Classification of Road Accidents using Artificial Neural Network (Case Study: Tehran-Pardis Freeway)

Occurrence of road accidents in terms of severity and loss of life and property in various environmental and geometric conditions has always been studied and considered. In this study, a multilayer perceptron model with different architectures including one and two hidden layers with different numbers of neurons has been used to estimate the number of road accidents. In modelling of independent variables (input data) including horizontal arc radius, time of the accident, age of the driver causing the accident, gender, having a driving license, type of vehicle, weather conditions, average daily traffic, ambient light, cause of accident, severity and location of accident used. The dependent variable (model output), the number of accidents was divided into four classes (first class equal to one accident, second class equal to 2 accidents, third class equal to 3 accidents, and fourth class equal to 4 accidents). In this regard, the reports of accidents on the Tehran-Pardis freeway between January 2017 and August 2019, which were received from the Tehran-Pardis freeway company, have been used. The number of accidents that occurred during this period was equal to 368 cases, of which 44 cases were excluded due to lack of recorded information and the neural models were constructed using 324 reports. The results show that the hidden two-layer neural network model with 14 neurons each with an accuracy of 83.3% for training data and a total accuracy of 83.3% showed the best performance in estimating the number of accidents.