A method and device for automatic registration of a BIM model and an AR scene

Abstract

The application discloses a kind of BIM model and the automatic registration method and device of AR scene, belong to augmented reality and building information model fusion technical field;The method comprises: obtaining the BIM coordinate information carried by visual mark;Visual mark is detected in AR scene and corresponding AR coordinate is obtained;According to the number of different visual marks detected, automatically select corresponding registration strategy;The registration transformation from BIM coordinate system to AR coordinate system is calculated by executing the selected strategy;BIM model is superimposed to AR scene according to registration transformation;The present application suppresses random noise by multi-frame observation weighted fusion and outlier rejection, distinguishes the differentiated contribution of vertical plane and horizontal plane to orientation constraint, and has automatic fallback mechanism for degenerate scene, realizes centimeter-level registration accuracy, and is suitable for BIM real model 1:1 comparison in indoor and outdoor construction acceptance scene.

Description

Technical Field

[0001] This invention belongs to the field of augmented reality (AR) and building information modeling (BIM) fusion technology, specifically relating to an automatic registration method and device for BIM models and AR scenes. Background Technology

[0002] In the field of integrated application of Building Information Modeling (BIM) and Augmented Reality (AR) technologies, accurately overlaying the BIM digital model onto the real building scene at a 1:1 scale is a core technical step for applications such as construction quality acceptance, project progress comparison, and visualization of hidden works. This process is called "BIM-AR registration," which involves calculating the spatial transformation relationship between the BIM model coordinate system and the AR scene coordinate system to precisely align the virtual model with the real scene in terms of position, orientation, and scale.

[0003] The current mainstream solutions for overlaying BIM models onto real-world AR scenes include the following categories: Option A: Manual Placement Method. After detecting a plane in the AR scene, the user places the BIM model onto the plane by tapping the screen, and then performs coarse adjustments by dragging, rotating, and scaling using gestures. This method relies on the user's subjective judgment, has low positioning accuracy (typically deviating by tens of centimeters to several meters), and cannot meet the engineering acceptance requirements for 1:1 comparison with the BIM physical model.

[0004] Option B: GPS / RTK positioning method. This method utilizes high-precision satellite positioning to obtain the geographical coordinates of the equipment and then performs registration by combining them with the geographical reference coordinates of the BIM model. However, this method suffers from severe GPS signal attenuation or unavailability in indoor scenarios, making it unsuitable for typical BIM-AR scenarios such as indoor construction acceptance.

[0005] Option C: Visual SLAM Global Localization. This method utilizes a visual-inertial SLAM system to calculate the device's 6DoF pose in the world coordinate system and directly aligns it with the global coordinates of the BIM model. This method requires the AR system and the BIM model to use the same global coordinate system, and is limited by SLAM drift accumulation errors, resulting in a significant decrease in positioning accuracy after long-term tracking.

[0006] The existing solutions described above have the following main technical defects: Defect 1: Insufficient registration accuracy. The manual placement method relies on the user's subjective judgment and cannot achieve sub-centimeter level registration; the GPS method is not usable indoors; the SLAM method is affected by drift and is difficult to maintain high accuracy for a long time.

[0007] Defect 2: Lack of a progressive registration mechanism. Existing solutions typically require all constraints to be provided at once, and do not support a gradual registration process of "coarse to fine". In actual construction sites, it is often impossible to set up multiple marker points at once, so a working mode that supports scanning and calibration as needed is required.

[0008] Defect 3: Lack of systematic handling of coordinate system transformation. BIM modeling software (such as Revit) generally uses the Z-up coordinate system (vertical axis is Z), while AR engines such as ARKit use the Y-up coordinate system (vertical axis is Y). Existing solutions do not handle coordinate system transformation systematically enough, especially when transforming rotation (yaw), which is prone to direction reversal due to differences in chirality.

