1.1. Background and Meaning

As robot technology and computer technology are advancing speedily, recent years have seen a certain degree of development in construction robot industry [1, 2]. At construction sites, construction robot replacing manual labour to finish complex work has already become a trending [3–5] such as construction fabrication robot [6], climbing robot [7], floor-tiling robot [8], collaborative robots [9], and measurement robot [10]. For example, floor-tiling robot [8] completes the automatic floor tiling with high-accuracy robot through control strategy based on visual measurement feedback and finite-state machine. Collaborative robot [9] controlled continuously by a BCI-based system robot to capture workers’ brainwaves acquired from a wearable electroencephalogram (EEG) device and interpret them into robotic commands with 90% accuracy. As the feasibility of setting construction robot into practical use in construction process has been enhancing by artificial intelligence and robot technology, the adaptation of computer algorithm and robotics in civil engineering is drawing wider and wider attention among relative science researchers [10–12].

At present, indoor measurement robot is typical of construction robots, and mainstreaming indoor measurement robot is an automatic machine featuring TLS [13, 14]. Differing from traditional manual measurement possesses low-efficiency, high human cost, and high health risk since it needs workers to be exposed to construction environment and to finish work with many tools like guiding rule and feeler gauge [15, 16]. The machine with TLS, which quickly collects 3D information of a room with LIDAR scanning, may be more efficient and be able to attain more comprehensive information of a 3D room.

However, existing equipment with TLS might lead to a waste of work, time, and finance because it is high cost and inevitably produces a tremendous size of point cloud data that usually consumes a large amount of time to process, while the house types and indoor structures of most constructions are actually not complicated and diversified, which means that, in most cases, indoor measurement may not need the possibly excessive collection and processing of information of an indoor room from robot with traditional TLS. Therefore, this paper proposes a self-adaptation TLS system that targets indoor measurement of quantity by combining relevant computer technology and robot technology.

1.2. Related Work

1.3. Innovation of This Article

3.1. Indoor Quantity Measurement Process Analysis

3.2. Adaptive Measurement Scanning Strategy

3.2.1. Scanning Motion Mechanism of Novel TLS in This Paper

The whole laser scanning process is generated by the motion of a two-degree-of-freedom motor platform. In this paper, a motion mode of “equal angle” is proposed: in a fixed range of motion, we rotate according to the equal angle interval and stay at this position for distance measurement with laser module. The novel TLS will automatically scan the entire room by repeating the motion mode of “equal-angle” several times. The scanning motion of the novel TLS is based on a two-degree-of-freedom motor platform, with a “horizontal-motion motor” and a “vertical-motion motor.” In order to complete the indoor laser scanning more efficiently, the laser module has multilasers in the equal angle distribution (see Figure 2) (more details about the novel TLS equipment can be watched in a video by clicking on the link https://youtu.be/h7SavWCaPvc).

The entire scanning process comprises many such motor-motion cycles (see Figure 3). The “vertical pitching-motion motor” finishes one “equal-angle” (not greater than 360 degree) rotation, with the “horizontal spinning-motion motor” keeping static and repeating rotations in the same pattern many times, and the laser scanning will have a whole-room coverage.

In order to achieve self-adaptation planning for scanning, namely, self-P4S (planning for scanning), with a certain scanning density preset by operators in specific measuring projects for different-size rooms, the novel TLS fulfils to finish scanning in shortest time.

3.2.2. Maximum Distance Interval Scanning in Planning Model

The distribution of point cloud under multiprocess scanning is irregularity. In order to plan the scanning density of indoor-scanning point cloud, we propose the maximum distance interval of point cloud in the planning model “s₀,” which is the distance between scanning paths of the largest plane in the hexahedral room (see Figure 4). Referring to the ACI standard [31], we define the maximum standard s_max = 330 mm, and the relationship between s₀ and s_max is as follows:

$$ (1)

3.2.3. Minimum Time Cost Programming Model

In actual measurement, not only the density of point cloud but also the efficiency of measurement should be considered. Therefore, this paper proposes the “minimum time cost programming model.”

