Intelligent Construction Monitoring Method for Large and Complex Steel Structures Based on Laser Point Cloud

Abstract

Large and complex steel structures play a vital role in building construction. However, deviations between the design model and the actual construction state are inevitable, which seriously affects the quality and safety of building construction. In our study, an intelligent construction monitoring method for large and complex steel structures based on laser point cloud is proposed. Firstly, three-dimensional laser scanning technology is introduced to capture accurate and complete spatial information on steel structures. Then, considering the inconsistency of the coordinate system between the design model and the laser point cloud, the building information model (BIM) is converted into the point cloud model, and the datum unification of the two types of the point cloud is achieved by adopting a coarse-to-fine registration strategy. Finally, the spatial information of steel structures is extracted from the laser point cloud based on the as-designed model, and the distance deviation between the two models is analyzed to reflect the actual construction state. To demonstrate the applicability of the proposed method, the steel structures’ point cloud of the stadium and the high-speed railway station is captured by the terrestrial three-dimensional laser scanner. The experimental results demonstrate that the method can extract the deviation between the design model and the actual construction, to provide accurate data sources for the intelligent fine construction of steel structures.

1. Introduction

Large and complex steel structures widely appeared in contemporary architecture, such as stadiums [1], airports [2], bridges [3], and high-speed railway stations [4], because of their lightweight, high strength, and toughness. Typically, the steel structures are manufactured in industrial factories and transported to the site for installation after passing dimensional quality tests [5,6]. However, in the process of steel structure manufacturing, due to dimensional errors of construction conditions, welding deformation, and raw materials, there is an inevitable dimensional deviation between the actual components and the design models, which seriously affects the on-site construction operation and process [7,8]. Hence, a robust and efficient method is necessary to assess the dimensional accuracy and structural performance of large and complex steel structures.

Traditionally, 2D drawings have been regarded as the primary design data source, and the manual measurement of the monitoring points is generally carried out using total stations or tape measures, which is difficult to monitor and record the inspection results as doing so is time-consuming, labor-intensive, and error-prone [9,10]. The deviation information cannot be accurately identified and visually fed back to the construction personnel [11]. More importantly, conventional construction monitoring methods are not working for complex steel structures because the surface and spatial performance of these elements cannot be accurately measured [7].

In recent years, building information modeling (BIM) has gained increasing interest in the field of intelligent construction. It can provide accurate information for the design, manufacture, and installation of building components due to its better completeness, higher accuracy, multidimensional visualization, and robustness to design changes [8,12,13]. However, some deviations between BIM data and the construction state are inevitable. Among these methods that have been introduced to capture actual construction states, the dense point cloud captured by the terrestrial 3D laser scanning (TLS) technology makes up for the shortcomings of traditional deformation monitoring methods, effectively avoiding the locality and one-sidedness of previous results based on sparse monitoring points [14,15,16]. The construction monitoring advantages of combined BIM and TLS technology have also been demonstrated in the application of large-scale inspection, such as buildings [17,18,19,20,21], civil infrastructure [22], plant piping systems [23], reinforced concrete [24], tunnels [25], and sheet pile [26].

To be able to accurately monitor the construction progress and the quality of the steel structures, the coordinate datum unification and comparison analysis between as-designed data (BIM) and as-built data (point cloud) models are two key problems. Some scholars have carried out a lot of research on this. Specifically, Bosché carried out a dimensional quality inspection of steel structures based on BIM and point cloud models, but the alignment of the design model and laser point cloud had a low degree of intelligence [27]. Kim et al. achieved the datum unification between the as-built point cloud model and the corresponding as-designed BIM data via the principal component analysis algorithm. Then, the dimensional and surface quality of precast concrete elements were evaluated based on the BIM and 3D laser point cloud [24]. Considering that the building has obvious corner points, Sheik et al. estimated the matching degree of the corresponding corner points extracted from the point cloud and BIM. Then, the optimal transformation parameters were evaluated, which solved the problem of building point cloud and BIM registration [28]. Apart from registration, the work of Laefer and Truong-Hong proposed a method for identifying steel structural members from the point cloud, which generated the corresponding geometric shapes compatible with BIM [29]. In addition, an improved dimensional inspection method of the piping system based on point cloud and BIM was proposed by Nguyen and Choi, and the distance and geometric parameters’ deviation were calculated for assessing the condition state [30]. Liu et al. presented a new approach to evaluating the dimensional accuracy and structural performance for free-form structure elements based on the point cloud. A series of data processing, such as laser point cloud registration, reverse modeling, and finite element analysis, were carried out for calculating and comparing the differences between the reverse model and the design model [7]. In the work of Maalek et al., for monitoring the construction progress and controlling dimensional compliance, the common structural elements, such as columns, rebars, and slabs, were extracted from the point cloud of the regular rectangular reinforced concrete structures. Then, the as-built schedule and dimensional discrepancies were automatically identified by comparing the extracted objects and the as-designed BIM [31]. Moreover, a bidirectional interaction mechanism between BIM and the multisource point cloud is shown in the work of Jia et al. Specifically, in the forward interaction, the parametric BIM provided essential information for fine construction. In the feedback interaction, for ensuring a fit between the as-designed model and the actual construction state, the laser point cloud model provided feedback to adjust the BIM. Furthermore, the resulting up-to-date BIM model can instruct the subsequent construction processes. As mentioned above, the high efficiency and feasibility of 3D laser scanning in quality detection have been demonstrated in these practices [8].

