Error Distribution Pattern Analysis of Mobile Laser Scanners for Precise As-Built BIM Generation

Abstract

Point clouds acquired by mobile laser scanners (MLS) are widely used for generating as-built building information models (BIM), particularly in indoor construction environments and existing buildings. While MLS offers fast and efficient scanning through SLAM technology, its accuracy and precision remains lower than that of terrestrial laser scanners (TLS). This study investigates the potential to improve MLS-based as-built BIM accuracy by analyzing and utilizing error distribution patterns inherent in MLS point clouds. Based on the assumption that each MLS device exhibits consistent and unique error distribution patterns, an experiment was conducted using three MLS devices and TLS-derived reference data. The analysis employed iterative closest point (ICP) registration and cloud-to-mesh (C2M) distance measurements on mock-ups with closed shapes. The results revealed that error patterns were stable across scans and could be leveraged as correction factors. In other words, the results indicate that when using MLS for as-built BIM generation, robust fitting methods have limitations in obtaining realistic object dimensions, as they do not account for the unique error patterns present in MLS point clouds. The proposed method provides a simple and repeatable approach for enhancing MLS accuracy, contributing to improved dimensional reliability in MLS-driven BIM applications.

1. Introduction

Point clouds are widely used to generate as-built building information models (BIM). These point clouds can be acquired using various technologies, such as laser scanners or stereo cameras [1,2,3]. Among these technologies, terrestrial laser scanners (TLS) have been actively employed in Scan-to-BIM applications due to their high accuracy [4,5]. However, acquiring comprehensive point clouds with TLS is both labor-intensive and time-consuming, as it requires multiple setup positions to capture the entire construction site. Indoor environments, in particular, demand a greater number of scanning positions than outdoor settings, further increasing the acquisition time.

To address these limitations, mobile laser scanners (MLS) have been adopted for capturing indoor construction sites [6,7]. Unlike TLS, MLS can perform continuous scanning at a higher speed by leveraging simultaneous localization and mapping (SLAM) technology. Despite this advantage, MLS systems generally offer lower accuracy, as SLAM-based scanning is inherently less precise than TLS-based methods [7,8]. These inaccuracies can significantly affect various construction applications, including inspection, quantity estimation, augmented/virtual reality (AR/VR), digital twin development, automation, and robotics.

To overcome these limitations, several enhanced MLS devices have been developed by integrating TLS and SLAM technology [9], or incorporating GNSS technology and highly precise IMU sensor [10]. However, these devices remain expensive, and GNSS use is constrained indoor environments. Yuan et al. [7] used MLS to evaluate rebar spacing and quality. The device used was the GeoSLAM ZEB-Horizon, which has a relative error up to 6 mm [11]. Ibrahimkhil et al. [12] used the GeoSLAM ZEB-Revo for as-built BIM generation, with a reported relative accuracy of ± 10–30 mm [13]. Cui et al. [14] also used GeoSLAM ZEB-Revo for 3D reconstruction. Sgrenzaroli et al. [15] generated as-built BIM from MLS point clouds captured using Heron MX Twin Color scanner, which has a global accuracy of less than 20 mm [16]. Consequently, MLS products widely adopted for indoor construction scanning still deliver lower accuracy than TLS.

These studies demonstrate the expanding application of MLS in Scan-to-BIM. However, most of the cited MLS devices exhibit relatively low accuracy. Moreover, existing studies utilizing MLS point clouds in BIM applications have paid limited attention to the inherent inaccuracies in MLS point clouds. Accordingly, a thorough characterization of these inaccuracies, along with the development of effective mitigation strategies, could play a critical role in ensuring their reliable application.

Nonetheless, previous research on MLS accuracy has rarely addressed such error distribution patterns. Furthermore, studies employing MLS for the generation of as-built BIM models have typically overlooked these device-specific biases in their methodologies. In other words, current approaches often apply algorithms such as random sample consensus (RANSAC) without accounting for these deviations, which may significantly impact the overall geometric accuracy of the resulting as-built BIM.

Meanwhile, for the correction of MLS point clouds, previous studies have explored approaches that utilize geometric features for self-calibration during real-time generation of complete MLS point clouds [17], as well as machine learning-based methods [18]. These studies focus on the development of more advanced MLS. While they contribute to improving the accuracy of MLS, they do not address the cumulative error distribution patterns that may occur during actual use.

Therefore, we designed a method to analyze error distribution patterns of MLS point clouds and conducted experiments for three MLS devices using this method. Additionally, this study verified that the obtained error pattern could serve as a correction factor to improve geometric accuracy of as-built BIM models.

This study is based on two key assumptions. First, MLS point clouds obtained from the same device exhibit similar error distribution patterns, and these patterns remain consistent across multiple scans of the device. Second, if the error distribution pattern of each MLS device can be measured, it can be used as a correction factor to generate more accurate as-built BIMs from MLS-acquired point clouds.

In order to validate these assumptions, the authors designed an experiment to identify error distribution patterns of MLS point clouds and designed a simple method to apply the patterns for improving the accuracy of as-built BIM. The experiment was based on the principle of iterative closest point (ICP) registration and involved the use of mock-ups with a closed shape in the xy-plane. The cloud-to-mesh (C2M) distance measurement method was employed to measure the distances between each point of the MLS point clouds and the reference planes. These reference planes were obtained from point clouds collected using TLS. The experiment was conducted using three MLS devices, resulting in a total of 11 datasets collected from an experimental room. The error distribution patterns for each device were analyzed, and their potential for dimensional error correction was demonstrated through the proposed method and corresponding test.

This paper is organized as follows: Section 2 reviews previous research on MLS analysis and the measurement of building component dimensions in point clouds. Section 3 describes the experimental design and the proposed analysis method. Section 4 presents the experimental results on MLS error distribution patterns. Section 5 discusses the proposed analysis method, the experiment, and its results. Finally, Section 6 provides a summary of the key findings and conclusions of this study.

