Methods and systems for measuring the attitude of structural surfaces
By capturing a single image with a mobile terminal and combining it with acceleration and magnetic field data for coordinate transformation, the problems of high safety risks, complex equipment, and magnetic interference in structural surface attitude measurement are solved. This enables rapid and accurate attitude measurement, reduces costs, and improves data reliability.
Patent Information
- Application Number
- CN202611319840.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-28
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies for measuring the attitude of structural surfaces present several challenges, including safety risks due to limited accessibility for personnel, high and complex equipment costs, lack of non-contact coverage of the entire process, unknown point cloud orientation, and significant heading deviations due to magnetic interference.
A single image is captured using a mobile terminal, and acceleration and magnetic field data are collected simultaneously. A point cloud is output using a feedforward 3D reconstruction model. A rotation matrix is constructed using acceleration and magnetic field data to perform coordinate transformation. Combined with semantic recognition and attitude correction, the orientation of the structural surface is directly calculated, avoiding contact measurement and specialized equipment.
It enables rapid, simple, and accurate measurement of structural surface orientation in complex environments, reducing safety risks and equipment costs, improving the non-contact nature of measurements and data reliability, and reducing heading deviations under magnetic interference.
Smart Images

Figure CN123111184A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of engineering geology and three-dimensional measurement technology, specifically to a method and system for measuring the attitude of structural surfaces. Background Technology
[0002] Structural planes are the general term for geological interfaces such as joints, bedding planes, and fault planes in rock masses. Their occurrence is characterized by strike, dip, and dip angle, and they are the basic parameters for rock mass quality classification and slope stability analysis.
[0003] Structural surface orientation has long relied on on-site measurements using geological compasses, requiring personnel to reach the target location and hold the compass close to the structural surface for readings—a contact-based measurement method. This is difficult to access in areas such as steep slopes, unstable rock masses, and the roof of mined-out areas, and carries the risk of rockfalls and collapses. While 3D laser scanning, photogrammetry, and UAV oblique photography can acquire 3D data non-contactly, the equipment is expensive, the fieldwork is cumbersome, and post-processing relies on specialized software, making it difficult to provide immediate results on-site.
[0004] Patent CN109738440B involves setting up horizontal targets on the rock surface and measuring their attitude using a mobile phone electronic compass. A point cloud is then generated from multiple photos containing the targets to complete the orientation. Patent CN108830317B uses a compass to measure the attitude of a smooth structural surface as a azimuth reference, and then establishes a separate benchmark to restore the scale. Both patents rely on manual placement of objects and contact measurement with the compass; if personnel cannot reach the surface, the placement itself is infeasible. Patent CN118152714A calculates the attitude solely using a mobile phone accelerometer and magnetometer, requiring the phone to be in contact with the measured surface. Patent CN116758133B uses a total station to measure three non-collinear points to solve for the attitude, requiring specialized instruments to be on-site for station setup. None of these methods allow for non-contact rapid surveying.
[0005] The feedforward 3D reconstruction model developed in recent years can directly output dense 3D geometric information from a single image through one forward inference, without the need for image matching and bundle adjustment. However, the orientation of its output point cloud is unknown and it lacks absolute physical scale, so it is generally considered difficult to use directly for engineering measurement and has not yet been used for obtaining the attitude of structural surfaces.
[0006] In summary, the existing technology has the following drawbacks: (1) Geological compass measurements require contact with the target structural surface, which is difficult for personnel to reach and poses safety risks; (2) Three-dimensional laser scanning, photogrammetry and UAV oblique photography require special equipment, which are costly, complex in post-processing, and difficult to measure in real time on site; (3) Existing mobile phone imaging solutions require the deployment of targets or benchmarks and the use of compasses to measure the orientation of reference surfaces, and non-contact methods do not cover the entire process; (4) The orientation of the point cloud output by the feedforward 3D reconstruction model is unknown, and it is impossible to directly calculate the attitude with geographical reference significance; (5) Magnetic interference sources such as steel arch frames and anchor bolts are common at the engineering site. The heading calculated by relying solely on magnetic sensors is inaccurate, and there is a lack of pre-judgment of data quality and correction of heading deviation. Summary of the Invention
[0007] In view of this, the present invention provides a method and system for measuring the attitude of structural surfaces to solve the following problems existing in the prior art: contact measurement is limited by personnel accessibility and operational safety; three-dimensional laser scanning and photogrammetry rely on specialized equipment and have cumbersome post-processing, making it difficult to obtain the attitude in real time on site; smartphone-based imaging solutions still require setting up targets and contacting the measurement reference surface, failing to achieve non-contact operation of the entire process; the geodetic spatial orientation and absolute physical scale of the point cloud output by the feedforward three-dimensional reconstruction model are unknown, making it impossible to directly calculate angle quantities with geographical reference significance; and there is a lack of pre-judgment of the quality of acquired data and correction methods for heading deviation under magnetic interference environments.
[0008] To address the aforementioned technical problems, this invention provides a method for measuring the attitude of a structural surface, comprising: capturing a single image containing the target structural surface using a mobile terminal, and simultaneously acquiring acceleration and magnetic field data of the mobile terminal at the time of capture; inputting the single image into a feedforward 3D reconstruction model, and inferring and outputting a 3D point cloud located in the image coordinate system, wherein the absolute physical scale of the 3D point cloud is unknown; transforming the 3D point cloud to the device coordinate system according to the mapping relationship between the image coordinate system and the device coordinate system of the mobile terminal, wherein the mapping relationship is an orthogonal mapping matrix corresponding to the orientation of the mobile terminal when capturing the single image, and the orthogonal mapping matrix represents the relationship between the coordinate axes of the image coordinate system and the coordinate axes of the device coordinate system. The following steps are performed: A gravity vector is obtained from the acceleration data, a geomagnetic vector is obtained from the magnetic field data, and a rotation matrix from the device coordinate system to the world coordinate system is constructed solely from the gravity vector and the geomagnetic vector. The world coordinate system uses east, north, and celestial axes as coordinate axes. A rotation transformation is applied to the three-dimensional point cloud located in the device coordinate system using the rotation matrix to obtain a three-dimensional point cloud located in the world coordinate system. A point set representing the target structural surface is extracted from the three-dimensional point cloud located in the world coordinate system. A plane normal vector is obtained by fitting the point set to a plane, and the attitude of the target structural surface is calculated using the plane normal vector. The attitude includes at least one of strike, dip, and dip angle. Absolute physical scale restoration is not performed on the three-dimensional point cloud throughout the entire process.
[0009] The feedforward 3D reconstruction model referred to in this invention is a neural network model that takes a single image as the sole image input and directly outputs the dense 3D geometry of the scene after a single forward inference. Its inference process does not include iterative optimization steps such as image feature matching, bundle adjustment, or multi-view stereo matching. Based on the output format, this type of model has two implementation routes: one is pixel-by-pixel 3D point map regression, where the model directly regresses a 3D coordinate for each pixel in the image, and the output is a 3D point cloud corresponding to each pixel; the other is dense depth map regression, where the model regresses a depth value for each pixel, and then backprojects the depth map into a 3D point cloud according to camera intrinsic parameters. This route can also further regress surface normal vectors simultaneously. Both of these models can be used in this invention.
[0010] The term "unknown absolute physical scale" in this invention refers to the absence of a known, definite proportional relationship between the dimensions of the coordinates of each point in the 3D point cloud and the units of length in the real world. In other words, the point cloud as a whole can differ by an unknown positive real multiple. This invention does not solve for this proportional relationship, nor does it rely on any reference objects, baselines, or ranging methods of known size for reconstruction. This is because the attitude of structural surfaces is characterized by three angular quantities: strike, dip, and dip angle. These angular quantities remain unchanged when the point cloud is scaled proportionally, exhibiting scale invariance. Furthermore, the direction of the normal vector obtained from plane fitting does not change due to proportional scaling. Therefore, in the application scenario of this invention, the lack of an absolute physical scale does not affect the calculation results of the attitude, and it eliminates the need for reference object setup and corresponding calculations required for scale reconstruction.
[0011] The world coordinate system referred to in this invention has its X-axis pointing due east, its Y-axis pointing due north, and its Z-axis pointing vertically upward, i.e., the Northeast-Northeast coordinate system. The azimuth angle referred to in this invention is always measured clockwise from north as positive, and its value is normalized to a range not less than 0 degrees and less than 360 degrees.
[0012] Preferably, the feedforward 3D reconstruction model is a neural network model that directly regresses a pixel-by-pixel 3D point map from a single image through one forward inference, or a neural network model that directly regresses a dense depth map from a single image through one forward inference and backprojects the dense depth map into a 3D point cloud according to camera intrinsic parameters; the feedforward 3D reconstruction model outputs the confidence score of each point in the 3D point cloud while outputting the 3D point cloud, and removes points in the 3D point cloud whose confidence score is lower than a set confidence score threshold.
[0013] Preferably, the mapping relationship is determined as follows: the component of the acceleration data in the screen plane of the mobile terminal is used as the gravity projection; the rotation angle of the gravity projection relative to a set reference direction in the screen plane is calculated; and according to the angle interval to which the rotation angle belongs, one set is selected from multiple sets of mapping matrices that correspond one-to-one with various holding orientations of the mobile terminal as the mapping relationship; the rotation matrix is constructed as follows: the cross product of the gravity vector and the geomagnetic vector, normalized, is used as the east vector; the cross product of the east vector and the gravity vector, normalized, is used as the north vector located in the horizontal plane; the opposite vector of the gravity vector is used as the celestial vector; and the east vector, the north vector, and the celestial vector constitute the rotation matrix.
