Modeling accident hotspots to locate roadside equipment based on intelligent transportation system

Published in “Proceedings of the 16th International Conference on Location Based Services (LBS 2021)”, edited by Anahid Basiri, Georg Gartner and Haosheng Huang, LBS 2021, 24-25 November 2021, Glasgow, UK/online.Road transport has always attracted immense attention in Iran’s planning as one of the main transportation system and economical infrastructures. High number of accidents and road fatalities in Iran reveals the weak safety in Iran’s roads, therefore in the era which is called digital era, information technology is a necessary and efficient tool to help the management of transportation and increasing the road safety. Geospatial information system and intelligent transportation system are among the branches of information technology that are used in transportation management. Intelligent transportation system is composed of various components with different applications can be used in all of the transportation systems. On the other hand installing and setting up the components of this system is very expensive, justifying the accurate and proper location determination for these facilities. To the best of our knowledge, Empirical Bayesian Kriging Regression Prediction (EBKRP) and Forest-based Classification and Regression to model the accident prone areas and predict the hotspots has not been undertaken so far. In addition, preparing, preprocessing and exploring the impact of input variables which varies in number and nature in ArcGIS Pro software has not been done. Contribution of this research is in predicting high risk areas as an appropriate place for the installation of intelligent speed bumps using machine learning methods and data mining based on artificial intelligence in a locational intelligence field. The data used in this study consist of the official data of the traffic accidents in the period 2018 to 2019 which are available in the accidents and road transport system and has been obtained using programming in the web environment, intelligent descriptive information obtained from smart cameras of the video surveillance in the context of locational information system, non-intelligent descriptive information obtained from the Ministry of Roads and Urban Development related to the characteristics of suburban roads of Mazandaran Province in the North of Iran and also TanDEM digital elevation model with a spatial resolution of 12.5 meters. In order to predict high traffic accident risk areas, first the area of Mazandaran Province was divided into a number of hexagons to reduce the error effect of the data fusion process. The accidents data and the surrounding land uses and land covers have been extracted from the images acquired from the smart transportation monitoring cameras. For predicting the dependent variable and estimating the coefficients of significance considering the available data uncertainty, an automatic method has been proposed in this research. The method is based on heuristic regression, ordinary least squares regression and spatial rhythmic regression by considering distance as an independent variable and regression and forest-based classification with a combination of raster, vector and artificial data were used as independent variables. In addition, a new method of predicting EBK regression with raster format was proposed and implemented in this paper. Heat mapping tools have been used to convert vector variables into raster format. The integration of DEM as a variable containing ground height information with other inputs of the EBKRP method was also employed. Furthermore, the combination of digital elevation layer was used as a variable containing information about the earth with other inputs of the EBKRP method. The results showed that the information variable obtained from smart images in the training regression process and forest-based classification methods are among the effective variables in the modeling and predicting high traffic accident risk areas. In addition, it is shown that the residuals obtained from the spatial statistics employed methods have a random distribution. On the other hand, based on the validation performed for each of the implemented methods, it was found that the adjusted coefficient of determination (Adjusted-R2) for spatial rhythmic regression method has been increased compared to those of the normal least squares regression, regression method and forest-based classification. 02% of the data were selected to validate the results and the mean square error (MSE) was estimated to be 0.012. The Geostatistics toolkit in the two cases has been used in terms of time. The cross-validation method employed showed that in the case of considering the digital layer of height in the modeling process, the accuracy of the model prediction process has been improved.117123