The automatic detection of subway tunnel lining disease is realised primarily by using industrial cameras and deep learning. However, due to the uniqueness of subway tunnel environments, industrial cameras can be subject to excess interference and most of the methods are not based on disease characteristics, leaving much room for improvement. This paper proposes a method using point cloud data and a mask region-based convolutional neural network. Using two types of diseases— that is, lining water leakage and dropblocks—as the research objects, the reflection intensity and three-dimensional space information of laser point cloud data were respectively used to generate grayscale and depth maps. Based on the MASK R-CNN learning framework, the proposed method realises the simultaneous detection of two types of diseases, and performs pixel-by-pixel analysis on the feature map through the mask branch to generate a binarised mask. Experiments showed that the disease identification method proposed in this paper could achieve a global accuracy rate of more than 95%, the mAP index value being 0.4. The prediction boxes obtained could cover a more complete disease area. The proposed method combined the advantages of a grayscale map, depth map and mask R-CNN network to achieve simultaneous object detection and instance segmentation for water leakage and drop blocks, and achieved high recognition accuracy and excellent mask results—that is, the area value of the mask could be used as the basis for judging the severity of the disease, providing a certain reference for subsequent maintenance and actual operational application.
|