[0009] Defect 4: The impact of plane type on registration constraints is ignored. When a ray hits a vertical plane (wall), the horizontal tangent direction of the wall provides a reliable yaw constraint; however, when a ray hits a horizontal plane (ground), the tangent direction is arbitrary and cannot provide an effective yaw constraint. The existing scheme does not distinguish between plane types, resulting in yaw noise introduced by horizontal plane sampling, which reduces registration accuracy.

[0010] Defect 5: Lack of fusion and denoising mechanisms for multi-frame observations. In actual scanning, multiple frames of observation data are generated continuously for the same location, and the ray hit position and direction in each frame contain random noise. Existing solutions usually only use single-frame data or simple averaging, failing to effectively suppress noise and remove outliers. Summary of the Invention

[0011] This invention addresses the technical shortcomings of existing BIM-AR registration schemes, such as insufficient registration accuracy, lack of progressive registration mechanism, lack of systematic processing for coordinate system transformation, neglect of the influence of plane type on registration constraints, and lack of fusion and denoising mechanism for multi-frame observations. It provides an automatic registration method and device for BIM models and AR scenes, which can effectively suppress yaw noise introduced by the horizontal plane; and through an automatic back-off mechanism for degraded scenes, it prevents numerical divergence and improves the robustness of the system in various actual construction scenarios.

[0012] The technical solution for implementing the present invention is as follows: An automatic registration method for BIM models and AR scenes, comprising: s1. Obtain coordinate information carried by at least one visual marker, the coordinate information indicating the position of a target point in the BIM model in the BIM coordinate system; s2. Detect the visual marker in the AR scene and obtain the corresponding position of the target point in the AR coordinate system; s3. Based on the number of different visual markers detected, automatically select a target registration strategy corresponding to the number from multiple candidate registration strategies, wherein different numbers correspond to different registration strategies; s4. Execute the target registration strategy and calculate the registration transformation from the BIM coordinate system to the AR coordinate system; s5. Overlay the BIM model onto the AR scene according to the registration transformation described above.

[0013] Furthermore, when the number is one, the target registration strategy includes: Calculate the translation amount to align the target point in the BIM coordinate system with the corresponding position in the AR coordinate system; When the number of fused samples of the same label reaches a preset threshold and the plane where the label is located is a vertical plane, the horizontal component length of each basis vector is calculated from the two columns of basis vectors of the transformation matrix of the plane. The direction of the basis vector with the larger horizontal component length is taken as the tangent direction along the wall. The rotation angle around the vertical axis is calculated according to the tangent direction to obtain a five-degree-of-freedom registration that includes translation and rotation around the vertical axis; otherwise, only a three-degree-of-freedom registration of translation is performed.

[0014] Furthermore, when the number is two, the target registration strategy includes: Calculate the angle between the first direction vector of two marker points in the BIM coordinate system and the second direction vector in the AR coordinate system on the horizontal plane, and use it as the rotation angle about the vertical axis; Calculate the ratio of the distance between two marker points in the AR coordinate system to the distance in the BIM coordinate system, use it as a scale factor, and truncate the scale factor to a preset safety range; Calculate the translation amount based on the centroid alignment of the two marked points; Construct a registration transformation matrix based on the rotation angle, the scale factor, and the translation amount; When the projection distance between two marker points on the horizontal plane is less than a first preset threshold, the system reverts to the single marker point registration strategy described in claim 2.

[0015] Furthermore, when the number is three or more, the target registration strategy includes: Calculate the first centroid of the marker point set in the BIM coordinate system and the second centroid of the marker point set in the AR coordinate system, and decentroid both sets of point sets respectively; Construct a covariance matrix, which is the sum of the products of the centroid-decentered BIM point set coordinates and the transpose of the AR point set coordinates; Perform singular value decomposition on the covariance matrix and solve for the rotation matrix based on the decomposition results; The scaling factor is calculated based on the rotation matrix, and the scaling factor is truncated to a preset safe range; The translation amount is calculated based on the rotation matrix, the scale factor, and the centroid relationship; When the distribution range of all marker points on the horizontal plane is less than the second preset threshold, the rotation matrix is ​​replaced with an identity matrix; when the determinant of the rotation matrix obtained by singular value decomposition is negative, the decomposition matrix is ​​corrected to make the determinant of the rotation matrix positive.