2. Literature Review

2.1. Analysis of MLS Point Clouds

Most of the related studies analyzed MLS devices based on comparison between MLS point clouds and reference data. Kalvoda et al. [19] evaluated the performance of vehicle-mounted MLSs using precisely measured feature points, applying metrics of absolute errors and standard deviations to assess performance. Sepagozar et al. [20] assessed the applicability of MLSs for architectural components like doors and windows, using manually measured dimensions as reference data and comparing them with dimensions extracted from point clouds. Lin et al. [21] evaluated the performance of vehicle-mounted MLSs by calculating the area of targets, such as windows on walls. The area was then compared to the outline of the target. The outline was generated by applying the alpha-shape algorithm to the point clouds obtained from MLSs. These studies used precisely measured feature points or areas of targets for comparative analysis.

Several studies have used TLS data as reference data. The ICP registration algorithm is commonly employed to align MLS point clouds (comparison data) with TLS point clouds (reference data). In this process, the MLS point clouds were translated and rotated to minimize the root mean square error (RMSE) between them and the TLS point clouds. Toschi et al. [22] evaluated the accuracy of the RIEGL VMX-450 by aligning photogrammetry and TLS-based reference point clouds using ICP and analyzing the error distribution with non-parametric statistics to account for non-Gaussian characteristics. Al-Durgham et al. [23] segmented MLS point clouds into temporal slices and applied planar feature-based registration to TLS data. Yiğit et al. [24] conducted object-based comparisons between MLS and TLS, including Pegasus Two and Backpack systems, using TLS point clouds as references. Tucci et al. [25] assessed discrepancies between MLS point clouds and TLS references using both multi model-to-model cloud comparison (M3C2) and cloud-to-mesh (C2M) distances. Maset et al. [26] compared robotic and handheld MLS modes against TLS benchmarks by analyzing point distribution and repeatability in GNSS-denied environments. Moyano et al. [27] quantified discrepancies between the BLK360 and Riegl VZ400i at both object and structural levels using TLS data as the baseline. Subsequently, deviations between the aligned TLS and MLS point clouds were either measured using cloud-to-cloud (C2C) distances [23,24,26,27], or analyzed based on semantic features of the objects [22]. Additionally, Tucci et al. [25] performed a comparative analysis of feature points following ICP registration. Furthermore, TLS point clouds, total station measurements [8,28], and synthetic reference models derived from as-planned BIMs [29] were also used for evaluating MLS performance. Teo and Yang [30] used MLS point clouds as references to evaluate the feasibility of implementing a scan-to-BIM workflow using an iPad LiDAR sensor. They extracted planar elements from MLS point clouds and calculated C2M distances and comparing them with measurements obtained from manually created as-built BIMs.

To summarize, previous studies have performed analysis of MLS point clouds using TLS point clouds, precisely measured feature points, target areas, or benchmark point clouds generated from as-planned BIMs. These studies analyze MLS point clouds using distance error-based metrics which are measured using C2C or C2M distances. However, previous studies present challenges to understanding the error distribution pattern in MLS point clouds. In addition, previous analyses using ICP registration are often applied to either a single surface of the analysis target or the global point clouds. Consequently, the unique error distribution patterns of MLS point clouds may be overlooked. Therefore, it is necessary to develop a method that analyzes the error distribution pattern in MLS point clouds.

2.2. Measurement of Object Dimensions Using Point Clouds

Measuring object dimensions from point clouds is a critical component of the Scan-to-BIM process, as it directly impacts the geometric accuracy of the resulting as-built BIMs. One of the most widely used and robust methods for extracting geometric primitives, such as lines and planes, is the RANSAC algorithm. Tang et al. [31] proposed a BIM generation method using a morphological approach. They used RANSAC to determine planar primitive segmentation to extract vertical and horizontal planes of each room region. Singh et al. [32] identified roof bolts in MLS point clouds which captured the mine roadway. RANSAC algorithm was adopted to extract cylinder fitted mesh of roof bolts. Pouraghdam et al. [33] used RANSAC to obtain representative lines of wall components in 2D projected point clouds. In addition, RANSAC algorithm was utilized to obtain representative planes or lines of objects from point clouds [34,35,36,37,38,39,40,41]. Thus, RANSAC has been widely adopted in Scan-to-BIM studies, as it fits the optimal lines or planes while accounting for outliers in the input point clouds.

Several studies have employed approaches that differ from the conventional RANSAC algorithm. Nguyen et al. [42] proposed a novel method for planar surface detection in MLS point clouds. This method utilizes scan profile patterns and planarity values between neighboring scan profiles to detect and segment planar surfaces. It demonstrated higher performance compared to RANSAC. Wang et al. [6] extracted representative object lines by leveraging normal vector-based tangent plane estimation and region-growing facet generation from 3D point clouds. Jung et al. [43] extracted indoor structure planes and lines using morphological processing with a 2D binary map and skeletonization. Xiao et al. [44] extracted representative planes and lines of indoor structures by coupling point cloud completion with surface connectivity inference, leveraging truncated signed distance function (TSDF) octree-based visibility estimation and normal-based region growing. Wu et al. [45] proposed an advanced method using modified ring-stepping clustering to extract lines from the xy-plane of main building components.

In summary, RANSAC has been the most frequently used and reliable method for extracting object lines and planes from point clouds. Additionally, approaches such as neighboring scan profiles, normal vectors, 2D binary maps, and surface connectivity inference have also been explored. These methods effectively extract representative lines or planes from objects captured in point clouds. While these methods extract accurate lines or planes, they overlook the potential bias in the point cloud distribution, which may significantly impact the geometric accuracy of as-built BIMs. In contrast, little research has explicitly addressed the analysis of error distribution patterns in MLS point clouds. Therefore, this study aims to bridge this gap by providing an in-depth analysis of the error distribution patterns in MLS-derived point clouds.

3. Experiment Design and Analysis Method

3.1. Laser Scanners

In this study, a commercial TLS and three handheld MLSs (MLS-01, MLS-02, and MLS-03) are used for experiments. TLS-01 was used for generating reference data, and the detailed method will be described in Section 3.4.1. MLS-02 and MLS-03 are the same model, but MLS-02 has been in use for over two years, whereas MLS-03 is newly purchased.

