Abstract:The performance of an artificial neural network (ANN) in forecasting crash risk is shown in this paper. To begin, some traffic and weather data are acquired as raw data. This data is then analyzed, and relevant characteristics are chosen to utilize as input data based on additional tree and Pearson correlation. Furthermore, crash and non-crash time data are separated; then, feature values for crash and non-crash events are written in three four-minute intervals prior to the crash and non-crash events using the average of all available values for that period. The number of non-crash samples was lowered after calculating crash likelihood for each period based on accident labeling. The proposed CNN model is capable of learning from recorded, processed, and categorized input characteristics such as traffic characteristics and meteorological conditions. The goal of this work is to forecast the chance of a real-time crash based on three periods before events. The area under the curve (AUC) for the receiver operating characteristic curve (ROC curve), as well as sensitivity as the true positive rate and specificity as the false positive rate, are shown and compared with three typical machine learning and neural network models. Finally, when it comes to the error value, AUC, sensitivity, and specificity parameters as performance variables, the executed model outperforms other models. The findings of this research suggest applying the CNN model as a multi-class prediction model for real-time crash risk prediction. Our emphasis is on multi-class prediction, while prior research used this for binary (two-class) categorization like crash and non-crash.

Multi-class real-time crash risk forecasting using convolutional neural network: Istanbul case study