Although various effective methods exist for different scanning scenes based on point cloud processing, the complexity of the large steel structures makes it infeasible to use a single algorithm that works perfectly in every case, and existing research on the use of TLS and BIM for the dimensional inspection of large steel structures is limited. Therefore, in our study, an intelligent construction monitoring method for large and complex steel structures based on the laser point cloud and BIM is proposed. The main contributions are as follows:

(1) The coordinate unification between the point cloud and BIM is the prerequisite for construction monitoring. The combination strategy of local feature extraction and global optimization is proposed to achieve accurate alignment from coarse to fine.

(2) For monitoring the construction quality of large and complex steel structures, the deformation information is extracted from the point cloud and BIM, which provides accurate guidance for the intelligent construction of the steel structure.

The assessment framework and the methodology for large and complex steel structures based on BIM and the point cloud obtained from 3D laser scanning technology are described in Section 2. Section 3 presents the case studies and evaluates the results to determine the flexibility and applicability of the proposed method. Section 4 summarizes the main conclusion of our study.

2. Materials and Methods

Figure 1 describes the flowchart for the intelligent construction monitoring method of large and complex steel structures based on the laser point cloud and BIM. Firstly, the BIM of the steel structures was transformed into the mesh and point cloud model, which is defined as the as-designed point cloud. The laser point cloud collected by terrestrial laser scanning technology was simplified based on a pass-through filtering and voxel down-sampling algorithm, which is defined as the as-built point cloud. Secondly, a coarse-to-fine registration strategy was developed to transform the as-built point cloud into the coordinate system, where the as-designed point cloud is located. Thirdly, the corresponding grids were searched by constructing spatial indexes to extract the steel structure point cloud according to the as-designed model. Finally, the deviation between the as-built and as-designed point clouds was analyzed to monitor the quality of construction and extract its deformation information.

2.1. Data Preprocessing

2.1.1. BIM Model Processing

BIM provides extensive prior knowledge for intelligent digital dimensional inspection and pre-assembly of large and complex steel structures due to it covering geometric, physical, and spatial information [32,33]. However, it is usually represented by body structures and does not contain real geographical coordinates. Therefore, it is necessary to convert the original BIM into a data format where the coordinate information of the vertices can be read, such as a mesh or 3D point cloud model, which is not only used to extract parameters of BIM components but also used for registration and fusion between the laser point cloud and the BIM.

(a) In order to obtain BIM geometry information, the BIM object elements were traversed and converted into entities.

(b) For each entity, the function Faces was used to obtain the faces of the entity.

(c) For each face, the function Triangulate was used to triangulate the faces.

(d) Each triangle was gridded according to the preset grid size. The intersection points of the grid and the intersection point between the grid and the triangle side were taken as the output data, to obtain the as-designed point cloud.

