Title: CrashNet-9D: A Deep Dense Neural Network for Large-Scale Traffic Accident Severity Classification
Authors: ALI M. A. BARHOOM, Prof. DR. HABIBOLLAH BIN HARON, Prof. DR. SAMY S. ABU-NASER, Prof. DR. NORIZAH BINTI MOHAMAD
Volume: 10
Issue: 1
Pages: 48-55
Publication Date: 2026/01/28
Abstract:
Road traffic accidents remain a major global public safety challenge, causing substantial human and economic losses each year. Accurate classification of accident severity is essential for effective traffic safety management, emergency response planning, and evidence-based policy formulation. This paper proposes CrashNet-9D, a deep nine-layer dense neural network designed for large-scale traffic accident severity classification using structured tabular data. The model is evaluated on a comprehensive United States traffic accident dataset comprising more than three million records collected between 2016 and 2021. A robust preprocessing pipeline is employed, including data cleaning, feature engineering, categorical encoding, feature standardization, and Synthetic Minority Oversampling Technique (SMOTE) to address severe class imbalance. The proposed model is compared against widely used machine learning classifiers, including Random Forest, Support Vector Machine, and Gradient Boosting models, using accuracy, precision, recall, F1-score, and computational efficiency metrics. Experimental results demonstrate that CrashNet-9D achieves superior and stable performance, attaining a macro F1-score exceeding 96% after class balancing, while maintaining strong recall for minority severity classes. The findings confirm the effectiveness of deep dense architectures for large-scale structured traffic data and highlight the practical potential of the proposed approach for intelligent transportation systems.