Prefabricated component site virtual pre-assembly method based on single-station scanning and augmented reality

Abstract

This invention discloses a method for virtual pre-assembly of precast components on-site based on single-station scanning and augmented reality. The method includes: acquiring point cloud data of the steel reinforcement ends and sleeve ends of the precast concrete component using a single-station laser scanner; performing semantic segmentation on the point cloud data of the steel reinforcement ends using a lightweight point cloud segmentation network to obtain point cloud sets of the steel reinforcement and the concrete main body; extracting assembly features from the steel reinforcement point cloud and sleeve point cloud; and, based on the steel reinforcement assembly features and sleeve assembly features, using a virtual pre-assembly algorithm that considers steel reinforcement position correction to achieve global pose alignment of the steel reinforcement and sleeve, obtaining a global assembly feasibility result, transmitting it to an AR device, and using spatial anchor point technology to align the 3D holographic projection of the assembled state with the on-site physical component, achieving a virtual-real fusion visualization interaction. This invention overcomes the time-consuming nature of traditional virtual pre-assembly and the information gap caused by detachment from the actual on-site situation, significantly improving the detection efficiency of precast concrete component assembly.

Description

Technical Field

[0001] This invention relates to the field of information technology in building construction and precast concrete component assembly technology, specifically to a method for on-site virtual pre-assembly of precast components based on single-station scanning and augmented reality. Background Technology

[0002] During on-site assembly of precast components, inaccurate matching often occurs due to manufacturing processes, transportation deformation, or geometric deviations in the dimensions of reinforcing bars and sleeves. If these problems are not discovered until the hoisting stage, it will lead to serious project delays, rework, and economic losses.

[0003] To mitigate these risks, two main methods are currently employed: physical pre-assembly and virtual pre-assembly. Physical pre-assembly requires significant manpower, space, and time, and is highly susceptible to damaging the components themselves. Traditional virtual trial assembly methods based on point clouds and Building Information Modeling (BIM) are mostly dedicated to analyzing the entire structure or large-scale scenes. The extremely high hardware computing requirements mean that point cloud acquisition and registration typically take 20 to 40 hours. This asynchronous computational lag not only fails to meet tight construction schedules but also results in extremely low efficiency for subsequent on-site component correction. Furthermore, most existing virtual pre-assembly methods neglect the physical correction of rebar positions by workers using wooden formwork before concrete pouring. This leads to the invalidation of virtual pre-assembly judgments based on initial scan data when deformation has not fully recovered. Additionally, the presentation of existing 3D deviation results is mostly confined to computer screens, making it difficult for on-site quality inspectors to intuitively map deviation data to actual physical components.

[0004] Therefore, developing a near-real-time virtual pre-assembly method for precast components that is suitable for construction sites, takes into account the physical correction of rebar positions, and has intuitive visual interaction capabilities, and which breaks through the bottleneck of existing offline virtual assembly by combining lightweight models with augmented reality technology, has important application value. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a virtual pre-assembly method for prefabricated components on-site based on single-station scanning and augmented reality. This invention significantly shortens the point cloud acquisition and processing cycle through single-station cloud acquisition, a lightweight segmentation model, and a high-precision feature extraction framework for single-station clouds. It enables near real-time point cloud processing on construction edge devices with limited computing power. Furthermore, it innovatively integrates mechanical deflection correction and spatial anchor point display technologies, greatly improving the realism of assembly inspection and the efficiency of human-computer interaction, thereby perfecting the closed-loop quality control of prefabricated components.

[0006] The objective of this invention is achieved through the following technical solution: The first aspect of this invention provides a method for on-site virtual pre-assembly of prefabricated components based on single-station scanning and augmented reality, comprising the following steps: (1) Use a laser scanner to collect single-station three-dimensional point cloud data of the steel bar end and sleeve end in the precast concrete component at the construction site; (2) Input the single-station 3D point cloud data of the steel bar end into the trained lightweight point cloud segmentation network for semantic segmentation, separate the steel bar from the concrete body, and obtain the point cloud set of the steel bar and the concrete body. (3) Perform assembly feature extraction on the segmented rebar point cloud to obtain the radius and center three-dimensional coordinates of each rebar section; (4) Extract the assembly features of the sleeve from the single-station three-dimensional point cloud data of the sleeve end, and obtain the cross-sectional radius and center three-dimensional coordinates of each sleeve; (5) Based on the rebar assembly features and sleeve assembly features, the ideal assembly coordinates after rebar correction are calculated using a virtual pre-assembly algorithm that considers rebar position correction. The global pose alignment of the rebar and sleeve is achieved through coordinate system translation and rotation matrices to obtain the global assembly feasibility results. (6) Transmit the global assembly feasibility results to the augmented reality device, and use spatial anchor point technology to align the three-dimensional holographic projection of the assembly state with the physical components on site to achieve virtual-real fusion visualization interaction.

[0007] Furthermore, the lightweight point cloud segmentation network adopts a lightweight point cloud Transformer network for single-site point cloud semantic segmentation. The lightweight point cloud segmentation network is designed based on PointTransformer v3 and specifically includes an input embedding module, a lightweight Transformer encoding module, a feature propagation and decoding module, and a point-by-point classification head. The input embedding module receives the voxelized, normalized, and tensorized single-site point cloud tensor of the rebar end and maps the spatial coordinates, color, or reflection intensity attributes of each point to high-dimensional point features, which serve as the input to the lightweight Transformer encoding module. The lightweight Transformer encoding module consists of several levels of sparse convolutional conditional location encoding layers, local attention mechanism layers, and multilayer perceptron feedforward layers, and is configured according to channel scaling ratios. The number of channels in each layer is compressed; the feature propagation decoding module is used to upsample the multi-scale point features output by the lightweight Transformer encoding module step by step and restore them to the original point resolution; the point-by-point classification head includes a linear layer, a normalization layer and a Softmax layer, which are used to output the probability of each point belonging to the category of steel reinforcement or concrete body.

