چكيده به لاتين
Attaining a suitable method for predicting the crashes and investigating the underlying factors affecting them, has been one of the main goals in numerous studies in recent decades. This is due to the fact that the ability of predicting a crash can play a vital role in avoiding its occurrence and decreasing its costs. Traditional statistical modellings have been used frequently in the past to predict the crashes, however, because of their flaws, researchers have been trying to implement modern methods. On the other hand, development of new intelligent models and machine learning methods made it possible to predict the traffic crashes with a higher accuracy.
In this study, two intelligent methods of neural networks, feed-forward multi-layer neural network and recurrent neural network, have been used to predict the severity of crashes. Data used in this study is for Tehran-Saveh freeway (March 2011 to March 2016) which has been recorded and categorized under three severity classes: property damage, injury and fatality. The input variables used in this study for modeling were factors related to: time of crash, road’s surface situation in the time of crash, type of collision, human effect, vehicle effect, place of crash, road effective factor, eyesight barrier, type of shoulder, existence of longitudinal slope and horizontal curve.
Finally, based on the results, among the achieved network architectures, in each method one network is presented as the suggested model. Comparing the chosen models shows that the recurrent neural network on the aforementioned data with overall accuracy of 84.82 % and average accuracy of 50.22% is the method with the highest accuracy for prediction of crashes. Regarding the confusion matrix of the designated models, recurrent neural network has a superior performance concerning the detection of death crashes.
Keywords: Traffic crashes; Traffic crash severity prediction; Intelligent Models; Machine learning; Multi-layer neural network; Recurrent neural network