[0016] Further, in step s2, for the same visual marker, multiple frames of observation samples are collected within a preset time window, and the multiple frames of observation samples are fused to obtain the corresponding position of the target point in the AR coordinate system; the fusion includes: The weighted mean position of all samples is calculated using the confidence level of each sample as the weight. The confidence level is determined based on the distance from the sampling point to the AR camera, the angle between the forward direction of the AR camera and the normal of the plane where the mark is located, and the plane type. The weight of vertical plane samples is greater than that of horizontal plane samples. Calculate the distance of each sample to the weighted mean position, and remove outlier samples whose distance exceeds a preset threshold; The weighted mean was recalculated for the retained inlier samples to obtain the final fusion position; For attitude angles, the circular mean algorithm is used to handle the periodicity of the angles.

[0017] Furthermore, the registration transformation from the BIM coordinate system to the AR coordinate system includes coordinate system transformation, wherein the BIM coordinate system adopts the Z-up coordinate system and the AR coordinate system adopts the Y-up coordinate system; the mapping relationship of the coordinate system transformation is: the x-coordinate in the AR coordinate system is equal to the x-coordinate in the BIM coordinate system, the y-coordinate in the AR coordinate system is equal to the z-coordinate in the BIM coordinate system, and the z-coordinate in the AR coordinate system is equal to the inverted y-coordinate in the BIM coordinate system; when the coordinate information contains a horizontal orientation angle, the horizontal orientation angle is inverted and corrected.

[0018] Furthermore, the coordinate information also includes length unit information; after step s1, the method further includes: normalizing the three-dimensional coordinates to a preset unit system based on the length unit information; when the coordinate information does not contain length unit information, heuristic inference is made based on the maximum value of the components of the three-dimensional coordinates: if the maximum value of the components is greater than a preset value, it is inferred to be millimeter unit, otherwise it is inferred to be meter unit.

[0019] Furthermore, whenever the number of different visual markers detected changes, steps s3 to s5 are re-executed to update the registration transformation.

[0020] An automatic registration device for BIM models and AR scenes includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0021] Compared with the prior art, the present invention has the following advantages: (1) The present invention automatically selects the corresponding registration strategy according to the number of different visual markers detected, realizing a progressive registration mechanism of "scanning and calibrating and refining step by step", which does not require all markers to be arranged at once, and is adapted to the operation mode of arranging markers in batches on the actual construction site. Existing solutions do not support progressive calibration.

[0022] (2) This invention effectively suppresses random noise in continuous multi-frame observations of the same marker point through a two-round weighted fusion and outlier removal mechanism, and limits the registration scale factor to a narrow range close to 1.0 to prevent scaling offset caused by ray noise or coordinate anomalies, ensuring a 1:1 ratio comparison between the BIM model and the real scene. In contrast, the manual placement method relies on subjective judgment, has a large deviation, and cannot meet the requirements of engineering acceptance.

[0023] (3) The present invention distinguishes the differential contribution of vertical and horizontal planes to the registration constraint. Vertical plane samples are given higher weight because they can provide reliable orientation constraints, while horizontal plane samples are given lower weight because they only provide position constraints. This makes the multi-frame fusion process naturally favor vertical plane samples and effectively suppress the orientation angle noise introduced by the horizontal plane. However, the existing scheme does not distinguish the plane type, which leads to a decrease in registration accuracy.

[0024] (4) This invention provides a complete mapping rule from the Z-up coordinate system to the Y-up coordinate system and performs chiral inversion correction on the horizontal orientation angle, eliminating the registration deviation and direction reversal caused by the coordinate system difference between the BIM modeling software and the AR engine.

