Fig. 1. Mirror reflection and image development

Fig. 2. Effect of mirror reflection on stripe extraction. (a) Mirror reflection at headband; (b) binarization result of Fig. 2(a); (c) mirror reflection on polished rail; (d) binarization result of Fig. 2(c)

Fig. 3. Gray-level distribution characteristics of laser stripe. (a) Laser stripe image; (b) gray-level distribution of laser stripe; (c) local zoom of Fig. 3(a); (d) local zoom of Fig. 3(b)

Fig. 4. Extraction process of light stripe center line based on sub-region multi-template matching

Fig. 5. Image segmentation performance statistics for different depth network structures [14]

Fig. 6. Rail profile data labeling

Fig. 7. Section partition of rail profile laser stripe

Fig. 8. Gradient direction of pixels.(a) Gradient direction solved for whole image; (b) gradient direction solved only on segmented light stripe

Fig. 9. Direction template corresponding to each sub-range of segmentation

Fig. 10. Extraction of pixel-level light stripe center based on direction template

Fig. 11. Extraction of sub-pixel level light stripe center based on gray-level centroid method

Fig. 12. Image rotation and scaling

Fig. 13. Comparison of light stripe segmentation results. (a) Original rail profile image; (b) fixed threshold binarization; (c) dynamic threshold binarization; (d) deep learning segmentation results

Fig. 14. Convolution features of light stripe image. (a) Convolution feature of first layer with size of 16×256×256; (b) convolution feature of second layer with size of 64×128×128; (c) convolution feature of third layer with size of 128×164×164

Fig. 15. Experimental results and corresponding local zoom for different laser stripe center extraction methods. (a) Gray-level centroid method[17]; (b) direction-template method [18]; (c) Steger method [19]; (d) proposed method

Fig. 16. Precision comparison of different laser stripe center extraction methods

Fig. 17. Execution time comparison of different laser strip center extraction methods