Comparison of the Results of Statistical Models and Neural Networks Models in Prediction of the Number of Accidents at Intersections

Traffic accidents are among the important factors of fatalities and injuries that bring indispensable harms to social, cultural, and economic aspects of human societies. Also the increase in trip and traffic movements in the city provides economic improvements and social welfare but in return, the rate of accidents in the city increases proportionally. On the basis of well-accomplished studies, great part of accidents happens at intersections. The main factor of traffic accidents at intersections is convergence of traffic streams at one point. In many countries numerous studies have been accomplished to construct a statistical model for prediction of traffic accidents. In this paper, for the first time in Iran, in addition to construction of statistical model, neural network was employed to obtain the model for accidents prediction at intersections through some parameters such as traffic volume, geometric design and characteristics of traffic control devices for an intersection. Finally the result of statistical models and neural networks model are examined and compared. Results of the statistical and neural network models show that:1. The correlation coefficient is %95 for statistical models and %97 for neural network models. The interesting point is that the %95 has been obtained by omitting one of the variables. It means that considering the omitted variable from the model, the correlation coefficient will be reduced. Therefore for more accuracy of the prediction model, the accuracy of both models responses is satisfactory and the neural model responses are even more accurate.2. For construction of the statistical model, a specific function and relation that are compatible with the frequency of data should be used. In case of non- compatibility of the models, the constructed statistical model is not valid. But constructing of the neural model does not require description of a specific function, and the neural network is self-trained by contribution of some data. Therefore modeling with neural networks is more comfortable and easier.3. Utilization of statistical model is more comfortable in a way that considering various weights for each of the independent variables, the dependent variable is obtained easily. At the other hand the obtained coefficients in statistical models are more comprehended, compared to neural network models, but the neural networks are suitable tools for applications where the responses are important factors compared to the comprehension of the model. It means that for the latter, the obtained coefficients are not illustrative of the significance of variables, and therefore studies should be implemented for such applications. For application of neural network models the inputs should be applied to MATLAB Software so that the satisfactory output is obtained.