The model takes the minimum time consumption as the objective and the minimum distance interval of scanning as the constraint condition. The specific formula is as follows:

$$ (2)

In the objective function, n_A represents the total number of times the laser module has been executed by the device during the movement. T_A represents the overall elapsed time for the device to perform an indoor 3D scanning. T_s represents the time spent by the laser module in a single scanning. T_y represents the total elapsed time of the device during the movement. The constraints of the objective function are expressed as follows:

$$ (3)

θ_1A represents the total range of motor rotation angles in a single pass of the horizontal motor. θ_2A represents the total range of motor rotation angles in a single pass of the pitch motor. Δθ₁ and Δθ₂ represent the angular interval values of the “equal angle” motion of the horizontal motor and the elevation motor, respectively, in a single process. θ_c (shown in Figure 5) is calculated as follows:

$$ (4)

In the above formula, L_a, L_b, and L_c, respectively, represent interior net story height, length of longer wall, and length of shorter wall.

3.3. Converting from Scanning Data to Point Cloud Data

Some parameters may be obtained from the novel absolute angles of the “horizontal spinning-motion motor” and the “vertical pitching-motion motor.” The poses of laser scanning points (the TLS machine as the origin in the coordinate system) can be attained on these parameters. In order to convert poses of laser scanning points into point clouds, this paper makes use of “modified DH modelling” in robotics and realizes it in this way.

The following is an illustration of the converting process of the data from one single laser scanning point. Based on the novel TLS equipment and “modified DH modelling,” this paper builds four coordinate systems marked as {0}, {1}, {2}, and {3}. The origin of {0} system is the intersection point of medial axis of the horizontal motor and the medial axis of the vertical motor; the origin of {1}, with the horizontal spinning-motion motor as its origin, represents a coordinate system overlapped with {0}; {2} is the coordinate system with the vertical pitching-motion motor as its origin; and {3} is established with the single laser scanning point as its origin. In modified D-H method, the coordinate frame i is attached to the near (proximal) end of link i as shown in Figure 6. The link and joint parameters are defined as D-H parameters and are summarized in Table 1.

In practice, because measuring with single-laser laser module is not efficient enough, this paper presents a laser module with four lasers. The four lasers array around the centre of the laser module, which is coaxial with the vertical pitching-motion motor, with an angular distance of 90 degree between every two of them. Such laser module may be referred as multi-laser-ranging module.

As for all lasers in the laser module, the link and joint parameters are defined as D-H parameters and are summarize in Table 1. In the table, θ_i represents the angular separation of laser i from the initial laser. L_i represents the measured value of laser i.

4.1. Overview of Point Cloud Processing Procedures

According to text above, 3D point clouds and poses of 3D point clouds are presented by converting laser scanning data through modified DH modelling. Next, the novel TLS will enter phase of point cloud processing. As shown in the flow chart (see Figure 7), the data obtained from previous scanning activity will go through point cloud processing, including point cloud filtering, point cloud segmentation, and point cloud fitting. Experiencing the three procedures of processing, the processed point cloud data can be directly used to calculate quantities of the measured construction since every interior surface of the construction can be represented by a mathematical equation.

4.2. Indoor Conditional Point Cloud Filtering

Because refraction occurs when laser beam hits on glasses and it will cause the laser ranging value to be greater than the actual value, point clouds extracted around door and window may usually present abnormality in practical measuring. Abnormal point clouds need to be eliminated during point cloud processing.

This paper utilizes conditional point cloud filter and distributing characteristics of indoor point clouds to identify abnormal point clouds. Indoor point clouds roughly distribute in form similar to hexahedron, and the centre of TLS equipment is basically located at the centre of measuring room. This paper defines the distance between the centre of the TLS equipment and the farthest point relative to the centre of the TLS equipment as “maximum polar radius”; in this situation, distance from every point cloud to the centre of the equipment is not supposed to surpass the “maximum polar radius”; and any point cloud not in conformity with the criterion above should be identified as an abnormal point cloud.

