Zero Fatality Roads: Leveraging Deep and Transfer Learning on CCD for Automated Crash Detection

Abstract

Over a million people have died in road accidents every year, making them one of the world's top sources of injury and death. Human error and delayed incident response are major contributors to road deaths, even with improvements in vehicle design and road infrastructure. This study proposes a robust deep-learning framework for automated road accident detection by harnessing the synergy of transfer learning and car crash data. Addressing the limitations of scarce annotated accident imagery, we employed a publicly available Kaggle dataset called the Car Crash Dataset (CCD), which contains real dashcam videos of traffic accidents, mostly collected from YouTube, to train and evaluate state-of-the-art convolutional neural network models. Using ImageNet-pretrained architectures such as MobileNetV2, VGG19, ResNet50, and EfficientNetB7, we adapted the classification layers for binary classification while leveraging their pre-learned visual features. Our experimental results show that ResNet50 achieved the highest detection performance, with a precision of 97.83%, a recall of 98.78%, an F1-score of 98.05%, and an MCC of 0.955, outperforming the other models. MobileNetV2 also demonstrated strong results, with a precision of 92.84%, a recall of 94.60%, an F1-score of 93.71%, and an MCC of 0.781, making it suitable for deployment in resource-constrained environments. EfficientNetB7 and VGG19 performed moderately but lagged behind ResNet50 and MobileNetV2 in accuracy and correlation measures. This comparison offers essential guidance for choosing CNN architectures that balance accuracy and efficiency for real-time applications.