Construction tunnel laser scanning point cloud attitude correction method based on multi-feature fusion

Abstract

The invention relates to the technical field of tunnel monitoring, in particular to a construction tunnel laser scanning point cloud attitude correction method based on multi-feature fusion, and the method comprises the steps: S1, synchronously extracting a ground baseline feature, a haunch symmetry axis feature and a vault circle center feature from a tunnel section laser scanning point cloud; s2, constructing confidence quantitative models for the three types of features respectively; s3, calculating the weight of each feature through a normalized exponential function based on the confidence of the three types of geometric features obtained in the step S2; setting a confidence threshold value, and forcibly returning to zero weight to shield failure features when the confidence threshold value is lower than the threshold value; when all the weights return to zero, activating a degradation protection mechanism to give equal weights; and finally, weighting and fusing the horizontal rotation angles of the three types of features according to weights, and outputting a correction angle. According to the method, the problem of systematic deviation caused by interference of a single reference is solved, and the correction precision and robustness in a complex construction environment are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of tunnel monitoring technology, specifically to a method for attitude correction of laser scanning point clouds in construction tunnels based on multi-feature fusion. Background Technology

[0002] With the accelerated pace of urbanization, the scale of highway and railway transportation infrastructure has grown rapidly, especially as transportation networks extend into complex, low-quality, and topographically challenging areas, leading to a continuous increase in the scale of new tunnel construction projects. In the construction and excavation of geotechnical tunnels, monitoring the initial support deformation is directly related to the effectiveness of preventing collapse disasters. In accordance with the guidance of the Ministry of Transport, it is clearly required to construct a tunnel structural safety risk monitoring network to achieve intelligent management, strengthen the structural disaster early warning capabilities of extra-long tunnels and tunnel groups, and improve emergency response and intelligent maintenance systems. However, the high-risk dynamic environment of tunnels during construction presents challenges to monitoring, including complex spatial constraints limiting the optimal deployment of sensors, high accuracy requirements for deformation sensing, and stringent real-time data requirements for disaster early warning.

[0003] In practical applications of tunnel monitoring, commonly used methods include total station analysis, close-range photogrammetry, and 3D laser scanning. While total station automatic monitoring is reliable and widely used in monitoring tunnels under construction, it suffers from bottlenecks such as low efficiency, inability to monitor cross-sectional deformation across the entire area, and difficulty in achieving real-time early warning. Close-range photogrammetry, due to its over-reliance on control point deployment and sensitivity to construction dynamics, has limited engineering applicability. In contrast, laser scanning technology, with its high efficiency, rapid response, and high-precision data acquisition capabilities, is more adaptable to monitoring tunnels under construction. Although laser scanning has significant advantages in accuracy, efficiency, and real-time performance, the deviation in scanning profile posture caused by blasting and vibration interference remains one of the core bottlenecks restricting monitoring accuracy. In the practice of calibrating the attitude deviation of tunnels under construction, there are several challenges. The accuracy of the benchmark is insufficient due to the lack of complete point cloud features and reliance on a single geometric benchmark. The feature extraction is difficult due to significant environmental interference. Furthermore, the lack of a dynamic evaluation mechanism for feature credibility makes it difficult to adaptively select the optimal calibration benchmark. These three challenges constitute multiple technical challenges that urgently need to be addressed.

[0004] Therefore, this invention proposes a confidence-weighted multi-feature fusion correction method. By integrating three complementary geometric elements in the tunnel cross-section—ground reference, arch waist symmetry features, and arch crown geometric constraints—a multi-source geometric feature collaborative mechanism is constructed to avoid systematic deviations caused by interference with a single feature. Furthermore, an adaptive confidence assessment model is designed to dynamically quantify feature reliability and optimize fusion weights, achieving robust solution of the correction angle and significantly improving the attitude correction accuracy and stability under complex working conditions. Summary of the Invention

[0005] The purpose of this invention is to provide a method for attitude correction of laser scanning point cloud in construction tunnels based on multi-feature fusion. By simultaneously extracting three types of geometric features—ground baseline, arch waist axis of symmetry, and arch crown center—a dynamic confidence evaluation model is constructed. Based on adaptive weighted fusion using a normalized exponential function, the method solves the problem of systematic deviation caused by interference with a single feature, and significantly improves the correction accuracy and stability under complex working conditions.

[0006] To achieve the above-mentioned technical objectives and effects, the present invention is implemented through the following technical solution: A method for attitude correction of laser scanning point clouds in construction tunnels based on multi-feature fusion includes the following steps: S1: Multi-source geometric feature extraction: Simultaneously extracting three types of geometric features from the laser scanning point cloud of the tunnel cross-section: Ground baseline features: Identify continuous straight segments on the ground through robust piecewise fitting and calculate the horizontal rotation angle; Arched waist symmetry axis features: Optimize the position of the vertical symmetry axis by point cloud mirror transformation and nearest neighbor matching, and calculate the horizontal rotation angle; Features of the arch center: The center of the arch arc is fitted using the RANSAC algorithm, and the horizontal rotation angle is calculated by combining the spatial relationship of the midpoint of the ground baseline. S2: Dynamic confidence assessment: Constructing quantitative confidence models for each of the three types of features: Ground baseline confidence model: based on the joint calculation of fitting residuals and effective straight line length; Confidence model for arched waist symmetry: based on hierarchical symmetry deviation statistics and penalty factor correction calculation; Confidence model of the center of the dome: calculated based on the radius consistency distribution and large deviation penalty mechanism; S3: Based on the confidence scores of the three types of geometric features obtained in step S2, calculate the weight of each feature using the normalized exponential function (Equation 12); set a confidence threshold, and force the weight to be zero when it is below the threshold to shield the failed features; when all weights are zero, activate the degradation protection mechanism to give equal weights; finally, the horizontal rotation angle of the three types of features is fused according to the weight (Equation 11), and the correction angle is output.

