Deformation Detection of Mining Tunnel Based on Automatic Target Recognition

Abstract

Mining tunnels have irregular and diverse cross-sectional shapes. Structural deformation detection using mobile laser measurement has some problems, such as the inconvenient positioning of the deformation, difficulties in unifying the multiphase data, and difficulties in solving the section parameters. To address these problems, this paper proposes a mining tunnel deformation detection method based on automatic target recognition. Firstly, a mobile tunnel laser detection scheme combined with the target layout is designed. Secondly, a preview image of the tunnel lining is generated using the mobile laser point cloud data, and the index relationship between the image and point cloud is established. The target recognition accuracy of the You Only Look Once version 4 (YOLOv4) model is optimized by integrating the prediction confidence threshold, target spatial position, and target gray scale rule. Based on target recognition and positioning, the chord length and vault net height of the mining tunnel are calculated using gross error elimination and curve fitting. Finally, the engineering application of the model and algorithm is realized using ML.NET. The research method was verified using the field measurement data of the mining tunnel. The target recognition accuracy reached 100%, and the repeated deviations of the chord length and net height of the arch crown were 1.7 mm and 1.4 mm, respectively, which established the effectiveness and high accuracy of the research method.