Rail Fastener Status Detection Based on MobileNet-YOLOv4

As an important part of track inspection, the detection rail fasteners is great significance to improve safety train operation. Additionally, fastener belongs small-target detection. The YOLOv4 algorithm relatively fast in and has some advantages Therefore, used for status However, still suffers from following two problems First, features extracted by original feature extraction network are rough, which not conducive crack anomaly on fasteners. In addition, traditional convolutional neural a larger number parameters calculations, difficult run embedded system with low memory processing power. To effectively solve those problems, this paper proposes based MobileNet-YOLOv4 (M-YOLOv4). edge texture very detection, CSPDarknet53 cannot extract MobileNet replace algorithm, can subtle reduce calculations algorithm. experimental results show that M-YOLOv4 high accuracy resource consumption false-alarm rate (FAR), missed-alarm (MAR), error (ER) were 5.71%, 1.67%, 4.24%, respectively, speed reached 59.8 fps. Compared YOLOv4, reduced about 80.75% 83.20%, respectively.