Inspección dimensional de carriles de tren utilizando visión por computador

Résumé

In order to ensure the safety and quality of rail transportation, there are a number of national and international standards that impose tight tolerances on different properties of rails. This thesis studies laser triangulation techniques for quality control and inspection, as well as camera calibration and self-assessment algorithms suitable for a laser triangulation system. To this end, a case of study is proposed: a quality inspection system for railway rails, called the profile measurement gauge (PMG). Among all the properties of the rails, the PMG is concerned with the distances between different regions of the rail cross-section, also called rail dimensions. These include the rail height, and its width or thickness at various locations. In order to detect dimensional defects, the PMG uses laser triangulation techniques to reconstruct a cross-sectional profile of rails shortly after they are manufactured. Different profile dimensions are then measured, and the measurements are compared with the nominal values. These differences are in turn compared with the allowed tolerances for the applicable standards, and out-of-tolerance dimensions are flagged as such. The main goal of this work is to address the question of how the accuracy and reliability of laser triangulation systems can be improved and quantified. For the most part, improvements can be made in two main areas: calibration and self-assessment. In this work, all the required steps for a quality inspection laser triangulation system are explored: image acquisition, laser line extraction, coordinate translation, profile generation, profile measurement and output. The main algorithms involved are described in detail: extracting laser lines and generating point clouds from them, matching point clouds to a rail model, decomposing the model into primitives, and fitting those primitives to point cloud points. Then, the general principles of camera calibration are examined to determine how best to perform camera calibration for a system such as the PMG. Both intrinsic and extrinsic calibration parameters are considered. In order to do this, different extrinsic calibration algorithms are proposed that are designed for a special calibration target suitable for the PMG and similar systems. These algorithms are then compared with each other, and with a wellknown sheet-of-light calibration procedure. It is shown that the accuracy of the best algorithm among those proposed is similar to that of the sheet-of-light procedure. A way for human operators to assess the quality of a given calibration is also proposed. Finally, the possibility of implementing autonomic computing features for the PMG is evaluated. A number of different features are considered that may be used as indicators of the health of the system, and techniques to implement sensors to measure such features are examined. Different rules for triggering alarm events are proposed and evaluated by applying them to the results of running the sensors on historical rail data (obtained at the ArcelorMittal factory in Gijón). In this way, this work has contributed the design and evaluation of different measurement and calibration algorithms, as well as self-assessment techniques. These contributions may be used to aid in maintenance and therefore protect system accuracy and reliability, achieving a great degree of robustness which is required in an industry environment.