[0025] (5) In the degenerate scenarios such as the horizontal projection of double marker points coincides and the horizontal distribution of multiple marker points is insufficient, the present invention automatically reverts to the low degree of freedom registration strategy and automatically corrects the chirality of the rotation matrix obtained by singular value decomposition when it is abnormal, effectively preventing the model from drastic changes caused by numerical divergence, while the existing scheme does not have the function of degeneracy protection. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0027] like Figure 1 As shown, this invention discloses an automatic registration method for BIM models and AR scenes, which can achieve accurate and adaptive overlay matching between BIM models and real-world AR scenes at construction sites. The overall execution process is clear and progressive, and the detailed implementation process is as follows: Step 1: Collect BIM coordinate information carried by visual markers This method uses QR codes as visual markers. Each QR code encodes the location information of a single target point in the BIM model, carrying the complete location data of that target point in the BIM coordinate system. The marker information includes core parameters such as a unique identifier, 3D coordinates, horizontal orientation, and length unit. The unique identifier is used to distinguish different visual markers, the 3D coordinates are the location data of the target point in the vertical coordinate system of the BIM, the horizontal orientation is an optional parameter representing the horizontal deflection angle of the target point in the BIM coordinate system, and the corresponding length unit is also indicated.

[0028] In actual construction applications, after the QR code visual markers are printed, they are posted in prominent positions on the vertical walls indoors and outdoors at the construction site. The posting height is controlled between 1.2 meters and 1.8 meters, while avoiding areas with strong light and severe glare to avoid affecting the recognition and detection accuracy of subsequent AR scenes.

[0029] Step 2: AR scene detection of visual markers and completion of coordinate transformation and data optimization Throughout the AR device's session, the system captures AR camera frames at fixed time intervals and uses a visual inspection framework to identify QR code visual markers in the images in real time. Upon successful marker detection, the system automatically parses the data contained within the marker and extracts the corresponding BIM target point coordinates. To improve parsing adaptability and robustness, the system supports both standard structured text and permissive object formats for data parsing. Simultaneously, it quickly extracts various parameters using preset matching rules, effectively adapting to visual markers with different encoding formats.

[0030] Because the BIM coordinate system and the AR device coordinate system have different axial rules, the system can automatically convert between the two coordinate systems, uniformly transforming the BIM vertical coordinate system data into the AR device-compatible vertical coordinate system data. The specific conversion rules are as follows: the horizontal coordinates of the AR scene are consistent with the horizontal coordinates of the BIM model; the vertical coordinates of the BIM model correspond to the vertical coordinates of the AR scene; and the horizontal depth coordinates of the BIM model are converted inversely to ensure that the overall coordinate system meets the right-hand rule standard. Simultaneously, due to the chirality reversal of the coordinate system, the horizontal orientation angle of the markers needs to be synchronously corrected in reverse to ensure accurate attitude angle matching.

[0031] The system supports automatic normalization of length units and can directly identify the length units explicitly marked in the labeled data. If the unit is not specified in the label, the system will automatically determine it through heuristic rules, distinguish between millimeter and meter units based on the absolute value of the maximum component of the coordinate value, and complete the unified conversion to provide standard unit data for subsequent registration calculations.

[0032] To avoid random errors in single-frame detection data and improve positioning accuracy, the system collects multiple frames of observation data for the same QR code marker within a set short time window, and optimizes the coordinate and attitude data through a two-round weighted fusion algorithm. The first round of fusion uses the overall confidence level of a single-frame sample as the weight to calculate the weighted average of the coordinates of all valid samples, obtaining the initial positioning center. Then, outlier removal is performed, calculating the distance between each sample and the initial center, removing samples exceeding a preset distance threshold, and filtering out noisy data. The second round of fusion, for the filtered valid sample set, again calculates the weighted average to finally obtain accurate and stable marker position data.

