Traffic Accident Severity Classification Using ResNet-18

The paper describes a system of automated classification of the severity of traffic accidents based on a fine-tuned ResNet-18 architecture. To meet the requirement of quick emergency reaction in low-resource urban areas, we pre-selected a dataset of about 50000 images, integrating the CADP and UCF-Crime datasets with rescue images of the area. The images were processed and augmented in order to rectify the imbalance of the classes. Transfer learning was used to train the model using Focal Loss and AdamW optimizer. Testing on a held-out test set shows a total accuracy of 98.54 and a precision of 0.99 on severe incidents and an F1-score of 0.98. The system maximizes the edge deployment, which provides low-latency inference applicable to real-time municipal surveillance. Comparative analysis shows that ResNet-18 offers a superior trade-off between accuracy and computational efficiency compared to deeper architectures..