Rail Fastener Status Detection Based on MobileNet-YOLOv4

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