Abstract: During intelligent shotcrete operation in tunnels, accurately extracting the position of steel arches on complex tunnel walls is challenging. To address this issue, an extraction method for steel-arch frames in TBM tunnels based on laser point-cloud feature template matching is proposed. First, a convolutional kernel sensitive to gradient directions was specifically designed to accurately capture the feature differences of steel-arch structures in different gradient directions, effectively enhancing the distinguishability between the steel-arch targets and the uneven tunnel wall background. Second, a geometric constraint template, which utilizes the inherent geometric features of steel-arch frames, was constructed to strictly filter nontarget areas in images. Through the synergistic effect of the convolutional kernel and the geometric constraint template, accurate positioning and extraction of steel-arch frames in the uneven tunnel wall environments is realized. Finally, to verify the performance of the algorithm, it was applied to an experimental test in a tunnel under construction in southwest China. The results show that, without manual assistance, in three typical tunnel scenarios, the comprehensive accuracy, recall rate, and F-score reach 91.18%, 93.33%, and 92.22%, respectively, and the computation time of the algorithm is on the order of seconds. Compared with traditional recognition methods, the proposed method demonstrates stronger adaptability in complex environments, more stable and reliable recognition performance, and significantly outperforms traditional algorithms, providing robust technical support for the efficient recognition of steel-arch frames in intelligent shotcreting.

Key words: feature extraction; , point-cloud segmentation, template matching, three-dimensional reconstruction, TBM tunnel, intelligent shotcreting, steel-arch frame recognition