[0014] Preferably, before inferring and outputting the 3D point cloud, a sensor reliability evaluation is further included: the magnitude of the magnetic field data exceeding the reference geomagnetic intensity range of the shooting location is taken as the magnetic field deviation; when the magnitude falls within this range, the magnetic field deviation is zero; the magnetic field reliability is the ratio of 1 minus the magnetic field deviation to a set magnetic field deviation tolerance, and the magnetic field reliability is zero when this ratio is greater than 1; the absolute value of the difference between the magnitude of the acceleration data and the gravitational acceleration is taken as the acceleration deviation; the acceleration reliability is the ratio of 1 minus the acceleration deviation to a set acceleration deviation tolerance, and the acceleration reliability is zero when this ratio is greater than 1; multiple poses continuously collected by the mobile terminal before shooting are evaluated. The attitude rotation matrix is used to calculate the average rotation matrix. The relative rotation matrix of each attitude rotation matrix with respect to the average rotation matrix is used as the attitude deviation matrix. The mean of the matrix distances between each attitude deviation matrix and the identity matrix is used as the attitude discrepancy. The attitude stability reliability is calculated as 1 minus the ratio of the attitude discrepancy to the set maximum allowable attitude change. When this ratio is greater than 1, the attitude stability reliability is zero. The overall reliability is obtained by weighting the magnetic field reliability, the acceleration reliability, and the attitude stability reliability with three weights, and the sum of the three weights is 1. When the overall reliability is lower than the set minimum reliability threshold, a re-acquisition prompt is output and subsequent steps are stopped.
[0015] Preferably, after obtaining the 3D point cloud located in the world coordinate system and before extracting the point set of the target structural surface, the method further includes orientation correction: performing semantic recognition on the single image to obtain a pixel mask corresponding to the scene object with known spatial pose; mapping the pixel mask to a reference point set located in the world coordinate system according to the correspondence between pixels and 3D points established by the feedforward 3D reconstruction model; performing plane fitting on the reference point set to obtain a reference plane normal vector; constructing a correction rotation matrix based on the deviation between the reference plane normal vector and the known spatial pose of the scene object, and applying a secondary rotation transformation to the 3D point cloud located in the world coordinate system with the correction rotation matrix, and replacing the 3D point cloud located in the world coordinate system with the 3D point cloud after the secondary rotation transformation; wherein, the scene object includes at least one of an object with a semantic recognition result of a horizontal plane and an object with a semantic recognition result of a vertical plane, the known spatial pose of the object with a horizontal plane is that its normal vector is along the vertical direction, and the known spatial pose of the object with a vertical plane is that its normal vector has a known azimuth angle in the world coordinate system, and this known azimuth angle is denoted as the true azimuth angle.
[0016] Preferably, when the scene object includes objects whose semantic recognition result is a horizontal plane, the correction rotation matrix obtained from the horizontal plane object is a tilt correction matrix. The tilt correction matrix is constructed using the cross product of the reference plane normal vector corresponding to the horizontal plane object and the celestial unit vector, normalized, as the rotation axis, and the angle between the reference plane normal vector and the celestial unit vector as the rotation angle. Furthermore, the tilt correction matrix only changes the pitch and roll components without changing the yaw component. When the scene object includes objects whose semantic recognition result is a vertical plane, the correction rotation matrix obtained from the horizontal plane object is a tilt correction matrix. The correction rotation matrix obtained for the object on the vertical plane is the heading correction matrix: the azimuth angle of the projection of the reference plane normal vector corresponding to the object on the vertical plane into the horizontal plane is taken as the measured azimuth angle, and the difference between the true azimuth angle and the measured azimuth angle is taken as the heading deviation. The heading correction matrix is a rotation matrix that rotates the heading deviation around the celestial axis, and the heading correction matrix only changes the heading component without changing the pitch and roll components. Simultaneously, the amount by which the magnitude of the magnetic field data exceeds the reference geomagnetic intensity range of the shooting location is taken as the magnetic field deviation. When the magnitude falls within this range... The magnetic field deviation is zero. The magnetic field reliability is calculated as 1 minus the ratio of the magnetic field deviation to a set magnetic field deviation tolerance. If this ratio is greater than 1, the magnetic field reliability is zero. The fusion weight is determined by the magnetic field reliability: if the magnetic field reliability is not higher than a set lower threshold, the fusion weight is zero; if the magnetic field reliability is not lower than a set upper threshold, the fusion weight is 1; if the magnetic field reliability is between the lower threshold and the upper threshold, the difference between the magnetic field reliability and the lower threshold is multiplied by the difference between the upper threshold and the lower threshold. The ratio of the difference between the lower limit thresholds is used as the fusion weight; the measured azimuth angle is weighted by the fusion weight, and the true azimuth angle is weighted by the difference between 1 and the fusion weight. The two are weighted and summed to obtain the fusion azimuth angle, and the difference between the fusion azimuth angle and the measured azimuth angle is used to replace the heading deviation to construct the heading correction matrix; when both the objects on the horizontal plane and the objects on the vertical plane are identified, the correction rotation matrix is the product of the heading correction matrix and the tilt correction matrix, and the tilt correction matrix is applied to the three-dimensional point cloud before the heading correction matrix.
[0017] Preferably, extracting the point set of the target structural surface includes: receiving a point selection operation applied to the region where the target structural surface is located on the single image; using the 3D point corresponding to the point selection operation as a seed point; using the normal vector of the plane fitted by the neighborhood point set of each point in the 3D point cloud as the local normal vector of the point; using the color of the pixel corresponding to the point in the single image as the color vector of the point; performing region growing starting from the seed point, such that the angle between the local normal vector and the local normal vector of the seed point is less than a set normal angle threshold, and the angle between the color vector and the color vector of the seed point is small. The neighboring points of the set color angle threshold are incorporated into the point set; the calculation of the attitude includes: taking the inverse cosine of the ratio of the absolute value of the celestial component of the plane normal vector to the magnitude of the plane normal vector as the dip angle; taking the negative celestial component of the plane normal vector and the opposite vector of the plane normal vector as the lower normal vector; taking the eastward component of the lower normal vector as the first independent variable and the northward component as the second independent variable, the arctangent of the four quadrants is normalized to a range of not less than 0 degrees and less than 360 degrees as the dip; and taking the dip minus 90 degrees and normalized to a range of not less than 0 degrees and less than 180 degrees as the orientation.
[0018] In another aspect, the present invention provides a structural surface attitude measurement system, including an image acquisition module, a three-dimensional reconstruction module, a coordinate mapping module, an attitude calculation module, a point cloud orientation module, a structural surface extraction module, and an attitude calculation module.
[0019] The image acquisition module is used to capture a single image containing the target structure surface using a mobile terminal, and simultaneously acquires the acceleration data and magnetic field data of the mobile terminal at the time of capture. The 3D reconstruction module is used to input the single image into the feedforward 3D reconstruction model and infer and output a 3D point cloud located in the image coordinate system. The absolute physical scale of the 3D point cloud is unknown. The coordinate mapping module is used to transform the three-dimensional point cloud to the device coordinate system according to the mapping relationship between the image coordinate system and the device coordinate system of the mobile terminal. The mapping relationship is an orthogonal mapping matrix corresponding to the holding orientation of the mobile terminal when capturing the single image. The orthogonal mapping matrix represents the correspondence between the coordinate axes of the image coordinate system and the coordinate axes of the device coordinate system. The attitude calculation module is used to obtain a gravity vector from the acceleration data, a geomagnetic vector from the magnetic field data, and construct a rotation matrix from the device coordinate system to the world coordinate system using only the gravity vector and the geomagnetic vector, wherein the world coordinate system uses the east, north, and sky directions as coordinate axes. The point cloud orientation module is used to apply a rotation transformation to the three-dimensional point cloud located in the device coordinate system using the rotation matrix to obtain a three-dimensional point cloud located in the world coordinate system; the structure surface extraction module is used to extract the point set of the target structure surface from the three-dimensional point cloud located in the world coordinate system, and perform plane fitting on the point set to obtain a plane normal vector. The attitude calculation module is used to calculate the attitude of the target structural surface from the plane normal vector. The attitude includes at least one of strike, dip, and dip angle. The absolute physical scale recovery of the three-dimensional point cloud is not performed throughout the process.
[0020] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-described method for measuring the orientation of structural planes. The computer-readable storage medium includes a read-only memory, a random access memory, flash memory, a magnetic disk, or an optical disk.
[0021] The present invention also provides an electronic device, including a camera, an accelerometer, a magnetometer, a processor, and a memory. The memory stores a computer program, which, when executed by the processor, implements the aforementioned structural plane attitude measurement method. The electronic device constitutes the mobile terminal in the structural plane attitude measurement method. The electronic device is a smartphone or a tablet computer.