The model names of the scanners are intentionally withheld; however, they are MLS devices that are widely used in the field. Our goal is to present a repeatable method for analyzing error distribution patterns in MLS point clouds and to confirm whether each device exhibits a unique pattern. The proposed analysis and experiments are reproducible regardless of the scanner model. However, accuracy-related specifications play a crucial role in our analysis of error distribution patterns; therefore, we have provided them in

Table 1. Specifications of terrestrial laser scanner (TLS) and mobile laser scanners (MLSs) used in this study.

Specifications	TLS	MLSs
	TLS-01	MLS-01	MLS-02 and MLS-03
Point (range) accuracy	±1 mm	10 mm	10–30 mm
3D accuracy	2 mm @ 10 m, 3.5 mm @ 25 m	-	-
Relative accuracy	-	≤10 mm	Up to 6 mm

.

Table 1 summarizes the key metrics related to accuracy. Point (range) accuracy refers to the systematic distance error of a single laser pulse, indicating how far the average measurement deviations from the actual surface. 3D accuracy incorporates angular errors in addition to range errors, representing the total 3D positional deviation that may occur at a given distance, particularly in the context of TLS performance. Relative accuracy refers to the final point cloud accuracy of MLS, encompassing all accumulated error throughout the SLAM process. This study primarily focuses on relative accuracy. Rather than analyzing internal SLAM components or individual error stages specified by the manufacturer, our interest lies in how the accumulated error manifests as bias in the final MLS point clouds and how accurately it reproduces actual surfaces.

3.2. Mock-Ups

The objective of the experiment is to analyze error distribution patterns in MLS point clouds. To achieve this, analysis targets were installed in the experimental room [15,16,17]. In this experiment, column-shaped objects (Column A and Column B) were used as analysis targets. Figure 1 presents the dimensional design diagrams of the Columns. The dimensions of Column A are 497 mm × 505 mm × 1500 mm, while those of Column B are 500 mm × 500 mm × 1500 mm. Both Columns were installed in the experimental room with their vertical alignment adjusted to match the gravity direction (z-axis) as closely as possible.

The objective of this study is to identify error distribution patterns that are biased from the actual surface of objects, and to examine whether the resulting error pattern can contribute to improving the accuracy of BIM reconstruction using MLS point clouds. Columns are one of the primary building components found in indoor construction environments. In addition, they provide four surfaces with sufficient area, making them suitable for our experiment, which involves extracting and analyzing data at the planar-segment level. For these reasons, column-type mock-ups were used in our experiments as the analysis targets.

Due to the room’s height and the limitations of handheld MLS devices, the ceiling and floor planes of these columns were not captured in the resulting MLS point clouds. Consequently, the MLS point clouds of the Columns appear as closed shapes in the xy-plane. This characteristic plays a crucial role in the proposed method for analyzing the error distribution patterns in MLS point clouds. The detailed analysis procedure is described in Section 3.4.

3.3. Data Acquisition

MLS acquires point cloud data using a SLAM-based approach that accounts for scanner movement. The data acquisition method of MLS generally includes a closed-loop algorithm, which reduces accumulated positional errors by performing feature matching between the start and end points of the scan and correcting the overall trajectory. Therefore, handheld MLS devices used in this study have difficulty obtaining stable point cloud data from a fixed position, unlike TLS. Therefore, the present experiment adopted a data acquisition approach where MLS point clouds were collected by traveling along a predefined path within a laboratory space to ensure the stable acquisition of handheld MLS data.

To ensure precise analysis, several constraints were applied during the acquisition of MLS point clouds. The MLS operator followed a predefined scanning path while capturing the MLS point clouds. Figure 2 illustrates the top view of the experimental room, depicting the installation of Column A and Column B, the MLS scanning path, and the scan positions of the TLS for reference data acquisition.

The MLS scanning path was designed to maintain a two-meter distance between the Columns and the MLSs. A temporary wall was installed to prevent data acquisition from excessive distances. These constraints were implemented to reduce the impact of distance-induced errors. The TLS scan positions were set at six locations to ensure comprehensive data acquisition for all surfaces of the Columns. The TLS point clouds obtained from these six positions were precisely registered using commercial software associated with TLS device and utilized as reference data. The registration error between the six TLS point clouds was measured at 0.5 mm. This 0.5 mm indicates that the average distance between corresponding point pairs during the registration of six point clouds was measured to be 0.5 mm. Such registration errors are inevitable in point cloud registration, particularly due to variations in point density and the characteristics of space-level data registration. Therefore, the 0.5 mm value is considered sufficiently reliable for use as reference data.

The acquisition of MLS point clouds was performed four times for each MLS device (MLS-01, MLS-02, and MLS-03). These four scans enable verification of whether the MLS point clouds consistently exhibit the same error distribution pattern across multiple scans. However, one dataset from MLS-01 was missing, so only three datasets were used in the actual analysis. Further details are provided in Section 4.1. Two scans were initiated from the left-hand scanning point as shown in Figure 2, while the other two started from the right scanning start position. This approach considers the possibility that the error distribution pattern may vary depending on the specific moment when the analysis target is captured during MLS scanning. Figure 3 presents a photograph of the experimental room where the actual Columns were installed.

The MLS point clouds were processed using the software provided with each MLS device. As a result, procedures such as SLAM algorithms and loop closing techniques, which are built into the device and software, are applied automatically to generate complete MLS point clouds. However, users typically cannot access the detailed implementation of these proprietary algorithms. For this reason, in our study, we applied only the minimum procedures necessary to produce complete point clouds and did not include any additional noise filtering or post-processing specific to each software.

3.4. Analysis Method of Error Distribution Patterns in MLS Point Clouds

This study proposes an analysis method for understanding the error distribution pattern in MLS point clouds. The proposed method was developed based on the following two key characteristics: (1) the use of the ICP registration, which minimizes the RMSE during the alignment of point clouds, and (2) the geometric properties of the mock-ups, which enable structured comparison and pattern identification. Reference planes were generated from TLS point clouds using the RANSAC algorithm. Figure 4 illustrates the overall procedure of the proposed analysis method.

