A Hybrid Online Method for Accident Prediction Non-Parametric Approach

This study aims to address a model to improve the road safety based on the spatial and temporal features of accident occurrence. The spatial positioning of the accident distributionfollows a probability density function (PDF) which can be estimated with parametric functions. Estimating the PDF of accident would help the road specialist to recognize theaccident prone locations and improve the road safety through countermeasure decisions. Inthis study, the PDF of accidents is estimated by Gaussian Mixture Models (GMMs). Gaussian models possess the local and density information about the accident distribution.For the predicting step, the parameters of Gaussian models are predicted by the means of Recursive Least Square (RLS) approach. Huge databases, such as traffic and accident databases, inevitably contain some noisy information. Thus, a hybrid model is proposed toincorporating the available prior knowledge of accidents occurrence in the prediction process. In accident databases, the prior knowledge can be defined as the most probable points which an accident had happened in earlier years and is closely correlated to its upcoming years. Maximum a Posterior (MAP) estimator is used as the prior knowledge module. The major feature of this study is that the proposed model can accurately predictthe physical points where the accident are more likely to happen in the upcoming year. Moreover, because of the non-stationary feature of accidents, the proposed model is online, implying that it would update itself annually. The range of errors obtained from the results of cross validation on the existing data is a very good indication of the accuracy of the proposed model.

کلیدواژه ها:

Road Safety ، Gaussian Mixture Model ، Recursive Least Square ، Prior Knowledge ، Maximum a Posterior Estimator