A Road Accident Detection Method Utilizing Deep Learning and Fast Fourier Transform

This paper presents the development of a real-time crash detection method aimed at improving the efficiency of emergency assistance to the location of the accident. The analysis involved reviewing existing classical and neural network-based crash detection approaches, focusing on architectures such as EfficientNet-B1, EfficientNet-B7, MobileNetV2, and ConvNeXtV2. A dedicated dataset consisting of 12,426 crash-related image frames was specifically compiled for this study, combining previous datasets with self-collected images. The performance of accident detection models was evaluated on this dataset, leading to the development of a new crash detection method. The ConvNeXtV2-Femto model was selected as the core architecture for the proposed system, which was modified employing Fast Fourier Convolution (FFC) to improve its performance. Comparative analysis demonstrated that the proposed model achieved a 94% accuracy, outperforming existing approaches in all metrics, including also precision, recall, and F1-score.