[0007] Furthermore, in step S1, the ground baseline feature extraction is segmented and screened based on the slope continuity constraint of adjacent points to identify a potential set of straight line segments; a robust segmented fitting strategy is adopted: the longest continuous and stable segment is screened by limiting the slope fluctuation of adjacent points, and the least squares method is applied to complete the straight line fitting, finally extracting the anti-interference cross-sectional ground baseline; the accurate acquisition of the horizontal calibration angle is achieved through highly robust fitting, specifically including the following sub-steps: S1.1A: Calculate the local slope and quantify the trend of height change between adjacent points.

[0008] (1) Where m i Let be the local slope of the i-th point, i.e., the slope from point i to point i+1; N is the number of effective point clouds after filtering. S1.2A: Detects continuous stable line segments and identifies the longest continuous straight line segment that conforms to the characteristics of the ground baseline.

[0009] (2) Where S is the set of candidate line segment points. The average slope of the line segment. Where L is the slope tolerance threshold and L is the line segment length requirement; S1.3A: Fit the baseline and rotate it to determine the optimal ground baseline and calculate the rotation angle; S1.3A.1: Select the longest line segment that meets the stability requirements.

[0010] (3) S1.3A.2: Least Squares Fitting (4) in The fitting result can be obtained as follows: ; S1.3A.3: Calculate the rotation angle .

[0011] Furthermore, the feature extraction of the arch waist symmetry axis in step S1 includes the following sub-steps: S1.1B: Define the axis of symmetry and generate mirror points to construct a mirror image of the point cloud about the axis of symmetry.

[0012] S1.1B.1: Define the axis of symmetry. Take the vertical axis that is perpendicular to the ground baseline. The axis that varies within ±5° of the centroid of the overall point cloud data is the preset axis of symmetry.

[0013] S1.1B.2: For any point Its mirror point .

[0014] (5) in Let be the directed distance from the point to the axis of symmetry, and let a, b, and c be the general form parameters of the line representing the axis of symmetry.

[0015] S1.2B: Calculate the symmetry score to quantify the degree of matching between the original point cloud and the mirror point cloud.

[0016] S1.2B.1: Randomly sample M points from the original point cloud, construct a KD tree as shown in the formula, and accelerate the nearest neighbor matching between the mirror point and the original point cloud as shown in formula (6).

[0017] (6) in This is the index of the nearest neighbor in the original point cloud. is the Euclidean distance from the mirror point to its nearest neighbor.

[0018] S1.2B.2: Symmetry Scoring (7) Where K is the number of valid mirror points.

[0019] S1.3B: Optimize the search for the axis of symmetry with the highest symmetry score and calculate the optimal rotation angle. .

[0020] Furthermore, step S1, the extraction of the arch center feature, includes directly constructing a point cloud horizontal correction angle calculation model by solving the spatial positional relationship between the fitted center and the midpoint of the baseline; specifically, it includes the following sub-steps: S1.1C: Segment and select the cloud data of the arch apex, fit the arc of the arch apex, and calculate the theoretical center of the circle.

[0021] S1.1C.1: Extract raw point cloud data It can extract the 70th percentile of the y-coordinate. That is, select the arch area points where the y-coordinate value is the largest 30%.

[0022] S1.1C.2: Fitting a circle using RANSAC, the process is as follows: Figure 1 As shown, calculate the actual center of the circle. .

[0023] S1.2C: Calculate the midpoint of the ground baseline under tilt conditions based on the overall point cloud data. .

[0024] S1.3C: Based on the theoretical center Midpoint of baseline Calculate the rotation angle .

[0025] Furthermore, the ground baseline confidence model in step S2 is as follows: (8) in, The baseline least squares fitting residuals reflect the fitting accuracy. The effective baseline is the number of consecutive points, and the length is calculated after excluding interference points; K is the length sensitivity coefficient.

[0026] Furthermore, the confidence model for the arch waist symmetry in step S2 is as follows: The angle calculation of the arch waist symmetry axis method is based on the geometric symmetry evaluation principle of the tunnel cross-section profile. The position of its symmetry axis is directly determined by the spatial distribution balance of the left and right arch waist points. Based on this, the confidence evaluation model adopts a hierarchical point-to-point matching mechanism. The sampling data groups are divided at equal intervals along the longitudinal direction of the tunnel. The left and right horizontal profile points are extracted at the symmetry axis position of each group and the distance deviation from the symmetry axis is calculated. First, the proportion of groups that meet the requirements is statistically analyzed based on the deviation threshold as the basic confidence level, reflecting the overall symmetry level. Second, abnormal data groups that significantly exceed the deviation tolerance are identified, and the suppression of unreliable data is strengthened through the penalty factor. The final confidence level is composed of the product of the basic confidence level and the penalty factor, as shown in equation (9).