3.4.1. Intentions of the Analysis Method

In the proposed analysis method, reference planes were obtained from the TLS point clouds, and the test data were acquired from the MLS point clouds. ICP registration was applied to align the MLS point clouds with the reference planes. This algorithm iteratively rotates and translates the MLS scan data to minimize the RMSE, thereby aligning the two datasets. Under such conditions, the expected result of ICP registration between a reference plane and a point cloud would place the reference planes near the region with the highest point density in the point clouds. In other words, when performing ICP registration between a reference plane and the point clouds, the alignment process tends to ignore the underlying error distribution pattern in the data due to its focus on minimizing RMSE. To address this, the mock-ups used in this experiment were designed in the form of columns, which have a closed-shape in xy-plane, ensuring that the influence of point clouds on all four sides is evenly reflected in the RMSE calculation, thereby allowing the inherent error distribution patterns of the MLS to emerge. Based on the error distribution pattern inherent in the MLS point clouds, the expected outcome of ICP registration in this configuration is shown in Figure 5. The Figure 5a-crepresent cases where the point clouds are primarily distributed behind the surface, near the surface, and in front of the surface of the object, respectively.

If the aforementioned conditions and assumptions are satisfied, the RMSE calculations for opposing planes during ICP registration become complementary. When considering the MLS point clouds corresponding to two opposing planes, it is advantageous for the RMSE calculation to position the densest regions of each plane as close as possible to the reference plane. However, when the densest region of one plane is aligned with the reference plane, the RMSE value for the opposite plane increases significantly. Accordingly, the error distribution patterns of opposing planes should be similar. Therefore, the authors believe that the error distribution patterns of MLS point clouds are observed throughout the proposed analysis method. Additionally, the distances from each point of MLS point clouds to reference planes play a representative role in quantifying error values.

3.4.2. Reference Data Generation

The reference data are mesh-type plane datasets generated from TLS point clouds. The TLS-01 device can acquire point clouds with high accuracy and automatically align the z-axis with the gravity direction using its built-in IMU sensor. Therefore, the z-axis values of the TLS point clouds are directly used as height values without additional processing. First, point clouds corresponding to the Columns are extracted from the TLS point clouds. Then, the RANSAC method is applied to generate the initial planes of the Columns. RANSAC, a method for fitting optimal lines or planes while accounting for outliers, has been robustly utilized in Scan-to-BIM studies [24,25,26,27,28,38]. Based on the sides of the columns, equations for four planes can be obtained, with each plane defined by the general form in Equation (1):

$$ax+by+cz+d=0$$

(1)

The edges of the columns are determined by the intersection lines formed by orthogonal planes. By solving the equations of two intersecting planes and assigning fixed $z$-values, the eight vertices that represent each Column are obtained. In this study, the fixed $z$-values, in accordance with the design specifications of the Columns, are set to $z=0$ and $z=1.5$. Additionally, the coordinate systems of the two Columns are defined separately, with the origin of each set at the centroid of the four vertices derived from the bottom section at $z=0$. This coordinate adjustment ensures consistency in the directionality of the C2M distance measurement performed later and provides intuitive coordinates for the eight vertices. Finally, these eight vertices are used to redefine four enclosed planes, which then serve as the reference data for subsequent analysis.

3.4.3. Test Data Preprocessing

The test data consisted of MLS point clouds acquired from the Columns. According to the assumptions of the proposed analysis method, equal point density across all four planes is required. To achieve this condition, the MLS point clouds of the Columns were manually segmented into individual planes. This plane-level segmentation is crucial for verifying the consistency of the error distribution pattern within a single MLS point cloud. Moreover, segmentation prevents inaccurate C2M distance measurements that may result from incorrect normal vector calculations near the edges of the Columns.

Subsequently, random down-sampling was applied to each segmented point cloud. The number of down-sampled points was set to match the smallest number of points among the four planes. This parameter was determined to ensure sufficient data for analysis while maintaining an equal number of points across all planes. Therefore, the random down-sampling parameter was determined independently for each Column to maximize the number of points used in the analysis.

3.4.4. Measurement Process

In the measurement process, ICP registration, C2M distance measurement, and Gaussian distribution fitting were conducted. For ICP registration, the compared MLS point clouds were translated and aligned with reference planes. After ICP registration, the C2M distance measurement algorithm was employed to measure the distance from each point to the reference planes.

The C2M distance represents the shortest distance from a point to the nearest reference mesh (plane). In this study, the C2M distance is applied after ICP registration to quantify how MLS point clouds deviate from the reference planes, which represent the actual surfaces of the Columns. The C2M distance is obtained by calculating the perpendicular distance from a point to a plane, using the plane’s normal vector to establish the directionality of the distance. The C2M distance can be mathematically represented as follows:

$$d=n·(p-p_0)$$

(2)

where $d$ is the distance; n is the normal vector of the plane; $p$ is a point in the point cloud; and $p_0$ is a point on the plane. The sign of d indicates whether a point is inside (negative) or outside (positive) relative to the plane. C2M distance measurement is performed for each reference plane and its corresponding segment of the compared MLS point clouds. This process prevents inaccurate C2M distance measurements caused by incorrect normal vector $\left(n\right)$ calculations near the edges of the Columns. By analyzing the measured C2M distances, the error distribution pattern in MLS point clouds can be identified.

Additionally, the proposed analysis method fits a Gaussian distribution to the measured C2M distance. The Gaussian distribution is mathematically defined as follows:

$$f\left(x\right)={\displaystyle \frac{1}{\sigma \sqrt{2\pi }}}{e}^{-{\displaystyle \frac{{(x-\mu )}^2}{2{\sigma }^2}}}$$

(3)

where $x$ represents the C2M distance values, $\mu$ is the mean, and $\sigma$ is the standard deviation. The mean value ($\mu$) represents the systematic bias in MLS point clouds, while the standard deviation ($\sigma$) quantifies the variability of the errors. By analyzing the Gaussian distribution of the C2M distances for each plane of the Columns, the error distribution pattern in MLS point clouds can be comprehensively analyzed. The purpose of applying Gaussian distribution fitting is to mathematically represent the error distribution patterns in the MLS point clouds. In this study, the derived error distribution patterns are utilized in a simplified form for testing purposes. However, in future research, these patterns are expected to be used for dimensional correction in the measurement of building components, where a more precise numerical representation will be required. Accordingly, Gaussian fitting is conducted in this study to examine whether the derived distribution functions adequately represent the full set of C2M distances measured on each plane.