[0022] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention combines a feedforward 3D reconstruction model with the accelerometer and magnetometer built into a mobile terminal, enabling the acquisition of structural surface attitude without contact with the target structural surface throughout the entire operation process, and allowing results to be obtained directly on-site. Compared to traditional measurement methods that require personnel to approach and align a compass with the structural surface, this invention eliminates the need for personnel to enter steep slopes, unstable rock masses, tunnel arches, and goaf roofs, reducing the safety risks of working at height. Compared to 3D laser scanning, photogrammetry, and existing mobile imaging solutions, this invention requires no specialized equipment, no targets or markers on the rock surface, no external measured benchmarks for spatial orientation, and no post-processing steps such as image matching, bundle adjustment, or multi-view stereo matching. Only a single image captured by the mobile terminal and synchronously acquired sensor data are used to complete all calculations from image acquisition to attitude output through a single forward inference. The operation is simplified to a single person taking a single shot, reducing equipment purchase and maintenance costs, as well as the time required for field data acquisition and internal processing, facilitating rapid on-site surveys.
[0023] 2. This invention utilizes the property that attitude remains unchanged as an angular quantity under proportional scaling of 3D point clouds. It avoids absolute physical scale restoration of the point cloud output from the feedforward 3D reconstruction model throughout the entire process. Without introducing reference objects, it can still correctly calculate the direction of the plane normal vector. Such models lack absolute physical scale limitations and therefore do not affect the attitude calculation, while also eliminating the need for scale restoration. Furthermore, before attitude calculation, this invention evaluates and weights the magnetic field, acceleration, and attitude stability to obtain a comprehensive reliability score. This score determines the usability of the acquired data and prompts for re-acquisition if necessary, preventing gross errors from entering subsequent calculations. For magnetic interference sources such as steel arches, anchor bolts, and magnetic mineral rock masses in engineering sites, this invention corrects the pitch and roll components using horizontal plane objects obtained through semantic recognition, and corrects the heading component by weighting and fusing the true and measured azimuth angles of vertical plane objects according to magnetic field reliability. This reduces the deviation in heading calculation under magnetic interference environments, ensuring that the obtained attitude remains consistent with the geodetic coordinate system in such sites. Attached Figure Description
[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art 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.
[0025] Figure 1 This is a flowchart of the structural plane orientation measurement method of the present invention; Figure 2 This is a schematic diagram illustrating the definition of the image coordinate system, device coordinate system, and world coordinate system involved in this invention, as well as the principle of point cloud spatial relocation. Figure 3 This is a geometric diagram of the orientation correction of the present invention, wherein (a) is tilt correction based on the horizontal plane, (b) is heading correction based on the vertical plane, and (c) is series correction when both are identified. Figure 4 This is a block diagram of the structural surface attitude measurement system of the present invention. Detailed Implementation
[0026] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.
[0027] The structural plane attitude measurement method and system provided by this invention are applicable to the acquisition of structural plane attitude and rock mass quality evaluation in areas difficult for personnel to access, such as steep slopes, deep foundation pit slopes, unstable rock masses and landslides, reservoir drawdown zone slopes, tunnel arches, tunnel faces, and goaf roofs in water conservancy, hydropower, transportation, mining, and municipal engineering projects. The spatial attitude of the structural plane is the basis for surrounding rock stability classification, slope slip mode identification, and support scheme formulation. The acquisition of attitude involves three stages: spatial geometric information acquisition, attitude spatial orientation, and angle calculation. Among these, the accuracy of image 3D reconstruction, the reliability of the calculation of the heading component in the equipment attitude, and the numerical stability of point cloud plane fitting are interrelated. Errors in any stage will be transmitted to the final strike, dip, and dip angle. Traditional compass measurements require surveyors to approach the target structural surface, where personnel safety is difficult to guarantee. 3D laser scanning and photogrammetry rely on specialized equipment and post-processing software, resulting in high field acquisition costs and long processing times, making it difficult to obtain data on-site in real time. While existing mobile phone imaging solutions introduce 3D image reconstruction, they still require setting up targets and contacting the measurement reference surface, and their non-contact nature does not cover the entire workflow. Attitude calculations based on magnetic sensors suffer from significant heading deviations in environments with magnetic interference, such as steel arches, anchor bolts, and magnetic mineral rock masses. Furthermore, there is a lack of pre-assessment of the quality of acquired data and methods for correcting heading deviations, making it difficult to assess the reliability of the obtained heading information.
[0028] This invention utilizes a mobile terminal's built-in camera, accelerometer, and magnetometer as its hardware foundation. It obtains a locally dense point cloud of structural surfaces from a single image using a feedforward 3D reconstruction model. A rotation matrix from the device coordinate system to the world coordinate system is constructed from acceleration and magnetic field data to achieve overall point cloud orientation. Furthermore, leveraging the scale invariance of angle quantities under proportional scaling of the point cloud, the absolute physical scale is not restored throughout the process. On this basis, a pre-evaluation of sensor reliability and scene object orientation correction based on semantic recognition are added to improve operational safety and data reliability in complex engineering environments. The technical solution of this invention will be further described below with reference to specific embodiments.
[0029] In this embodiment of the invention, a single image is captured by the built-in camera of the mobile terminal, and data from the accelerometer and magnetometer are read simultaneously. A feedforward 3D reconstruction model is deployed on the mobile terminal inference framework using Python or C++, and the mobile terminal inference framework can be TensorFlow Lite or Core ML. Combined with the data from the built-in sensors of the terminal, spatial orientation from the image coordinate system to the world coordinate system is completed through quaternion or rotation matrix operations, and all calculations from image acquisition to state output are completed locally on the mobile terminal. No target deployment, special equipment, or restoration of absolute physical scale are required, and it has the ability to pre-identify magnetic interference and semantic auxiliary correction.
[0030] This embodiment uses the measurement of the structural surface attitude of a hydraulic tunnel face as an example to fully illustrate the implementation process of the method of the present invention. The overall process of the method is as follows: Figure 1 As shown.
[0031] The hardware platform is an intelligent mobile terminal with a built-in rear multi-camera system and a high-precision inertial measurement unit. Its image sensor has a resolution of 3072 pixels by 4096 pixels and is equipped with a three-axis accelerometer and a three-axis magnetometer.
[0032] S1 image acquisition and sensor synchronous recording: The operator held a mobile terminal approximately 3 meters above the working face and took a single, vertical photograph of the area containing the target structure. To ensure that the image and sensor data strictly corresponded to the terminal's pose at the same moment of shooting, the mobile terminal timestamped and aligned the image, acceleration, and magnetic field data using the same system clock at the same instant the shutter was triggered, with a time deviation of no more than 5 milliseconds. The data acquired in this study were: image resolution 3072 pixels by 4096 pixels; three-axis acceleration (0.2, -9.7, -1.3); three-axis magnetic field (16.4, -40.3, -0.2).
[0033] S2 sensor reliability evaluation: Before inputting the images into the reconstruction model, the reliability of the sensor data collected in this study is evaluated.
[0034] The first item is the magnetic field reliability. The magnitude of the magnetic field data was calculated to be 43.51 microtesla. The test site was in Hubei Province. Using the latitude and longitude of the shooting location, the international geomagnetic reference field model was consulted, and the reference range for the local total geomagnetic intensity was found to be 40 to 50 microtesla. The magnetic field reliability was determined by the following formula:
[0035] In the formula, B represents the reliability of the magnetic field data, and B represents the magnitude of the magnetic field data. and These are the lower and upper limits of the reference range, respectively. To account for the magnetic field deviation tolerance, a value of 15 microtesla is used. In this case, B is 43.51, falling within the reference range, therefore the deviation is zero. It equals 1.
[0036] The second item is the acceleration reliability. The magnitude of the acceleration data is calculated to be 9.789 meters per second squared. The acceleration reliability is determined by the following formula:
[0037] In the formula, To assess the reliability of the acceleration, A represents the acceleration data. The acceleration due to gravity is taken as 9.80 meters per second squared. To account for acceleration deviation from the tolerance, we take 1.0 m / s². Substituting this into the equation, we get... It equals 0.989.
[0038] The third item is the attitude stability reliability. Within 2 seconds of the operator entering the camera interface and clicking the shutter, the mobile terminal continuously acquires N frames of attitude rotation matrices; in this embodiment, N equals 3. First, the arithmetic mean of these N frames of rotation matrices is calculated and orthogonalized to obtain the average rotation matrix. Then, the relative rotation matrix of each frame's rotation matrix with respect to the average rotation matrix is used as the attitude deviation matrix. Finally, the matrix distance between each attitude deviation matrix and the identity matrix is calculated.
[0039] In the formula, Let be the matrix distance of the i-th frame. Let I be the attitude deviation matrix of the i-th frame, where I is a third-order identity matrix and the subscript F denotes the Frobenius norm. The matrix distance can also be expressed as the spectral norm or the maximum element difference; this embodiment uses the Frobenius norm. The matrix distances for the three frames are 0, 0.0023, and 0.0023, respectively, with a mean of 0.00153. The attitude stability reliability is determined by the following formula:
[0040] In the formula, To represent the attitude stability reliability, d is the mean of the matrix distance across frames. The maximum permissible attitude change is set to 0.05. Substituting this into the equation yields... It equals 0.969.