[0027] (9) in, The basic confidence factor has N sets of samples. This is the allowable deviation threshold. Let H be the symmetric distance deviation of the k-th group, and H() be the unit step function; For the penalty correction factor term, To punish the lower limit, The amount of punishment per unit This is the deviation penalty threshold.

[0028] Furthermore, the confidence model for the center of the dome in step S2 is as follows: The distance from each point in the arch region to the fitted circle center is calculated, and the proportion of points whose deviation from the radius is less than the allowable threshold is used as the basic confidence level. In addition, to suppress the significant local interference, a large deviation penalty mechanism is introduced simultaneously, that is, when the deviation of a point exceeds the penalty threshold, the penalty intensity is quantified on a point-by-point basis. The final confidence level model is shown in Equation (10).

[0029] (10) in, Based on the basic confidence factor, we have Group sampling number, This is the allowable threshold for radius deviation. The radius deviation of the i-th group; For the penalty correction factor term, To punish the lower limit, The amount of punishment per unit This is the deviation penalty threshold.

[0030] Furthermore, step S3 specifically includes the following sub-steps: S3.1: Weighting Coefficient Calculation: Based on the ground baseline confidence, arch waist symmetry confidence, and arch crown center confidence output in step S2, the weighting coefficients corresponding to each feature are calculated using the normalized exponential transformation function, satisfying the following mathematical relationship: (12) in The confidence level for each feature is calculated separately based on the corresponding feature, where λ is the weight concentration adjustment coefficient, and the weight coefficients satisfy the following conditions: This represents the contribution weight of each feature in the fusion process; S3.2: Feature Activation and Masking: Setting the Feature Activation Function Activated by a confidence threshold, if the confidence of a feature is lower than the confidence threshold, its weight coefficient is forced to zero to mask the invalid feature. S3.3: Degradation protection mechanism: When all feature weights are reduced to zero, each feature is automatically assigned an equal weight; S3.4: Correction Angle Fusion Output: The horizontal rotation angles of the three types of features are weighted and summed according to the weight coefficients to output the final correction angle. (11).

[0031] The beneficial effects of this invention are: This invention establishes a spatial complementary constraint mechanism by simultaneously extracting three types of geometric features: ground baseline, arch waist symmetry axis, and arch crown center. The ground baseline is robustly piecewise fitted based on the slope continuity of adjacent points (Equations 1-4), the arch waist symmetry axis is accelerated by point cloud mirror transformation and KD tree matching (Equations 5-7), and the arch crown center is fitted using the RANSAC algorithm. When any type of feature fails due to disturbance, the remaining features still provide effective geometric constraints, thus mitigating the risk of systematic bias from relying on a single benchmark.

[0032] This invention innovatively constructs a differentiated confidence model for interference scenarios such as ground debris and vehicle disturbance: the ground baseline confidence (Equation 8) couples the fitting residual with the effective length to enhance resistance to local interruptions; the arch waist symmetry confidence (Equation 9) uses hierarchical deviation statistics and penalty factors to accurately quantify asymmetry; the arch crown center confidence (Equation 10) designs a large deviation penalty term based on the radius consistency distribution to suppress the influence of outliers. This transforms feature quality into computable indicators to support fusion decision-making.

[0033] This invention dynamically allocates feature weights using a normalized exponential function (Equation 12), controls the dominance of high-confidence features using an adjustment coefficient, sets a confidence threshold to mask failed features, and innovates a weight degradation protection mechanism to assign equal weights under extreme conditions, ensuring the system continuously outputs effective correction values. This framework forms a thresholded weighted state estimator, guaranteeing algorithm stability under complex disturbances.

[0034] The confidence-weighted fusion mechanism (Equation 11) of this invention suppresses low-quality feature noise, significantly improving correction accuracy under typical interference conditions. Relying on efficient algorithms such as RANSAC fitting and KD-tree matching, it meets the millisecond-level response requirements of embedded scanning devices, breaking through the real-time bottleneck of traditional methods. Dual-range laser testing verifies its adaptability to point cloud density, directly supporting real-time early warning targets for tunnel construction disasters.

