Abstract: In order to quantitatively identify and predict traffic risk points of non-motor vehicle conflict in intersections, visibility graph analysis method and random forest model were adopted to establish a spatial model to predict conflict points with their severity. Then, existing quantitative conflict indexes were improved so that cross conflicts, lateral scraping conflicts and accidents could be identified quantitatively, and avoidance characteristic parameter was adjusted according to avoidance trend of non-motor vehicles. Finally, verification was conducted with four typical intersections in Nanjing city as cases. The results show that the improved index can effectively complement missing values of current index in non-motor vehicle conflict quantization. And the prediction model based on visibility graph analysis and random forest model can better realize quantitative prediction and error evaluation of conflict risk points. The overall error between prediction area and actual conflict area of cases is 6.76%, which indicates a great prediction accuracy.

Key words: urban intersection, non-motor vehicle, traffic risk points, spatial prediction model, space syntax