3.4.5. Analysis Metrics

The primary metrics analyzed in this study are the mean and standard deviation of the cloud-to-mesh (C2M) distances measured from the MLS point clouds to the reference planes derived from TLS point clouds. The representative error is defined as the mean C2M distance calculated between the reference planes and their corresponding MLS point clouds, and it is hereafter referred to as the mean error for clarity. This mean error corresponds to the peak of the Gaussian distribution fitted to the entire set of measured distances, indicating the most densely concentrated region of the MLS point cloud for a given plane.

The standard deviation, on the other hand, quantifies the spread of C2M distances from the mean within each plane. It serves as an indicator of data precision, functionally analogous to the “relative accuracy” reported in Table 1. Therefore, analyzing the standard deviation offers additional insight into both the precision of MLS point cloud measurements.

4. Results

4.1. Data Acquisition and Preprocessing Results

Figure 6 presents the acquired TLS point clouds (Figure 6a) and the generated reference planes (Figure 6b,c). Additionally, the compared MLS point clouds were captured using three different MLS devices, with each MLS device scanning the entire experimental room four times. However, during the data extraction process, the first dataset from MLS-01 was missing due to an unexpected issue. The cause is presumed to be either a malfunction of the internal SLAM algorithm or an error during data transmission. Nevertheless, the remaining three datasets from MLS-01 were successfully extracted without issue.

In our experiments, a total of 11 MLS point clouds were collected: three from MLS-01 and four each from MLS-02 and MLS-03. The acquired MLS point clouds were processed according to the procedures described in Section 3.3. As a result, data for 22 Columns were generated, producing a total of 88 reference planes.

Table 2. The number of points in each Column used for analysis.

MLS Point Clouds		Title 2		Title 3
		Column A		Column B
		Individual Plane	Total (Four Planes)	Individual Plane	Total (Four Planes)
MLS-01	01	13,000	52,000	9000	36,000
	02	25,000	10,0000	13,900	55,600
	03	20,993	83,972	13,994	55,976
MLS-02	01	20,000	80,000	8995	35,980
	02	16,000	64,000	13,000	52,000
	03	23,000	92,000	15,000	60,000
	04	18,000	72,000	11,000	44,000
MLS-03	01	28,000	112,000	18,000	72,000
	02	32,000	128,000	21,000	84,000
	03	28,000	112,000	13,000	52,000
	04	21,000	84,000	10,000	40,000

presents the number of points for the 22 Columns used in the analysis. In this study, random down-sampling was applied to the four planes of each Column. Therefore, the “Individual plane” values in Table 2 represent the down-sampling parameters, while the “Total” values are four times the “Individual plane” values.

The smallest plane segment contains approximately 9000 points. A dataset with more than 9000 data points is sufficient to represent one plane of the column used in this experiment. Additionally, the number of points is sufficient for applying a Gaussian distribution for detailed analysis. Meanwhile, as shown in Table 2, despite constraints of the MLS scanning speed and distance, the number of points varies across planes. This result can be attributed to the constraint in maintaining equal exposure times across the surfaces, even under constrained MLS scanning speed and distance. This variation may introduce sensitivity issues when applying the proposed analysis method, particularly during random downsampling. Therefore, we evaluated the sensitivity of the analysis with respect to random downsampling. Figure 7 illustrates examples of Columns acquired from the three MLS devices, viewed in the xy-plane. C2M distance measurements were performed for each segment against its corresponding reference plane.

4.2. Sensitivity Analysis of Random Downsampling

The proposed analysis method employs the random option of the subsampling tool in the open-source software CloudCompare. To evaluate the sensitivity of random downsampling, one plane was selected for each MLS device. The selected planes correspond those with the lowest sampling ratios among the datasets acquired by each MLS device. Specifically, Plane 02 of Column A in Dataset 01 was selected for MLS-01, Plane 02 of Column A in Dataset 02 for MLS-02, and Plane 01 of Column A in Dataset 01 for MLS-03. The respective sampling ratios were 0.10, 0.14, and 0.15.

For each selected plane, random downsampling was performed 30 times, and sensitivity was analyzed based on the coordinate values in the direction normal to the plane, which was the primary focus of the analysis. Two indicators were used for comparison: the relative variation of the mean $(∆\mu )$ and relative variation of the standard deviation $(∆\sigma )$, as defined in Equations (4) and (5). In addition, a one-way ANOVA test was conducted. The null hypothesis was that there is statistically significant difference in mean values between the raw and sampled groups. The ANOVA results with $p$-value above 0.05 support the conclusion that the mean geometry of the raw point cloud remains statistically unaffected by the random downsampling process.

$$∆\mu ={\displaystyle \frac{\left|{\mu }_{sampling}-{\mu }_{raw}\right|}{\left|{\mu }_{raw}\right|}}$$

(4)

$$∆\sigma ={\displaystyle \frac{\left|{\sigma }_{sampling}-{\sigma }_{raw}\right|}{\left|{\sigma }_{raw}\right|}}$$

(5)

As a result of the analysis, the maximum relative variations in mean and standard deviation were $∆\mu =0.007\%$ and $∆\sigma =1.131\mathrm{\%}$, respectively. For MLS-02, they were $∆\mu =0.008\mathrm{\%}$, and $∆\sigma =1.076\mathrm{\%}$, and for MLS-03, $∆\mu =0.003\mathrm{\%}$ and $∆\sigma =0.819\mathrm{\%}$. All maximum $∆\mu$ values remained below 0.01%, and all maximum $∆\sigma$ values below 1.2%. Additionally, the ANOVA results yield $p$-values of 0.727 for MLS-01, 0.281 for MLS-02, and 0.532 for MLS-03, indicating that none of the differences were statistically significant ($p>0.05$). The overall sensitivity analysis results are summarized in

Table 3. Sensitivity analysis of random downsampling.