[0033] The confidence level of the samples is calculated comprehensively based on multiple factors to ensure data reliability. First, there is the distance attenuation factor, which follows the principle that closer proximity equates to higher confidence, employing an exponential decay rule where the greater the distance between the camera and the marked sampling point, the lower the sample weight. Second, there is the incident angle factor, where the closer the camera's shooting direction is to the normal direction of the wall where the mark is located, and the more directly the shooting angle is directed towards the wall, the higher the sample confidence level. Third, there is the plane type weight, prioritizing samples from vertical walls and assigning lower weights to samples collected from horizontal surfaces such as the ground and ceiling, ensuring that the data fusion results naturally adapt to the layout scenario of wall markings on the construction site.

[0034] For attitude data fusion with marked horizontal orientation angles, the system abandons the simple linear averaging algorithm and adopts a circular mean algorithm to handle the periodicity of angles, effectively solving the problems of angle calculation jumps and bias distortion, and ensuring the continuity and accuracy of attitude data.

[0035] Step 3: Based on the number of different visual markers detected, automatically select the target registration strategy corresponding to the number from multiple candidate registration strategies, where different numbers correspond to different registration strategies; The system can count the number of successfully detected and parsed independent visual tags in the current scene in real time. Based on the number of tags, it automatically matches the corresponding dedicated registration algorithm, forming a progressive registration mechanism of "calibrating as you scan and gradually optimizing accuracy." Different registration strategies correspond to different numbers of tags. When there is one valid tag, the system uses a single-code registration strategy; when there are two valid tags, it automatically switches to a dual-code registration strategy; when there are three or more valid tags, it calls the singular value decomposition rigid alignment registration strategy.

[0036] Step 4: Execute the target registration strategy and calculate the transformation matrix from the BIM coordinate system to the AR coordinate system. (a) Single marker point: Single code registration strategy Single-marker registration is based on coordinate translation alignment to achieve positional matching between BIM target points and corresponding points in the AR scene. The system determines the number of sample collections and the planar attributes of the marker layout to achieve differentiated registration: when the number of effectively fused samples reaches a preset threshold and the detection ray hits the vertical wall, the system can extract the horizontal tangent direction from the wall plane transformation data, calculate the precise horizontal rotation angle, and complete a five-degree-of-freedom accurate registration of translation plus vertical axis rotation; if the above conditions are not met, the system automatically reverts to a three-degree-of-freedom basic registration of pure translation to ensure the stability of the basic registration effect.

[0037] The extraction of the horizontal rotation angle employs rigorous judgment and calculation logic. The system extracts the basis vector data of the wall plane transformation, projects it onto a horizontal plane, and distinguishes between the tangent and normal directions of the wall by comparing the lengths of the horizontal components of the two sets of basis vectors. Then, it calculates the precise horizontal rotation angle using the arctangent function. Simultaneously, a vertical plane verification mechanism is implemented, performing the angle extraction operation only when the angle between the plane normal and the vertical direction meets a threshold requirement, thus avoiding posture deviations caused by non-vertical plane data.

[0038] (ii) Dual marker points: Dual code registration strategy Dual-marker registration enables precise registration across all six degrees of freedom, while simultaneously performing translation, rotation, and scaling corrections. The system calculates the overall horizontal rotation angle based on the BIM model's direction vector and the AR scene's direction vector from two marker points in the horizontal plane, using the angle difference. Simultaneously, it calculates the distance between the two marker points in both the BIM model and the AR scene, determines the scaling factor through the distance ratio, and applies interval truncation constraints to limit the scaling factor within a safe range, thus preventing scaling offset issues caused by scene detection noise.

[0039] The registration translation amount is calculated by aligning the centroids of the two marker points. Finally, the translation, rotation, and scaling parameters are integrated to construct a complete coordinate system transformation matrix. To avoid geometric degradation issues, the system adds a degradation detection mechanism. If the projection distance between the two marker points on the horizontal plane is less than a preset threshold and they are approximately on the same vertical line, the system will automatically degrade to a single-code registration strategy, avoiding the problem of rotation angle calculation divergence caused by geometric data failure.