In the formula above, β₁, β₂, and β₃ are correction factors within the range (1,1.5). The correction factor is of significance since only under ideal condition, the centre of the novel TLS equipment may be located just at the centre of a room.

4.3. Segmentation of Indoor Point Clouds Based on Structural Characteristic

Segmentation of point clouds is a core procedure of data processing, and this paper proposes two algorithms for segmenting point clouds based on structural characteristics. Namely, this paper makes full use of point clouds and poses of them and spatial geometric features of room to create two novel algorithms.

4.3.1. On-Boundary Segmentation Algorithm for Indoor Point Cloud

In this paper, point clouds extracted from the novel self-adaptive TLS equipment present following features: 1. in each “process one,” the angle distance between the original point cloud and the final point cloud from the same single laser will present 90 degree. 2. Similarly, in one “process two,” the angle separation between the original point cloud and the final point cloud from a single laser is sure to be 180 degree. 3. Point clouds distribute in a form similar to a hexahedron. 4. Original state of the novel TLS equipment is with four lasers in the laser module separately vertical to four different interior surfaces, and the centre of the equipment is put close to the centre of the room.

This paper names this novel segmentation algorithm as “on-boundary segmentation for indoor point cloud,” and the whole process is consisted of two rounds of segmentation. The first round is segmenting out the ceiling and floor, and the second round is segmenting out four interior walls of measured room (see Figure 8).

(1) First Round of Segmentation. Here, only one laser L_i is taken to illustrate distributing characteristics and dataset of point clouds from one single laser after the whole scanning process. When the entire scanning phase of measurement is over, it is manifested that all point clouds extracted from laser L_i are placed on Wall_Datai, Wall_{Data(i + 1)}, Wall_{Data(i + 2)}, and a plane (ceiling or floor). All point clouds from laser L_iform a point cloud dataset $$, in which the “L_i” represents point clouds from the laser, and “i” and “j” represent, respectively, the first-conducted “one-step process one” in first “process one” and first-conducted “one-step process two” in first “process two.” In addition, it is known that single “process one” contains n₁ times of “one-step process one” and single “process two” contains n₂ times of “one-step process two.”

With the point cloud dataset, distributing characteristics explained above and the concept “offset point” mentioned before, the following will demonstrate the first round of segmentation in detail. Firstly, all the point cloud data from first “one-step process two” gather into a dataset $$, anddata in it may be present as $$. The coordinates of every $$ may be marked as $$. In that, the borderline segmentation algorithm requires the distance from point clouds to the “offset point” to identify the boundary of surfaces, and the coordinates of the “offset point” need to be used in calculation (see Figure 9). As for the coordinates of “the offset point” $$, this paper has to acquire them by classified discussion and the formula is as follows:

$$ (8)

The equation represents the distance between every point cloud and the offset point, which is as follows:

$$ (9)

Next, this paper identify the maximum value of the dataset $$:

$$ (10)

This paper puts serial numbers on every point clouds in sequence according to clockwise direction, and the $$ is numbered as L_ii1_max. The L_ii1_max can be the dividing point of this set of point clouds so that all the point clouds can be separated into ceiling surface or wall surface, and they may be present as follows:

$$ (11)

The end of the entire set of procedures in the first round of segmentation signifies that the segmentation of point clouds from laser L_i in a “one-step process one” is over. When the set of procedures is repeatedly executed n₂ times with n₂ sets of point clouds to be put into the algorithm, all point clouds of laser L_i that are obtained from the entire scanning process and that are also distributed on ceiling or floor surfaces will be segmented out from other point cloud clusters.