[0035] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0036] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 A schematic diagram of the process for calculating the center of a circular arc fitted to RANSAC. Figure 2 A schematic diagram for verifying the contour correction effect; Figure 3 This is a schematic diagram of the multi-feature fusion calibration angle process; Figure 4 This is a schematic diagram of experimental data acquisition. Figure 5 A contour plot of the data collected for the experiment; Figure 6 This is a schematic diagram of the relative error of the fusion method. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] Example 1 The method for attitude correction of laser scanning point clouds in construction tunnels based on multi-feature fusion described in this embodiment includes the following steps: S1: Multi-source geometric feature extraction: Simultaneously extracting three types of geometric features from the laser scanning point cloud of the tunnel cross-section: Ground baseline features: Identify continuous straight segments on the ground through robust piecewise fitting and calculate the horizontal rotation angle; Arched waist symmetry axis features: Optimize the position of the vertical symmetry axis by point cloud mirror transformation and nearest neighbor matching, and calculate the horizontal rotation angle; Features of the arch center: The center of the arch arc is fitted using the RANSAC algorithm, and the horizontal rotation angle is calculated by combining the spatial relationship of the midpoint of the ground baseline. S2: Dynamic confidence assessment: Constructing quantitative confidence models for each of the three types of features: Ground baseline confidence model: based on the joint calculation of fitting residuals and effective straight line length; Confidence model for arched waist symmetry: based on hierarchical symmetry deviation statistics and penalty factor correction calculation; Confidence model of the center of the dome: calculated based on the radius consistency distribution and large deviation penalty mechanism; S3: Based on the confidence scores of the three types of geometric features obtained in step S2, calculate the weight of each feature using the normalized exponential function (Equation 12); set a confidence threshold, and force the weight to be zero when it is below the threshold to shield the failed features; when all weights are zero, activate the degradation protection mechanism to give equal weights; finally, the horizontal rotation angle of the three types of features is fused according to the weight (Equation 11), and the correction angle is output.

[0040] Example 2 Point cloud correction feature extraction method Feature extraction of ground baseline Ideally, the ground baseline of the tunnel cross-section is parallel to the horizontal reference, and its tilt angle can be directly used as the reference for point cloud rotation correction. Considering the continuity of the 2D laser-scanned point cloud profile, segmentation screening is performed based on the slope continuity constraint between adjacent points to identify a candidate set of potential straight line segments. Simultaneously, considering the significant noise and frequent local distortions commonly found in tunnel ground point clouds, a robust segmented fitting strategy is adopted: the longest continuous and stable segment is selected by limiting the slope fluctuation between adjacent points, and the least squares method is applied to complete the straight line fitting, ultimately extracting the interference-resistant cross-section ground baseline. This process achieves accurate acquisition of the horizontal calibration angle through highly robust fitting; the specific steps are as follows: S1: Calculate the local slope to quantify the height change trend between adjacent points.

[0041] (1) Where m i Let be the local slope of the i-th point, i.e., the slope from point i to point i+1. N is the number of effective point clouds after filtering.

[0042] S2: Detect continuous stable line segments and identify the longest continuous straight line segment that conforms to the characteristics of the ground baseline.

[0043] (2) Where S is the set of candidate line segment points. The average slope of the line segment. Where is the slope tolerance threshold, and L is the line segment length requirement.

[0044] S3: Fit the baseline and rotate it to determine the optimal ground baseline and calculate the rotation angle.

[0045] S3.1: Select the longest line segment that meets the stability requirements.

[0046] (3) S3.2: Least Squares Fitting (4) in The fitting result can be obtained as follows: .

[0047] S3.3: Calculate the rotation angle .

[0048] Feature extraction for arch symmetry assessment The core spatial geometric feature of the arched structure is the height axis symmetry of its left and right contours. Under ideal conditions, the point clouds on the left and right sides of the arched structure exhibit precise mirror-symmetric distribution. To quantify this symmetry, a point cloud mirror transformation coupled with nearest-neighbor matching is employed. By maximizing the matching degree between the original point cloud and the mirrored point cloud, the optimal position of the vertical axis of symmetry is determined.

[11] The horizontal rotation angle calculated based on this axis of symmetry can be directly used for point cloud coordinate system correction. The specific steps are as follows: S1: Define the axis of symmetry and generate mirror points to construct a mirror image of the point cloud about the axis of symmetry.

[0049] S1.1: Define the axis of symmetry. Take the vertical axis that is perpendicular to the ground baseline. The axis that varies within ±5° of the centroid of the overall point cloud data is the preset axis of symmetry.

[0050] S1.2: For any point Its mirror point .

[0051] (5) in Let be the directed distance from the point to the axis of symmetry, and let a, b, and c be the general form parameters of the line representing the axis of symmetry.

[0052] S2: Calculate the symmetry score to quantify the degree of matching between the original point cloud and the mirror point cloud.

[0053] S2.1: Randomly sample M points from the original point cloud, construct a KD tree as shown in the formula, and accelerate the nearest neighbor matching between the mirror point and the original point cloud as shown in formula (6).

[0054] (6) in This is the index of the nearest neighbor in the original point cloud. is the Euclidean distance from the mirror point to its nearest neighbor.

[0055] S2.2: Symmetry Scoring (7) Where K is the number of valid mirror points.

[0056] S3: Optimize the search for the axis of symmetry with the highest symmetry score and calculate the optimal rotation angle. .

[0057] Feature extraction of the positioning center of the dome The circular surface of the dome, regardless of its tilt, produces a highly intact circular arc point cloud. The RANSAC algorithm can fit the theoretical center of this circle. [12, 13] The geometric constraints of a typical semi-circular arch tunnel cross-section are as follows: the center of the upper theoretical circular arc and the center point of the lower rectangular baseline should lie on the same vertical line in a horizontal state. Based on this prior knowledge, a point cloud horizontal correction angle calculation model can be directly constructed by solving the spatial relationship between the fitted center and the midpoint of the baseline. This method deeply integrates geometric priors and point cloud fitting techniques. The specific implementation process is as follows: S1: Segment and select the cloud data of the arch apex, fit the arc of the arch apex, and calculate the theoretical center of the circle.