[0040] (iii) Three or more markers: Singular value decomposition rigid alignment strategy When the number of valid marker points reaches three or more, the system adopts the classic rigid alignment algorithm to complete the global optimal registration, adapting to the high-precision matching requirements of multi-marker point scenarios. The specific calculation process is as follows: First, the centroids of the BIM model point set and the AR scene point set are calculated respectively, and the two sets of point sets are decentroided to eliminate the influence of the overall position deviation; then, the covariance matrix of the two sets of point sets is constructed, and singular value decomposition is performed on the matrix; the rotation matrix is ​​solved based on the decomposition result, and the determinant value of the rotation matrix is ​​checked. If the value is negative, the matrix parameters are corrected to ensure that the rotation direction is compliant.

[0041] Based on this, the system calculates the global scale scaling factor by the ratio of the sum of the product of the point set coordinates to the sum of the squares of the modulus, and also performs safe interval truncation to avoid abnormal scale offset; finally, by combining the centroid of the AR scene point set, the scaling factor and the rotation matrix, the final translation amount is solved, and the rigid transformation parameters from the BIM coordinate system to the AR coordinate system are obtained in full.

[0042] To ensure algorithm stability, the system adds a horizontal distribution degradation protection mechanism. If the distribution range of all marker points in the horizontal plane is less than a preset threshold, the geometric distribution of the point set is too concentrated, which will cause the rotation calculation value of the singular value decomposition to be unstable. At this time, the system automatically adopts the unit rotation matrix, retains only the translation registration effect, and is consistent with the single code registration logic, effectively avoiding the problem of calculation failure.

[0043] Step 5: Achieve overlay and fusion of BIM model and AR scene based on registration transformation The system integrates the registration parameters obtained from the above strategies into a fourth-order homogeneous transformation matrix, which is directly used as the anchor point transformation parameters of the AR scene model, driving the BIM model to complete accurate overlay and placement in the AR real-world scene. As the number of effective visual markers detected on-site continues to increase, the system will recalculate and update the registration transformation matrix in real time, continuously optimizing the model overlay accuracy, and achieving a progressive registration effect of "loading the model with a single scan and automatically calibrating after the deviation accumulates and is scanned again".

[0044] Once the registration operation is complete and the model overlay is stable, the system automatically closes the new user guide interface for the AR scene and opens the model transparency adjustment function. Users can control the slider through the interface to freely adjust the transparency within the range of 5% to 100%. Through the semi-transparent overlay effect, users can intuitively compare the matching deviation between the real construction site scene and the BIM model, realizing visual verification and acceptance.

[0045] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An automatic registration method for BIM models and AR scenes, characterized in that, include: s1. Obtain coordinate information carried by at least one visual marker, the coordinate information indicating the position of a target point in the BIM model in the BIM coordinate system; s2. Detect the visual marker in the AR scene and obtain the corresponding position of the target point in the AR coordinate system; s3. Based on the number of different visual markers detected, automatically select a target registration strategy corresponding to the number from multiple candidate registration strategies, wherein different numbers correspond to different registration strategies; s4. Execute the target registration strategy and calculate the registration transformation from the BIM coordinate system to the AR coordinate system; s5. Overlay the BIM model onto the AR scene according to the registration transformation described above.

2. The method according to claim 1, characterized in that, When the number is one, the target registration strategy includes: Calculate the translation amount to align the target point in the BIM coordinate system with the corresponding position in the AR coordinate system; When the number of fused samples of the same label reaches a preset threshold and the plane where the label is located is a vertical plane, the horizontal component length of each basis vector is calculated from the two columns of basis vectors of the transformation matrix of the plane. The direction of the basis vector with the larger horizontal component length is taken as the tangent direction along the wall. The rotation angle around the vertical axis is calculated according to the tangent direction to obtain a five-degree-of-freedom registration that includes translation and rotation around the vertical axis; otherwise, only a three-degree-of-freedom registration of translation is performed.

