基于改进YOLOv4算法的铁路扣件检测

扣件的健康状态是保障轨道车辆正常运行的关键。当前人工检测轨道扣件效率较低,具有缺陷性。针对这一问题,提出了基于改进YOLOv4算法的轨道扣件与检测。在YOLOv4网络中,利用CSPDarknet53第二个残差块嵌入conv卷积结构与YOLO头部结构,增加输出端,并进行网络中的上采样与下采样。与YOLOv4原算法模型相比,提升了准确率与检出率。将使用改进YOLOv4的方法,实现对有砟轨道与无砟轨道上扣件的状态检测。试验结果表明：基于改进YOLOv4算法检出率和准确率比原YOLOv4算法分别提升4.65%和4.88%,并且YOLOv4模型体积与其他模型相比更小,适用于轨道扣件检测。

Detection of Track Fastener Based on Improved YOLOv4 Algorithm

The state of fasteners is the key to the operation of railway vehicles. At present, manual detection of track fasteners is inefficient and defective. To solve this problem, the track fastener location and detection were proposed based on the improved YOLOv4 algorithm. In the YOLOV4 network, the second block of CSPDarknet53 was used to embed the conv convolution structure and the YOLO head structure, add output terminals, and carry out up sampling and down sampling in the network. Compared with the original YOLOv4 algorithm, the accuracy and detection rate are improved. The improved YOLOv4 method was used to detect the state detection of fasteners, and realize the state detection of fasteners on ballast track and ballast free track. The experimental results show that the detection rate and accuracy of the improved YOLOv4 algorithm are 4.65% and 4.88% higher than the original YOLOv4 algorithm. And the volume of YOLOv4 model is smaller than other models, which is suitable for the detection of track fastener.