A lightweight detection method, device and storage medium for building components based on OneFormer3D

Abstract

The present invention discloses a lightweight detection method, equipment and storage medium for building components based on OneFormer3D, which uses a ground-based three-dimensional laser scanner to collect three-dimensional point cloud data inside the building. The point cloud data inside the building is denoised and annotated to complete data preparation. A deep learning model based on OneFomer3D network structure optimization is used to obtain instance information of building components. Building components with instance information are detected using machine learning methods and compared with relevant specifications to achieve lightweight detection. The present invention can quickly and accurately complete the detection task of building components through a deep learning method based on point cloud data. At the same time, the method has the advantages of high degree of automation and low cost, is suitable for quality control and maintenance of large-scale construction projects, and has broad application prospects.

Description

Technical Field

[0001] The present invention relates to a OneFormer3D-based lightweight building component detection method, device and storage medium, and belongs to the technical field of civil engineering and artificial intelligence interaction. Background Art

[0002] Currently, the detection of construction dimensional deviations in engineering projects mostly relies on manual measurement using traditional tools such as steel tape measures, straightedges, feeler gauges, theodolites, and levels. Although automated equipment such as laser rangefinders and laser levels have improved measurement efficiency to a certain extent, their point-by-point measurement mode still relies on manual operation, limiting further efficiency improvements.

[0003] To improve the efficiency and accuracy of construction quality inspections, 3D laser scanning technology has been widely used due to its ability to quickly and accurately capture the surface shape of objects and convert it into point cloud data. Research has already achieved intelligent dimensional quality inspection for precast concrete components, steel structures, and cast-in-place concrete components. However, existing point cloud-based inspection technologies still fall short of the requirements of engineering applications. While some research teams have made progress in inspecting the flatness of precast concrete components and the size and position of rebar, the accuracy still falls short of engineering requirements.

[0004] Deep learning algorithms offer the advantage of broad applicability in component identification. Networks such as PointNet++ and PointConv have been applied to component classification and identification, but their accuracy rates generally fall below 40%. While some research has achieved promising detection results, existing studies primarily focus on planar components such as floor slabs and wall panels, largely ignoring the complex environment of construction sites. Deep learning algorithms also suffer from uneven classification accuracy, requiring further refinement to enhance their practical application.

[0005] Therefore, in order to further improve the efficiency and accuracy of construction quality inspection, it is urgent to research more intelligent and efficient construction dimension quality inspection technology. Summary of the Invention

[0006] Objective: To overcome the shortcomings of the existing technology, the present invention provides a lightweight detection method, equipment and storage medium for building components based on OneFormer3D, which can automatically identify building components, reduce labor costs and significantly improve engineering efficiency.

[0007] Technical solution: To solve the above technical problems, the technical solution adopted by the present invention is:

[0008] First, a lightweight detection method for building components based on OneFormer3D, specifically comprising:

[0009] Step S1: Acquire point cloud data inside the building.

[0010] Step S2: De-noise and simplify the point cloud data inside the building to obtain pre-processed point cloud data.

[0011] Step S3: Classify the pre-processed point cloud data by region, and label the building component types in the point cloud data in each region, so as to obtain the building components in each region and the labels of each building component as the point cloud data set.

[0012] Step S4: Use the deep learning model based on the Onefomer3D network structure optimization to train the point cloud dataset. By setting different training hyperparameter modules for multiple trainings, the best training hyperparameter module is obtained. The deep learning model is configured with the best training hyperparameter module for training to obtain a trained deep learning model. The point cloud data to be predicted is input into the trained deep learning model to obtain the instance information point cloud data inside the building.

[0013] Step S5: performing post-processing optimization on the instance information point cloud data to obtain optimized instance information point cloud data.

[0014] Step S6: Using a multi-plane fitting method based on random consistency and a directed envelope box algorithm based on principal component analysis on the optimized instance information point cloud data, the size information and spatial position information of the building components corresponding to the optimized instance information point cloud data are obtained.

[0015] Step S7: Compare the size information and position information of the building components with the design specifications to complete the inspection of the building components.