[0058] S1.1: Extract raw point cloud data It can extract the 70th percentile of the y-coordinate. That is, select the arch area points where the y-coordinate value is the largest 30%.

[0059] S1.2: Fit a circle using RANSAC, the process is as follows: Figure 1 As shown, calculate the actual center of the circle. .

[0060] S2: Calculate the midpoint of the ground baseline under tilt conditions based on the overall point cloud data. .

[0061] S3: Based on the theoretical center Midpoint of baseline Calculate the rotation angle .

[0062] Feature extraction performance verification To verify the effectiveness of the ground baseline method, symmetry axis method, and center-of-circle positioning method in horizontal correction of tunnel point clouds, an ideal contour point cloud dataset was constructed based on the design parameters of the initial support section of a semi-circular arch tunnel. Simulating equipment installation deviation conditions, counterclockwise rotation transformations of -10°, 10°, and 20° were applied to the ideal point cloud, respectively. Figure 2 As shown in (a). The verification results are as follows. Figure 2 As shown in (b), the calculated ground baseline, axis of symmetry, and fitted circle center are consistent with the theoretical geometric features, and the horizontal correction results of each feature are in line with the theoretical expectations, thus verifying the effectiveness of the ground baseline method, axis of symmetry method, and circle center positioning method.

[0063] Multi-feature fusion level correction method Confidence assessment In actual tunnel construction sites, random disturbances such as obstruction of ventilation pipes at the arch crown, vehicle movement at the arch waist, and ground debris can easily lead to characteristic distortions such as inaccurate center fitting, symmetry loss, or baseline interruption. [4] By establishing a mathematical mapping relationship between confidence level and feature integrity, the degree of interference with features can be identified, and the reliability of the level correction results can be evaluated.

[14] .

[0064] Confidence level of ground baseline method To quantify the synergistic effect between baseline fitting quality and effective geometric scale, a method is proposed to construct a fitting quality evaluation term by coupling the fitting residuals and the effective straight line length. This allows the confidence level to increase as the residuals decrease and the length increases, and the confidence gain of the effective length term is nonlinearly modulated based on the length sensitivity coefficient, as shown in Equation (8).

[0065] (8) in, The baseline least squares fitting residuals reflect the fitting accuracy. The effective baseline is the number of consecutive points, and the length is calculated after excluding interference points; K is the length sensitivity coefficient.

[0066] Confidence level of arched waist symmetry method The angle calculation of the arch waist symmetry axis method is based on the geometric symmetry evaluation principle of the tunnel cross-section profile. The position of its symmetry axis is directly determined by the spatial distribution balance of the left and right arch waist points. Based on this, the confidence evaluation model adopts a hierarchical point-to-point matching mechanism. The sampling data groups are divided at equal intervals along the longitudinal direction of the tunnel. The left and right horizontal profile points are extracted at the symmetry axis position of each group and the distance deviation from the symmetry axis is calculated. First, the percentage of groups that meet the requirements is statistically analyzed based on the deviation threshold as the basic confidence level, reflecting the overall symmetry level. Second, abnormal data groups that significantly exceed the deviation tolerance are identified, and the suppression of unreliable data is strengthened by the penalty factor. The final confidence level is composed of the product of the basic confidence level and the penalty factor, as shown in Equation (9).

[0067] (9) in, The basic confidence factor has N sets of samples. This is the allowable deviation threshold. Let H be the symmetric distance deviation of the k-th group, and H() be the unit step function; For the penalty correction factor term, To punish the lower limit, The amount of punishment per unit This is the deviation penalty threshold.

[0068] Confidence level of dome center alignment method Based on the principle of radius consistency, a confidence assessment method for the center of the arch is proposed: the distance from the standard arc point of the tunnel to the center should be approximately equal, so the radius deviation distribution is used as the core evaluation index. The method selects a point set in the arch area to calculate the distance from each point to the fitted center, and statistically analyzes the proportion of points with radius deviations less than the allowable threshold as the basic confidence level. In addition, to suppress the significant local interference, a large deviation penalty mechanism is introduced simultaneously, that is, when the deviation of a point exceeds the penalty threshold, the penalty intensity is quantified on a point-by-point basis. The final confidence model is shown in Equation (10).

[0069] (10) in, Based on the basic confidence factor, we have Group sampling number, This is the allowable threshold for radius deviation. The radius deviation of the i-th group; For the penalty correction factor term, To punish the lower limit, The amount of punishment per unit This is the deviation penalty threshold.

[0070] Multi-feature fusion framework In practical tunnel point cloud rotation correction, environmental interference such as ventilation duct obstruction at the arch crown, vehicle activity at the arch waist, and ground debris can easily lead to the failure of single features. Therefore, this study proposes a multi-feature fusion framework based on dynamic confidence weighting, the process of which is as follows: Figure 3 As shown.