2.1.2. Laser Point Cloud Processing

The huge amount of point cloud data obtained from 3D laser scanning technology inevitably contains redundant and noisy points, such as points that are outside the scanning setting range, do not belong to the object in the study area, and are caused by external environments (vibration, wind, and temperature), which not only increases the amount of point cloud data but also seriously affects the efficiency and accuracy of the modeling [34]. Therefore, in our study, the constraint conditions were set through the coordinate information of the steel structures, and the points within the setting range were obtained by passthrough filtering. Formula (1) shows the constraint conditions:

$$\left\{\begin{array}{c}{x}_{\min }\le x\le x_{\max }\\ y_{\min }\le y\le y_{\max }\\ z_{\min }\le z\le z_{\max }\end{array}\right.$$

(1)

where the ranges in $x$, $y$, and $z$ directions were set manually according to different scanning scenes. Based on the above steps, most of the redundant points could be eliminated. Furthermore, considering that point cloud resampling can improve the processing efficiency under the condition of meeting accuracy requirements, the voxel grid was constructed for point cloud down-sampling, which divided the point cloud space and selected the closest point from the center of the grid to replace all the points in the grid. It is characterized by a uniform distribution of sampling points, which does not move the point cloud and is more accurate.

2.2. Coarse-to-Fine Registration

To be able to compare the as-built point cloud obtained by 3D laser scanning technology with the as-designed point cloud transformed by BIM, both models need to use the same reference system, which often is not the case. Therefore, transforming the coordinate system of the as-built point cloud into the as-designed point cloud is a specialist process and is a prerequisite for automatic construction monitoring. In our study, the coarse-to-fine strategy was introduced to achieve the automatic registration of the two data sources.

As both point cloud models are in separate coordinate systems with large differences in the initial position, the registration algorithms based on 3D normal distribution transformation (3D-NDT) have difficulty in achieving correct convergence. Therefore, the rough alignment was carried out by extracting four-point congruent sets from two-point cloud models.

(a)

Construction of affine invariants

In the affine transformation, there are four points $a$, $b$, $c$, and $d$ on the surface $S_1$ coplanar, and the line segments $ab$ and $cd$ intersect at the point $e$. Defining the line segment ratio:

$$\left\{\begin{array}{l}{r}_1=\left(a-e\right)/\left(a-b\right)\\ r_2=\left(c-e\right)/\left(c-d\right)\end{array}\right.$$

(2)

If the points $a^{\prime }$, $b^{\prime }$, $c^{\prime }$, and $d^{\prime }$ are the corresponding points of $a$, $b$, $c$, and $d$ on the surface $S_2$, the $a^{\prime }$, $b^{\prime }$, $c^{\prime }$, and $d^{\prime }$ are also coplanar points and satisfy Formula (3):

$$\left\{\begin{array}{l}{r}_1=\left(a^{\prime }-e^{\prime }\right)/\left(a^{\prime }-b^{\prime }\right)\\ r_2=\left(c^{\prime }-e^{\prime }\right)/\left(c^{\prime }-d^{\prime }\right)\end{array}\right.$$

(3)

(b)

Searching for the four-point congruent sets

Three non-collinear points $p_a$, $p_b$, and $p_c$ were randomly selected from the reference point cloud $P$, a fourth point $p_d$ that is approximately coplanar with them was found. The selection process should ensure that the distance between the coplanar points is as large as possible and does not exceed the overlap area. These points form a wide-area point base $B\left(p_a,p_b,p_c,p_d\right)$. Then, the affine invariants $r_1$ and $r_2$ were determined using connected point pairs, and four points $q_a$, $q_b$, $q_c$, and $q_d$, that are approximately congruent with the point base within a threshold, were searched in the target point cloud $Q$. The constraint conditions are as follows:

$$\left\{\begin{array}{c}{e}_1=q_{\mathrm{a}}+r_1\left(q_b-q_a\right)\\ e_2=q_a+r_2\left(q_b-q_a\right)\end{array}\right.$$

(4)

If there are two pairs of points in the target point cloud $Q$ that meet the above conditions, the intermediate points $e_1$ of one pair approximately coincide with the intermediate points $e_2$ of the other pair within a threshold. These two pairs may be affine counterparts of the reference point cloud.

However, the set of congruent points obtained by the above steps was not unique. Considering that the point-to-point distances remained constant in the rigid body transformation, it was necessary to consider the distance between pairs of points in relation to the point base within the error range, i.e., adding the distance constraint between point pairs based on the invariant affine transformation ratio to obtain approximately congruent, four-point sets. The distance constraint condition is as follows:

$$\left\{\begin{array}{c}{d}_1=‖p_a-p_b‖\approx ‖q_a-q_b‖=d_1^{\prime }\\ d_2=‖p_c-p_d‖\approx ‖q_d-q_e‖=d_2^{\prime }\end{array}\right.$$

(5)

(c)