[0041] The fourth item is the overall credibility, which is obtained by weighted summation of the above three items:
[0042] In the formula, Q represents the overall credibility. , , The weights for magnetic field reliability, acceleration reliability, and attitude stability reliability are respectively assigned, and the sum of the three is always equal to 1. The weight values are determined based on the degree of influence of each item on the final attitude accuracy: magnetic field data only affects the heading component, which directly determines the dip and azimuth, and magnetic interference is most common in engineering sites, therefore it has the highest weight; attitude stability affects all three attitude components, so its weight is second; acceleration data is extremely stable when the terminal is stationary, so its weight is the lowest. This embodiment uses... It equals 0.5. Equals 0.2, It equals 0.3. Substituting this into the equation, we get Q equals 0.9886.
[0043] Set minimum credibility threshold The ideal confidence threshold is 0.60. The value is 0.90. In this case, Q equals 0.9886, which is higher than the ideal confidence threshold, indicating good data quality. The subsequent steps will continue. If Q falls below the minimum confidence threshold of 0.60, a prompt to re-acquire data will be displayed to the user, and the subsequent steps will be terminated. If Q is not lower than 0.60, the subsequent steps will continue. In this case, the weighting of the heading and geometric constraints calculated by the sensor in the subsequent orientation correction will be determined by the magnetic field confidence level. The method described in Example 3 is used for adaptive determination. The lower the value, the greater the weight of the geometric constraint. This time... When the value is 1, the heading is determined entirely by the sensor calculations and no geometric correction is required.
[0044] S3 Single-Image 3D Reconstruction: The image acquired by S1 is input into the feedforward 3D reconstruction model. This embodiment employs a pixel-by-pixel 3D point map regression approach, meaning the model directly regresses a 3D coordinate for each pixel in the image. The model simultaneously outputs the confidence score for each point. A confidence score threshold of 0.4 is set, and points with confidence scores below this threshold are removed, resulting in a 3D point cloud of approximately 240,000 points. The confidence score threshold can be set between 0.3 and 0.6: a lower threshold retains more points but increases noise in edge regions, while a higher threshold reduces the number of points but increases reliability. Comparison shows that at 0.3, the point cloud contains approximately 310,000 points with a plane fitting residual of 0.008; at 0.4, the point cloud contains approximately 240,000 points with a residual of 0.005; and at 0.6, the point cloud contains approximately 150,000 points with a residual of 0.005, but holes appear at the edges of the structural surfaces. Therefore, this embodiment uses 0.4.
[0045] At this point, the point cloud is located in the image coordinate system, which is defined as follows: the X-axis extends horizontally to the right along the image, the Y-axis extends vertically downwards along the image, and the Z-axis is perpendicular to the image plane and points inwards. Figure 2 As shown. The absolute physical scale of the point cloud is unknown, and no scale recovery is performed on it in this embodiment. Taking a point located on the target structure surface in the point cloud as an example, denoted as point A, its coordinates in the image coordinate system are (0.064, -0.201, 0.090).
[0046] S4 point cloud mapping from image coordinate system to device coordinate system: The device coordinate system is defined as follows: the X-axis points to the right side of the mobile terminal body, the Y-axis points to the top of the body, and the Z-axis is perpendicular to the screen and points outwards. Figure 2 As shown.
[0047] Since the user's holding posture cannot be restricted in actual operation, and the image after automatic system rotation always points to the right on the X-axis and downwards on the Y-axis, it is necessary to first determine the holding orientation and then select the corresponding mapping matrix. The determination method is as follows: take the two components of acceleration data in the screen plane as gravity projection, and use the direction pointing to the top of the camera body in the screen plane as the set reference direction to calculate the rotation angle.
[0048] In the formula, φ is the rotation angle. and These are the components of the acceleration data on the X and Y axes of the device coordinate system, respectively. The output range of the four-quadrant arctangent is greater than -180 degrees and not greater than 180 degrees.
[0049] Based on the angle interval to which φ belongs, select one of the four mapping matrices. The correspondence between the four angle intervals and the four aircraft orientations is as follows: when φ is between -45 degrees and 45 degrees, the top of the fuselage is vertically upward; when φ is between 45 degrees and 135 degrees, the top of the fuselage is horizontally to the right; when φ is between 135 degrees and 180 degrees or between -180 degrees and -135 degrees, the top of the fuselage is vertically downward; when φ is between -135 degrees and -45 degrees, the top of the fuselage is horizontally to the left. The boundary lines of each interval lie on the angle bisectors of the rotation angles corresponding to the two adjacent aircraft orientations. The four mapping matrices are as follows:
[0050] In the formula, , , , These are the mapping matrices corresponding to the four hand-held orientations mentioned above. The determinant of each of the four is 1, and they are all orthogonal rotation matrices.
[0051] In this embodiment, Equals 0.2, Equals -9.7, substituting gives Equal to -1.18 degrees, falling within the range of -45 degrees to 45 degrees, it is determined that the top of the fuselage is vertically upward, and is selected. As a mapping relationship, after transforming the point cloud from the image coordinate system to the device coordinate system, the coordinates of point A are (0.064, 0.201, -0.090).
[0052] It should be noted that the orientation of the device determined here is only used to roughly determine the correspondence between the image axis and the body axis, and does not require the terminal to be strictly vertical or horizontal; any tilt or deflection within a certain range will be fully compensated in the rotation matrix of S5.
[0053] As an alternative implementation of this step, when the operating procedures require the user to shoot in portrait mode, the determination of the camera's orientation can be omitted, and the factory-preset fixed mapping matrix can be used directly. As a mapping relationship, it is not necessary to read the acceleration data components in the screen plane at this time. Implementation details of this alternative method are given in Example 4.
[0054] Orientation of S5 point cloud from device coordinate system to world coordinate system: The world coordinate system is defined as follows: the X-axis points due east, the Y-axis points due north, and the Z-axis points vertically upwards, as shown below. Figure 2 As shown.
[0055] First, normalize the acceleration data to obtain the gravity vector, resulting in (0.020, -0.991, -0.133); then normalize the magnetic field data to obtain the geomagnetic vector, resulting in (0.377, -0.926, -0.005). Since the geomagnetic vector points north but is not horizontal due to magnetic inclination, it cannot be directly used as the north vector. An orthogonal basis must be constructed through two cross products: the cross product of the gravity vector and the geomagnetic vector, normalized, becomes the east vector; the cross product of the east vector and the gravity vector, normalized, becomes the north vector. In this case, the north vector must lie within the horizontal plane. The opposite direction of the gravity vector is used as the celestial vector. The calculation process and results are as follows:
[0056] In the formula, g is the gravity vector, m is the geomagnetic vector, e, n, and u are the east vector, north vector, and celestial vector, respectively, and R is the rotation matrix from the device coordinate system to the world coordinate system. All vectors are represented in the device coordinate system. Substituting these values in this embodiment, we obtain the east vector as (-0.314, -0.132, 0.940), the north vector as (0.949, -0.023, 0.314), and the celestial vector as (-0.020, 0.991, 0.133). Verification shows that these three vectors are pairwise orthogonal and all have a magnitude of 1, and the resulting matrix satisfies orthogonality.
[0057] Apply the rotation transformation to the point cloud in the device coordinate system using this rotation matrix to obtain the point cloud in the world coordinate system. The world coordinate system coordinates of point A are (-0.131, 0.028, 0.186).
[0058] S6 structural surface extraction: First, calculate the local normal vector for each point in the world coordinate system point cloud: take the k nearest neighbors of the point to form a neighborhood point set. In this embodiment, k is set to 30. Perform plane fitting on the neighborhood point set, and use the normal vector of the fitted plane as the local normal vector of the point. Equivalently, principal component analysis can be performed on the covariance matrix of the neighborhood point set, and the eigenvector corresponding to the smallest eigenvalue can be taken. At the same time, according to the correspondence between pixels and 3D points established by the feedforward 3D reconstruction model, the red, green, and blue channel values of the corresponding pixels in the image are used as the color vector of the point.
[0059] The operator clicks on the target structural surface area on the mobile terminal screen. The system receives this selection and uses the corresponding 3D point as the seed point. In this embodiment, the seed point is the aforementioned point A. Region growth is performed starting from the seed point, incorporating neighboring points that simultaneously meet the following two conditions into the structural surface point set: the angle between the local normal vector of the neighboring point and the local normal vector of the seed point is less than a set normal angle threshold; the angle between the color vector of the neighboring point and the color vector of the seed point is less than a set color angle threshold. In this embodiment, both the normal angle threshold and the color angle threshold are set to 10 degrees. Both thresholds can be between 5 and 15 degrees: a value of 5 degrees results in conservative growth, producing a pure point set but potentially undersegmented; a value of 15 degrees results in aggressive growth, potentially crossing structural surface boundaries and incorporating points from adjacent surfaces. In this embodiment, a value of 10 degrees yields a structural surface point set of 841 points.
[0060] A plane fit is performed on the point set. In this embodiment, outliers are first removed using a random sampling consensus algorithm, and then the interior points are refined using the overall least squares method to obtain the plane equation and the plane normal vector:
[0061] In the formula, A, B, C, and D are the coefficients of the plane equation, and n is the plane normal vector. The plane equation obtained in this embodiment is 0.510x minus 0.750y minus 0.420z plus 0.166 equals 0, and the plane normal vector is (0.510, -0.750, -0.420), with a magnitude of 0.9995.
[0062] S7 Occurrence Calculation The angle of inclination is calculated using the following formula:
[0063] In the formula, α is the tilt angle, C is the celestial component of the plane normal vector, and the symbol ‖n‖ represents the magnitude of the plane normal vector. Substituting these values, we get α equal to 65.15 degrees, but we take 65.2 degrees.