MLS	Dataset	Column	Plane	$\mathbf{Max}∆\mathit{\mu }$	$\mathbf{Max}∆\mathit{\sigma }$	$\mathit{p}$-Value
01	01	A	02	0.007%	1.131%	0.727
02	02	A	02	0.008%	1.076%	0.281
03	01	A	01	0.003%	0.819%	0.532

.

Therefore, the random downsampling applied in this study preserves the distributional characteristics of the raw point clouds at the plane level and maintains the integrity of the internal error distribution patterns.

4.3. Error Distribution Patterns in MLS Devices

In this section, the experimental results derived from the complete dataset were presented. A total of 88 planes were analyzed, and the analysis was primarily based on the mean and standard deviation values of the C2M distances measured on each plane. As previously mentioned, the mean C2M distance for each plane represents the central value of the error distribution pattern, while the standard deviation reflects the relative accuracy.

Eighty-eight planes were analyzed using the proposed analysis method. Figure 8 presents representative results organized at the Column level. According to Figure 8, each MLS device exhibits distinct error distribution patterns and shows biases relative to the reference planes. MLS-01 appears to have more outer points than inner points, while MLS-02 exhibits this tendency more prominently. In contrast, MLS-03 shows a slightly greater number of inner points compared to other MLS devices.

To obtain a representative overall mean and Gaussian distribution for each MLS device, the C2M distance measurements of all points acquired from the analyzed planes were aggregated. Based on these aggregated data, the overall mean and corresponding histograms were generated for each MLS device as shown in Figure 9. Figure 10 shows a box plot generated from the mean error values of each MLS device.

Table 4. Overall mean error and standard deviation for each MLS device.

MLS Type	Mean Error (mm)	Standard Deviation (mm)
MLS-01	4.3	7.6
MLS-02	9.2	10.8
MLS-03	−2.0	9.3

presents overall mean and standard deviation for each MLS device. Based on these results, the error distribution patterns of each MLS device can be summarized as follows:

The measured overall mean of MLS-01 is 4.3 mm, while the median is 4.4 mm, as shown in Figure 10. This suggests that MLS-01 exhibits a biased error distribution pattern relative to the reference planes, with a representative offset of 4.3 mm. Additionally, Figure 9a shows that the aggregated error distribution of MLS-01 forms a bell-shaped curve, indicating that a Gaussian distribution effectively characterizes the pattern. The standard deviation is measured at 7.6 mm, implying that approximately 68% of the data fall within one standard deviation from the mean, i.e., within the range of 4.3 ± 7.6 mm, consistent with the characteristics of a normal distribution illustrated in Figure 9a.

The measured overall mean and median of MLS-02 are both 9.2 mm, as shown in Figure 10. This suggests that MLS-02 exhibits a more biased error distribution pattern relative to the reference planes, with a representative offset of 9.2 mm. Compared to MLS-01, MLS-02 shows a greater degree of bias. Figure 9b illustrates that the aggregated error distribution of MLS-02 forms a bell-shaped curve, indicating that a Gaussian distribution reasonably characterizes the pattern. The standard deviation is measured at 10.8 mm, implying that approximately 68% of the data fall within one standard deviation from the mean, i.e., within the range of 9.2 ± 10.8 mm, consistent with the characteristics of a normal distribution illustrated in Figure 9b. As a result, MLS-02 not only exhibits a more biased error pattern but also shows lower data precision compared to MLS-01.

The measured overall mean of MLS-03 is –2.1 mm, while the median is –2.0 mm, as shown in Figure 10. This suggests that MLS-03 exhibits a more biased error distribution relative to the reference planes, with a representative offset of –2.1 mm in the inward direction. Compared to MLS-01 and MLS-02, MLS-03 shows an opposite bias pattern. Figure 9c shows that the aggregated error distribution of MLS-03 forms a bell-shaped curve, indicating that a Gaussian distribution reasonably characterizes the pattern. The standard deviation is 9.3 mm, implying that approximately 68% of the data fall within one standard deviation from the mean, i.e., within the range of –2.1 ± 9.3 mm, consistent with the characteristics of a normal distribution illustrated in Figure 9c. As a result, MLS-03 exhibits an inward bias pattern opposite to those of MLS-01 and MLS-02, with data precision higher than MLS-02 but lower than MLS-01.

Figure 11, Figure 12 and Figure 13 present the error distribution curves of MLS-01, MLS-02, and MLS-03, respectively. According to Figure 12 and Figure 13, MLS-02 and MLS-03 exhibit consistent error patterns across all columns within each of the four datasets, particularly when considering the entire set of planes. In contrast, the error distribution curve of MLS-01 shows less consistency across planes within individual columns compared to those of MLS-02 and MLS-03. In other words, MLS-01 shows relatively unstable error patterns and point distribution consistency within a single dataset. However, from the perspective of the overall analysis, MLS-01 still demonstrates a clearly biased error distribution, and the measured overall mean error is considered sufficient to serve as a correction factor. This aspect is further examined in Section 4.4.

As a result, our findings reveal that each MLS device exhibits a unique and consistent biased error distribution pattern relative to the actual surface of the objects. These findings are clearly distinguished from those of previous studies. Sepagozar et al. [20], who shared a similar objective with our study, measured the dimensions of architectural components such as doors and windows using MLS point clouds. Their study reported that a handheld MLS device exhibited a mean error of 7.8 mm and a standard deviation of 4.3 mm. This mean error was presented as an absolute value, making it difficult to reflect the directional bias, which is the central focus of our study. This limitation is commonly found in other related works, as reviewed in Section 2.1. Meanwhile, Teo and Yang [30] presented positive and negative dimensional errors for BIM elements generated from an iPad LiDAR device. However, their results represent proportional deviations in overall object size, not the internal error patterns in point clouds that our study aims to investigate.

