Increasing the Accuracy of a Deep Learning Model for Traffic Accident Severity Prediction by Adding a Temporal Category

Artificial Intelligence become a tool widely used in the context of urban mobility and road safety applications. This paper focuses on predicting the severity of traffic accidents, from the point of view of the need for assistance, using general features that can be easily and quickly collected. We propose a deep learning model based on a convolutional neural network that compares, in performance, to several machine learning models for predicting the severity of traffic accidents. Our proposal modifies a previous model by refining the categorizing of the accident, implementing an area filter to address the imbalance data, reorganizing dataset into different features based on their nature, and discretizing the time of accidents using sine and cosine functions. This work demonstrates superior performance over six machine learning models, achieving an important improvement in the prediction of the two categories analyzed (accidents with and without requiring assistance). Datasets from two cities in the United Kingdom were analyzed, obtaining an improvement in the F1-score of \(4.6\%\) and \(13.2\%\) for attended and unattended accidents in the Liverpool dataset and \(3.1\%\) and \(17.2\%\) in the Southwark dataset.