[0016] As a preferred solution, the point cloud data inside the building in step S1 meets the following requirements: the data density of the point cloud data contains at least 100 points per square meter, and shall not be less than 50 points per square meter under special circumstances or requirements; the spatial range covers the entire target area, including the area within 30 meters outside the boundary line; the file storage format is E57 file format, the scanning resolution is 8000 points per meter, and the ranging error is ≤±1mm at 25 meters; the point cloud data contains RGB color information.

[0017] As a preferred solution, the denoising process uses Gaussian filtering, and the simplification uses an Octree tree structure to perform 12-level downsampling simplification on the point cloud data.

[0018] As a preferred solution, the annotated objects of the point cloud dataset also include on-site objects.

[0019] As a preferred embodiment, the deep learning model specifically includes: the output of the downsampling module is respectively connected to the second input of the mask module, the second input of the instance module, and the input of the superpoint extraction module; the output of the superpoint extraction module is respectively connected to the first input of the mask module, the first input of the decoder, and the first input of the instance module; the output of the instance module is connected to the second input of the decoder, the output of the mask module is connected to the first input of the prediction head, and the output of the decoder is connected to the second input of the prediction head.

[0020] The downsampling module includes a first fully connected layer with a dimension of 6x32, a farthest point sampling layer, a K nearest neighbor layer, a second fully connected layer with a dimension of 32x16, and a maximum pooling layer, which are connected in sequence.

[0021] The farthest point sampling layer is used to iteratively calculate the distance between each point in the current point set and the center of the existing point cloud model, and select the N points with the farthest distance as a new subset.

[0022] The K nearest neighbor layer is used to capture the local structural information of the point set.

[0023] The superpoint extraction module is used to voxelize the input point cloud, convert it into a sparse three-dimensional convolution format, and extract point-by-point features.

[0024] The mask module is used to integrate the residual mask feature vector output by the downsampling module and the feature vector output by the superpoint extraction module to perceive the global feature.

[0025] The instance module is used to obtain a query vector with instance features from the downsampling module and the superpoint extraction module.

[0026] The decoder is used to convert the high-dimensional query vector into a low-dimensional feature representation.

[0027] The prediction head is used to convert the output mask feature perception information and instance mask into a final prediction result and evaluate the quality of each prediction.

[0028] As a preferred solution, the training hyperparameter module configuration includes: number of threads, learning rate, training batch size, number of iterations, weight decay, optimizer and learning rate scheduler.

[0029] As a preferred solution, the post-processing optimization adopts a statistical filtering algorithm and a cluster analysis method.

[0030] As a preferred solution, step S6 specifically includes:

[0031] S6-1: Building components are divided into three-dimensional components and planar components according to their point cloud scanning features.

[0032] S6-2: For a three-dimensional component with multiple planes, a multi-plane fitting method based on random consistency is used to fit each plane and calculate the distance between parallel planes to obtain the size information of the three-dimensional component.

[0033] S6-3: For a planar component having only a single plane, obtain the size information of the planar component by measuring the size and center coordinates of the directed bounding box.

[0034] In a second aspect, a computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements a lightweight detection method for building components based on OneFormer3D as described in any one of the first aspects.

[0035] According to a third aspect, a computer device includes:

[0036] Memory, used to store instructions.

[0037] The processor is configured to execute the instructions so that the computer device performs an operation according to any one of the OneFormer3D building component lightweight detection methods described in the first aspect.

[0038] Beneficial Effects: This invention provides a lightweight building component inspection method, device, and storage medium based on OneFormer3D. By leveraging deep learning methods based on point cloud data, this method can quickly and accurately inspect building components. Furthermore, this method boasts high automation and low cost, making it suitable for quality control and maintenance in large-scale construction projects and possessing broad application prospects.

[0039] Compared with existing technologies, its advantages are as follows:

[0040] 1. This invention fills the gap in the existing point cloud dataset in the field of building completion.

[0041] 2. This invention uses 3D laser scanners and deep learning technologies, such as the OneFormer3D model, to rapidly capture and process point cloud data of building interiors. Compared to current measurement methods, this automated process significantly reduces data collection and analysis time, significantly improving the efficiency of construction dimensional inspections.

[0042] 3. The present invention enhances the ability to identify complex structures through precise size detection and efficient data processing, ensuring higher data accuracy and consistency of detection results.

