The method for the fast, accurate and reliable extraction of laser stripe centers in the train-running environments is studied. The multi-section fast segmentation of laser stripe is achieved based on the ENet deep learning model. The histograms along the gradient direction of light stripes in each section are counted to determine the normal dominant direction and construct the corresponding direction template. The gray-level centroid method with sub-region multi-template matching is proposed for the extraction of sub-pixel coordinates of light stripe center. The research results show that the proposed method can effectively overcome the influences of various types of interference information on laser stripe center extraction in the outdoor train-running environment. The light stripe extraction time is only 2.1 ms for a single rail profile, the mean error is about 0.082 pixel and the standard deviation is 0.047 pixel. The method has good time-effectiveness and high accuracy for laser stripe center extraction.