[0071] The core idea of ​​this method is to utilize the spatial constraints of three geometric features in three-dimensional space—the ground baseline, the arch waist symmetry axis, and the actual center of the circle—to construct a joint optimization model through their inherent geometric consistency. The fusion equation is defined as: (11) (12) The weight coefficients satisfy the following: The weights represent the contribution of each feature to the fusion process. Each weight is calculated using a normalization exponent as shown in equation (12), where... The confidence level for each feature is calculated based on the respective feature. The activation function is a feature activated by a confidence threshold. This calculation model implements a differential amplification mechanism, where the weight of a single feature approaches 1 when its confidence level is significantly dominant. Simultaneously, a weight degradation protection mechanism is set up, automatically assigning equal weights when all weights return to zero. Furthermore, the degree of weight concentration is controlled by adjusting the coefficient λ, increasing interpretability.

[0072] Example 3 Experimental setup The experiment used a self-developed two-dimensional laser contour scanning device, such as... Figure 4 As shown, compared with traditional 3D laser scanning equipment, this device has significant engineering advantages such as real-time dynamic monitoring and low economic cost, while ensuring millimeter-level measurement accuracy.

[0073] The experiment selected the Daxiangshu Tunnel section under construction on the Mouding-Yuanmou Expressway in Yunnan Province, and collected point cloud data of the initial support profile of four continuous cross-sections spaced 5 meters apart. To eliminate the interference of the laser range measuring instrument's range on the point cloud quality, data was collected simultaneously using 40M and 80M dual-range laser probes. The four typical cross-section interference scenarios were as follows: the arch of cross-section 1 was obstructed by ventilation ducts and the arch waist was disturbed by vehicles; the arch waist of cross-section 2 was disturbed by vehicles; the arch waist of cross-section 3 was disturbed by vehicles and there were debris distributed on the ground; cross-section 4 showed a clean profile with no significant interference from the arch, arch waist and ground compared to the other cross-sections.

[0074] The outline of the collected data is as follows Figure 5 As shown, a progressive test sample set is constructed, ranging from complex to ideal. During the data acquisition process, a horizontal rangefinder was used to measure the deviation angle between the initial pose of the laser scanning device and the horizontal reference, as shown in Table 1. This measured value will serve as the benchmark for subsequent verification of the point cloud rotation correction accuracy.

[0075] Table 1 Measured Angles Section 1 Section 2 Section 3 Section 4 40M（°） 15.37 16.40 15.83 16.46 80M（°） 15.09 16.04 15.79 16.2 Feature extraction and confidence verification Based on the contour feature analysis of the four sections and the geometric principles of horizontal calibration angle calculation, it can be inferred that the three individual methods have certain limitations. The center-of-circle positioning method has low accuracy in sections 1 and 2 where the ventilation duct at the arch is interfered with; its arch center fitting process suffers from systematic bias due to obstructions. The symmetry axis method also has low confidence in sections 1 and 2. The confidence calculation mechanism of this method relies on the symmetry evaluation of the arch waist, but data interference at the arch waist leads to structural asymmetry, rendering it unreliable. The ground baseline method performs poorly in section 3, where ground debris and personnel activity are present. Its confidence stems from the quantitative assessment of baseline continuity and effective length; however, interference disrupts the integrity of the ground point cloud, resulting in low reliability of the angle calculation results.

[0076] Table 2 Calculation Table of Rotation Angle and Confidence Level As shown in Table 2, the ground baseline method (b), the arch waist symmetry axis method (s), and the arch crown center alignment method (c) all exhibit clear limitations in calculating rotation angles under interference scenarios. Regardless of whether a 40M or 80M range laser probe is used, all three methods show low confidence levels when interference exists at the arch crown, arch waist, or ground. For example, the confidence levels of the symmetry axis method (s) and the center alignment method (c) are as low as 0.0189 and 0.3021, respectively, in section 1 where ventilation ducts and vehicles coexist. The ground baseline method (b) suffers from zero confidence and cannot complete the fitting calculation due to the lack of an effective continuous baseline caused by ground debris interference in section 3. The calculation results verify that the horizontal calibration method relying on a single geometric feature has significant limitations, specifically manifested in insufficient reliability of the angle calculation results.

[0077] Confidence fusion calculation Based on the confidence-driven weight allocation mechanism designed in the aforementioned multi-feature fusion framework, when the confidence level calculated by any method is lower than the 0.5 threshold, the weight of the corresponding calibration angle is reset to zero, that is, the low-confidence data is actively excluded from the fusion calculation. This strategy essentially ensures the stability of the fusion result by dynamically filtering out the noise input caused by feature failure. Table 3 shows the calibration angle table for the fusion calculation and its corresponding relative error.