[0008] Furthermore, the training method for the lightweight point cloud segmentation network specifically includes: The single-station 3D point cloud data of the rebar end is voxelized, normalized, and converted into a single-station point cloud tensor of the rebar end; The single-site point cloud tensor of the rebar end is input into a lightweight point cloud segmentation network, and the output is a point-by-point semantic label probability matrix that corresponds one-to-one with the input point. Phase 1: Construct an augmentation model based on a lightweight point cloud segmentation network; calculate the semantic segmentation loss of the micro-branch and the augmentation branch based on the predicted semantic label probabilities output by the lightweight point cloud segmentation network and the augmentation model and their corresponding real semantic labels, and obtain the joint loss by weighted summation; adjust the model parameters of the lightweight point cloud segmentation network with the optimization objective of minimizing the joint loss until the preset number of training rounds is reached, and obtain a pre-trained lightweight point cloud segmentation network; The second stage involves using the pre-trained optimal PointTransformer v3 as the teacher model and the initially trained lightweight point cloud segmentation network as the student model. Distillation loss is calculated based on the outputs of the teacher and student models. Semantic segmentation loss is calculated based on the predicted semantic label probabilities and corresponding ground truth semantic labels output by the student model. The total loss is calculated based on the semantic segmentation loss and distillation loss of the student model. The model parameters of the student model are adjusted with the goal of minimizing the total loss until the preset number of training epochs is reached, resulting in the final trained lightweight point cloud segmentation network.

[0009] Furthermore, the formula for calculating the joint loss is as follows:

[0010] In the formula, Denotes the joint loss function. This represents the semantic segmentation loss of the augmentation branch. This represents the semantic segmentation loss for micro-branches. This represents the loss weighting coefficient for the first stage; The formula for calculating the distillation loss is as follows:

[0011] In the formula, Indicates distillation loss, Indicates the output distribution of the teacher model With student model output distribution The KL divergence loss between them, where T represents the distillation temperature coefficient, This represents the temperature scaling compensation factor; The formula for calculating the KL divergence loss is as follows:

[0012] In the formula, This represents the Softmax soft label probability distribution output by the teacher model. This represents the predicted probability distribution output by the student model. Let represent the i-th point in the input point cloud, N represent the number of points in the input point cloud, c represent the c-th semantic category, and C represent the number of semantic categories. Represents the teacher model for points The predicted probability of belonging to class c. Represents student model points The predicted probability of belonging to class c. Represents the natural logarithm; The formula for calculating the total loss is as follows:

[0013] In the formula, Indicates the total loss. This indicates that the semantic segmentation loss of the student model is calculated based on manually labeled hard tags. This represents the loss weighting coefficient for the second stage.

[0014] Furthermore, step (3) specifically includes: A density-based noisy spatial clustering algorithm is used to cluster the rebar point cloud into instances and separate independent rebar instances. For a single rebar, the initial principal axis of the Z-axis is extracted using the principal component analysis algorithm to achieve coarse alignment of the Z-axis; A pixel-level alignment algorithm is applied, and a small rotation matrix is ​​constructed by introducing a small rotation angle. The centroid of the rebar point cloud after coarse alignment along the Z-axis is used as the rotation center. The rotated point cloud is calculated, and its X and Y coordinates are taken as the XY plane projection coordinates. The projection coordinates are discretized into pixel indices according to the pixel resolution, and the set of non-repeating occupied pixels is counted based on the pixel indices. The coverage of the two-dimensional mapping area is minimized by Powell's derivative-free optimization method to determine the optimal rotation angle, so that the rebar axis is parallel to the Z-axis, thus achieving pixel-level fine alignment. Random sampling and uniform circle fitting are performed on the precisely aligned two-dimensional projection points to extract the effective radius of the reinforcing bars and the center coordinates of the top and root sections.

[0015] Furthermore, step (4) specifically includes: For the single-station 3D point cloud data of the sleeve end, after Z-axis alignment of the sleeve point cloud by principal component analysis, the point cloud is rotated in the XY plane so that the longest side of the convex hull is parallel to the X-axis, thus achieving orthogonal axis alignment; the Z-axis coordinates are discarded and the two-dimensional projection points of the sleeve point cloud on the XY plane are obtained. An adaptive pixel mapping strategy is introduced to rasterize the projection point coordinates into a two-dimensional binary image according to the resolution; Image denoising is performed on two-dimensional binary images by applying Gaussian blur, as well as morphological opening and closing operations. Based on the denoised image, a speckle detection algorithm is used to identify sleeve candidate regions, and the identified sleeve candidate regions are back-located to the original point cloud coordinate system according to the rasterization mapping relationship to form sleeve candidate point cloud regions. For the candidate point cloud region of the sleeve, an annular mask is generated by combining the sleeve center and inner and outer radius constraints, and a density-based noisy spatial clustering algorithm is used at the edge of the annular mask to cluster the independent sleeve instances. For each sleeve instance, its boundary feature points are selected by applying polar coordinate angle binning; The boundary feature points are fitted with a circle using a random sampling consensus algorithm to extract the cross-sectional radius and center three-dimensional coordinates of each sleeve.

[0016] Furthermore, in step (5), the virtual pre-assembly algorithm considering the correction of the reinforcement position specifically includes: The center coordinates of the root of the rebar are regarded as fixed support and the top is regarded as free end. Based on the assumption of linear elastic small deflection, the lateral displacement correction of the rebar at each height in the X and Y directions is calculated in combination with the rebar bending deformation equation to correct the original detected center coordinates of the rebar and obtain the ideal assembly coordinates of the rebar after correction. Determine the lower left corner reference instance in the rebar group and sleeve group respectively, and extract the center coordinates of the corresponding reference instance to calculate the translation vector. Based on the translation vector, translate the entire rebar group so that the center of the lower left corner reference rebar coincides with the center of the lower left corner reference sleeve, thus achieving rough assembly. By introducing the coordinates of the upper left and lower right corners, the direction vectors of the reinforcement group and the sleeve group are calculated, and the rotation matrix is ​​calculated based on the direction vectors of the two. The reinforcement group is then transformed into the sleeve group coordinate system based on the rotation matrix to achieve fine assembly. After fine assembly, calculate the center deviation, axial angle deviation, and radial clearance deviation between each steel bar and the corresponding sleeve, and compare each deviation value with the preset assembly allowable deviation threshold. If all deviation values ​​are within the assembly allowable deviation threshold range, the overall assembly is deemed feasible; otherwise, it is deemed that there is an assembly risk or that assembly is not possible, thus obtaining the overall assembly feasibility result.

[0017] Furthermore, the specific steps of aligning the assembled 3D holographic projection with the physical components on site using spatial anchoring technology include: Plane fitting and surface cleaning are performed on the point cloud set of the concrete main body classified in step (2) to generate a three-dimensional prior anchor point model; The 3D prior anchor point model is input into the Azure Object Anchor service of the augmented reality device. The feature-based model-sensor alignment is performed by combining the depth data collected by the augmented reality device with the SLAM map to restore the six-degree-of-freedom pose of the physical components in the augmented reality world coordinate system. Instantiate a spatial anchor point at a six-degree-of-freedom pose and lock the virtual pre-assembly deviation information and the corresponding holographic world obtained in step (5) onto the surface of the actual component; wherein, the virtual pre-assembly deviation information includes the center deviation, axial angle deviation and radial clearance deviation between the steel bar and the corresponding sleeve, as well as the global assembly feasibility result obtained based on the deviation information.

