Computer vision for railroad track inspection

Railroad Track Inspection;Computer Vision;Periodic Images;Multiple Signal Classification;Activity Recognition;Periodic Motion in Video

Resendiz, Esther I.

关键词: Railroad Track Inspection;    Computer Vision;    Periodic Images;    Multiple Signal Classification;    Activity Recognition;    Periodic Motion in Video;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/18558/Resendiz_Esther.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

In railroad track inspection, the inspection images contain periodically occurring components. Computer vision has recently been applied to several railroad applications due to its potential to improve the efficiency, objectivity, and accuracy when analyzing large databases of acquired video and images.We utilize those promising results to develop a more general method to detect and segment any periodically occurring objects in an image.The techniques used to analyze the periodically occurring track components could be used to analyze a broader class of images which contain periodically repeating objects that are similar, but not identical.We demonstrate how spectral estimation-based methods can be used to extract periodically repeating components in track inspection video.Periodically occurring activities occur in many videos.Particularly in biological applications, activities tend to be formed from one or two characteristic poses that move in a repetitious manner.We introduce a signal-processing based method for periodic activity detection and segmentation that utilizes a unified spatial-frequency approach.The spectral estimation technique that we used requires a one-dimensional signal as input.In images and video, one-dimensional signals are created.We demonstrate how the more general technique of frequency estimation, object localization, and iterative decomposition using the frequency domain can be used to analyze images with periodically occurring components, video of translating images, and videos containing periodic activities.Additionally, a method is introduced that quantifies the perceptual quality reduction in distorted images.Humans perceive distortion in images more prominently when it occurs in perceptually salient regions.This is similar to detecting periodically occurring objects, since humans will notice periodically occurring objects.Objects that occur in a periodic fashion, and whose photometric properties result in more saliency, will be more observable from a human's perspective.We demonstrate our signal processing-based methods on railroad track inspection images, which were our primary motivation.We provide more experimental evidence of its generalization beyond this specific application.