Visual abnormality detection framework for train-mounted pantograph headline surveillance

Computer vision enhanced automatic routing inspection and monitoring for railway pantograph headlines is a promising technical orientation to reduce manual operations. The vision based approach, which comply exactly with human cognition, draws many researcher´s attention because of its informative nature. However, due to the dimension collapse in photogenic process that is essentially an ill-posed problem, to automatically detect and locate the abnormal visual pattern and structure in railway inspection videos is still a challenging task. We propose a visual abnormality detection framework by combining the appearance, scale and location of visual patterns via a probabilistic Bayesian approach. The poles and supporting arms of power supply lines are detected firstly to yield the region of interest for detailed processing. Then the hypothesis of potential abnormal pattern are collected from the endpoints extracted from local curve and line segments, so as to strengthen the perceptual difference. After that, the abnormality model and background model are implemented to classify the candidates in hypothesis. Finally, promising experimental results demonstrate the potentials of the proposed abnormal detection method with respect to various insignificant patterns and cluttered backgrounds.