[0078] Table 3. Fusion Angle and its Relative Error Section 1 Section 2 Section 3 Section 4 40M Angle (°) 16.6 15.75 15.82 15.86 Relative error (%) 3.6 0.1 4 8 80M Angle (°) 16.1 16.59 15.77 15.72 Relative error (%) 3 0.1 3.4 6.7 like Figure 6As shown, angle outputs with confidence levels below the validity threshold are considered invalid data, and their relative errors are uniformly equivalent to 100%. In the unobstructed section 4 test, data from the 40M range laser probe shows that the relative errors of the corrected angles for the ground baseline method (b), the arch waist symmetry axis method (s), and the arch crown center method (c) are 1.6%, 4.9%, and 5.1%, respectively, showing an increasing trend. These errors are negatively correlated with the decreasing trend of their confidence levels (0.8410, 0.7792, and 0.7164). For the 80M range probe, the errors for methods b, s, and c are 3.1%, 7.4%, and 5.3%, respectively, showing a convex distribution. These errors are negatively correlated with the confidence levels of the concave responses (0.7432, 0.6964, and 0.7114). This result is consistent with the confidence-driven fusion principle, i.e., the error and confidence level have a strong coupling relationship, and the confidence level can effectively characterize the reliability of angle calculations, providing a quantitative decision-making basis for multi-feature fusion.

[0079] From the perspective of error stability, the multi-feature fusion method demonstrates excellent reliability in cross-section scenarios 1-3, where ventilation duct obstruction, vehicle activity, and ground debris interference coexist. Its relative error in the horizontal calibration angle is consistently controlled below 8.0%, proving the method's strong adaptability to complex interference environments. More importantly, this method possesses the advantage of systematic accuracy. Under all test conditions, the relative error of the fusion result is significantly lower than the maximum relative error of any independent method—the ground baseline method, the arch waist symmetry axis method, and the arch crown center method. This characteristic verifies that the confidence-driven fusion mechanism fundamentally ensures the accuracy and robustness of the angle output by dynamically suppressing failure features.

[0080] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for attitude correction of laser scanning point clouds in construction tunnels based on multi-feature fusion, characterized in that, Includes the following steps: S1: Multi-source geometric feature extraction: Simultaneously extracting three types of geometric features from the laser scanning point cloud of the tunnel cross-section: Ground baseline features: Identify continuous straight segments on the ground through robust piecewise fitting and calculate the horizontal rotation angle; Arched waist symmetry axis features: Optimize the position of the vertical symmetry axis by point cloud mirror transformation and nearest neighbor matching, and calculate the horizontal rotation angle; Features of the arch center: The center of the arch arc is fitted using the RANSAC algorithm, and the horizontal rotation angle is calculated by combining the spatial relationship of the midpoint of the ground baseline. S2: Dynamic confidence assessment: Constructing quantitative confidence models for each of the three types of features: Ground baseline confidence model: based on the joint calculation of fitting residuals and effective straight line length; Confidence model for arched waist symmetry: based on hierarchical symmetry deviation statistics and penalty factor correction calculation; Confidence model of the center of the dome: calculated based on the radius consistency distribution and large deviation penalty mechanism; S3: Based on the confidence scores of the three types of geometric features obtained in step S2, calculate the weight of each feature using a normalized exponential function; set a confidence threshold, and force the weights to zero when the scores are below the threshold to shield the failed features; activate the degradation protection mechanism to assign equal weights when all weights are zero; finally, fuse the horizontal rotation angles of the three types of features according to their weights and output the correction angle.

2. The method for attitude correction of laser scanning point clouds in construction tunnels based on multi-feature fusion as described in claim 1, characterized in that: In step S1, ground baseline feature extraction is performed by segmenting and screening based on the slope continuity constraint between adjacent points to identify a candidate set of potential straight line segments. A robust segmented fitting strategy is adopted: the longest continuous and stable segment is selected by limiting the slope fluctuation between adjacent points, and the least squares method is applied to complete the straight line fitting, finally extracting the anti-interference cross-sectional ground baseline. The accurate acquisition of the horizontal calibration angle is achieved through highly robust fitting, specifically including the following sub-steps: S1.1A: Calculate the local slope and quantify the height change trend between adjacent points; （1） Where m i Let be the local slope of the i-th point, i.e., the slope from point i to point i+1; N is the number of effective point clouds after filtering. S1.2A: Detect continuous stable line segments and identify the longest continuous straight line segment that conforms to the characteristics of the ground baseline; （2） Where S is the set of candidate line segment points. The average slope of the line segment. Where L is the slope tolerance threshold and L is the line segment length requirement; S1.3A: Fit the baseline and rotate it to determine the optimal ground baseline and calculate the rotation angle.

3. The method for attitude correction of laser scanning point clouds in construction tunnels based on multi-feature fusion as described in claim 2, characterized in that: Step S1.3A specifically includes the following sub-steps: S1.3A.1: Select the longest line segment that meets the stability requirements; （3） S1.3A.2: Least squares fitting; （4） in The fitting result can be obtained as follows: ; S1.3A.3: Calculate the rotation angle .