[0018] A second aspect of the present invention provides a virtual pre-assembly device for prefabricated components based on single-station scanning and augmented reality, comprising one or more processors and a memory, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the above-described virtual pre-assembly method for prefabricated components based on single-station scanning and augmented reality.

[0019] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, is used to implement the above-described method for on-site virtual pre-assembly of prefabricated components based on single-station scanning and augmented reality.

[0020] Compared with the prior art, the beneficial effects of the present invention are:

[0021] (1) The present invention proposes a lightweight point cloud segmentation network TPNet model with only 0.73M parameters (about 1 / 64 of the mainstream baseline model), but can achieve high-precision segmentation with an average accuracy (mAP) of 94.63%, and shorten the single inference time to 4.54s, which significantly shortens the point cloud acquisition and processing cycle, realizes near real-time processing of point clouds, and is perfectly adapted to field edge devices with limited computing power.

[0022] (2) The present invention improves the assembly feature extraction algorithm and introduces a rebar pose correction based on pixel alignment and a sleeve boundary extraction strategy based on angle binning. The average absolute error (MAE) of the rebar radius extracted by the improved assembly feature extraction algorithm is reduced to 0.53 mm and the MAE of the sleeve radius is reduced to 0.77 mm. It also shows an order-of-magnitude improvement in extraction speed and noise robustness.

[0023] (3) This invention innovatively integrates the mechanical deflection model (linear elastic deflection theory) for correcting the position of reinforcing bars into the virtual pre-assembly method. The predicted error of the corrected reinforcing bar coordinates is only 1.8mm, eliminating the data distortion problem caused by on-site manual correction and outputting evaluation results that are more in line with the actual engineering situation.

[0024] (4) This invention utilizes the Azure Object Anchors service to convert abstract point cloud deviation data into high-fidelity holographic images and achieve unmarked spatial alignment with physical components; this interaction mode effectively reduces the cognitive load of construction personnel and achieves seamless data traceability and on-site decision guidance. Attached Figure Description

[0025] Figure 1 This is a flowchart of the on-site virtual pre-assembly method for prefabricated components based on single-station scanning and augmented reality according to the present invention; Figure 2 This is a semantic segmentation effect diagram of the lightweight point cloud segmentation network of the present invention; Figure 3 This is a visual flowchart of the steel reinforcement assembly feature extraction of the present invention; Figure 4 This is a visual flowchart of the sleeve assembly feature extraction of the present invention; Figure 5 This is a schematic diagram of the virtual pre-assembly deflection analysis and registration method considering the correction of the reinforcement position according to the present invention; Figure 6 This is a physical image of the augmented reality holographic projection alignment pipeline based on spatial anchor points according to the present invention; Figure 7 This is a schematic diagram of a prefabricated component on-site virtual pre-assembly device based on single-station scanning and augmented reality according to the present invention. Detailed Implementation

[0026] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0027] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular forms “a,” “the,” and “the” used in this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0028] It should be understood that although the terms first, second, third, etc., may be used in this invention to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first information may also be referred to as second information without departing from the scope of this invention, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0029] The present invention will now be described in detail with reference to the accompanying drawings. Unless otherwise specified, the features of the following embodiments and implementations can be combined with each other.

[0030] See Figure 1 The present invention provides a method for on-site virtual pre-assembly of prefabricated components based on single-station scanning and augmented reality, which specifically includes the following steps:

[0031] (1) At the construction site, a laser scanner was used to collect single-station three-dimensional point cloud data of the steel bar end and sleeve end of the precast concrete component.

[0032] (2) Input the single-station 3D point cloud data of the rebar ends collected in step (1) into the trained lightweight point cloud segmentation network (TPNet) for semantic segmentation to separate the rebar from the concrete body and obtain the point cloud set of the rebar and the concrete body. The semantic segmentation effect is as follows: Figure 2 As shown.

[0033] In this embodiment, TPNet employs a lightweight point cloud Transformer network for single-site cloud semantic segmentation. TPNet is designed based on PointTransformer v3 (PTv3) and specifically includes an input embedding module, a lightweight Transformer encoding module, a feature propagation and decoding module, and a point-by-point classification head. The input embedding module receives the voxelized, normalized, and tensorized single-site cloud tensor of the rebar end and maps the spatial coordinates, color, or reflection intensity of each point to high-dimensional point features, which serve as the input to the subsequent lightweight Transformer encoding module. The lightweight Transformer encoding module consists of several levels of sparse convolutional conditional position coding (xCPE) layers, local attention mechanisms, and multilayer perceptron (MLP) feedforward layers, scaled according to channel ratios. (like The number of channels in each layer is compressed to build a lightweight point cloud Transformer network, reducing the channel size of each layer and thus reducing the number of model parameters and computational cost. The feature propagation decoding module is used to upsample the multi-scale point features output by the lightweight Transformer encoding module and restore them to the original point resolution. The point-by-point classification head includes a linear layer, a normalization layer and a Softmax layer, which are used to output the probability of each point belonging to the category of steel reinforcement or concrete.

[0034] In this embodiment, the lightweight semantic segmentation of TPNet is based on a computing workstation (15 vCPU Intel Xeon Platinum 8474C, 80 GB RAM, NVIDIA RTX 4090D 24GB) to train a lightweight point cloud segmentation network. Before training TPNet, data augmentation strategies can be used to enhance the single-station 3D point cloud data of the rebar ends collected in step (1) to expand the training data samples and enhance the generalization of the trained TPNet. Among them, the data augmentation strategies include 3D random rotation of the entire spatial axis, random scaling at a ratio of 0.9-1.1, horizontal / vertical flipping, and intensity jitter. (like In addition, chromatic aberration perturbation and random spherical clipping are performed. Then, the TPNet is trained using the following TPNet training method to obtain a trained TPNet.