Determination of the transformation parameters

Although most of the non-corresponding point pairs could be eliminated by the affine scale invariance and the point pair’s distance constraints, the correspondence point sets were still not unique. It was necessary to calculate the optimal rigid body transformation matrix $T_i$ between the wide-area point base $B$ and the correspondence point base $U_i$, and the alignment accuracy was determined by the largest common point (LCP) sets between the two point clouds. That is, to verify the validity of the optimal transformation matrix $T_i$, it was necessary to apply it to the target point cloud $Q$ to obtain $Q^{\prime }$, and count the number of points whose distance between the reference point cloud $P$ and target point cloud $Q^{\prime }$ was less than the threshold. The transformation matrix with the highest number of corresponding points is the optimal solution.

The algorithm was based on a global search strategy to find the corresponding points under constraints and iteratively compute the optimal transformation matrix with better alignment accuracy. However, its time complexity is high in the search processing of consistency point sets, reaching $O(n^2)$. For this reason, Nicolas et al. proposed to extract the corresponding point pairs by finding a sphere with a center $q_i$ and radius $d_1+\kappa$ and $d_2+\kappa$ [35]. Specifically, by determining the point base $B$ in the reference point cloud $P$, the affine scale $r_1$, $r_2$, and the point-to-point distance $d_1$, $d_2$ can be determined. Then, point sets $S_1$ and $S_2$ in the target point cloud $Q$ were searched by (6), where the time complexity was $O(n)$:

$$\begin{array}{l}{S}_1=\left\{\left(q_i,q_j\right)|q_i,q_j\in Q,‖q_i-q_j‖\in \left[d_1-\kappa ,d_1+\kappa \right]\right\}\\ S_2=\left\{\left(q_i,q_j\right)|q_i,q_j\in Q,‖q_i-q_j‖\in \left[d_2-\kappa ,d_2+\kappa \right]\right\}\end{array}$$

(6)

(d)

Eliminating the wrong four-point sets

Considering that the angle of the line connecting two pairs of points also remained the same when performing rigid body transformations, it was also used as a constraint condition to determine the unique four-point sets. Then, the accuracy of the rigid body transformation was characterized by counting the number of points where the distance between the nearest point of the transformed point cloud and the reference point cloud was less than the threshold, and iterative calculations were carried out until the optimal one was found.

(e)

Fine registration

Based on the above steps, the two types of point cloud can approximately transform the unified coordinate system, providing a better initial position for the subsequent fine registration. In our study, the 3D-NDT algorithm was developed to achieve accurate datum unification [15]. Its core is to find the optimal transformation parameter between the as-designed and as-built point clouds by estimating the probability distribution function. A detailed description of the algorithm can be found in [36,37].

It should be noted that the selection of the cell size is the key to the 3D-NDT registration algorithm, which affects the time complexity and accuracy of the coordinate unification. If the cell size is large, it is easy to fall into a local optimum, making it difficult to achieve accurate convergence. Conversely, if the cell size is small, the speed of registration is low. Therefore, the cell size of the initial iteration was set according to the local point cloud density, and it was iteratively adjusted according to the registration result. That is, the transformation matrix generated by the last iteration was used as the initial transformation matrix for this iteration, and the iteration was repeated until it converged to the correct result.

2.3. Target Point Cloud Extraction

Although the non-object points in the scanning scenes could be removed by setting constraint conditions in the $x$, $y$, and $z$ directions, there were still redundant points around the steel structures. To eliminate the influence of redundant points on the deviation analysis, the octree index was constructed to search corresponding grids from the two registered point clouds.

Considering that the BIM design data only contain the point cloud of the steel structures, they were taken as the reference data, and the as-built point cloud was compared and analyzed with the BIM. If there was a corresponding grid in the as-designed models, the grid data in the as-built point cloud were retained; otherwise, they were regarded as redundant points and removed.

However, due to the complexity of the construction process, there was still noise attached to the surface of the steel structures only by searching corresponding grids from the two models. These points had no obvious geometrical characteristics and were difficult to remove using existing algorithms, so the human interaction approach was adopted to eliminate the noise and retain the realistic laser point cloud of the steel structures.

2.4. Intelligent Construction Monitoring

After acquiring the as-built and as-designed point clouds of steel structures, the distances between the corresponding points were calculated to analyze the construction quality, and the multiscale model-to-model cloud comparison (M3C2) algorithm was introduced in our study, which measures the displacement magnitude along the local surface normal and bypasses the uncertainties involved in mesh generation [14].