4. The method for attitude correction of laser scanning point clouds in construction tunnels based on multi-feature fusion as described in claim 1, characterized in that: Step S1, the extraction of the arch waist symmetry axis features, includes the following sub-steps: S1.1B: Define the axis of symmetry and generate mirror points to construct a mirror image of the point cloud about the axis of symmetry; S1.1B.1: Define the axis of symmetry. Take the vertical axis that is perpendicular to the ground baseline. The axis that varies within a range of ±5° with the centroid of the overall point cloud data as the center is the preset axis of symmetry. S1.1B.2: For any point Its mirror point ; （5） in Let be the directed distance from the point to the axis of symmetry, and let a, b, and c be the general form parameters of the line of symmetry. S1.2B: Calculate the symmetry score to quantify the degree of matching between the original point cloud and the mirror point cloud; S1.2B.1: Randomly sample M points from the original point cloud, construct a KD tree as shown in the formula, and accelerate the nearest neighbor matching between the mirror point and the original point cloud as shown in formula (6); （6） in This is the index of the nearest neighbor in the original point cloud. The Euclidean distance from the mirror point to its nearest neighbor; S1.2B.2: Symmetry Scoring （7） Where K is the number of valid mirror points; S1.3B: Optimize the search for the axis of symmetry with the highest symmetry score and calculate the optimal rotation angle. .

5. The method for attitude correction of laser scanning point clouds in construction tunnels based on multi-feature fusion as described in claim 1, characterized in that: Step S1, the extraction of the arch center feature, includes directly constructing a point cloud horizontal correction angle calculation model by solving the spatial positional relationship between the fitted center and the midpoint of the baseline; specifically, it includes the following sub-steps: S1.1C: Segment and select the cloud data of the arch apex, fit the arc of the arch apex, and calculate the theoretical center of the circle; S1.2C: Calculate the midpoint of the ground baseline under tilt conditions based on the overall point cloud data. ; S1.3C: Based on the theoretical center Midpoint of baseline Calculate the rotation angle .

6. The method for attitude correction of laser scanning point clouds in construction tunnels based on multi-feature fusion as described in claim 5, characterized in that: Step S1.1C specifically includes the following sub-steps: S1.1C.1: Extract raw point cloud data Extract the 70th percentile of the y-coordinate That is, select the arch area points where the y-coordinate value is the largest 30%; S1.1C.2: Fit the circle using RANSAC and calculate the actual center of the circle. .

7. The method for attitude correction of laser scanning point clouds in construction tunnels based on multi-feature fusion as described in claim 1, characterized in that: The ground baseline confidence model in step S2 is as follows: （8） in, The baseline least squares fitting residuals reflect the fitting accuracy. The effective baseline is the number of consecutive points, and the length is calculated after excluding interference points; K is the length sensitivity coefficient.

8. The method for attitude correction of laser scanning point clouds in construction tunnels based on multi-feature fusion as described in claim 1, characterized in that: In step S2, the confidence model for the symmetry of the arch waist is as follows: the sampling data groups are divided at equal intervals along the longitudinal direction of the tunnel, and the left and right horizontal contour points are extracted at the symmetry axis position of each group and the distance deviation from the symmetry axis is calculated; firstly, the proportion of groups that meet the requirements is statistically calculated based on the deviation threshold as the basic confidence level, reflecting the overall symmetry level; secondly, abnormal data groups that significantly exceed the deviation tolerance are identified, and the suppression of unreliable data is strengthened by the penalty factor; finally, the confidence level is composed of the product of the basic confidence level and the penalty factor, as shown in equation (9). （9） in, The basic confidence factor has N sets of samples. This is the allowable deviation threshold. Let H be the symmetric distance deviation of the k-th group, and H() be the unit step function; For the penalty correction factor term, To punish the lower limit, The amount of punishment per unit This is the deviation penalty threshold.

9. The method for attitude correction of laser scanning point clouds in construction tunnels based on multi-feature fusion as described in claim 1, characterized in that: The confidence model for the center of the arch in step S2 is as follows: The distance from each point to the fitted circle center is calculated by selecting a point set in the arch area. The proportion of points whose deviation from the radius is less than the allowable threshold is used as the basic confidence level. In addition, in order to suppress the significant local interference, a large deviation penalty mechanism is introduced simultaneously. That is, when the deviation of a point exceeds the penalty threshold, the penalty intensity is quantified by point. The final confidence level model is shown in Equation (10). （10） in, Based on the basic confidence factor, we have Group sampling number, This is the allowable threshold for radius deviation. The radius deviation of the i-th group; For the penalty correction factor term, To punish the lower limit, The amount of punishment per unit This is the deviation penalty threshold.

10. The method for attitude correction of laser scanning point clouds in construction tunnels based on multi-feature fusion as described in claim 1, characterized in that: Step S3 specifically includes the following sub-steps: S3.1: Weighting Coefficient Calculation: Based on the ground baseline confidence, arch waist symmetry confidence, and arch crown center confidence output in step S2, the weighting coefficients corresponding to each feature are calculated using the normalized exponential transformation function, satisfying the following mathematical relationship: （12） in The confidence level for each feature is calculated separately based on the corresponding feature, where λ is the weight concentration adjustment coefficient, and the weight coefficients satisfy the following conditions: This represents the contribution weight of each feature in the fusion process; S3.2: Feature Activation and Masking: Set feature activation function Activated by a confidence threshold, if the confidence of a feature is lower than the confidence threshold, its weight coefficient is forced to zero to mask the invalid feature. S3.3: Degradation protection mechanism: When all feature weights are reduced to zero, each feature is automatically assigned an equal weight; S3.4: Correction Angle Fusion Output: The horizontal rotation angles of the three types of features are weighted and summed according to the weight coefficients to output the final correction angle. （11）。

Images