Automatic 2D-3D vision based assessment of the attitude of a train pantograph

In this paper we propose an automatic visual based technique, integrated in a wayside monitoring system for train inspection, that allows to assess the attitude of the metal bow of a pantograph by combining a colour image captured by an RGB digital camera and a point cloud built from a range sensor scan. An efficient and fast template-matching procedure allows to detect the pantograph in the scene and associate a matching attitude, searching for the most similar model present in a database. The record of templates belonging to the database exploits a virtual rendering environment that allows to optimize the training stage in terms of computational load and time. During actual inspection the RGB image and point cloud of the pantograph are opportunely processed and aligned to the same reference frame. After the preliminary template-matching step, the point cloud is augmented with the virtual model of the matched template and the attitude angular values are refined by applying the iterative closest point (ICP) algorithm between the real object and the virtual one, with the aim of reducing eventual residual errors.