”Grievous” and ”Fatal” cases around roundabouts. Limitations included class imbalance and the absence of behavioural and environmental features. Future research could involve Gradient Boosting Machines (GBM) and Artificial Neural Networks (ANNs) for improved modeling. Integrating findings into Intelligent Transportation Systems (ITS) may enhance real-time alerts for drivers, supporting road safety in Malta.