[0064] The tendency and trend are calculated using the following formula:
[0065] In the formula, β represents the tendency and γ represents the direction. and These are the eastward and northward components of the lower normal vector, respectively. The lower normal vector is the negative of the celestial component of the plane normal vector and its opposite vector. In this embodiment, the celestial component of the plane normal vector is -0.420, which is less than zero, so it is itself the lower normal vector. Substituting these values, we get the dip β equals 145.8 degrees and the strike γ equals 55.8 degrees.
[0066] The final output for the attitude of this structural surface is: strike 55.8 degrees, dip 145.8 degrees, and dip angle 65.2 degrees. For comparison, the same structural surface, measured with a geological compass at a personnel-accessible location, yields a strike of 54 degrees, a dip of 144 degrees, and a dip angle of 63 degrees. The deviations between this method and the compass measurement are 1.8 degrees, 1.8 degrees, and 2.2 degrees, respectively.
[0067] Example 2 This embodiment takes the structural surface attitude measurement of a steep highway slope as an example to illustrate the second implementation route of the feedforward three-dimensional reconstruction model, the second case of machine orientation, and the tilt correction based on the horizontal plane.
[0068] The operating conditions in this embodiment are: the slope is about 30 meters high, the workers are located on the platform at the foot of the slope, and cannot reach the target structural surface in the middle and upper part of the slope.
[0069] Step S1: The operator takes a single image in landscape orientation with the top of the camera facing right. The synchronously acquired three-axis accelerations are (-9.72, 0.35, -1.15) and the three-axis magnetic fields are (41.8, 8.6, -9.4).
[0070] Step S2: The magnetic field modulus is 43.72 microtesla, falling within the local reference range of 40 to 50 microtesla. The magnetic field confidence level is... The value equals 1; the acceleration modulus is 9.80 m / s², consistent with gravitational acceleration, indicating high reliability of the acceleration. The value equals 1; before shooting, three consecutive frames of attitude rotation matrix were acquired, and the average distance between the matrices in each frame was 0.0031, indicating the attitude stability reliability. It equals 0.938. (Press...) Take 0.5, Take 0.2, With a weighting of 0.3, the overall credibility Q equals 0.981, which is higher than the ideal credibility threshold of 0.90, indicating good data quality.
[0071] Step S3: Input the image into the dense depth map regression model. The model regresses a depth value for each pixel and outputs a dense depth map. Then, according to the camera intrinsics of the mobile terminal, the depth map is back-projected into a 3D point cloud. The camera intrinsics include focal length and principal point coordinates, which are read from the metadata of the image's exchangeable image file format or pre-calibrated by a checkerboard calibration board. In this embodiment, the equivalent pixel focal length corresponding to the focal length is 3120 pixels, and the principal point is located at the center of the image. The model also outputs pixel-wise confidence. The confidence threshold is set to 0.5. After removing low-confidence points, a 3D point cloud of approximately 190,000 points is obtained. The point cloud obtained by this route also lacks absolute physical scale, and scale restoration is not performed in this embodiment.
[0072] Step S4: The acceleration data components on the X and Y axes of the device coordinate system are -9.72 and 0.35 respectively. Substituting these into the rotation angle calculation formula, φ equals 88.0 degrees, falling within the range of 45 to 135 degrees. Therefore, the top of the fuselage is determined to be horizontally facing right. As a mapping relationship.
[0073] Step S5: The gravity vector is (-0.992, 0.036, -0.117), and the geomagnetic vector is (0.956, 0.197, -0.215). Calculate the rotation matrix from the device coordinate system to the world coordinate system according to the construction order of the east vector, north vector, and celestial vector, and use it to orient the point cloud to the world coordinate system.
[0074] Orientation Correction: In this embodiment, the lower part of the image includes a concrete platform at the foot of a slope. Semantic recognition is performed on the image using a cue-guided segmentation model based on sample images. This model identifies the platform region and outputs a pixel-level binary mask without requiring network retraining. Semantic recognition can also be implemented using a convolutional neural network semantic segmentation model, a segmentation model based on a visual transformer, or an object detection model.
[0075] Based on the correspondence between pixels and 3D points, the mask is mapped to a set of reference points in the world coordinate system. A plane is fitted using a random sampling consensus algorithm, yielding a reference plane normal vector of (0.031, -0.048, 0.998). The semantic recognition result for this object is a horizontal plane, and its known spatial pose in the world coordinate system is that the normal vector is along the vertical direction, meaning the ideal value should be (0, 0, 1). The measured upward component of the normal vector is positive 0.998, which already satisfies the sign disambiguation condition and does not require inversion.
[0076] The tilt correction matrix is constructed according to the Rodriguez formula, such as... Figure 3 As shown in (a):
[0077] In the formula, Let K be the tilt correction matrix, k be the unit vector of the rotation axis (taken as the cross product of the reference plane normal vector and the yaw unit vector and normalized), θ be the rotation angle (taken as the angle between the reference plane normal vector and the yaw unit vector), and K be an antisymmetric matrix composed of k. In this embodiment, substituting these values, we obtain the rotation axis k as (-0.840, -0.543, 0) and the rotation angle θ as 3.28 degrees. Since the rotation axis is located in the horizontal plane, this matrix only changes the pitch and roll components without changing the yaw component.
[0078] A second rotation transformation is applied to the world coordinate system point cloud using the tilt correction matrix. After correction, the normal vector of the reference plane becomes (0, 0, 1), meaning the horizontal plane returns to horizontal. The corrected point cloud replaces the original point cloud for subsequent steps.
[0079] Steps S6 and S7: The operator clicks on the target structural surface on the slope, and extracts a set of 623 structural surface points using region growing. The normal angle threshold is set to 8 degrees, and the color angle threshold is set to 12 degrees. Plane fitting yields a plane normal vector of (-0.286, 0.638, -0.715), with a magnitude of 1.000. The dip angle is 44.4 degrees. The celestial component of this normal vector is -0.715, which is less than zero, indicating it is the lower normal vector. The dip is -24.2 degrees, which, after normalization, becomes 335.8 degrees. The strike is 245.8 degrees, which, after normalization to a range above 0 degrees and below 180 degrees, becomes 65.8 degrees. The final output is a strike of 65.8 degrees, a dip of 335.8 degrees, and a dip angle of 44.4 degrees.
[0080] Compared with the attitude before tilt correction, the tilt angle changed by 3.2 degrees and the dip changed by 0.5 degrees after tilt correction. This indicates that when there is a slight pitch deviation at the terminal, the semantic horizontal plane can effectively correct the pitch and roll components, with the correction of the tilt angle being more obvious.
[0081] Example 3 This embodiment takes the measurement of the structural surface orientation of a tunnel with strong magnetic interference as an example to illustrate the heading correction based on the vertical plane, the series correction when both the horizontal and vertical planes are identified, the adaptive weighted fusion based on the reliability of the magnetic field, the gross error elimination of multiple vertical planes, and the data validity verification.
[0082] The operating conditions for this embodiment are as follows: the tunnel has been equipped with a steel arch frame and system anchor bolts, and there is significant magnetic interference; the bottom of the tunnel is a hardened concrete inverted arch surface, and both sides of the tunnel wall and the tunnel face are within the field of view.
[0083] Steps S1 and S2: Capture a single image in portrait orientation and simultaneously acquire sensor data. The magnetic field modulus is 57.3 μT, exceeding the upper limit of the local reference range of 40 to 50 μT by 7.3 μT. Based on the magnetic field reliability calculation formula, assuming a magnetic field deviation from the tolerance of 15 μT... It equals 1 minus 7.3 divided by 15, which gives 0.51.
[0084] The magnetic field confidence level is 0.51. The acceleration modulus is 9.81 m / s². Equal to 0.99; Attitude stability reliability The overall confidence level Q is 0.5 x 0.51 + 0.2 x 0.99 + 0.3 x 0.95, which equals 0.738. This is higher than the minimum confidence threshold of 0.60, so we continue with the subsequent steps. Since the confidence level of the magnetic field is only 0.51, geometric constraints will play a major role in the subsequent orientation correction.
[0085] Steps S3 to S5: Same as in Example 1, a 3D point cloud of approximately 210,000 points is reconstructed and oriented to the world coordinate system through a two-stage transformation using a mapping matrix and a rotation matrix. Due to the interference of the magnetic field by the steel arch frame, the heading component of the point cloud is deviated at this time, while the pitch and roll components are unaffected because they are determined only by the gravity vector.
[0086] Orientation correction: Data validity verification.
[0087] Semantic recognition is performed on the image to obtain pixel masks of four objects with known spatial poses: the tunnel face, the left tunnel wall, the right tunnel wall, and the tunnel floor. First, the validity of the tunnel face is verified: its mask is mapped to a reference point set and fitted to a plane, resulting in the plane equation 0.9680x - 0.2510y + 0.0018z - 0.097 = 0. The reference plane normal vector is (0.9680, -0.2510, 0.0018). Since the tunnel face is generally vertical, its normal vector should be perpendicular to the direction of gravity; therefore, the reference value for the angle between this normal vector and the vertical direction is 90 degrees. The measured angle is calculated to be 89.90 degrees. The absolute value of the difference from the reference value is 0.10 degrees, which is less than the set verification threshold of 10 degrees, thus the data acquisition and reconstruction result is deemed valid. The set verification threshold can be between 5 and 15 degrees. If the difference is greater than the verification threshold, the acquisition is deemed invalid, and a prompt to re-acquire is output.
