Railway inspection oriented foreground objects detection and occlusion reasoning for locomotive-mounted camera video

Due to the informative and compatible nature of visual data, locomotive mounted video cameras are experimentally used in rail-line inspection of pantograph overhead line infrastructures. Computer vision aided semi-automatic inspection is considered as a promising technical trend to improve the efficiency and reduce the manpower burdens. However, it is still a challenging task for computers to automatically detect and segment key structures in surveillance videos under various specific environments. This paper propose a segmentation framework for utility poles and gantries by combining the appearance and motion patterns in a sequential Bayesian approach. Observing videos are firstly frame-to-frame aligned to suppress the rotational effects; Then the potential poles and supporting arms of power supply lines are extracted by depth warping to obtain region of interest. After that the motion hypothesis of foreground and background are estimated by edge flow for occlusion reasoning. Finally, the models of pole and background are utilized to classify the candidates in hypothesis. Promising experimental results demonstrate potentials of the proposed method with respect to various insignificant target structure and cluttered backgrounds.