(2) The Second Round of Segmentation. Points clouds from laser L_i on ceiling and floor are segmented out through the first round of segmentation, but those on three walls are left for segmenting. However, there are also edges between every two walls, and the borderline segmentation method is also applicable to segmenting point clouds from laser L_i on four walls, which are ready to be put into three dataset marked as Wall_Datai, Wall_{Data(i + 1)}, and Wall_{Data(i + 2)}. During the first round of segmentation of point clouds from laser L_i, in one “process one,” all point clouds not in dataset of ceiling or floor are retained in the form of an array and all point clouds are present in dataset $$ (m is the maximum number in one case process one in first round of segmentation).

$$ (12)

This paper selects point clouds numbered as the first one in every array to form a point cloud dataset $$for tagging, which will be used to identify the “tagging point on edges” (the point cloud nearest to the cross-point of the horizontal cross-lines and edges of two adjacent walls) (see Figure 10).

$$ (13)

Since there are two “tagging points on edges” in the point cloud set for tagging, this paper divides the whole dataset into two parts, $$ and $$.

$$ (14)

Actually, the two tagging points on edges are separately in the two datasets $$ and $$, and they are two point clouds that, respectively, have longest distance to the centre of the equipment in their dataset. The formulae of them are as follows:

$$ (15)

Identifying the two “tagging points on edges” is as follows:

$$ (16)

From this, the number of the two tagging points is obtained and they serve as dividing points to segment point clouds separately into three walls Wall_Datai, Wall_{Data(i + 1)}, and Wall_{Data(i + 2)}.

$$ (17)

After implementing the first and second rounds, the segmentation processing for all points clouds extracted from entire scanning of laser L_i comes to its end and all point cloud data are segmented into four datasets Plane_Data, Wall_Datai, Wall_{Data(i + 1)}, and Wall_{Data(i + 2)}.

When point clouds collected from all the four laser modules have finished the first and second rounds of segmentation algorithm processing, the results of segmentation of point clouds on all interior surfaces will be attained.

4.3.2. On-Angular-Distance Segmentation Algorithm for Indoor Point Cloud

Apart from the novel segmentation algorithm introduced above, this paper proposes a novel segmentation algorithm for indoor point clouds based on angular distance emerging during scanning process (Figure 11). The laser beam from laser L_i will hit on some points in the intersection line of two wall surfaces when the laser module on vertical pitching-motion motor moves in circle along with the horizontal spinning-motion motor. At the moment, an angular distance presents between two lines, one from current position of laser module to the centre of the equipment and another from original position of laser module to the centre. This angular distance is marked as $$, which is called “wall-to-wall angular distance” in this paper, and $$ is another “wall-to-wall angular distance.” (Figures 12(a) and 12(d)). $$ represents an angular separation occurs when the laser beam from L_i hits on the intersection line of wall and ceiling or floor, and one line starts from the intersection line to the spatial position of the laser module and another line from it to the centre of the TLS machine. $$ is named to be “wall-to-plane angle” (Figure 13), and the “j” means the times of conducting “process two.” However, accompanying implementation of all times of “process two,” there are four different formulae for calculating “wall to plane angle” in correspondence to four range of angular distance.

(1) Procedures of Segmentation Algorithm on Angle Characteristics. This segmentation algorithm on angle characteristics needs information, apart from point clouds, including interior net story height and the length of long and short wall of room. With this information, the procedures of segmentation algorithm on angle characteristics for point clouds collected from a single laser are demonstrated in Figure 12. As shown in Figure 12, after obtaining “wall-to-wall” or “wall-to-plane” angles, the algorithm processes point clouds in different formulae in accordance with which range the jΔθ₂ is divided into.

The procedures of algorithm processing are shown in Figure 11. Firstly, the point clouds and parameters of room size are loaded and the system will calculate the “wall-to-wall angle” and “plane-to-wall angle” on the given parameters. Then, it enters the branch structure to judge which formula to be invoked based on angular ranges of jΔθ₂.When the branch structure is terminated, point clouds extracted from one single laser Li during the entire scanning process will be segmented into four datasets, respectively, relative to four interior surfaces: Plane_Data, Wall_Datai, Wall_{Data(i + 1)}, and Wall_{Data(i + 2)}.

