Visual Detection Approach of Pantograph-Catenary Arcing Based on Dense Mesh Learning

Pantograph-catenary system is an important part of the railway train. The appearance of offline arcs at the contact point between pantograph and catenary will affect the stability of train current collection. Accurately identifying the occurrence of arcs is helpful to maintain the pantograph-catenary system and enhance the train’s safety and reliability. Aiming at the problems that the existing methods have poor detection effect on small arcs and the detection results are greatly influenced by the surrounding environment, this paper proposes a visual detection approach for pantograph-catenary arcing based on dense mesh learning. First, a dense mesh attention network based on Unet is used to extract the arcing features. Then, a feature pyramid network embedded in the context modulation module is applied to integrate the different levels features and highlight the small arcs’ feature in the deep network. Finally, a post-processing algorithm is utilized to reduce the false detection rate. Experiments on the data of railway operation pantograph-catenary system show that the mean Intersection over Union (mIoU) of the proposed method gets 0.8143, the F1-Measure obtains 0.8845, and the frame per second (FPS) achieves 18.7. The experimental results indicate that our proposed model has higher performance on mIoU and F1-Measure compared with several typical detection methods, while the detection speed is improved by 78.1% compared with BlendMask.