A combined modelling approach to predicting injury severity in rear-end collisions

Abstract

Rear-end collisions constitute the most prevalent category of urban road traffic accidents, resulting in severe traffic congestion, casualties, and substantial economic losses. To mitigate the impact of such accidents effectively, this study proposes a severity prediction model that integrates Convolutional Neural Networks (CNN) and Extreme Gradient Boosting (XGBoost). The model employs the U.S. Department of Transportation's Fatality Analysis Reporting System (FARS) accident dataset, which undergoes preliminary preprocessing. Subsequently, Principal Component Analysis (PCA) is applied to reduce the dimensionality of the influencing factors prior to their input into the combined model for classification. CNN is utilized to extract features, while XGBoost is responsible for classification. Experimental results demonstrate that the combined model achieves a classification accuracy of 96.2 %, with superior AUC and F1 scores compared to traditional models, indicating excellent predictive performance.

• •
  This paper proposes a hybrid CNN-XGBoost algorithm that combines the superior feature extraction capability of CNN with the powerful structured data processing and precise prediction ability of XGBoost, resulting in a significant performance improvement over traditional algorithms.