[0035] The training method of TPNet specifically includes: voxelizing the single-station 3D point cloud data of the rebar end collected in step (1), specifically by dividing the data into a grid (grid size set to 0.02 m) and then performing voxel downsampling; at the same time, normalizing the voxelized point cloud data and converting it into a tensor format single-station point cloud tensor of the rebar end. The single-station point cloud tensor of the rebar end... Input into TPNet, where This represents the feature vector of the i-th point. , These are the normalized x-axis, y-axis, and z-axis coordinates (i.e., three-dimensional coordinates). These represent the color information for the R, G, and B channels, respectively. Where N is the laser reflection intensity, and N is the total number of input points. When the point cloud does not contain color or reflection intensity, The features at the corresponding positions can be set to 0, using only 3D coordinates or available attributes as input features; finally, TPNet outputs a point-by-point semantic label probability matrix that corresponds one-to-one with the input points. Where C is the number of categories, and in this embodiment, C=2, corresponding to the steel reinforcement category and the concrete body category, respectively. The final semantic category of a point is determined by comparing the probabilities of each point in the semantic label probability matrix belonging to the steel reinforcement category and the concrete body category. Specifically, in the point-by-point semantic label probability matrix, the semantic label probability corresponding to the i-th input point is... and through The final semantic category of the point is determined, thus obtaining the reinforcement point cloud subset and the concrete point cloud subset, where Let represent the probability that the i-th point belongs to the category of rebar. Let represent the probability that the i-th point belongs to the category of the concrete body. Let represent the probability that the i-th point belongs to class c, where c is either rebar or concrete. Let c represent the final semantic category of the i-th point. Then, based on the output point-by-point semantic label probabilities, each loss function is calculated, and a two-stage training process is performed.

[0036] The first stage involves model enhancement: based on TPNet, the network expansion rate is randomly selected in each iteration. The model is expanded to build an enhanced model. The semantic segmentation loss of the micro-branch is calculated based on the predicted semantic label probability values ​​and the corresponding real semantic labels output by TPNet. The semantic segmentation loss of the enhanced branch is calculated based on the predicted semantic label probability values ​​and the corresponding real semantic labels output by the enhanced model. The semantic segmentation losses of the micro-branch and the enhanced branch are weighted and summed to obtain the joint loss. The model parameters of TPNet are adjusted with the goal of minimizing the joint loss until the preset number of training rounds is reached, resulting in a pre-trained TPNet.

[0037] It should be noted that in semantic segmentation tasks, the loss function is used to measure the difference between the pixel / voxel category predicted by the model and the true label. For the semantic segmentation task of this invention, the commonly used cross-entropy loss or weighted cross-entropy loss can be used as the semantic segmentation loss.

[0038] Furthermore, the formula for calculating the joint loss is:

[0039] In the formula, Denotes the joint loss function. This represents the semantic segmentation loss of the augmentation branch. This represents the semantic segmentation loss for micro-branches. This represents the loss weight coefficient for the first stage, used to adjust the contribution ratio of TPNet and the augmentation model in joint training. It starts at 0.9 and ends by smoothing to 0.5.

[0040] The second stage involves knowledge distillation: the pre-trained optimal PTv3 is used as the teacher model, and the pre-trained TPNet is used as the student model; the KL divergence loss is calculated based on the Softmax soft label probability distribution output by the teacher model and the predicted probability distribution output by the student model. The KL divergence loss measures the prediction difference between the teacher model and the student model, and the distillation loss is calculated based on the KL divergence loss; the semantic segmentation loss of the student model is calculated based on the predicted semantic label probability values ​​output by the student model and the corresponding real semantic labels; the semantic segmentation loss and distillation loss of the student model are weighted and summed to obtain the total loss; the model parameters of the student model are adjusted with the goal of minimizing the total loss until the preset number of training epochs are reached, resulting in the final trained TPNet.

[0041] It should be noted that since TPNet is designed based on PTv3, the network architectures of PTv3 and TPNet are the same, but their parameter counts differ; the latter is lightweight, while the former is not. Therefore, the pre-trained optimal PTv3 is also obtained using the voxelized, normalized, and tensorized single-site cloud tensor of the rebar end according to this invention. The parameters of PTv3 are optimized by minimizing the semantic segmentation loss of PTv3, ultimately obtaining the pre-trained optimal PTv3.

[0042] Furthermore, the formula for calculating the KL divergence loss is as follows:

[0043] In the formula, Indicates the output distribution of the teacher model With student model output distribution KL divergence loss between them This represents the Softmax soft label probability distribution output by the teacher model (i.e., the class probability distribution output by the teacher model for each point in the input point cloud, rather than a single class hard label manually labeled, which includes the probability information that the point belongs to the steel reinforcement category and the concrete body category). This represents the predicted probability distribution output by the student model. Let represent the i-th point in the input point cloud, N represent the number of points in the input point cloud, c represent the c-th semantic category, and C represent the number of semantic categories. Represents the teacher model for points The predicted probability of belonging to class c. Represents student model points The predicted probability of belonging to class c. It represents the natural logarithm.

[0044] Furthermore, the formula for calculating distillation loss is:

[0045] In the formula, This represents the distillation loss; T represents the distillation temperature coefficient, used to smooth the class probability distributions of the teacher and student models, and T is set to 1.0. This represents the temperature scaling compensation factor.

[0046] Furthermore, the formula for calculating the total loss is:

[0047] In the formula, Indicates the total loss. This indicates that the semantic segmentation loss of the student model is calculated based on manually labeled hard tags. This represents the loss weight coefficient for the second stage, used to adjust the contribution ratio of the student model and the teacher model in the second stage of training. Through inference, the semantically dense labels of the point cloud are recovered.

[0048] The above model parameter optimization process uses the AdamW optimizer, with a base learning rate of 0.006, weight decay of 0.05, and a learning rate of 0.006 for the lightweight Transformer encoder module. A OneCycleLR scheduler with a 5% warm-up ratio is used for 10,000 training cycles, resulting in the final trained TPNet.

[0049] (3) The assembly features of the segmented rebar point cloud from step (2) are extracted. Independent components are separated by density-based noisy spatial clustering (DBSCAN) algorithm. Principal component analysis (PCA) and pixel-level alignment algorithm are used to calculate and extract the radius and center three-dimensional coordinates of each rebar section, such as... Figure 3 As shown.