[0088] Orientation correction: tilt correction based on the horizontal plane.
[0089] The reference point set obtained by mask mapping at the bottom of the hole is fitted with a plane, and the normal vector of the reference plane is (-0.012, 0.019, 0.9997), and the semantic recognition result is a horizontal plane. The tilt correction matrix is constructed in the manner described in Example 2, and the rotation axis is (0.846, 0.533, 0), and the rotation angle is 1.29 degrees.
[0090] Orientation correction: Heading correction based on the vertical plane.
[0091] The semantic recognition result of the tunnel face is a vertical plane, with its normal vector along the tunnel excavation axis. The true azimuth of this vertical plane in the world coordinate system is determined by engineering survey data: in this embodiment, the design azimuth of the tunnel centerline for this section is found to be 88.50 degrees from the tunnel construction design drawings, meaning the true azimuth of the tunnel face normal vector is 88.50 degrees. The true azimuth can also be obtained from the as-built survey results table, the breakthrough survey record, or by measuring the tunnel axis direction from satellite imagery; for building walls and retaining walls, it can be measured from the architectural design master plan or satellite imagery.
[0092] The projection of the reference plane normal vector onto the horizontal plane is taken as (0.9680, -0.2510). Using its eastward component as the first independent variable and its northward component as the second independent variable, the arctangent is calculated in four quadrants, yielding a measured azimuth of 104.54 degrees. The geometric relationship between the measured azimuth, true azimuth, and heading deviation is as follows: Figure 3 As shown in (b).
[0093] Since the reliability of the magnetic field is only 0.51, directly using the difference between the true azimuth and the measured azimuth as the heading deviation would introduce all errors of the geometric constraints; if the sensor heading is completely trusted without correction, the magnetic interference error cannot be eliminated. Therefore, the weights of the two are adaptively allocated according to the reliability of the magnetic field. The fusion weights are determined by the following formula:
[0094] In the formula, w is the fusion weight. For the credibility of the magnetic field, and These are the lower and upper thresholds for magnetic field reliability, respectively, set to 0.50 and 0.90 in this embodiment. The physical meaning of this mapping is as follows: when the magnetic field reliability is not higher than the lower threshold, the sensor heading is unreliable, the fusion weight is zero, and the heading is completely determined by geometric constraints; when the magnetic field reliability is not lower than the upper threshold, the sensor heading is reliable, the fusion weight is 1, and no geometric correction is performed; when it is between the two, it is a linear transition.
[0095] The reliability of the magnetic field in this embodiment The value is 0.51, only slightly higher than the lower threshold of 0.50. Substituting this into the equation, the fusion weight w equals 0.025, meaning the heading is essentially determined by geometric constraints. The deviation between the fused azimuth and the final heading is calculated using the following formula:
[0096] In the formula, To integrate azimuth angles, For the measured azimuth angle, ΔΨ is the true azimuth, and ΔΨ is the heading deviation. Substituting the measured azimuth of 104.54 degrees and the true azimuth of 88.50 degrees, we get a combined azimuth of 88.901 degrees and a heading deviation of -15.639 degrees.
[0097] The heading correction matrix is the rotation matrix that rotates the heading deviation about the heading axis:
[0098] In the formula, This is the heading correction matrix. Since this matrix is a rotation about the heading axis, it only changes the heading component and not the pitch and roll components.
[0099] Orientation correction: elimination of gross errors from multiple vertical planes.
[0100] In this embodiment, in addition to the tunnel face, two vertical planes, the left and right tunnel walls, are identified within the field of view. Their true azimuths are -1.50 degrees and 178.50 degrees, respectively, meaning they are both parallel to the tunnel axis. The heading deviations for the three vertical planes are calculated as -15.639 degrees, -15.766 degrees, and -23.000 degrees. The median of this set of heading deviations is taken as -15.766 degrees. The absolute values of the differences between each heading deviation and the median are calculated as 0.127 degrees, 0 degrees, and 7.234 degrees, respectively. A discrete threshold of 3 degrees is set, therefore the third heading deviation is considered a gross error and discarded. The discrete threshold can be set between 1 and 5 degrees. Upon verification, the right tunnel wall mask was found to have a deflected fitting plane due to local over-excavation and shotcrete protrusions, a common source of gross errors. The arithmetic mean of the remaining two heading deviations is taken as the statistical center value, resulting in a final heading deviation of -15.702 degrees. A heading correction matrix is constructed using this value. The statistical center value can also be the median of the remaining heading deviations.
[0101] Orientation correction: Series correction.
[0102] This embodiment simultaneously identifies the horizontal plane (bottom of the tunnel) and the vertical plane (face of the tunnel and tunnel wall), therefore the correction rotation matrix is taken as the product of the heading correction matrix and the tilt correction matrix, and the tilt correction matrix is applied to the point cloud before the heading correction matrix, such as... Figure 3 As shown in (c). The reason is that tilt correction corrects the pitch and roll components, while heading correction corrects the heading component. If heading correction is performed first and then tilt correction is performed, the rotation introduced by tilt correction will cause the corrected heading to deviate again. However, after tilt correction is performed first, the sky axis of the point cloud coincides with the true vertical direction. At this time, the heading rotation around the sky axis no longer affects the pitch and roll, and the two levels of correction do not affect each other.
[0103] Steps S6 and S7: Extract the structural surface from the corrected point cloud and calculate the attitude, obtaining a strike of 112.4 degrees, a dip of 22.4 degrees, and a dip angle of 71.3 degrees. As a control, if orientation is directly determined using the rotation matrix calculated by the sensor without azimuth correction, the resulting dip is 38.5 degrees, differing from the corrected result by 16.1 degrees, and from the measured value of 20.5 degrees obtained using a geological compass during the shutdown window by 18.0 degrees. The azimuth-corrected dip differs from the compass measurement by only 1.9 degrees. This comparison demonstrates that, under strong magnetic interference, azimuth correction based on semantic plane geometric constraints has a significant impact on the accuracy of the dip.
[0104] Example 4 This embodiment illustrates a second method for determining the mapping relationship and other values for several parameters, in order to supplement other embodiments of the present invention.
[0105] Fixed mapping method: When the work procedure explicitly requires shooting in portrait mode, a fixed mapping matrix can be preset during the mobile terminal's factory settings or application initialization, as mentioned above. In the coordinate mapping step, this fixed mapping matrix is directly used, without reading the acceleration data components in the screen plane, calculating the rotation angle, or determining the angle range. The advantage of this method is lower computational load and a simpler process; the disadvantage is that it requires operators to strictly adhere to the agreed-upon machine-holding posture. This implementation method, together with the machine-holding orientation adaptive method described in Example 1, constitutes two ways to determine the mapping relationship, and the two can be switched via application settings.
[0106] Other methods for extracting structural surfaces: Besides region growing, structural surface point sets can also be extracted using normal vector clustering or supervoxel segmentation algorithms. When using normal vector clustering, the local normal vectors of all points are clustered on the orientation sphere, and the points in the cluster to which the seed point belongs are taken as the structural surface point set. When using supervoxel segmentation, the point cloud is first over-segmented into several supervoxels, and then adjacent supervoxels whose angle with the normal vector of the seed point's supervoxel is less than a threshold are merged. When the feedforward 3D reconstruction model does not output color information, or the color difference between the target structural surface and adjacent rock surfaces is not significant, only the local normal vector angle can be used as the growth criterion without using the color criterion. In this case, the normal angle threshold should be tightened to 5 to 8 degrees to suppress over-segmentation.
[0107] Other methods for plane fitting: Besides the random sample consensus algorithm and population least squares, plane fitting can also use ordinary least squares or principal component analysis of the neighborhood covariance matrix. When the surface of the structural plane has large undulations or local spalling exists, the random sample consensus algorithm should be given priority to suppress the influence of gross errors.
[0108] Other forms of matrix distance: The matrix distance between the attitude deviation matrix and the identity matrix, besides the Frobenius norm, can also be expressed as the spectral norm or the maximum element difference. When using the spectral norm, the maximum allowable attitude change should be adjusted to 0.03; when using the maximum element difference, it should be adjusted to 0.02. All three have similar magnitudes under small-angle perturbations and can characterize the dispersion of attitude across multiple frames.
[0109] Other values for weights and thresholds: In addition to the weights of 0.5, 0.2, and 0.3 used in this embodiment, the three weights for overall credibility can be adjusted to 0.3, 0.2, and 0.5 in environments with weak magnetic interference but strong vibration (such as near blasting faces) to improve sensitivity to camera shake; the sum of the three is always equal to 1. The minimum credibility threshold can be between 0.50 and 0.70, and the ideal credibility threshold can be between 0.85 and 0.95.
[0110] Other forms of attitude output: In addition to outputting strike, dip and dip angle simultaneously, it is also possible to output only dip and dip angle according to engineering practice, or only strike and dip angle; the three satisfy the relationship that dip is equal to strike plus 90 degrees, so any two can completely characterize the attitude.
[0111] Example 5 This embodiment provides a structural plane attitude measurement system to illustrate its composition, the hardware and software support and deployment locations of each module, and the data flow between modules. This system is used to execute the structural plane attitude measurement method described in any of Embodiments 1 to 4, and its composition is as follows: Figure 4 As shown.
