Abnormal Detection of Track Fastener Based on Improved YOLOV3 Algorithm

Abstract

In order to improve the low efficiency and insufficient accuracy of manual inspections in traditional railway track maintenance, an improved YoloV3 (you only look once V3)algorithm is proposed to detect abnormalities of track fasteners. First, the K-means++ algorithm was used to make some clustering analysis of the data set from the rail fastener, and determine the height-width ratio and the number of prediction boxes. Then, according to the characteristics of the track fastener size, the network multi-scale recognition was improved to solve the problem that YoloV3 can not recognize when the target box size is too small. For the last, the loss function, using GIOU(Generalized Intersection over Union)loss instead of the MSE(the Mean Square Error)loss, prevents abnormal change of the scale. Meanwhile, the imbalance between positive and negative samples is also settled by using Focal Loss instead of the cross entropy loss. The results show that the high rate of data detection with Yolo V3 algorithm can reach a speed of 37 frame/s and the average precision is as high as 93.51 %, which is 10.64% higher than that of original network. The result indicates the accuracy and efficiency of our method.