[0050] Specifically, for the cross-sectional features of the rebar point cloud, based on the lateral clear spacing constraints of mechanical connectors specified in the code, the DBSCAN clustering algorithm is used to cluster the rebar point cloud instances, separating independent rebar instances by setting a neighborhood radius of 25mm. For a single rebar, the PCA algorithm is used to extract the initial principal axis of the Z-axis to achieve coarse alignment along the Z-axis. A pixel-level alignment algorithm is applied, introducing a small rotation angle. Constructing a tiny rotation matrix The centroid of the rebar point cloud after coarse alignment along the Z-axis Using the rotation center, calculate the rotated point cloud. And take its X and Y coordinates as the XY plane projection coordinates. and ;in, This represents the coordinates of the i-th 3D point in the reinforcement point cloud. This represents the centroid of the rebar point cloud after coarse alignment along the Z-axis. This represents the initial rotation matrix obtained from principal component analysis (used to coarsely align the principal axes of the reinforcement point cloud to the Z-axis). This represents the small rotation angle of the secondary projection introduced on top of the coarse alignment. Indicates a small rotation angle The corresponding tiny rotation matrix, This represents the projection operator that projects three-dimensional points onto the XY plane. This represents the two-dimensional projected coordinates of the i-th point on the XY plane after coarse alignment and rotation. and They represent The projected coordinate components in the X and Y directions, with the superscript T denoteing the transpose of the vector. The projected coordinates are discretized into pixel indices according to the pixel resolution. and based on pixel index Statistical set of non-repeating occupied pixels ;in, This represents the pixel index of the i-th projection point in the discretized two-dimensional pixel domain. and These represent the minimum coordinate values ​​in the X and Y directions of the current projection point set (used to determine the lower left corner reference of the two-dimensional pixel domain). This indicates pixel discretization resolution or grid size. This represents a rounding operation, preferably rounding down, used to convert continuous projected coordinates into discrete pixel indices. The covered two-dimensional mapping region is minimized using Powell's derivative-free optimization method. To maximize its verticality, obtain Determine the optimal rotation angle This is to ensure that the axis of the reinforcing bar is parallel to the Z-axis; where, This represents the optimal rotation angle that minimizes the number of pixels occupied by the projection of the steel reinforcement point cloud. Indicates a small rotation angle The following is a set of non-repeating occupied pixels formed by all projected points. This indicates the number of non-repeating pixels; the Powell method is used to search within a preset angle range for... smallest The optimal rotation angle is determined by the fact that the closer the reinforcing bar axis is to the Z-axis, the closer its XY plane projection area is to a circular cross-section, occupying fewer pixels. Therefore, the highest verticality can be obtained based on this. Random Sampling Consensus (RANSAC) circle fitting is performed on the precisely aligned 2D projection points to extract the effective radius of the reinforcing bar and the center coordinates of the top and root sections.

[0051] (4) Extract the assembly features of the sleeve from the single-station 3D point cloud data of the sleeve end collected in step (1), and calculate and extract the cross-sectional radius and center 3D coordinates of each sleeve, such as... Figure 4 As shown.

[0052] Specifically, for the single-station 3D point cloud data of the sleeve end collected in step (1), after Z-axis alignment of the sleeve point cloud through principal component analysis, the point cloud is rotated in the XY plane so that the longest side of the convex hull is parallel to the X-axis, achieving orthogonal axis alignment; the Z-axis coordinates are discarded and the two-dimensional projection points of the sleeve point cloud on the XY plane are obtained. An adaptive pixel mapping strategy is introduced to rasterize the projection point coordinates into a two-dimensional binary image with a resolution of 1 pixel / 3 mm. Let the rasterized two-dimensional binary image be... The structuring element is B. For a two-dimensional binary image, Gaussian blur is first applied to smooth local discrete noise, followed by morphological opening and closing operations for image denoising. The opening operation represents erosion followed by dilation, used to remove isolated noise points and small burrs; the closing operation represents dilation followed by erosion, used to fill small holes and boundary breaks within the sleeve region. For the denoised image, a blob detection algorithm (such as SimpleBlobDetector) is then used to identify sleeve candidate regions. Based on the rasterization mapping relationship, the identified sleeve candidate regions are back-localized to the original point cloud coordinate system, forming sleeve candidate point cloud regions. Specifically, for the denoised binary image... Connectivity component labeling is performed, and each connected component is used as a candidate blob region. And calculate its area. Circumscribed rectangle dimensions and roundness and center pixel coordinates ,in The perimeter of the connected region boundary is defined. Based on the preset area range, aspect ratio range, and circularity threshold of the sleeve in the 2D projection, non-sleeve spots are screened out, and connected regions that meet the conditions are retained as sleeve regions. For the retained sleeve regions, according to the mapping relationship between pixel coordinates and planar coordinates during rasterization, the pixel region is reverse-mapped to a candidate point set in the XY plane, thus forming a sleeve candidate point cloud region. For the sleeve candidate point cloud region, a ring mask is further generated by combining the sleeve center and inner and outer radius constraints, i.e., retaining those that meet the conditions. The point cloud points are used to obtain the corresponding annular mask for the sleeve, where Indicates the inner radius of the sleeve. Indicates the outer radius of the sleeve. The coordinates of the sleeve's center point are represented. Clustering is performed at the edge of the annular mask using the DBSCAN algorithm, with a radius of 25mm consistent with the reinforcing bar, to separate independent sleeve instances. To avoid local anomalies introduced by random unevenness in the point cloud, polar coordinate angle binning is applied to select the boundary feature points for each sleeve instance. Specifically, the projection points of the sleeve cluster are converted to polar coordinates relative to the initial center, the polar angle and radial distance are calculated, and the cluster is divided into multiple angle bins according to a preset angle interval. The point with the largest or smallest radial distance in each non-empty angle bin is selected as the outer or inner boundary point of the sleeve, i.e., the boundary feature point, thus forming a boundary point set, which can be represented as:

[0053]

[0054]

[0055] In the formula, Let E represent the set of point indices within the k-th angle bin, j represent the point indices in the sleeve candidate point cloud, E represent the set of point indices for the current sleeve cluster or sleeve candidate point set, and K represent the set of points to be added. The total number of boxes obtained by dividing the angle range. This represents the polar angle of the j-th point relative to the center of the sleeve cluster. This represents the index of the point selected as the boundary feature point in the k-th angle bin. Represents the two-dimensional or three-dimensional coordinates of the j-th point. This indicates the center coordinates or centroid coordinates of the current sleeve cluster. This indicates the distance from the j-th point to the center of the sleeve cluster. The Euclidean distance (i.e., radial distance). This represents the coordinates of the final selected boundary feature point in the k-th angle bin.

[0056] Finally, the RANSAC algorithm is used to perform circle fitting on the extracted boundary feature points to extract high-precision geometric parameters for each sleeve, namely the cross-sectional radius and the three-dimensional coordinates of the center. Specifically, three non-collinear boundary points are randomly selected to determine candidate circle models, and the radial residual from the boundary point to the candidate circle is less than a threshold as the inlier determination criterion. Finally, the circle model with the most inliers and the smallest residual is selected as the sleeve fitting result, thus obtaining the cross-sectional radius and center coordinates of the sleeve.

[0057] (5) Based on the rebar assembly features extracted in step (3) and the sleeve assembly features extracted in step (4), the ideal assembly coordinates of the rebar after correction (i.e., after correcting bending deformation) are calculated using a virtual pre-assembly algorithm that considers rebar position correction. Global pose alignment of the rebar and sleeve is achieved through coordinate system translation and rotation matrices to obtain global assembly feasibility results, such as... Figure 5As shown.