(2) The Calculation of “Wall-to-Wall Angle.” Here, only a single laser L_i is taken as a sample to calculate “wall-to-wall angle.” As shown in Figures 12(a) and 12(d), this paper may calculate the angle $$ and $$ with given interior net story height and length of walls in the following formula:

$$ (18)

(3) The Calculation of “Wall-to-Plane Angle.” With loaded point cloud data and parameters of room size, the variable range of jΔθ₂ determines different formulae to calculate the “wall-to-plane angle,” which is relative to the position of the “horizontal spinning-motion motor” as the angular distance of it and its original position is equal to jΔθ₂. All four different formulae for calculating “wall-to-plane angle” (see Figures 12 and 13), also “$$,” are as follows:

$$ (19)

(4) Procedures of Data Segmentation. As known above, accompanying jΔθ₂ within different range, the calculation formula to get one certain degree value $$ diverges and so does the segmentation results of relevant point clouds. Here, it comes a $$ to help in segmenting point clouds into different datasets, and its equation is $$.

All the point clouds are loaded in a dataset $$, and it is comprised of many datasets. $$ is the only object to be processed in loop structure

$$ (20)

If for jΔθ₂, $$:

$$ (21)

If for jΔθ₂, $$:

$$ (22)

If forjΔθ₂, $$:

$$ (23)

If forjΔθ₂, $$:

$$ (24)

When the data segmentation above is finished, point clouds collected from a single laser after entire scanning process will be attached to three walls and one horizontal plane, ceiling, or floor.

4.3.3. Analysis of Point Cloud Segmentation in This Paper

As for the point cloud measured by the four lasers in the point cloud data, each laser contains four planes, including the point cloud data of three indoor walls and one plane (ceiling or floor). And the main idea of the above two algorithms is to segment the four planes of each one laser-scanning point cloud and then combine all the segmentation results from each laser to finally realize the segmentation of the point cloud of the six planes in the whole room. Both algorithms use the different scanning path’s characteristics of TLS in the hexahedron-type room. On-boundary segmentation uses edge features of indoor space, so two rounds of segmentation are required for point cloud data segmentation of single laser measurement data. One turn is for the edge feature of the intersection line between wall and plane, and the second turn is for the edge feature of the intersection line between wall and wall. On-angular-distance segmentation algorithm uses the angle features in space to segment point clouds. On the premise of knowing the size of indoor rooms, the angle features are used to realize the presegmentation of all point clouds on the walls. Thus, point cloud segmentation of four planes can be realized only through a cycle of data traversal. Therefore, the time complexity of on-angular-distance segmentation algorithm is lower.

To better understand the data process of the two algorithms, the principle of two algorithms can be understood through figure (see Figure 14). All point cloud can be regarded as a array in n₂ rows and n₁ columns. Due to the execution of “one-step process 2” for n₂ times, the arrays are combined with the n₂son-arrays. The son-array which size of 1 × n₁ represents the point cloud scanned in “Procedure 1.” In the point cloud segmentation scenario in this paper, the key of point cloud segmentation is to find the “cut-off point” for point cloud segmentation (the red areas of the picture represent the “cut-off points”). Although the on-angular-distance segmentation algorithm is better than the on-boundary segmentation in time complexity, the on-angular-distance segmentation algorithm needs to know get indoor room size, and the centre of TLS cannot be completely placed in the middle of the room during scanning, so the segmentation accuracy will be affected. On-boundary segmentation algorithm can segment point cloud without indoor house size. For regular hexahedral rooms, ideal point cloud segmentation can be achieved, but for indoor rooms with no obvious “edge features,” the segmentation accuracy will be affected. Therefore, the advantages and disadvantages of the two algorithms should be analysed through segmentation experiments in actual point cloud in reality.

4.4. Plane Fitting Algorithm for Indoor Point Clouds

5.1. Experiment Preparation

5.1.1. Preparation of TLS in This Paper

In order to verify the feasibility of the novel TLS and data processing system, we conducted scanning and processing in several rooms ready-to-move-in in a room shape that basically conforms with a hexahedron (see Figure 15(a)).