[0043] 4. The present invention helps the construction party and the supervisor to make decisions quickly during the acceptance stage, and also improves the quality of the entire project. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a flow chart of the lightweight detection method of building components based on OneFormer3D.

[0045] Figure 2 This is a schematic diagram of the Faro Focus Premium 150 terrestrial 3D laser scanner used for acquisition in an example.

[0046] Figure 3 The present invention is a schematic diagram of an on-site picture of a building component point cloud collected by a terrestrial 3D laser scanner under an example.

[0047] Figure 4 This is a schematic diagram of a building point cloud on a certain floor after denoising in an example.

[0048] Figure 5 This is a schematic diagram of the pre-processed point cloud after annotation in an instance. Figure 5 The middle picture (a) is an unlabeled point cloud. Figure 5 The middle figure (b) is the annotated point cloud.

[0049] Figure 6 This is a network diagram of a deep learning model based on Onefomer3D network structure optimization.

[0050] Figure 7 It is a schematic diagram of instance segmentation of point cloud area under an instance. Figure 7 The middle picture (a) is the point cloud before segmentation. Figure 7 The middle figure (b) is the point cloud after instance segmentation.

[0051] Figure 8 This is a schematic diagram of model hyperparameter configuration under an embodiment.

[0052] Figure 9 It is a schematic diagram of the fitting plane of the beams and columns in the point cloud area after multi-plane fitting in an example.

[0053] Figure 10 This is a schematic diagram of the ceiling, floor, and wall of the point cloud area after being processed with a directed envelope box in an example. DETAILED DESCRIPTION

[0054] The following is a clear and complete description of the technical solutions in the examples of the present invention, in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative work are within the scope of protection of the present invention.

[0055] The present invention will be further described below with reference to specific embodiments.

[0056] Example 1:

[0057] This embodiment introduces a lightweight building component detection method based on OneFormer3D, which relies on a deep learning neural network framework and point cloud processing technology to achieve lightweight building component detection. The method specifically includes the following steps:

[0058] Step S1: Use a terrestrial 3D laser scanner to collect point cloud data inside the building.

[0059] Step S2: De-noise and simplify the point cloud data inside the building to obtain pre-processed point cloud data.

[0060] Step S3: Classify the preprocessed point cloud data by region, and annotate the building component types in the point cloud data in each region to obtain the building components and annotations of each building component in each region as a point cloud dataset. The building components include but are not limited to walls, beams, columns, slabs, doors or windows.

[0061] Step S4: Build a deep learning model based on the Onefomer3D network structure optimization, train the point cloud dataset obtained in step S3, and perform multiple trainings by setting different training hyperparameter modules to obtain the best training hyperparameter module. Configure the deep learning model with the best training hyperparameter module for training to obtain a trained deep learning model, and save the best training hyperparameters. Input the point cloud data to be predicted into the trained deep learning model to obtain instance information point cloud data inside the building.

[0062] Step S5: Post-processing and optimizing the instance information point cloud data of step S5 to obtain optimized instance information point cloud data, thereby further improving data quality.

[0063] Step S6: For the optimized instance information point cloud data in step S5, a multi-plane fitting method based on random consensus (RANSAC) and an oriented bounding box (OBB) algorithm based on principal component analysis are used to obtain the size information and spatial position information of the building components corresponding to the optimized instance information point cloud data.

[0064] Step S7: By combining the size information and position information of the building components obtained in step S6 and comparing them with the design specifications, the lightweight detection of the building components can be completed.

[0065] Furthermore, the point cloud data collected in step S1 must ensure a data density of at least 100 points per square meter, and no less than 50 points per square meter in special circumstances or under specific requirements. The spatial range must cover the entire target area, including the area within 80 meters outside the edge guideline. The file storage format must be E57, the scanning resolution must be 8000 points per meter (pt / m), and the ranging error must be ≤ ±1mm at 25 meters. The point cloud data must include RGB color information to provide richer visual details and the possibility of subsequent analysis.

[0066] Furthermore, in step S2, Gaussian filtering is used to effectively remove noise. After denoising, the point cloud data is downsampled to 12 levels using an Octree tree structure, aiming to reduce the amount of data while maintaining its structural features.

