Unsupervised learning based target localization method for pantograph video

Localization of structural regions in pantograph videos is the core and key to solve the problem of pantograph state detection using image processing technology. In this paper, an unsupervised learning based target localization method for pantograph video is proposed. First, an unsupervised learning approach is used to learn and estimate the pixel nodes in the same frame and neighborhood video image sequences that have undergone superpixel segmentation to achieve initialized prediction of the pantograph region and localization of the target region based on the correlation and semantic information between the pixel nodes. Secondly, we segment the target region by constructing the minimum energy function of the CRF model for the localization results. Finally, we construct pantograph video data sets in various environments. The experimental results show that the method is able to obtain accurate localization and segmentation results for different pantograph video data in different complex scenes, and has obvious superiority and robustness compared with other algorithms.