We implemented 3D scanning to collect data of the room on the novel self-adaptation TLS equipment; as for software system for laser scanning, we designed and developed two software systems: one for data collecting and the another for point cloud processing. All data collection was carried out on the equipment (see Figure 15(c)), and then, the data were transferred to a cloud server through which the scanning data were downloaded into a computer terminal to implement processing. In this paper, we temporarily adopted TCP/IP on a local area network to transfer the data for the sake of short-term experimental stage.

When the TLS machine was appropriately located, we clicked on the “To Scan” to start 3D scanning of the entire room. After it finished, scanning points were transferred to computer where we proposed two novel point cloud segmentation algorithms: one based on edges of interior surfaces and one based on angle range. On data-processing interface, the process of the data point processing was visualized and the interior quantities of room were obtained by data processing (see Figure 15(b)).

5.1.2. Experiment Process of Data Analysis

Afterwards, the collected data will be processed on computer and the entire processing procedures are displayed in Figure 16. We collected indoor 3D point cloud data with the multilaser module on the self-adaptation novel TLS equipment, but there was a possibility for laser beam to hit outside the room or on some light-reflecting objects causing high measuring error of data and high error of the point clouds corresponding to the measured data, which were called noisy points. Hence, we needed the filter mentioned in the paper to filter out the noisy points and retain point clouds applicable to indoor point cloud processing. Then, retained indoor point clouds were segmented by the point cloud segmentation algorithms for point clouds. The segmented clusters of point clouds were fitted into different surfaces. The equations of fitted surfaces will be used in the analysis of the measuring accuracy of this measuring system with a traditional manual laser-ranging method.

5.2. Experiment Result and Discussion

(1) The Analysis of Effectiveness of 3D Scanning. The results of actual indoor 3D scanning data collected by the novel self-adaptation TLS equipment might basically reflect indoor constructional features, and the general configuration of windows and door in the real room was presented by visualizing 3D point clouds data (see Figure 17).

(2) The Result of Point Cloud Segmentation. As shown in Figure 18, for point clouds from hexahedral room with clear boundaries between surfaces, “on-boundary segmentation algorithm” will perform relatively effective segmentation but for that from rooms without clear boundaries of surfaces, and the segmentation results of “on-boundary segmentation algorithm” are not close to 6 indoor surfaces. The unclear boundaries of surfaces are shown in these two forms: (1) carvings made between ceiling and walls, and (2) relatively big internal and external corner existing on the edge zones of surfaces. However, “on-angular-distance segmentation algorithm” will be able to perform normal segmentation under these two kinds of conditions and it is noticeable that known room indoor sizes are the premise of performing high-accuracy segmentation.

We have implemented data scanning and collecting of the two kinds of rooms in different conditions and evaluation of segmentation accuracy of two types of segmentation algorithms mentioned in this paper (shown in Table 2). On house type of room 1, the performance of two kinds of segmentation algorithms is ideal, with the F1 index of point cloud clusters from “on-boundary segmentation algorithm” up to highest value of 0.9663 and the down to 0.7692 and the F1 index from “on-angular-distance segmentation algorithm” ranging from 0.6866 to 0.9789. As for house type of room 2, the segmentation accuracy of “on-angular-distance segmentation algorithm” is higher than that in another. Highest F1 index of point cloud clusters from “on-boundary-segmentation algorithm” is 0.7713 and the lowest one is 0.2969, while the F1 index from the another segmentation algorithm has highest value of 0.9781 and lowest one of 0.7824. From the perspective of data, we can draw a conclusion that these two kinds of segmentation algorithms are both able to attain relatively high-accuracy segmentation for indoor point clouds and the accuracy of segmentation is lower when the segmented objective is a lesser order of magnitudes of point cloud clusters.

Apart from that, viewing for the segmentation accuracy of room 2, whereas the distinction of surfaces in the house type as room 2 is less significant relative to that of room 1, the particular point clouds used for “on-boundary segmentation” are not easy to identify, and hence, the segmentation accuracy is not high. On the other hand, the accuracy of two kinds of segmentation algorithms for two rooms is basically close, attributing to the oneness of determinant for the segmentation accuracy of “on-angular-distance segmentation algorithm”—the accuracy degree of distance from equipment to every surfaces in a room.

