Predictive modeling for monocular vision based rail track extraction

Camera based approaches are the most widely analyzed applications of track extraction for railway systems. Finalizing the extraction with an analytic representation of the rail tracks is very useful both for representing the smooth characteristics of the railways well and for compensating the faulty image processing results. Moreover, polynomial or geometric definition of the rails creates parametric representations, which can be transformed to generate a dynamic ROI around the rails and can be used for applications like obstacle detection. However, the existing studies only use presumed knowledge and assumptions about the geometry of railway tracks. This paper provides a statistical analysis of railway track turns on various video records to quantify the margins of variations. Moreover, based on the results of those statistical analyses, this study introduces a method to predict the railway tracks by means of polynomial approximation followed by multilayer perceptron networks.