[0112] System Overall Loading and Deployment: In this embodiment, the image acquisition module, 3D reconstruction module, coordinate mapping module, attitude calculation module, point cloud orientation module, structural surface extraction module, and attitude calculation module are all deployed locally on the same mobile terminal as computer program modules. The mobile terminal is a smartphone or tablet computer, and its hardware includes a camera, accelerometer, magnetometer, processor, memory, and touchscreen. The processor may further include a neural network acceleration unit for accelerating neural network inference. The executable code of the above seven modules is stored in the memory as an application program and is loaded and executed by the processor; data between modules is transferred via in-process memory objects, without network transmission. Therefore, all calculations from image acquisition to attitude output are completed locally on the mobile terminal, without the need for a network connection, making it suitable for operating environments without public communication network signals, such as tunnels, deep foundation pits, and mountain slopes.
[0113] The image acquisition module is jointly implemented by the camera, the accelerometer, the magnetometer, and the acquisition control program running on the processor. When the shutter is triggered, this module timestamps and aligns the image data, acceleration data, and magnetic field data using the same system clock, with a time deviation of no more than 5 milliseconds, as described in S1. The module's output is divided into two paths: a single image is output to the 3D reconstruction module, and the acceleration and magnetic field data are output to the attitude calculation module.
[0114] The 3D reconstruction module is implemented by a mobile inference framework and its loaded feedforward 3D reconstruction model, running on the processor or the neural network acceleration unit. The mobile inference framework can use TensorFlow Lite or Core ML, and the model file is stored in the memory along with the application. This module receives a single image output by the image acquisition module, performs one forward inference, outputs a 3D point cloud in the image coordinate system and the confidence score of each point, removes points with confidence scores below a set confidence threshold, and then outputs the 3D point cloud to the coordinate mapping module, as shown in S3.
[0115] The coordinate mapping module is implemented by a coordinate transformation program running on the processor. It receives the 3D point cloud output by the 3D reconstruction module, transforms the 3D point cloud to the device coordinate system according to a mapping matrix corresponding to the orientation of the camera during shooting, and then outputs it to the point cloud orientation module. The process is described in S4. In this embodiment, the mapping matrix is pre-stored in the memory according to the fixed mapping method described in Embodiment 4. Therefore, this module does not require any other inputs besides the 3D point cloud.
[0116] The attitude calculation module is implemented by an attitude calculation program running on the processor. Its inputs are only the acceleration and magnetic field data output by the image acquisition module. This module obtains the gravity vector from the acceleration data and the geomagnetic vector from the magnetic field data. After two cross products and normalization, it obtains the east, north, and celestial vectors, which are then used to construct a rotation matrix from the device coordinate system to the world coordinate system. This rotation matrix is then output to the point cloud orientation module, as shown in S5. There is no data dependency between this module and the coordinate mapping module; both can be executed in parallel by the processor.
[0117] The point cloud orientation module is implemented by a rotation transformation program running on the processor. It has two inputs: one is the 3D point cloud located in the device coordinate system output by the coordinate mapping module, and the other is the rotation matrix output by the attitude calculation module. This module applies a rotation transformation to the 3D point cloud using the rotation matrix to obtain a 3D point cloud located in the world coordinate system and outputs it to the structure surface extraction module. The process is shown in S5.
[0118] The structural surface extraction module is implemented by a point cloud processing program running on the processor and the touch screen. This module receives a 3D point cloud in the world coordinate system output by the point cloud orientation module. The touch screen receives point selection operations applied to the region where the target structural surface is located on a single image to determine seed points. Region growth is performed according to the local normal vector and color vector to obtain the point set of the target structural surface. Then, plane fitting is performed on the point set to obtain the plane normal vector and output to the attitude calculation module. The process is shown in S6.
[0119] The attitude calculation module is implemented by the attitude solution program running on the processor and the touch screen. It receives the plane normal vector output by the structural surface extraction module, calculates the direction, dip and dip angle of the target structural surface, and displays it on the touch screen or writes it to the memory for storage. The process is shown in S7.
[0120] As an alternative deployment method in this embodiment, considering the high inference computing power requirements of the feedforward 3D reconstruction model, the 3D reconstruction module can also be deployed on an edge computing node or server connected to the mobile terminal via a wireless communication network. In this case, the image acquisition module uploads the single image to the node via the wireless communication network, and the 3D reconstruction module returns the 3D point cloud to the mobile terminal after completing one forward inference at the node. The remaining six modules are still deployed locally on the mobile terminal. Under this alternative deployment method, the functions of each module and the data flow between modules remain unchanged; only the data transfer method between modules changes from in-process memory object transfer to transmission via a wireless communication network.
[0121] In a preferred embodiment of this invention, the structural plane attitude measurement system further includes a reliability evaluation module and an orientation correction module. The reliability evaluation module is implemented by a reliability evaluation program running on the processor. It is located between the image acquisition module and the 3D reconstruction module in the data stream. Its inputs are the acceleration data, magnetic field data, and multiple attitude rotation matrices continuously acquired before shooting, all output by the image acquisition module. This module calculates the overall reliability as described in S2. When the overall reliability is lower than the minimum reliability threshold, a re-acquisition prompt is output via the touch screen, and the execution of subsequent modules is suspended. The orientation correction module is implemented by a semantic recognition program and a correction calculation program running on the processor or the neural network acceleration unit. It is located between the point cloud orientation module and the structural plane extraction module in the data stream. Its inputs are the 3D point cloud in the world coordinate system output by the point cloud orientation module, a single image output by the image acquisition module, and the correspondence between pixels and 3D points established by the 3D reconstruction module. This module constructs a correction rotation matrix as described in Embodiments 2 and 3, applies a secondary rotation transformation to the 3D point cloud, and then outputs the corrected 3D point cloud to the structural plane extraction module. In this preferred embodiment, the coordinate mapping module can further take the acceleration data output by the image acquisition module as input, and adaptively determine the orientation of the device by the components of the acceleration data in the screen plane as described in Embodiment 1, and select the corresponding mapping matrix.
[0122] Figure 4 Only the seven modules and their main data flows are shown; the optional modules and their data flows, the optional acceleration data input of the coordinate mapping module, and the data flow of the 3D reconstruction module providing the structure surface extraction module and the orientation correction module with the correspondence between pixels and 3D points are not shown in the figure.
[0123] The module division of the system described in this embodiment is merely an exemplary division. Those skilled in the art can merge the above modules or further divide any one of the modules into several sub-modules according to actual needs. As long as the merged or divided modules as a whole achieve the same function and have the same data flow relationship, they fall within the protection scope of this invention.
[0124] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.
Claims
1. A method for measuring the attitude of structural surfaces, characterized in that, include: A single image containing the target structure surface is captured using a mobile terminal, and acceleration and magnetic field data of the mobile terminal are collected simultaneously at the time of capture; The single image is input into a feedforward 3D reconstruction model, and the model outputs a 3D point cloud located in the image coordinate system. The absolute physical scale of the 3D point cloud is unknown. According to the mapping relationship between the image coordinate system and the device coordinate system of the mobile terminal, the three-dimensional point cloud is transformed to the device coordinate system. The mapping relationship is an orthogonal mapping matrix corresponding to the holding orientation of the mobile terminal when capturing the single image. The orthogonal mapping matrix represents the correspondence between the coordinate axes of the image coordinate system and the coordinate axes of the device coordinate system. The gravity vector is obtained from the acceleration data, the geomagnetic vector is obtained from the magnetic field data, and the rotation matrix from the device coordinate system to the world coordinate system is constructed only from the gravity vector and the geomagnetic vector, with the east, north and sky directions as coordinate axes of the world coordinate system. A rotation transformation is applied to the three-dimensional point cloud located in the device coordinate system using the rotation matrix to obtain a three-dimensional point cloud located in the world coordinate system; The point set of the target structural surface is extracted from the three-dimensional point cloud located in the world coordinate system, the point set is fitted with a plane to obtain a plane normal vector, and the attitude of the target structural surface is calculated from the plane normal vector. The attitude includes at least one of strike, dip and dip angle. Throughout the process, the absolute physical scale of the three-dimensional point cloud is not restored.
2. The method for measuring the attitude of structural planes according to claim 1, characterized in that, The feedforward 3D reconstruction model is either a neural network model that directly regresses a pixel-by-pixel 3D point map from a single image through a single forward inference, or a neural network model that directly regresses a dense depth map from a single image through a single forward inference and backprojects the dense depth map into a 3D point cloud according to camera intrinsic parameters; the feedforward 3D reconstruction model outputs the confidence score of each point in the 3D point cloud while outputting the 3D point cloud, and removes points in the 3D point cloud whose confidence score is lower than a set confidence score threshold.
3. The method for measuring the attitude of structural planes according to claim 1, characterized in that, The mapping relationship is determined in the following manner: Using the component of the acceleration data in the screen plane of the mobile terminal as the gravity projection, the rotation angle of the gravity projection relative to the set reference direction in the screen plane is calculated, and according to the angle interval to which the rotation angle belongs, one set is selected from multiple sets of mapping matrices that correspond one-to-one with various holding orientations of the mobile terminal as the mapping relationship. The rotation matrix is constructed as follows: the cross product of the gravity vector and the geomagnetic vector, normalized, is used as the east vector; the cross product of the east vector and the gravity vector, normalized, is used as the north vector located in the horizontal plane; the opposite vector of the gravity vector is used as the celestial vector; and the rotation matrix is composed of the east vector, the north vector, and the celestial vector.
