An Intelligent System for Automatic Evaluation and Reporting of Vehicle Accidents Based on EfficientNet Model and CBAM Attention Mechanism

Abstract

Road traffic accidents remain a critical challenge in road safety, where rapid detection and accurate severity assessment play a vital role in reducing fatalities. This study presents an intelligent system for the automatic evaluation of vehicle accidents based on the EfficientNet-B0 model enhanced with the Convolutional Block Attention Module (CBAM). The primary objective is to overcome the severe class imbalance issue, particularly in the “low-risk” class, and to improve detection accuracy under real-world conditions. The proposed architecture consists of two main modules: a multi-task convolutional neural network for accident detection and severity classification, and a YOLOv8-Nano module for real-time fire and smoke identification. To address data imbalance, a combined strategy involving external data augmentation, oversampling, and an eight-stage training pipeline was employed. Experimental evaluation on the Accident Images Analysis dataset demonstrated that adding the CBAM mechanism led to a 38% improvement in the F1-score of the minority class. The final system achieved an accuracy of 94% in accident detection and 74% in severity classification (with a Macro F1 score of 0.61). Comparison with Pashaei et al. (2020) on the public version of the dataset showed a 4.49% improvement in the Macro F1 metric. The results confirm that integrating attention mechanisms with efficient architectures and smart data management significantly enhances the performance of emergency response support systems.