WEATHER-RESILIENT ROAD INCIDENT MONITORING: A DEEP TRANSFER LEARNING APPROACH FOR ACCIDENT DETECTION IN LOW-VISIBILITY CONDITIONS

Abstract

Automated road accident detection is critical for rapid emergency response and traffic management. However, current vision-based systems often suffer from significant performance degradation under adverse weather conditions such as rain, snow, and fog. This paper proposes a weather-resilient framework for accident detection using Deep Transfer Learning. We utilize the MobileNetV3 architecture, pre-trained on ImageNet, and integrate a Contrast Limited Adaptive Histogram Equalization (CLAHE) preprocessing step to enhance feature extraction in low-visibility environments. The model was trained and validated on a large-scale dataset of 75,000 frames derived from the Car Crash Dataset (CCD). To further improve generalization, we employed data augmentation strategies including random horizontal flipping and color jittering. The experimental results demonstrate that the proposed CNN-based transfer learning model achieves a robust validation accuracy of 80.01%, with a training accuracy of 95.68%. The model effectively minimizes loss across 5 epochs, confirming that pre-trained weights can be successfully fine-tuned for accident detection tasks.