关键词: 铁路路基病害, 探地雷达, 卷积神经网络, YOLOv3

Abstract:

Railway subgrade defects are constantly increasing. Mud pumping and subgrade settlement are the most common defect, which serious threats railway safety operations. Vehicle-borne Ground Penetrating Radar（GPR） detection method has become the main method for railway subgrade defect detection. However, the recognition of subgrade defect by radar images is still relies on artificial recognition, requiring extensive expertise of experts. Due to the large-scale and the complex shape of defect, automatic recognition is a challenging task. In response to these problems, this paper constructs a railway subgrade defect data set by field detect of vehicle-borne ground penetrating radar, and presents a real-time intelligent detection method for railway subgrade defect（LS-YOLOv3）. This method designs the deep residual network to extract the railway subgrade defect features, and multi-scale prediction networks has been used to merge feature maps on four scales to form a real-time detection model for railway subgrade defect. The experimental results show that compared with the traditional HOG+SVM, two-stage algorithm Faster-RCNN, Cascade R-CNN, one-stagea lgorithm YOLOv3 and light weight algorithm Tiny YOLOv2, Tiny YOLOv3, algorithm proposed in this paper achieves the highest mean average accuracy（82.67%） and real-time detection（32.26 frames per second）on a computing platform equipped with a NVIDIA GeForce RTX 2080Ti GPU. This paper tries to provide a real-time new method for the detection of railway subgrade defect.

Key words: railway subgrade defect, ground penetrating radar, convolutional neural networks, YOLOv3