Road traffic accidents are a major concern worldwide, often resulting in serious injuries and fatalities. Rapid and accurate assessment of accident severity can significantly aid emergency response and resource allocation. This project presents a comprehensive, automated system for predicting the severity of road traffic accidents using both tabular (structured) and image (unstructured) data. For tabular data, models including Multilayer Perceptron (MLP), XGBoost and TabNet were implemented to classify severity levels. For image-based classification, deep convolutional neural networks such as MobileNetV2 and ResNet50 were used, leveraging transfer learning to enhance performance. An EfficientNetB7 model was employed to auto-label unlabelled accident images, improving dataset consistency and reducing manual labelling effort. The system was evaluated using accuracy, precision, recall and F1-score, with performance visualized through confusion matrices and bar charts. The results demonstrate high classification accuracy across all models. This intelligent, multi-input framework serves as a practical tool for real-time accident severity assessment, with potential applications in emergency response systems and smart traffic management.
Introduction
Road traffic accidents (RTAs) cause significant loss of life and property. Accurately predicting accident severity can greatly improve emergency response and traffic management. This project introduces a multimodal machine learning system that combines tabular data (e.g., time, location, weather) with image data (accident scene photos) to predict accident severity levels (low, medium, high).
Key Components & Methodology:
Tabular Data Models:
MLP, XGBoost, and TabNet were used for structured data.
XGBoost and MLP achieved perfect accuracy (1.000), while TabNet performed slightly below with added interpretability.
Image Data Models:
MobileNetV2 and ResNet50 were applied using transfer learning.
Accuracy was ~0.78–0.79, but precision dropped to ~0.60, indicating challenges in visual classification.
EfficientNetB7 was used for auto-labeling unlabeled images, improving dataset quality.
Data Preprocessing & Feature Engineering:
Tabular data was cleaned, encoded, and standardized.
Images were resized, normalized, and augmented.
Additional features (e.g., Accident_Hour) were engineered to enhance prediction.
Evaluation Metrics:
Accuracy, precision, recall, and F1 score were used. Visual comparisons via bar charts and confusion matrices showed that tabular models significantly outperformed image models.
Literature Review Insights:
Prior research emphasizes deep learning, AI in autonomous systems, and class imbalance challenges.
Techniques like SHAP values, GANs, and RNNs have been used for severity prediction and traffic anomaly detection.
XGBoost, CNNs, and LSTMs emerged as consistently strong models across different studies.
Conclusion
In conclusion, this project successfully developed an end-to-end, automated system for predicting the severity of road traffic accidents using both tabular and image data. By integrating multiple machine learning and deep learning modelsincluding MLP, XGBoost, TabNet, MobileNetV2 and ResNet50the system demonstrated strong performance, especially in structured data classification, with some models achieving near-perfect accuracy. The use of EfficientNetB7 for auto-labeling unlabeled accident images further streamlined the data preparation process and enhanced model training. Evaluation through accuracy, precision, recall and F1-score confirmed the effectiveness of the models, and comparative analysis highlighted the strengths of both data modalities. Looking forward, this framework holds immense potential for real-time deployment in smart traffic systems and emergency response platforms. Future enhancements may include integrating more advanced vision models such as Vision Transformers (ViTs), expanding datasets for better generalization, incorporating temporal or geospatial data for context-aware predictions, and deploying the system via web or mobile interfaces for live monitoring. Furthermore, combining this predictive system with IoT and edge computing devices could enable low-latency accident analysis and immediate alert generation, contributing significantly to safer and smarter transportation infrastructure.
References
In conclusion, this project successfully developed an end-to-end, automated system for predicting the severity of road traffic accidents using both tabular and image data. By integrating multiple machine learning and deep learning modelsincluding MLP, XGBoost, TabNet, MobileNetV2 and ResNet50the system demonstrated strong performance, especially in structured data classification, with some models achieving near-perfect accuracy. The use of EfficientNetB7 for auto-labeling unlabeled accident images further streamlined the data preparation process and enhanced model training. Evaluation through accuracy, precision, recall and F1-score confirmed the effectiveness of the models, and comparative analysis highlighted the strengths of both data modalities. Looking forward, this framework holds immense potential for real-time deployment in smart traffic systems and emergency response platforms. Future enhancements may include integrating more advanced vision models such as Vision Transformers (ViTs), expanding datasets for better generalization, incorporating temporal or geospatial data for context-aware predictions, and deploying the system via web or mobile interfaces for live monitoring. Furthermore, combining this predictive system with IoT and edge computing devices could enable low-latency accident analysis and immediate alert generation, contributing significantly to safer and smarter transportation infrastructure.