With the development of the laser scanning system hardware equipment, the large-volume and unstructured point cloud could be obtained within a short period, which led to a slow operating efficiency. In addition, the sampling point spacing expanded as the distance increased, and the point density distribution was closely related to the deviation analysis. Therefore, the notion of core points was introduced to significantly speed up data processing, which can be generated by setting a minimum point spacing [38].

Furthermore, the cylinder of radius $d/2$ whose axis goes through the core point $i$ and which is oriented along its normal vector was defined. The intercept of each point cloud with the cylinder defines two subsets of points of size $n_1$ and $n_2$. Projecting each of the subsets on the axis of the cylinder yielded two distributions of distances. The mean of the distribution yielded the average positions, $i_1$ and $i_2$, of the point cloud along the normal direction. The local distance between the two point clouds was given by the distance between $i_1$ and $i_2$.

3. Results and Discussion

3.1. Study Area

To demonstrate the applicability of the proposed method, we applied it to the steel structures’ construction monitoring for the stadium, which is shown in Figure 2a. The skylight structure of the stadium is welded indoors. As the structure parameters of the toughened glass need to match the actual model, the skylight structures needed to be scanned and analyzed in comparison with the design model to provide accurate data for its production and installation. In addition, the method was used in the construction inspection of the roof structure of a high-speed railway station, which is shown in Figure 2b. The point cloud of the roof structures was captured by terrestrial laser scanning technology and compared with the design model to assure the construction quality. The detailed descriptions of the test datasets are presented in

Table 1. Descriptions of the test datasets.

		Skyline Structures of the Stadium	Roof Structures of a High-Speed Railway Station
Number of scans		17	11
The average number of neighbors (radius = 1 cm)		6	3
Accuracy	Range (mm)	2
	Angle (μrad)	80

.

3.2. Skylight Structure Analysis

As shown in Figure 2a, the skylight structure is placed in the factory with less shelter surrounding it, and the bottom is supported on the ground using columns. According to the manual measurement results, the highest point is about 1.78 m above the ground, and the length and width are about 25.7 m and 18.2 m, respectively.

Figure 3 displays the results of the point cloud data processing, in which (a) is the diagram of the station layout. There were 17 stations in the experiment, but it was difficult to acquire the bottom of the steel structures due to their proximity to the ground. Figure 3b shows the registration results of the muti-view point cloud based on sphere targets, and the colors represent the object’s reflection intensity. Figure 3c presents the skylight structure point cloud after filtering by setting constraint conditions in the $x$, $y$, and $z$ directions.

To be able to compare the as-built point cloud captured by laser scanning technology to the BIM design data, both models need to use the same reference system. Therefore, the initial coordinates relative to the laser scanner were transformed into post-alignment coordinates relative to the design coordinates. In our study, the mesh model (Figure 4b) and point cloud model (Figure 4c) were acquired by transforming the BIM (Figure 4a) of the skylight structure, where the point cloud spacing was 2 cm.

Considering that the initial positions of the as-designed and as-built point clouds significantly differed, the coarse-to-fine alignment strategy was adopted to achieve the coordinate unification of the two point clouds. In addition, the common registration methods were conducted to compare with our proposed method. Figure 5a–e present the initial positions of the two point clouds, the coarse registration result, the result of the coarse registration combined with the iterative closest points (ICP) algorithm, the feature points-based registration result, and the registration result of our proposed method, where the red and green points represent the as-designed and as-built point clouds, respectively. In the coarse registration process, the overlap degree and the LCP distance between the two models were set at 0.5 and 0.05 m, and the maximum calculation time and the number of samples used for matching were set at 1000 s and 5000.

We can see from Figure 5 that the as-built point cloud can approximately transform the coordinate system where the as-designed point cloud is located, and the coarse registration accuracy was 0.261 m, which provides the initial position for fine registration based on the ICP and 3D-NDT algorithms. However, the ICP algorithm needs to search the nearest point of the two point clouds and is prone to local optimality, so the registration efficiency and accuracy were low, which is not suitable for large-volume steel structures. Although the registration approach-based feature points could achieve the alignment of the two types of point clouds, it was dependent on the shape of the steel structure and had a low level of automation. The 3D-NDT algorithm accurately aligned the two types of point clouds, and the registration accuracy was 7.2 mm. Furthermore, considering that there were still redundant points around the skylight structure after setting the constraint conditions in the $x$, $y$, and $z$ directions, the corresponding grids from the as-built and as-designed point clouds were searched by the octree index, and the skylight structure was extracted from the laser point cloud, which is shown in Figure 6a, where the octree resolution was 0.2 m. Then, the human–computer interaction approach was adopted to eliminate a small number of isolated points. The result is shown in Figure 6b, where the red and green points present the as-designed and as-built point clouds, respectively.

