Detection of accident situation by machine learning methods using traffic announcements: the case of metropol Istanbul

Abstract

Information about the reality of the traffic accident, the clearness of the roads and the status of the accident can be obtained from the traffic accident announcements. By using the words in the radio or telephone announcements, you can be informed about the status of the accident. Inferences can be made with machine learning methods using a large number of data. In this study, the accident situation was classified using three different machine learning methods using radio and telephone announcements in Istanbul in Turkey. The dataset contains 156.856 announcement data. Classifications were performed using Artificial Neural Network (ANN), k-Nearest Neighbor (kNN) and Decision Tree (DT) machine learning methods. Classification success was 92.1% in the classification made with the ANN model, 91% in the classification made with the kNN model, and 89.8% in the classification made with the DT model. Classification performances of the models were also analyzed with precision, recall, F-1 Score and specificity metrics. In addition, the estimation abilities of the models with ROC curves and AUC values were analyzed. In addition, the training and testing times of the models were also analyzed. It will be possible to use the suggested models to automatically detect the accident situation from the announcements. In this way, it is thought that the most accurate direction can be made by obtaining information about crew orientation, traffic jams and the size of the accident.