Hence, the key to enhance the accuracy of segmentation is to ensure the room size and type as-near to a standard hexahedron as possible and also the normative location of the TLS equipment at the centre of a room.

(3) The Accuracy of Point Cloud Fitting. After finishing the point cloud segmentation, this paper selects the segmented point cloud clusters of two planes with symmetrical distribution for plane fitting and plane distance calculation. And we calculated the actual distance result through the manual measurement, which was carried out by the method proposed above. Referring to relevant literature, in the fitting process, this paper selects different plane fitting algorithms, including TLS, LS, WTLS-D, RANSAC-LS, and the improved RANSAC-TLS proposed in this paper.

In the fitting experiment, we divided the segmented point cloud clusters into three categories: class 1 point cloud clusters, class 2 point cloud clusters and class 3 point cloud clusters. Among them, class 1 point cloud clusters were mainly the point cloud clusters distributed on the two opposite walls with the largest area of the blank part on the indoor surface; class 2 point cloud clusters were the point cloud clusters on the floor and ceiling; and class 3 point cloud clusters referred to the point cloud clusters on the wall with windows and opposite to the wall. We selected five house types of rooms for data scanning and point cloud segmentation, and three kinds of point cloud clusters were extracted for fitting accuracy experiment. The experimental results are shown in the Figure 19.

It can be seen from the figure that among class 1, class 2, and class 3 point cloud clusters, RANSAC-TLS algorithm is the closest to the results of manual measurement. The second is RANSAC-LS and WTLS-D algorithm. LS and TLS algorithm performed worst in fitting, and the changing trend of the two algorithms is similar in the calculation results of different rooms. Because there were some noise points of information in the whole segmented point cloud cluster, these noise points are composed of random error in the scanning process and systematic error of the equipment. Moreover, in the process of point cloud segmentation, because the segmented point cloud clusters are sparse point cloud clusters, and not all of the point cloud clusters are ideal point clouds in the flat wall area; on the contrary, there are often some outliers in the point cloud clusters, which become noise points in the sparse point cloud clusters and affect the fitting accuracy evaluation results.

According to the measurement results, LS and TLS do not filter all point cloud data in the fitting process, so these two algorithms are easy to introduce some noise data into the fitting process. Although WTLS-D algorithm also performed iterative fitting based on the distance to the TLS plane as the weight, its initial fitting plane was still the fitting result of TLS. Therefore, in the fitting process, although its fitting accuracy was improved, its robustness was relatively poor. Both RANSAC-LS and RANSAC-TLS belong to combined algorithms, which get the advantages of RANSAC algorithm. The ideal point cloud region of plane fitting was found by calling out a reasonable threshold. From the results, compared with RANSAC-LS, RANSAC-TLS better optimized the fitting weights of all point clouds in the filtered point cloud data, and finally performed better in the fitting results. For class 1 point cloud, it could be seen that there were relatively few noise points in its point cloud cluster, so all fitting algorithms were relatively stable in the measurement results. However, complex outliers would appear in the data scanning of class 2 and class 3 point cloud clusters. For example, the point cloud of objects such as lampshade often appeared in the ceiling data in class 2 point cloud clusters. In the three kinds of point cloud clusters, due to the measurement target of window, there would be more complex point cloud clusters. For the point cloud fitting algorithm without prescreening in the fitting process, the calculation results based on the fitting plane equation would have a large deviation from the actual measurement results.

Thus, in the fitting process, RANSAC can screen out noise and outliers and improve the fitting accuracy. From the overall fitting results, RANSAC-TLS algorithm performs best in fitting accuracy.

(4) Analysis of Scanning Parameters Threshold in Actual Measurement. To test measurement accuracy and efficiency of the proposed planning-scanning method in real indoor condition, we conducted one measurement experiment with the novel TLS equipment in a room, in shape of a hexahedron, ready to move in, in size of 4250 mm × 3510 mm × 3065 mm. According to the planning-scanning method proposed in Chapter 3, the relevant planning parameters in this house shape include D₀ = 142 mm, L_b = 4250 mm, L_c = 3510 mm, s_max = 330 mm, T_d = 0.5 s, and T_s = 1 s.