4. The method for measuring the attitude of structural planes according to claim 1, characterized in that, The sensor reliability evaluation is also included before the inference output of the 3D point cloud: The magnetic field deviation is defined as the amount by which the magnitude of the magnetic field data exceeds the reference geomagnetic intensity range of the shooting location. When the magnitude falls within this range, the magnetic field deviation is zero. The magnetic field reliability is defined as the ratio of 1 minus the magnetic field deviation to a set magnetic field deviation tolerance, and the magnetic field reliability is zero when this ratio is greater than 1. Similarly, the acceleration deviation is defined as the absolute value of the difference between the magnitude of the acceleration data and the gravitational acceleration. The acceleration reliability is defined as the ratio of 1 minus the acceleration deviation to a set acceleration deviation tolerance, and the acceleration reliability is zero when this ratio is greater than 1. The average rotation matrix is calculated for multiple attitude rotation matrices continuously collected by the mobile terminal before shooting. The relative rotation matrix of each attitude rotation matrix with respect to the average rotation matrix is used as the attitude deviation matrix. The mean of the matrix distance between each attitude deviation matrix and the identity matrix is used as the attitude discrepancy. The attitude stability reliability is calculated as 1 minus the ratio of the attitude discrepancy to the set maximum allowable attitude change. When the ratio is greater than 1, the attitude stability reliability is zero. The overall credibility is obtained by weighting the magnetic field credibility, the acceleration credibility, and the attitude stability credibility with three weights respectively, and the sum of the three weights is 1. When the overall credibility is lower than the set minimum credibility threshold, a prompt to re-collect data is output and subsequent steps are terminated.
5. The method for measuring the attitude of structural planes according to claim 1, characterized in that, After obtaining the 3D point cloud located in the world coordinate system and before extracting the point set of the target structural surface, orientation correction is also included: Semantic recognition is performed on the single image to obtain the pixel mask corresponding to the scene object with known spatial pose; Based on the correspondence between pixels and 3D points established by the feedforward 3D reconstruction model, the pixel mask is mapped to a set of reference points located in the world coordinate system. A reference plane normal vector is obtained by performing plane fitting on the reference point set; a correction rotation matrix is constructed based on the deviation between the reference plane normal vector and the known spatial pose of the scene object; and a second rotation transformation is applied to the three-dimensional point cloud located in the world coordinate system using the correction rotation matrix, so that the three-dimensional point cloud located in the world coordinate system is replaced by the three-dimensional point cloud after the second rotation transformation. The scene object includes at least one of the semantic recognition result of a horizontal plane and the semantic recognition result of a vertical plane. The known spatial pose of the horizontal plane object is that its normal vector is along the vertical direction. The known spatial pose of the vertical plane object is that its normal vector has a known azimuth angle in the world coordinate system. This known azimuth angle is denoted as the true azimuth angle.
6. The method for measuring the attitude of structural surfaces according to claim 4 or 5, characterized in that, When the scene object includes an object whose semantic recognition result is a horizontal plane, the correction rotation matrix obtained from the object on the horizontal plane is a tilt correction matrix. The tilt correction matrix is constructed by normalizing the cross product of the reference plane normal vector and the celestial unit vector corresponding to the object on the horizontal plane as the rotation axis and the angle between the reference plane normal vector and the celestial unit vector as the rotation angle. The tilt correction matrix only changes the pitch and roll components and does not change the heading component. When the scene object includes an object whose semantic recognition result is a vertical plane, the correction rotation matrix obtained from the vertical plane object is the heading correction matrix: the azimuth angle of the projection of the reference plane normal vector corresponding to the vertical plane object onto the horizontal plane is taken as the measured azimuth angle, and the difference between the true azimuth angle and the measured azimuth angle is taken as the heading deviation. The heading correction matrix is a rotation matrix that rotates the heading deviation around the celestial axis, and the heading correction matrix only changes the heading component without changing the pitch and roll components. Meanwhile, the amount by which the magnitude of the magnetic field data exceeds the reference geomagnetic intensity range of the shooting location is taken as the magnetic field deviation. When the magnitude falls within this range, the magnetic field deviation is zero. The magnetic field reliability is taken as 1 minus the ratio of the magnetic field deviation to the set magnetic field deviation tolerance. When the ratio is greater than 1, the magnetic field reliability is zero. The fusion weight is determined by the magnetic field reliability: when the magnetic field reliability is not higher than the set lower limit threshold of magnetic field reliability, the fusion weight is zero. When the magnetic field credibility is not lower than the set upper limit threshold of magnetic field credibility, the fusion weight is 1; When the magnetic field confidence level is between the lower threshold and the upper threshold, the ratio of the difference between the magnetic field confidence level and the lower threshold to the difference between the upper threshold and the lower threshold is used as the fusion weight. The measured azimuth angle is weighted by the fusion weight, and the true azimuth angle is weighted by the difference between 1 and the fusion weight. The two are weighted and summed to obtain the fusion azimuth angle. The difference between the fusion azimuth angle and the measured azimuth angle is used to replace the heading deviation to construct the heading correction matrix. When both the objects on the horizontal plane and the objects on the vertical plane are identified, the correction rotation matrix is the product of the heading correction matrix and the tilt correction matrix, and the tilt correction matrix is applied to the 3D point cloud before the heading correction matrix.
7. The method for measuring the attitude of structural planes according to claim 1, characterized in that, Extracting the point set of the target structural surface includes: Receive a point selection operation applied to the region where the target structure surface is located on the single image, and use the three-dimensional point corresponding to the point selection operation as a seed point; The normal vector of the plane fitted by the neighborhood point set of each point in the 3D point cloud is used as the local normal vector of that point, and the color of the pixel corresponding to that point in the single image is used as the color vector of that point; a region growing is performed starting from the seed point, and neighborhood points whose angle between the local normal vector and the local normal vector of the seed point is less than a set normal angle threshold, and whose angle between the color vector and the color vector of the seed point is less than a set color angle threshold are incorporated into the point set. The calculation of the attitude includes: using the inverse cosine of the ratio of the absolute value of the celestial component of the plane normal vector to the magnitude of the plane normal vector as the dip angle; using the negative celestial component of the plane normal vector and its opposite vector as the lower normal vector; using the eastward component of the lower normal vector as the first independent variable and the northward component as the second independent variable to calculate the arctangent of the four quadrants and normalize it to a range of not less than 0 degrees and less than 360 degrees as the dip; and subtracting 90 degrees from the dip and normalizing it to a range of not less than 0 degrees and less than 180 degrees as the strike.
8. A structural plane attitude measurement system, characterized in that, include: The image acquisition module is used to capture a single image containing the target structure surface using a mobile terminal, and simultaneously acquire the acceleration data and magnetic field data of the mobile terminal at the time of capture. The three-dimensional reconstruction module is used to input the single image into the feedforward three-dimensional reconstruction model and infer and output a three-dimensional point cloud located in the image coordinate system, wherein the absolute physical scale of the three-dimensional point cloud is unknown. The coordinate mapping module is used to transform the three-dimensional point cloud to the device coordinate system according to the mapping relationship between the image coordinate system and the device coordinate system of the mobile terminal. The mapping relationship is an orthogonal mapping matrix corresponding to the holding orientation of the mobile terminal when capturing the single image. The orthogonal mapping matrix represents the correspondence between the coordinate axes of the image coordinate system and the coordinate axes of the device coordinate system. The attitude calculation module is used to obtain the gravity vector from the acceleration data, obtain the geomagnetic vector from the magnetic field data, and construct the rotation matrix from the device coordinate system to the world coordinate system using only the gravity vector and the geomagnetic vector, wherein the world coordinate system has east, north and sky as coordinate axes; A point cloud orientation module is used to apply a rotation transformation to the three-dimensional point cloud located in the device coordinate system using the rotation matrix to obtain a three-dimensional point cloud located in the world coordinate system. The structural surface extraction module is used to extract the point set of the target structural surface from the three-dimensional point cloud located in the world coordinate system, and to perform plane fitting on the point set to obtain the plane normal vector; An attitude calculation module is used to calculate the attitude of the target structural surface from the plane normal vector, wherein the attitude includes at least one of strike, dip and dip angle; Throughout the process, the absolute physical scale of the three-dimensional point cloud is not restored.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the structural plane attitude measurement method as described in any one of claims 1 to 7.
10. An electronic device, characterized in that, The device includes a camera, an accelerometer, a magnetometer, a processor, and a memory, wherein the memory stores a computer program that, when executed by the processor, implements the structural plane attitude measurement method as described in any one of claims 1 to 7, and the electronic device constitutes the mobile terminal in the structural plane attitude measurement method.
Citation Information
Patent Citations
A Rapid and Refined Method for Determining Joint Attitude in Rock Mass Slopes of Open-Pit Mines Based on Digital Photogrammetry
CN108830317B
A non-contact measurement method for the attitude of rock mass structural surfaces based on smartphones
CN109738440B
A method for solving the occurrence of structural plane without interference of external magnetic field
CN116758133B
Stratum attitude attribute calculation method based on mobile phone sensor
CN118152714A
