Convolutional Neural Network Algorithms for Car Accident Classification

Road accident is one of the top contributors to the number of death tolls around the globe. Consequently, much effort has been put in to minimize road accidents, and not often are many lives lost due to delayed response from emergency medical teams, notably in secluded or congested areas in the event of road accidents. The implementation of numerous deep learning algorithms has been widely used in many research fields; road accident detection is no exception. Therefore, this paper highlights and compares 3 different classification models from convolutional neural networks (CNN), namely MobileNetV3-Large, EfficientNetV2L, and ResNet-50 trained with a dataset of Closed-Circuit Television (CCTV) scenes obtained from Kaggle. The evaluation metrics consist of four categories for each proposed model, such as accuracy, precision, recall, and f1-score from two distinct classes, which are accident and non-accident events. Based on our study, it is evidently proven that ResNet-50 gained the overall highest f1-score of 95.13% followed by EfficientNetV2L and MobileNetV3-Large with 83.24% and 75.56% respectively. Additionally, ResNet-50's result validates the highest point with an average of 95.19% in terms of the remaining evaluation metrics, then 83.49% and 75.62% for EfficientNetV2L and MobileNetV3-Large correspondingly. Ultimately, ResNet-50 serves as the most appropriate and reliable convolutional neural network model in detecting a road accident effectively, followed with the lowest risk of detection inaccuracies.