[0058] In this embodiment, the virtual pre-assembly algorithm considering rebar position correction specifically includes: First, the displacement of the rebar root and top caused by the on-site clamp adjustment is regarded as linear bending moment and tangential mechanical constraint. The center coordinate of the rebar root is regarded as a fixed support, and the top is regarded as a free end subjected to concentrated force and end bending moment. Based on the linear elastic small deflection assumption, the lateral displacement correction of rebars at each height in the X and Y directions is calculated in combination with the rebar bending deformation equation to correct the originally detected rebar center coordinates and obtain the ideal assembly coordinates after rebar correction. The rebar bending deformation equation can be expressed as:

[0059] In the formula, The x-coordinate represents the deflection at a point on the reinforcing bar (i.e., the y-coordinate), and L represents the total length of the reinforcing bar. This is the root deviation angle.

[0060] Then, the lower left corner reference instances in the rebar group and sleeve group are determined respectively, and the center coordinates of the corresponding reference instances are extracted. A translation vector is calculated based on the center coordinates of the two reference instances, and the entire rebar group is translated according to the translation vector so that the center of the lower left corner reference rebar coincides with the center of the lower left corner reference sleeve, thus achieving rough assembly. Specifically, the lower left corner reference instance refers to the rebar instance or sleeve instance closest to the minimum X and minimum Y directions in the XY plane projection of the unified assembly coordinate system. This is specifically determined by the coordinates of the center points of each instance (i.e., each rebar instance and each sleeve instance). Calculate the discriminant value in the lower left corner. and take The smallest instance determines the lower left reference instance of the reinforcement group and sleeve group; where, This represents the coordinates of the center point of the h-th instance. This represents the three-dimensional coordinates of the center point of the h-th instance. This represents the bottom-left discriminant value of the h-th instance. and These represent the minimum and maximum values ​​of the instance's X-axis coordinate, respectively. and These represent the minimum and maximum values ​​of the Y-direction coordinates of the instance, respectively. Let the center coordinates of the lower left corner reference reinforcement in the reinforcement group be... The center coordinates of the reference sleeve in the lower left corner of the sleeve group are: Then the translation vector is .

[0061] Subsequently, the coordinates of the upper left and lower right corners are introduced to calculate the direction vector of the reinforcement group. and the direction vector of the sleeve group And calculate the rotation matrix based on the direction vectors of the two. ,in , It has no practical significance. , According to the rotation matrix Transform the steel reinforcement group to the sleeve group coordinate system to achieve precise assembly.

[0062] Finally, the center deviation, axial angle deviation, and radial clearance deviation between each rebar and its corresponding sleeve after precision assembly are calculated. Each deviation value is then compared to a preset assembly allowable deviation threshold. If all deviation values ​​are within the allowable assembly deviation threshold range, the overall assembly is deemed feasible; otherwise, an assembly risk or inassembly is deemed impossible, thus obtaining the overall assembly feasibility result. The above process can be expressed by the following formula:

[0063]

[0064]

[0065]

[0066]

[0067]

[0068] In the formula, This represents the center point of the i-th rebar after rough assembly. Indicates the center point of the corresponding sleeve. Let K represent the rotation matrix and K represent the translation vector. Indicates the center point of the rebar after precision assembly. Indicates center distance deviation. Indicates the angular deviation of the axis. and These represent the axial directions of the reinforcing bars and the sleeve after precision assembly, respectively. and These represent the center deviation threshold and the angle deviation threshold, respectively. and These represent the radius of the reinforcing bar and the inner radius of the sleeve, respectively.

[0069] (6) Transmit the global assembly feasibility results obtained in step (5) to the augmented reality device, and use spatial anchoring technology to align the three-dimensional holographic projection of the assembly state with the physical components on site to achieve virtual-real fusion visualization interaction. The results are as follows: Figure 6 As shown.

[0070] Furthermore, spatial anchor point technology is used to align the 3D holographic projection of the assembled state with the physical components on site. Specifically, this includes: First, performing plane fitting and surface cleaning on the point cloud set of the concrete main body classified in step (2) to generate a 3D prior anchor point model. Specifically, statistical filtering and radius filtering are performed on the point cloud set of the concrete main body to remove outliers and floating noise points; the RANSAC plane fitting algorithm is used to extract the main surface plane from the concrete point cloud to obtain the plane equation, which is the fitting plane; the local coordinate system of the anchor point model is established according to the normal vector of the fitting plane, where the normal direction of the concrete surface is used as the local Z-axis, the edge direction of the concrete component or the main direction obtained by principal component analysis is used as the local X-axis, and the local Y-axis is determined by the right-hand coordinate system; the cleaned concrete surface point cloud, the fitting plane, the boundary contour points and the local coordinate system together constitute the 3D prior anchor point model. Then, the 3D prior anchor point model is input into the Azure Object Anchors service of the augmented reality device. Combined with the depth data collected by the augmented reality device and the SLAM (Simultaneous Localization and Mapping) map, feature-based model-sensor alignment is performed to restore the six-degree-of-freedom (6-DoF) pose of the physical component in the augmented reality world coordinate system. Subsequently, spatial anchor points are instantiated at the six-DoF pose, and the virtual pre-assembly deviation information and corresponding holographic world obtained in step (5) are locked onto the actual component surface. The virtual pre-assembly deviation information includes the center deviation, axial angle deviation, and radial clearance deviation between the reinforcing bar and the corresponding sleeve, as well as the global assembly feasibility result determined based on the deviation information. By locking the virtual pre-assembly deviation information of the reinforcing bar and sleeve obtained in step (5) and its global assembly feasibility result at a scaled-down real position within the device's field of view as 3D feature data, it is convenient for operators to review correction parameters and make on-site decisions without drawings. The three-dimensional feature data includes: the center coordinates of the reinforcing bar, the center coordinates of the sleeve, the radius of the reinforcing bar, the inner radius of the sleeve, the planar offset between the center of the reinforcing bar and the corresponding center of the sleeve, the radial clearance deviation between the radius of the reinforcing bar and the inner radius of the sleeve, the allowable deviation threshold corresponding to each deviation item, the global assembly feasibility result, and the deviation annotation position and color coding information used for augmented reality display.

[0071] In one embodiment, the global assembly feasibility result may include three states: qualified, warning, and out of tolerance. Qualified indicates that all deviation values ​​are within the allowable deviation range; warning indicates that at least one deviation value is close to but does not exceed the allowable deviation threshold; and out of tolerance indicates that at least one deviation value exceeds the allowable deviation threshold.

