Vision-based Foreign Objects Detection For High Speed Railway Catenary

High-speed railway EMU trains obtain electricity through the pantograph connected to the overhead catenary system(OCS)during operation.Therefore,the normal working conditions of the catenary are very important for the stable operation of the train.Various foreign objects sometimes hang on the OCS because it’s set up in the open air,which affects the power supply to the EMU trains,and causes various accidents such as train delays.Therefore,the technical research on the detection of foreign objects on the OCS has been highly valued by academia and industry.Among them,the non-contact detection method based on vision has been widely studied and applied with the development of computer vision technology.In the surveillance video data captured by the on-board camera,the visual features of foreign objects on the OCS are weak,and the complex background dynamically changes,which makes the existing algorithms unable to accurately capture the information of the foreign objects.Therefore,based on the dataset of foreign objects images on the OCS,this paper proposes an algorithmic solution for the foreign objects detection problem for the OCS,which contains two part:object detection in images and anomaly detection in videos.The main research contents of this paper include:(1)Build a dataset of foreign objects images on the OCS with surveillance video data from EMU Engineer Operation Analysis System(EOAS),including image annotation and augmen-tation.Aiming at the problem of motion blur in some images in the dataset,Wiener filtering is used to reduce image noise,thereby reducing interference information.Aiming at the problem of the weak features of foreign objects,a context-aware saliency detection method with prior knowledge is proposed to enhance the saliency of foreign objects(2)Object detection in images based on deep learning.Algorithm solutions are proposed for the two difficulties of foreign object detection in the OCS images.For the difficulty of complex dynamic background,an adaptive horizon segmentation algorithm is proposed,which crops the sky area by estimating the horizon position to reduce the interference of complex information on the ground.For the difficulty of weak feature of foreign objects,two inno-vative improvements of Single Shot MultiBox Detector(SSD)were proposed:improving the prior boxes and modifying the network structure to fuse visual saliency information,thereby strengthening the network’s ability to extract features of foreign objects.Experimental results show that the proposed algorithm achieves 88.04%mAP(mean Average Precision)and 15.67 FPS(Frames Per Second),and has the ability to detect foreign objects in the OCS images with high accuracy.(3)Anomaly detection in videos based on multi-instance learning.Considering the limita-tions of image target detection,a weakly supervised anomaly detection method based on multi-instance learning is proposed,which judges whether an abnormal situation occurs by scoring the video clips.This method achieves two innovative improvements:one is to use a stronger feature extraction network,and the other is to use attention mechanism by global context mod-ule.The two improvements strengthen the ability of the algorithm to extract features,thereby improving detection accuracy.Experimental results show that the proposed algorithm achieves 76.97%AUC(Area Under ROC Curve),and can detect anomaly events on the OCS,such as foreign objects invading.