To guide site construction, the geometry deviation was analyzed by calculating the distance of corresponding points from the as-designed and as-built point clouds, which is shown in Figure 7, where (a) is the distance deviation diagram represented by color, and (b) is the statistical distribution of the distance deviations.

As we can see from Figure 7, the mean deviation between the as-built and as-designed models was 2.4 cm. Specifically, the small deviation was mostly in the welding points of the skylight structure, which demonstrated that the construction length of each component conformed to the design model. On the other hand, large deviations were seen in the top and edge areas of the skylight structure up to 300 mm, indicating that the bend significantly deviated from the design model. Therefore, to meet the construction requirements, it was necessary to adjust the dimensions in the subsequent process.

3.3. Roof Structures of High-Speed Railway Station Analysis

In the construction of steel structures on top of the high-speed railway station, BIM data provide accurate structural parameters to ensure that the steel structure is constructed according to the design model. However, there are inevitable errors due to the influence of the construction process, environment, and other factors, which seriously affect the construction safety if the deviation exceeds a certain threshold. Therefore, in this study, we obtained the complete spatial information on the roof structures of high-speed railway stations by terrestrial 3D laser scanning technology, and the deviation was analyzed by calculating the distance between the as-built point cloud and the design model.

The as-built point cloud is represented in Figure 8, in which (a) is the registration results of the muti-view point cloud based on sphere targets, and the colors represent the object reflection intensity, and (b) is the steel structure point cloud after filtering by setting the constraint conditions in the $x$, $y$, and $z$ directions, where the point cloud is colored according to the elevation value.

Considering that the BIM and as-built point clouds are not in the same reference system, the BIM was transformed into the mesh and point cloud model, and the coarse-to-fine registration approach was applied to transform the as-built model into the coordinate system in which the as-designed model is located. Figure 9a–c represent the BIM, mesh model, and point cloud model of a roof structure of a high-speed railway station, respectively.

It can be counted from Figure 9c that the number of point cloud was 2,999,636, which provided detailed geometry information on the steel structure model. Figure 10a,b display the coarse and fine registration results between the as-designed and as-built point clouds, where the red and green points represent the as-designed and as-built point clouds, respectively.

We can see from Figure 10 that although the initial positions of the as-designed and as-built point clouds considerably differed, the coarse registration could transform the two sets of point clouds into a uniform coordinate system, and the registration error was 0.068 m. Furthermore, the fine registration was carried out by an improved 3D-NDT algorithm. After that, the corresponding grids from the as-built and as-designed point clouds were searched by the octree index, and the steel structure point cloud was extracted, which is shown in Figure 11, where the resolution was 3.0 m.

After obtaining the point clouds of the steel structure, the geometry deviation was analyzed by calculating the distance of corresponding points from the as-designed and as-built point clouds, which is shown in Figure 12, where (a) is the distance deviation diagram represented by color, and (b) is the statistical distribution of the distance deviations.

As shown in Figure 12, the deviations between the as-built and as-designed models were mostly in the range of −0.3 m to 0.1 m, and the number of points accounted for 91% of the overall point cloud. It can be concluded that there was a large error between the steel structure and the design model. Therefore, in the subsequent construction process, it will be necessary to build a BIM model based on the existing laser point cloud to adjust the construction plan.

4. Conclusions

An intelligent construction monitoring method for large and complex steel structures was proposed in our study, and two main improvements were implemented. On the one hand, to be able to compare the as-built point cloud obtained by 3D laser scanning technology to the BIM design data, the BIM was converted into the mesh and point cloud model, and the coarse-to-fine registration approach was presented to transform the as-built point cloud into the coordinate system where the as-designed point cloud was located. On the other hand, the redundant and noise points were removed by setting constraint conditions and building an octree index, and the distance deviation was analyzed by searching the corresponding points from the two models along the normal direction. Although the proposed method can monitor the construction process and quality, the use of construction deviation to guide subsequent fine construction needs to be further discussed.