[0072] In summary, this invention acquires 3D point cloud data of precast concrete components using a single-station laser scanner; constructs a lightweight point cloud segmentation network TPNet to perform semantic segmentation on the acquired single-station point cloud to separate the reinforcing steel from the concrete body; employs an improved assembly feature extraction method to extract the center and radius features of the reinforcing steel and sleeve respectively; constructs a virtual pre-assembly algorithm considering reinforcing steel position correction to detect the assembly results; and transmits the assembly detection results to an augmented reality (AR) device for 3D visualization projection on the actual site using spatial anchor points. This invention, through a lightweight network, an improved high-precision feature extraction algorithm, and the interactive presentation of AR devices, achieves near real-time on-site processing of single-station point clouds, effectively overcoming the time-consuming virtual pre-assembly of traditional large-scene point clouds and the information gap caused by detachment from the actual site conditions. This significantly improves the detection efficiency of precast concrete component assembly and the quality traceability throughout its lifecycle.

[0073] Corresponding to the aforementioned embodiments of the on-site virtual pre-assembly method for prefabricated components based on single-station scanning and augmented reality, the present invention also provides embodiments of an on-site virtual pre-assembly device for prefabricated components based on single-station scanning and augmented reality.

[0074] See Figure 7 The present invention provides a virtual pre-assembly device for prefabricated components based on single-station scanning and augmented reality, comprising one or more processors and a memory, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the virtual pre-assembly method for prefabricated components based on single-station scanning and augmented reality in the above embodiments.

[0075] The embodiment of the prefabricated component on-site virtual pre-assembly device based on single-station scanning and augmented reality of this invention can be applied to any device with data processing capabilities, such as a computer. The device embodiment can be implemented through software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 7 The diagram shown is a hardware structure diagram of any device with data processing capabilities, which is the prefabricated component on-site virtual pre-assembly device based on single-station scanning and augmented reality according to the present invention. Except for... Figure 7 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0076] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0077] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0078] This invention also provides a computer-readable storage medium storing a program that, when executed by a processor, implements the on-site virtual pre-assembly method for prefabricated components based on single-station scanning and augmented reality as described in the above embodiments.

[0079] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0080] 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.

Claims

1. A method for on-site virtual pre-assembly of prefabricated components based on single-station scanning and augmented reality, characterized in that, Includes the following steps: (1) Use a laser scanner to collect single-station three-dimensional point cloud data of the steel bar end and sleeve end in the precast concrete component at the construction site; (2) Input the single-station 3D point cloud data of the steel bar end into the trained lightweight point cloud segmentation network for semantic segmentation, separate the steel bar from the concrete body, and obtain the point cloud set of the steel bar and the concrete body. (3) Perform assembly feature extraction on the segmented rebar point cloud to obtain the radius and center three-dimensional coordinates of each rebar section; (4) Extract the assembly features of the sleeve from the single-station three-dimensional point cloud data of the sleeve end, and obtain the cross-sectional radius and center three-dimensional coordinates of each sleeve; (5) Based on the rebar assembly features and sleeve assembly features, the ideal assembly coordinates after rebar correction are calculated using a virtual pre-assembly algorithm that considers rebar position correction. The global pose alignment of the rebar and sleeve is achieved through coordinate system translation and rotation matrices to obtain the global assembly feasibility results. (6) Transmit the global assembly feasibility results to the augmented reality device, and use spatial anchor point technology to align the three-dimensional holographic projection of the assembly state with the physical components on site to achieve virtual-real fusion visualization interaction.

2. The method for on-site virtual pre-assembly of prefabricated components based on single-station scanning and augmented reality according to claim 1, characterized in that, The lightweight point cloud segmentation network employs a lightweight point cloud Transformer network for single-site point cloud semantic segmentation. Designed based on PointTransformer v3, the network specifically includes an input embedding module, a lightweight Transformer encoding module, a feature propagation and decoding module, and a point-by-point classification head. The input embedding module receives the voxelized, normalized, and tensorized single-site point cloud tensor of the rebar end and maps the spatial coordinates, color, or reflection intensity attributes of each point to high-dimensional point features, serving as the input to the lightweight Transformer encoding module. The lightweight Transformer encoding module consists of several levels of sparse convolutional conditional location encoding layers, local attention mechanism layers, and multilayer perceptron feedforward layers, scaled according to channel ratios. The number of channels in each layer is compressed; the feature propagation decoding module is used to upsample the multi-scale point features output by the lightweight Transformer encoding module step by step and restore them to the original point resolution; the point-by-point classification head includes a linear layer, a normalization layer and a Softmax layer, which are used to output the probability of each point belonging to the category of steel reinforcement or concrete body.

3. The method for on-site virtual pre-assembly of prefabricated components based on single-station scanning and augmented reality according to claim 1, characterized in that, The training method for the lightweight point cloud segmentation network specifically includes: The single-station 3D point cloud data of the rebar end is voxelized, normalized, and converted into a single-station point cloud tensor of the rebar end; The single-site point cloud tensor of the rebar end is input into a lightweight point cloud segmentation network, and the output is a point-by-point semantic label probability matrix that corresponds one-to-one with the input point. Phase 1: Construct an augmentation model based on a lightweight point cloud segmentation network; calculate the semantic segmentation loss of the micro-branch and the augmentation branch based on the predicted semantic label probabilities output by the lightweight point cloud segmentation network and the augmentation model and their corresponding real semantic labels, and obtain the joint loss by weighted summation; adjust the model parameters of the lightweight point cloud segmentation network with the optimization objective of minimizing the joint loss until the preset number of training rounds is reached, and obtain a pre-trained lightweight point cloud segmentation network; The second stage involves using the pre-trained optimal PointTransformer v3 as the teacher model and the initially trained lightweight point cloud segmentation network as the student model. Distillation loss is calculated based on the outputs of the teacher and student models. Semantic segmentation loss is calculated based on the predicted semantic label probabilities and corresponding ground truth semantic labels output by the student model. The total loss is calculated based on the semantic segmentation loss and distillation loss of the student model. The model parameters of the student model are adjusted with the goal of minimizing the total loss until the preset number of training epochs is reached, resulting in the final trained lightweight point cloud segmentation network.