According to Figure 11, Figure 12 and Figure 13, as well as Appendix A, the mean errors measured from opposing planes were highly similar. This suggests that RMSE minimization during ICP registration balances error across opposing planes. In addition, the plane-level mean errors were consistently distributed within a narrow range, indicating stable error characteristics. If the error distribution had varied significantly in direction across planes within the same dataset, such consistency would not have been observed. These findings further differentiate our work from previous studies.

Moreover, the earlier works paid limited attention to the influence of point density differences during registration, and many used overly broad datasets [23,24,27], which may have increased registration errors between datasets. Accurate alignment with reference data is a prerequisite in the analysis of MLS point clouds, as inaccurate registration can significantly increase uncertainty in the evaluation process. Therefore, we believe that the findings presented in this study reflect more through consideration of ICP registration characteristics compared to previous studies, resulting in a more reliable analysis.

4.4. Verification of Dimensional Correction Using Error Distribution Patterns

The goal of this study is to identify the error distribution patterns of MLS point clouds and enhance the dimensional accuracy of as-built BIMs generated from MLS data. In this section, we verify whether the representative mean error derived for each MLS device improves the dimensional accuracy of 3D modeling for actual building components.

For verification, the same method used to generate reference planes in previous experiments was applied to the MLS point clouds. RANSAC was employed to extract planes, and fixed z-values were assigned to the intersection lines of these planes to obtain eight vertices (MLS-RANSAC). The eight derived vertices were compared with the reference vertices obtained from TLS point clouds using RANSAC, and the errors were measured. The average distance measured at these eight vertices was defined as the error value of the column.

The representative mean error of each MLS device was used to translate the RANSAC planes accordingly. From the corrected plane equations, intersection lines were obtained, and fixed z-values were assigned to derive eight corrected vertices. Finally, the verification was conducted by comparing the errors of the eight vertices derived from MLS-RANSAC with those obtained using MLS-RANSAC with correction.

The verification results demonstrate that applying the representative mean error derived in this study significantly improved the dimensional accuracy of 3D modeling.

Table 5. Error measurement results of MLS-RANSAC and MLS-RANSAC with correction methods.

MLS Type	Dataset	Column	MLS-RANSAC (mm)	MLS-RANSAC with Correction (mm)	Accuracy Improvement (mm)
MLS-01	01	A	10.7	5.0	5.7
		B	7.0	4.1	2.9
	02	A	5.8	5.2	0.6
		B	5.5	4.8	0.8
	03	A	9.0	4.5	4.5
		B	4.7	2.9	1.8
MLS-02	01	A	11.8	5.8	6.0
		B	12.8	2.7	10.1
	02	A	13.4	4.9	8.5
		B	12.3	2.9	9.4
	03	A	13.3	2.3	11.0
		B	12.1	1.9	10.2
	04	A	14.3	3.0	11.3
		B	13.2	2.3	10.9
MLS-03	01	A	5.4	3.2	2.2
		B	6.3	3.7	2.6
	02	A	4.7	2.6	2.1
		B	4.2	2.0	2.3
	03	A	2.0	2.4	−0.4
		B	2.6	2.0	0.6
	04	A	3.0	2.9	0.1
		B	4.9	3.9	1.0

presents the errors measured for all datasets, while Figure 14 shows the error reduction for each MLS device. Except for Column A in Dataset 03 of MLS-03, error reduction was observed in all Columns. According to Figure 14, the error reduction for MLS-01, MLS-02, and MLS-03 was 2.7 mm, 9.7 mm, and 1.3 mm, respectively. Notably, MLS-02 exhibited the most significant improvement. This result suggests that although MLS-02 had larger error values, its error distribution pattern was more consistent than that of MLS-01, leading to a more effective correction. These findings indicate that even if an MLS device exhibits substantial errors, a consistent and well-defined error distribution pattern enables accurate dimensional correction of building components.

Despite employing a general approach, this verification demonstrated a remarkable improvement in dimensional accuracy. Therefore, if more precise methods for error distribution measurement and correction are applied in the future, this approach will significantly contribute to inspection processes and the high-precision generation of as-built BIMs using MLS point clouds.

5. Discussion

5.1. Appropriateness of the Experimental Design and Analysis

The experiments were conducted in a controlled 220 m² laboratory environment, designed to effectively simulate real-world point cloud acquisition conditions. Two scanning start positions were selected based on the locations of the columns; however, the results indicate that the influence of starting position on the measurements was negligible.

The sensitivity analysis was conducted to evaluate the impact of random downsampling. The results confirmed that the error distribution patterns in the raw point clouds were preserved even after applying random downsampling. This finding supports the validity of our experiment in appropriately analyzing the inherent error patterns in the MSL point clouds.

The proposed method integrates robust algorithms, including random down-sampling, ICP registration, and RANSAC. Apart from the precise fabrication and installation of the mock-ups, the overall process is significantly simpler than conventional manufacturer calibration procedures. While manufacturer calibration remains the most accurate approach for error correction, it is time-consuming and often impractical to perform frequently in real-world settings.

In contrast, the proposed method enables efficient and repeatable assessment of MLS error distribution patterns with minimal setup, requiring only the installation of basic analysis targets. Therefore, this approach can meaningfully contribute to improving the accuracy of as-built BIM models generated from MLS point clouds, particularly in construction contexts where geometrical precision is critical.

5.2. Directional Bias in MLS Devices

The results indicate that each MLS device exhibits a unique error distribution pattern specific to the device. Previous studies have primarily focused on the absolute magnitude of errors in MLS point clouds, with limited attention to their directional distribution. In contrast, this study confirms that MLS point cloud errors may deviate either inward or outward relative to the actual surface. The analysis results are summarized as follows:

• MLS-01 point clouds exhibit an outward bias in error relative to the Column surface, with a representative mean error of +4.3 mm, standard deviation of 7.6 mm.
• MLS-02 point clouds exhibit an outward bias in error relative to the Column surface, with a representative mean error of +9.2 mm, standard deviation of 10.8 mm.
• MLS-03 point clouds exhibit an inward bias in error relative to the Column surface, with a representation mean error of −2.0 mm, standard deviation of 9.3 mm.

