Image processing techniques for the detection and characterisation of features and defects in railway tracks

Johnson, CI (2013) Image processing techniques for the detection and characterisation of features and defects in railway tracks. Doctoral thesis (PhD), Manchester Metropolitan University.

Preview

Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (11MB) | Preview

Abstract

This thesis describes the research that led to the development of a machine vision system in collaboration with TATA, UK and Sheffield Supertram. This was part of a European initiative for Predictive Maintenance employing non-intrusive inspection and data analysis known as PM’n’Idea. The hardware and software design, construction, and evaluation of a prototype for predictive maintenance are presented. The prototype was tested on Sheffield and Warsaw’s tram systems. The prototype has been designed with due account of a specified set of environmental constraints such as a high level of vibrations and space restrictions of the target trams. Special computer vision techniques have been specifically developed to be used with the prototype. Various image processing techniques and algorithms have been evaluated for the purpose of detection and characterisation of a series of rail abnormalities and faults. The system described in this thesis makes use of a number of standard and modified image processing techniques, not only to alleviate the requirements for manual inspections, but also to allow continuous monitoring and tracking of any defects or abnormalities in a rail track. Currently, detecting defects in their earlier stages can only be achieved by using close visual inspection i.e. line walking. Extensive testing and evaluation of the performance of the prototype inspection system at Sheffield Supertram indicated that the system was able to detect abnormalities with a resolution down to 0.1 mm. Evidence of the classification rates for the standard and modified algorithms that are implemented in the system are presented in this thesis. The algorithms developed show an average success rate of 88.9% in detecting surface bound abnormalities.