3. The method according to claim 2, characterized in that, When the number is two, the target registration strategy includes: Calculate the angle between the first direction vector of two marker points in the BIM coordinate system and the second direction vector in the AR coordinate system on the horizontal plane, and use it as the rotation angle about the vertical axis; Calculate the ratio of the distance between two marker points in the AR coordinate system to the distance in the BIM coordinate system, use it as a scale factor, and truncate the scale factor to a preset safety range; Calculate the translation amount based on the centroid alignment of the two marked points; Construct a registration transformation matrix based on the rotation angle, the scale factor, and the translation amount; When the projection distance between two marker points on the horizontal plane is less than a first preset threshold, the system reverts to the single marker point registration strategy described in claim 2.

4. The method according to claim 1, characterized in that, When the number is three or more, the target registration strategy includes: Calculate the first centroid of the marker point set in the BIM coordinate system and the second centroid of the marker point set in the AR coordinate system, and decentroid both sets of point sets respectively; Construct a covariance matrix, which is the sum of the products of the centroid-decentered BIM point set coordinates and the transpose of the AR point set coordinates; Perform singular value decomposition on the covariance matrix and solve for the rotation matrix based on the decomposition results; The scaling factor is calculated based on the rotation matrix, and the scaling factor is truncated to a preset safe range; The translation amount is calculated based on the rotation matrix, the scale factor, and the centroid relationship; When the distribution range of all marker points on the horizontal plane is less than the second preset threshold, the rotation matrix is ​​replaced with an identity matrix; when the determinant of the rotation matrix obtained by singular value decomposition is negative, the decomposition matrix is ​​corrected to make the determinant of the rotation matrix positive.

5. The method according to claim 1, characterized in that, In step s2, for the same visual marker, multiple frames of observation samples are collected within a preset time window, and the multiple frames of observation samples are fused to obtain the corresponding position of the target point in the AR coordinate system; the fusion includes: The weighted mean position of all samples is calculated using the confidence level of each sample as the weight. The confidence level is determined based on the distance from the sampling point to the AR camera, the angle between the forward direction of the AR camera and the normal of the plane where the mark is located, and the plane type. The weight of vertical plane samples is greater than that of horizontal plane samples. Calculate the distance of each sample to the weighted mean position, and remove outlier samples whose distance exceeds a preset threshold; The weighted mean was recalculated for the retained inlier samples to obtain the final fusion position; For attitude angles, the circular mean algorithm is used to handle the periodicity of the angles.

6. The method according to claim 1, characterized in that, The registration transformation from the BIM coordinate system to the AR coordinate system includes coordinate system transformation, wherein the BIM coordinate system adopts the Z-up coordinate system and the AR coordinate system adopts the Y-up coordinate system; the mapping relationship of the coordinate system transformation is: the x-coordinate in the AR coordinate system is equal to the x-coordinate in the BIM coordinate system, the y-coordinate in the AR coordinate system is equal to the z-coordinate in the BIM coordinate system, and the z-coordinate in the AR coordinate system is equal to the inverted y-coordinate in the BIM coordinate system; when the coordinate information includes a horizontal orientation angle, the horizontal orientation angle is inverted and corrected.

7. The method according to claim 1, characterized in that, The coordinate information also includes length unit information; after step s1, the method further includes: normalizing the three-dimensional coordinates to a preset unit system based on the length unit information; when the coordinate information does not include length unit information, heuristic inference is made based on the maximum value of the components of the three-dimensional coordinates: if the maximum value of the components is greater than a preset value, it is inferred to be millimeter unit, otherwise it is inferred to be meter unit.

8. The method according to any one of claims 1-7, characterized in that, Whenever the number of different visual markers detected changes, steps s3 to s5 are re-executed to update the registration transformation.

9. An automatic registration device for BIM models and AR scenes, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method of any one of claims 1 to 8.

Images