Geometry constraints-based visual rail track extraction

Abstract

Rail track extraction from the image can be used to determine the position of the rails ahead of a train, which is one of the fundamental tasks for vision-based driver support systems in railways. This paper introduces a method that extracts rails by matching the edge features of the real image to the candidate rail patterns which is parameterically modeled, and the geometric constraints of the rail pattern are taken into consideration during the pattern generation. To address the challenge posed by the open environment, we assume the top surfaces of the pair rails are located in a virtual plane, then a homography matrix can be used to describe the geometry relationship between the virtual plane and the camera image plane. Then based on inverse projective mapping (IPM), the whole image is divided into several sections along the vertical direction. In each section, the rail pair can be approximated by two parallel line segmentations. Based on the prior geometric constraints, candidate rail patterns are generated and the rail track recognition is modeled as a 2D searching process. The rail track extraction in the whole IPM image is obtained by integrating the results of each section. This proposed method combines several advantages at each processing level to improve the robustness. During this integration stage, a curve fitting is applied and statistics of some parameters, such as the direction and position of the line segments can be adopted to remove the noise results from each section. Experiments show the performance of the proposed method.