A notable finding is that MLS-02 and MLS-03, despite being the same model, exhibit opposite error distribution patterns. This suggests that such errors cannot be fully characterized based solely on manufacturer-provided specifications (e.g., Table 1). From this perspective, identifying and understanding the error distribution patterns of MLS point clouds is not only relevant to Scan-to-BIM applications, but also critical in fields that require high-precision 3D modeling, such as quality inspection, and digital twin development.

5.3. Discussions on the Impact of Downsampling Methods on Error Pattern

The proposed analysis method adopted random downsampling for each plane-level point cloud. As presented in Section 4.2, sensitivity analysis confirmed that random downsampling preserves the error distribution patterns inherent in the raw point clouds. Therefore, the error patterns considered to reliably represent those of the original point clouds.

Meanwhile, applications that utilize point clouds in Scan-to-BIM contexts commonly adopt various downsampling methods, such as voxel-based and space-based downsampling, rather than relying solely on random downsampling. Accordingly, it is necessary to discuss the potential risk that the error patterns derived through random downsampling may be altered or disregarded when other downsampling methods are applied.

Voxel-based downsampling methos are robust approaches that generate fixed size of voxels and select one point for each voxel. The error pattern investigated in this study is not considered during voxel generation. However, the main concern is how to select one point from each voxel. For example, methods that regenerate geometrical center point for each voxel have high potential to disregard the error pattern. By contrast, strategies that pick either a random point or the densest point within each voxel are more likely to preserve the inherent error pattern in point clouds. The same argument applies to octree-based downsampling methods.

Therefore, the findings of this study are expected to be highly applicable to a wide range of density-aware downsampling methods, which account for local point distributions. While these results may be less suitable for traditional schemes that uniformly regenerate geometrical central points within voxels, they nonetheless offer valuable insights for advanced sampling strategies. Further systematic evaluation is needed to verify how robust error patterns are preserved across various downsampling strategies.

5.4. Contributions and Implications for BIM Applications

This study investigates the error distribution patterns of MLS point clouds, focusing on the direction and consistency of error occurrence, which have not been adequately addressed in previous related studies. The proposed analysis method adopts an approach that aims to preserve the inherent error patterns in MLS point clouds, a consideration often overlooked in earlier research. As a result, the findings include not only the magnitude of errors but also their directional characteristics. Compared to previous works, this study provides a more in-depth analysis of MLS point clouds, with specific consideration of their relevance to Scan-to-BIM applications.

The identified error pattern has the potential to enhance the geometric accuracy of as-built BIM generated from MLS point clouds. Previous studies on Scan-to-BIM and reconstruction have commonly adopted robust fitting methods such as RANSAC [32,33,34,35,36,37,38,39]. However, these approaches are proposed to extract representative lines or planes from the given point cloud. Therefore, if the input point cloud contains a biased error pattern, the resulting geometric primitives are likely to be generated at biased positions. In this regard, the error pattern identified in this study can play a critical role in improving geometry reconstruction accuracy in MLS-based Scan-to-BIM. This potential improvement is supported by the results presented in Section 4.4.

Our study contributes by identifying the internal error patterns inherent in MLS point clouds, a topic that previous studies have overlooked. The experimental results show that these error patterns are device-specific and may negatively affect the dimensional accuracy of components when MLS point clouds are used in BIM applications. These findings provide practical insights into the integration of MLS in BIM applications.

6. Conclusions

This study proposed and validated an analysis method to identify the error distribution patterns in point clouds acquired from mobile laser scanning (MLS) devices. Three MLS devices were evaluated using terrestrial laser scanning (TLS) data as a reference. Wooden columns were fabricated as analysis targets, taking advantage of their closed shape in the XY-plane. The analysis method integrated the characteristics of iterative closet points (ICP) registration and cloud-to-mesh (C2M) distance measurement. Random consensus sampling (RANSAC) was used for generating reference planes and verification of dimensional correction using error distribution patterns. The results demonstrate that the proposed method effectively revealed the unique error distribution patterns of each MLS device. The standard deviation of the measured C2M distances allowed a comparative assessment of the data precision and consistency among devices. Furthermore, the application of representative mean error values showed potential for improving dimensional accuracy in as-built modeling.

Through this study, the two assumptions presented in the introduction were validated, and the following conclusions were drawn:

• Point clouds from each MLS device exhibited distinct error distribution patterns, with errors biased either inward or outward relative to the actual surface. Notably, two devices of the same model demonstrated opposite error directions (Assumption 1 is satisfied).
• Representative values derived from the identified error patterns were effective in enhancing the dimensional accuracy of reconstructed BIM components (Assumption 2 is satisfied).

This study contributes to addressing the limitations associated with MLS point cloud inaccuracies in as-built BIM generation. By developing an optimized correction method based on the identified error patterns, the accuracy of as-built BIMs generated from MLS data can be significantly improved. This approach is not limited to specific devices or scenarios and holds potential for generalization across a wide range of MLS-based applications in construction management. Furthermore, these findings highlight the importance of accounting for MLS-specific error patterns in point cloud processing, particularly in tasks that require high geometric accuracy in as-built BIM generation.

7. Limitations and Future Work

In this study, Column A and Column B were constructed from wood, chosen for ease of fabrication and dimensional control. During installation, efforts were made to align the columns vertically with the gravitational axis; however, minor deviations and tilting were observed in reference vertices.

Column-shaped mock-ups were used as the analysis targets. As described in Section 3.2, these column forms were appropriate for our experiment, which is based on planar surface analysis. However, additional experiments have to be considered with curved-shaped targets, as such forms are also frequently found in construction sites. Furthermore, future work should investigate whether the error pattern varies depending on the material properties of the objects.

Furthermore, the study was limited to three MLS devices. To generalize the findings, further experiments involving a wider range of MLS devices are required. If similar error distribution patterns are observed across different devices, the robustness of the proposed method will be further validated. Lastly, future research should focus on applying the derived error distribution-based corrections to entire MLS point clouds. In addition, there is a need to develop more advanced methods that can generate accurate as-built BIMs even when using MLS, by leveraging the identified point cloud error distribution patterns.