Computer Vision Based Health Monitoring in Railway Infrastructure


Student thesis: Doctoral Thesis

View graph of relations


Related Research Unit(s)


Awarding Institution
  • Zijun ZHANG (Supervisor)
  • Kwok Leung TSUI (Co-supervisor)
Award date27 Aug 2020


Railway transport is considered as one of efficient and sustainable modes of intrametropolitan and intermetropolitan transportation. The worldwide growth of developing railway networks has been witnessed in the past few decades. Meanwhile, the capacity expansion and the speed increase of rolling stock bring new challenges to railway operation. To ensure a safe and reliable railway service, health monitoring of railway infrastructure is expected to be conducted periodically. Although monitoring the condition of railway infrastructure based on non-destructive testing methods has been introduced in the literature, the performance of reported methods is bound to the quality and quantity of installed sensors. Thus, it is meaningful and valuable to further investigate more advanced techniques for health monitoring of railway infrastructure.

Recent achievements of computer vision based techniques have been widely discussed, which provide the potential performance in the non-contact inspection. Moreover, the flexible deployment of digital cameras on test trains has enabled the collection of images/videos and subsequently offers an unprecedented opportunity to improve the railway maintenance efficiently via image/video analyses. Some existing commercial systems for the visual inspection of railway infrastructure are not robustness enough. Thus, computer vision based health monitoring in railway infrastructure is further investigated in this thesis. Two types of image datasets, sets from the China railway high-speed system and sets from a commercial metro system in China, are employed to develop computer vision based methods for health monitoring of railway infrastructure.

In this dissertation, health monitoring of railway infrastructure focuses on two directions, inspecting rail surface condition and detecting track components. In health monitoring of the rail surface, a novel object detection based two-stage data-driven framework and an image segmentation powered framework are proposed to detect cracks and their boundaries from rail images. To conduct in-depth investigation of rail surface flaws, various flaws appearing on the rail surface are considered in the inspection. A deep learning powered framework for the rail surface multi-flaw inspection is developed. The health monitoring of track components includes the detection of critical components, rails and fasteners, as well as the recognition of missing fasteners. The developed computer vision based methods have been validated with collected datasets to demonstrate their effectiveness and advantages in health monitoring of railway infrastructure.