For the planning model of indoor scanning, s_max is a reference value, and value s₀ determines the distributing density of the entire laser scanning point cloud data. To make TA a relatively small value, which is an expectation, s_max is usually taken from s₀, and k is taken as large as possible. In this experiment, for value taking, we set different s₀ values for actual data collecting. Besides, we obtained segmented point cloud clusters and the point-cloud-cluster-plane equation based on the proposed point-cloud-processing algorithm.

Referring to the calculation method presented in the relevant literature [33], the flatness and verticality of four indoor walls of the measured room are calculated with different scanning parameter. In addition, we obtained the same two indexes by manual measurement with guiding rule. Based on the above operation, the mean error of flatness (mF.E) and mean error of verticality (mV.E) can be achieved by first attaining errors between manually measured values and variable values coming from the novel TLS equipment with varying scanning parameters.

Next, the calculation is executed in method proposed in this paper alike with changing scanning parameter, and the results of the actual manual measurement, respectively, find the error of the measurement value and manual measurement value of the equipment based on different scanning parameters, and calculate the mF.E of the wall and the mV.E of the wall.

Five groups of real quantities of measured indoor walls were taken from the entire experiment (see Table 3). As shown in the below table, S₀ ranges from 200, 330, 400, and 630. When S₀ and k are both in a relatively-low value, T_A will reach an excessively large value. In this case, Δθ₂ was once set at 10° by controlling the value of k. Additionally, with s₀ in value of 330 and meanwhile Δθ₂ at 5°, a scanning was executed for contrasting (shown in Table 4) presents the specific experiment results.

From the calculation results with different parameters, it should be known that although the scanning scheme 2 takes the longest time, the measurement results have the smallest mean error by being compared with manual measurement results. The scanning scheme 5 takes the shortest time, but the measurement mean error is the largest in the same comparison pattern. Walls 3 and 4 are two walls further away from the novel TLS equipment than the other two, corresponding to point cloud clusters 5 and 6. Associating with the point cloud segmentation results, it can be concluded that point clouds distributed on the two walls are less so that the wall point cloud data are less and the measurement error emerges greater, which is less, and the measurement error is greater. Because of lower mV.E value in scheme 2, it might be true that minimizing Δθ₂ is more efficient in enhancing measurement accuracy. However, scheme 2 has one flaw that the time consumption of scanning in this scheme is even larger than traditional manual measurement method.

5.3. Comparison Result between the Novel TLS and the Existing TLS

(1) Comparison of the Device Parameters. Most existing TLS adopted in indoor measurement is nothing more than FARO Focus 3D laser scanner. In Table 5, the parameters of the two TLS devices are compared, and the existing TLS devices have higher efficiency and density in environment-sensing data collection. However, the existing TLS equipment and the novel TLS equipment can both achieve ranging accuracy ranging within ±1 mm. In another word, in the perspective of scanning range, the mechanical structure of the novel TLS equipment proposed in this paper has no scanning blind area.

(2) Comparison of the Actual Measurement Result. In order to verify the further application and practical potential of the novel TLS equipment in indoor actual measurement, we compare the measurement accuracy, efficiency, and cost of FARO focus [21] and novel TLS in similar experimental conditions, taking the mF.E results of flatness in the indoor slab measurement [21] as accuracy-evaluating standard. Next, contrasting FARO equipment and the novel TLS equipment in two aspects are presented in Table 6: scanning time and amount of point cloud data in actual measurement.

As seen in Table 6, in scanning time consumption, the scanning scheme 1 of the novel TLS equipment proposed in this paper is 1 minute slower than FARO. The measurement accuracy in scanning scheme 2 is 0.8 mm, near to FARO. However, the cost of novel TLS equipment is much lower than FARO. Besides, in the aspect of file size of scanning data, the novel TLS equipment features lower cost of data processing due to its smaller size of collected scanning data than existing TLS equipment.