Foreign Body Detection in the Electrified Area of Urban Rail Trains Using Improved Yolov3 Algorithm

Abstract

Foreign body invade the electric receiving area of urban rail train, interfere with the operation of electric equipment on the roof, and affect the normal operation of urban rail traffic. Aiming at the problems of the traditional non-contact foreign body detection in the electric area of urban rail train, such as slow detection speed and poor detection accuracy of small target foreign body,An improved YOLOV3 (You Only Look Once) network model based on PAN feature pyramid structure and adaptive spatial feature fusion is proposed. By improving the main body of the YOLOv3 network model, it can alleviate the problem that the network prediction size map is too large and the experience field is too small. The features of different levels of foreign objects are initially fused with PAN’s feature pyramid to extract strong location information and strong semantic information of the foreign objects, then the method of adaptive spatial feature fusion was used to learn the spatial weights of the fusion of feature maps at various scales, obtaining more effective prediction feature maps at different scales after fusion and improving the detection ability of small targets. The improved k-means clustering algorithm is used to obtain the size of anchor and match it to the corresponding feature layer, which can mark the position of foreign body more accurately. Experimental results show that the detection accuracy of the improved YOLOV3 reaches 95.7%, which is 5.1% higher than the detection effect of the original network. It can accurately and quickly identify the different size of intrusive foreign body in the electric area of the roof of the urban rail train.