4. The method for on-site virtual pre-assembly of prefabricated components based on single-station scanning and augmented reality according to claim 3, characterized in that, The formula for calculating the joint loss is: In the formula, Denotes the joint loss function. This represents the semantic segmentation loss of the augmentation branch. This represents the semantic segmentation loss for micro-branches. This represents the loss weighting coefficient for the first stage; The formula for calculating the distillation loss is as follows: In the formula, Indicates distillation loss, Indicates the output distribution of the teacher model With student model output distribution The KL divergence loss between them, where T represents the distillation temperature coefficient, This represents the temperature scaling compensation factor; The formula for calculating the KL divergence loss is as follows: In the formula, This represents the Softmax soft label probability distribution output by the teacher model. This represents the predicted probability distribution output by the student model. Let represent the i-th point in the input point cloud, N represent the number of points in the input point cloud, c represent the c-th semantic category, and C represent the number of semantic categories. Represents the teacher model for points The predicted probability of belonging to class c. Represents student model points The predicted probability of belonging to class c. Represents the natural logarithm; The formula for calculating the total loss is as follows: In the formula, Indicates the total loss. This indicates that the semantic segmentation loss of the student model is calculated based on manually labeled hard tags. This represents the loss weighting coefficient for the second stage.

5. The method for on-site virtual pre-assembly of prefabricated components based on single-station scanning and augmented reality according to claim 1, characterized in that, Step (3) specifically includes: A density-based noisy spatial clustering algorithm is used to cluster the rebar point cloud into instances and separate independent rebar instances. For a single rebar, the initial principal axis of the Z-axis is extracted using the principal component analysis algorithm to achieve coarse alignment of the Z-axis; A pixel-level alignment algorithm is applied, and a small rotation matrix is ​​constructed by introducing a small rotation angle. The centroid of the rebar point cloud after coarse alignment along the Z-axis is used as the rotation center. The rotated point cloud is calculated, and its X and Y coordinates are taken as the XY plane projection coordinates. The projection coordinates are discretized into pixel indices according to the pixel resolution, and the set of non-repeating occupied pixels is counted based on the pixel indices. The coverage of the two-dimensional mapping area is minimized by Powell's derivative-free optimization method to determine the optimal rotation angle, so that the rebar axis is parallel to the Z-axis, thus achieving pixel-level fine alignment. Random sampling and uniform circle fitting are performed on the precisely aligned two-dimensional projection points to extract the effective radius of the reinforcing bars and the center coordinates of the top and root sections.

6. The method for on-site virtual pre-assembly of prefabricated components based on single-station scanning and augmented reality according to claim 1, characterized in that, Step (4) specifically includes: For the single-station 3D point cloud data of the sleeve end, after Z-axis alignment of the sleeve point cloud by principal component analysis, the point cloud is rotated in the XY plane so that the longest side of the convex hull is parallel to the X-axis, thus achieving orthogonal axis alignment; the Z-axis coordinates are discarded and the two-dimensional projection points of the sleeve point cloud on the XY plane are obtained. An adaptive pixel mapping strategy is introduced to rasterize the projection point coordinates into a two-dimensional binary image according to the resolution; Image denoising is performed on two-dimensional binary images by applying Gaussian blur, as well as morphological opening and closing operations. Based on the denoised image, a speckle detection algorithm is used to identify sleeve candidate regions, and the identified sleeve candidate regions are back-located to the original point cloud coordinate system according to the rasterization mapping relationship to form sleeve candidate point cloud regions. For the candidate point cloud region of the sleeve, an annular mask is generated by combining the sleeve center and inner and outer radius constraints, and a density-based noisy spatial clustering algorithm is used at the edge of the annular mask to cluster the independent sleeve instances. For each sleeve instance, its boundary feature points are selected by applying polar coordinate angle binning; The boundary feature points are fitted with a circle using a random sampling consensus algorithm to extract the cross-sectional radius and center three-dimensional coordinates of each sleeve.

7. The method for on-site virtual pre-assembly of prefabricated components based on single-station scanning and augmented reality according to claim 1, characterized in that, In step (5), the virtual pre-assembly algorithm that considers the correction of the position of the reinforcing bars specifically includes: The center coordinates of the root of the rebar are regarded as fixed support and the top is regarded as free end. Based on the assumption of linear elastic small deflection, the lateral displacement correction of the rebar at each height in the X and Y directions is calculated in combination with the rebar bending deformation equation to correct the original detected center coordinates of the rebar and obtain the ideal assembly coordinates of the rebar after correction. Determine the lower left corner reference instance in the rebar group and sleeve group respectively, and extract the center coordinates of the corresponding reference instance to calculate the translation vector. Based on the translation vector, translate the entire rebar group so that the center of the lower left corner reference rebar coincides with the center of the lower left corner reference sleeve, thus achieving rough assembly. By introducing the coordinates of the upper left and lower right corners, the direction vectors of the reinforcement group and the sleeve group are calculated, and the rotation matrix is ​​calculated based on the direction vectors of the two. The reinforcement group is then transformed into the sleeve group coordinate system based on the rotation matrix to achieve fine assembly. After fine assembly, calculate the center deviation, axial angle deviation, and radial clearance deviation between each steel bar and the corresponding sleeve, and compare each deviation value with the preset assembly allowable deviation threshold. If all deviation values ​​are within the assembly allowable deviation threshold range, the overall assembly is deemed feasible; otherwise, it is deemed that there is an assembly risk or that assembly is not possible, thus obtaining the overall assembly feasibility result.

8. The method for on-site virtual pre-assembly of prefabricated components based on single-station scanning and augmented reality according to claim 1, characterized in that, The method of aligning the assembled 3D holographic projection with the physical components on site using spatial anchoring technology specifically includes: Perform plane fitting and surface cleaning on the point cloud set of the concrete main body classified in step (2) to generate a three-dimensional prior anchor point model; The 3D prior anchor point model is input into the Azure Object Anchor service of the augmented reality device. The feature-based model-sensor alignment is performed by combining the depth data collected by the augmented reality device with the SLAM map to restore the six-degree-of-freedom pose of the physical components in the augmented reality world coordinate system. Instantiate a spatial anchor point at a six-degree-of-freedom pose and lock the virtual pre-assembly deviation information and the corresponding holographic world obtained in step (5) onto the surface of the actual component; wherein, the virtual pre-assembly deviation information includes the center deviation, axial angle deviation and radial clearance deviation between the steel bar and the corresponding sleeve, as well as the global assembly feasibility result obtained based on the deviation information.

9. A virtual pre-assembly device for prefabricated components based on single-station scanning and augmented reality, comprising one or more processors and a memory, characterized in that, The memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the on-site virtual pre-assembly method for prefabricated components based on single-station scanning and augmented reality as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, It stores a program that, when executed by a processor, is used to implement the on-site virtual pre-assembly method for prefabricated components based on single-station scanning and augmented reality, as described in any one of claims 1-8.

Images