Traffic Injury Prevention Techniques Using STATS19 Road Safety Data: A Model-comparison Approach

In this thesis, five models, which are based on artificial neural network (ANN) and support vector machine (SVM), were proposed for prediction of personal injury severities. The models were examined by two case studies using STATS19 road safety data that occurred in the city of London and Cambridge. The main purpose of the first case study was to identify the group most in need of road safety intervention by predicting the severities sustained by all road users. Using Radial Basis Function (RBFNN), different factors and areas of concern contributing to direct actual influences in both case studies were identified and ranked. In more detail to the first case study, non–motorised road users were recognised to benefit from the interventions. Therefore, the second case study aims to predict cyclist injury severities. Furthermore, most of the key factors in both case studies were in connection with busy junctions and poor turn / manoeuvres, here, truly protected junctions might be the best answer to create space for everybody. On focus to the two-wheeled group, there were limited crossing facilities near to where they cycled. Importantly for Britain's everyday cycling capital, narrow bike lane defenders are needed to provide a fully segregated solution where road width is too limited. In order to increase prediction accuracies, key factors were applied to multi–layer perceptron neural network (MLPNN) and SVM in both case studies. The models were selected as the benchmark due to their popularity in prediction modelling. Although the results of the predictions are encouraging, the models were not able to overcome incorrect predictions for ‘fatal’ and ‘serious injury’ severities due to limited data for those classes. In response to this, the two well-known models were combined as a hybrid MLPNN-SVM, and a learning vector quantization neural network (LVQNN) was improved for the first time ever to verify the best-fit model. Following this, different comparisons were made to evaluate the performance of the models in different classes. In addition, the models’ fitting results were presented and discussed, suggesting that all proposed models have ability to achieve satisfactory predictions, nevertheless, the improved LVQNN model performed better than others and was properly able to solve the incorrect predictions. This thesis concludes by identifying evidence-based road safety intervention options to mitigate the identified concerns. The general conclusion that can be drawn from this study is that most of the factors directly blame some kind of human error with high injury concentration being linked to junction actions. Therefore, in to crack down on bad driving / cycling, besides the road engineering interventions, it is recommended to deliver innovative road safety education and broadcast promotional messages for the recognised groups. Keywords: cyclist, driver, injury severity prediction, road safety intervention, STATS19.