The quality of tunnel engineering is crucial factor for ensuring traffic operation safety. In this regard, tunnel construction analysis, tunnel deformation monitoring, and tunnel disaster prediction and early warning can be realized by using three-dimensional laser scanner to scan the tunnel regularly or irregularly to generate a point cloud model and then analyze the point cloud data. The segmentation of tunnel wall and target sphere from scanned tunnel laser point cloud scenes is an important part of tunnel 3D reconstruction and is the key technology for realizing the automatic monitoring of tunnel scenes. However, the tunnel 3D point cloud obtained by laser scanning often contains noise points and outliers, a high proportion of which is attributable to the ground points, which are connected with the tunnel wall, in the tunnel point cloud scenes. Direct processing of the tunnel point cloud data affects the extraction and recognition of the target sphere and tunnel wall. Considering the challenge of tunnel ground filtering and the lack of application platforms directly usable to segment the tunnel wall and extract the point cloud of the tunnel target ball, a ground filtering algorithm suitable for tunnel scenes and extraction method of the tunnel wall and target ball are proposed herein.

Given that extant point cloud filtering algorithms are not suitable for tunnel point cloud scenes, this paper proposes a ground filtering algorithm based on a combination of RANSAC plane fitting and pass-through filtering based on normal evaluation. A normal estimation process is added to the plane based on RANSAC fitting to ensure that the fitting optimal plane model is on the tunnel ground level. Then, the points below the optimal plane level are filtered out in conjunction with the pass-through filtering to obtain the tunnel ground points. Considering that it is challenging to extract the tunnel wall and target sphere, a segmentation method for laser point cloud on the tunnel wall (from coarse to fine) and a target sphere extraction method are proposed herein. First, the region of interest is extracted from the tunnel feature points, and the noise point clusters of the non-tunnel wall are then filtered out based on the DBSCAN method under constraints so as to obtain the tunnel wall model and possible distribution area of the target ball. The DBSCAN fine segmentation of the possible distribution area of the target ball under constraints is performed to obtain the target ball point clusters, and the nonlinear least squares (NLS) fitting is employed to obtain the ball center coordinates and related parameters of the target ball.

This paper also presents the verification of the effectiveness of the proposed ground filtering algorithm and target ball and tunnel wall extraction method in two tunnel scenes. Table 3 indicates that the proposed ground filtering method is superior to the conventional CSF filtering, slope filtering, and ground filtering methods in terms of the regional growth and elevation change for three types of errors, Kappa coefficient values, and time efficiency values. Thus, the effectiveness and advantages of the proposed filtering algorithm in tunnel scenes are demonstrated. Table 3 and Fig.7 (b) show that the Kappa coefficient value of the proposed filtering algorithm is the highest, indicating its high robustness. As indicated by the results of the fusion comparison experiment, as compared to the RANSAC plane fitting, the fusion of normal estimation, pass‐through filtering, and downsampling has a better effect on the ground filtering, and the comprehensive filtering effect is stronger than that of the comparison methods. This paper presents a comparison of the target sphere cluster (DBRTS) obtained by the DBSCAN condition constraint method and the manually intercepted target sphere cluster (Manual) by using ball fitting experiments. The final ball center error and fitting rate verify the effectiveness of the method of obtaining the target sphere and fitting method proposed herein. It is evident from the comparison in Table 4 that among the three target balls, the fitting effect of the proposed method is not as good as that of the NLS fitting. This is because the manually intercepted target ball point cloud retains more target points, and the more the details, the better the fitting effect. However, the proposed method can be employed to automatically obtain the target ball point cloud; this is more applicable and robust than the method of manually intercepting the target ball. Moreover, the spherical center error of DBRTS on the second and third fitting target balls is smaller than that of the NLS fitting method of manually intercepting the target ball; this indicates that the spherical center coordinates of the target ball fitted by this method are closer to the true values. In general, the target sphere point cloud obtained with our method and sphere fitting method proposed in this paper have evident advantages and are applicable in tunnel point cloud scenes.

Considering the 3D point cloud data of tunnel scenes as the research object, this paper first proposes a ground filtering algorithm that entails a combination of the RANSAC plane fitting method based on normal evaluation and pass‐through filtering. This method dynamically adjusts the threshold according to the point cloud scene and realizes adaptive filtering, which has good applicability and robustness. Second, a target sphere and tunnel wall extraction scheme based on DBSCAN constraint conditions and NLS sphere fitting is also proposed. The clustering segmentation is mainly realized using the density property of point cloud. The tunnel wall and target sphere point clusters are extracted by filtering the noise point clusters in the tunnel by standardizing the constraint conditions, and the relevant parameters and distribution are then obtained by fitting the target sphere point clusters based on the NLS sphere. The target sphere point cloud obtained by this method is compared with the manually intercepted target sphere point cloud. The experimental results indicate the effectiveness of this method and also highlight the advantages of the NLS sphere fitting, which completed the target sphere extraction task in tunnel scenes. Although the proposed algorithm achieves good results, a few deficiencies still exist. For example, regarding large-scale laser point cloud scenes, adoption of DBSCAN clustering segmentation is time-consuming, and some methods to improve the efficiency of clustering optimization algorithm can be studied in the future.