Comparison of deep learning and analytic image processing methods for autonomous inspection of railway bolts and clips.

• Implementation of a low-cost optical inspection system of railway clips and bolts based on RGB cameras. • Development of image processing algorithms to inspect the positioning of clips and bolts. • Evaluation and characterization of errors and computation time of the proposed solutions. In this work, different methods are proposed and compared for autonomous inspection of railway bolts and clips. A prototype of an autonomous data acquisition system was developed to automatically obtain information of the state of the railway track using LiDAR and camera sensors. This system was employed in a testing railway track installed in the facilities of the University of Vigo to obtain the images used in this work. Then, the images were further processed using analytic image segmentation algorithms as well as a neural network to detect the bolts and clips. Once these elements are detected, their relative position is computed to evaluate if there is any missing component. Finally, the orientation of the clips is computed to ensure that all the bolts are correctly placed. Four different methods were implemented, and their performance was evaluated using the segmentations provided by the analytical methods and the neural network. [ABSTRACT FROM AUTHOR]