[0067] Furthermore, the point cloud dataset produced in step S3 is divided into three parts: training set, validation set, and test set in a ratio of 8:1:1 or 7:2:1. The annotated objects are building components and on-site items, such as stacked steel bars, masonry and other building materials, discarded construction waste and garbage, construction tools, etc. The point cloud dataset is saved in the format of the S3DIS instance segmentation dataset.

[0068] Furthermore, the number of point clouds in each region in step S3 does not exceed five million points.

[0069] Furthermore, the deep learning model based on the Oneformer3D network structure optimization described in step S4 uses sparse 3D convolution as the backbone, uses superpoint pooling, does not use query selection, and uses a single instance query. The backbone is pre-trained on the S3DIS dataset, the Structured3D dataset, and the ScanNet dataset.

[0070] Furthermore, the deep learning model optimization described in step S4 includes the introduction of a downsampling module and a residual connection mechanism. These improvements not only reduce computational overhead but also ensure the information integrity of the deep neural network, effectively preventing training anomalies caused by the vanishing gradient phenomenon, and improving training efficiency and accuracy.

[0071] The working principle of the entire network structure is described as follows:

[0072] Input the Nx6 point cloud dataset into the downsampling module.

[0073] Where N represents the number of input point cloud data, and 6 represents the dimensions of the point cloud data, including: x, y, z, R, G, and B. x represents the value of the x-axis in the spatial coordinate system, y represents the value of the y-axis in the spatial coordinate system, z represents the value of the z-axis in the spatial coordinate system, R represents the red value of the point cloud data, G represents the green value of the point cloud data, and B represents the blue value of the point cloud data.

[0074] The downsampling module includes a first fully connected layer with a dimension of 6x32, a farthest point sampling layer, a K nearest neighbor layer, a second fully connected layer with a dimension of 32x16, and a maximum pooling layer, which are connected in sequence.

[0075] The farthest point sampling layer is used to iteratively calculate the distance between each point in the current point set and the center of the existing point cloud model, and select the N points with the farthest distance as a new subset.

[0076] The K nearest neighbor layer is used to capture the local structural information of the point set, thereby enhancing the expressiveness and robustness of the features.

[0077] The downsampling module extracts deep features through two fully connected layers and selects a representative subset using farthest point sampling, reducing the number of points while preserving global spatial information. Subsequently, local features are aggregated using K-nearest neighbors and then mapped to a lower-dimensional space through a fully connected layer to reduce computational overhead while preserving the integrity of the contextual point cloud.

[0078] The output of the downsampling module is connected to the second input of the mask module, the second input of the instance module, and the input of the superpoint extraction module. The output of the superpoint extraction module is connected to the first input of the mask module, the first input of the decoder, and the first input of the instance module. The output of the instance module is connected to the second input of the decoder, the output of the mask module is connected to the first input of the prediction head, and the output of the decoder is connected to the second input of the prediction head.

[0079] The superpoint extraction module voxelizes the input point cloud, converts it into a sparse 3D convolutional format, and extracts point-by-point features. This module consists of a U-net-like backbone composed of sparse 3D convolutions and average pooling layers. After processing by this module, the output feature matrix undergoes various linear transformations to generate three key vectors: query vector (Query), key vector (Key), and value vector (Value). By converting the input point cloud of millions of points into just a few hundred superpoints, the computational cost of subsequent processing is significantly reduced.

[0080] The mask module is used to integrate the residual mask feature vector output by the downsampling module with the feature vector output by the superpoint extraction module to perceive global features. This module uses a multi-layer perceptron (MLP) consisting of four fully connected layers to process the outputs of these two modules. This method effectively integrates and transforms the input features, generating richer and more detailed mask feature perception information and achieving better generalization performance.

[0081] The instance module is used to obtain the query vector with instance features from the downsampling module and the superpoint extraction module. The module consists of multiple linear transformation matrices, position encoding matrices and nonlinear activation functions. The feature vectors from different modules are multiplied by the corresponding matrices to obtain the query vector; the instance query vector obtained after processing dynamically calculates and integrates the relevant information from the instance residual, so that the model can more accurately capture the complex structure and pattern in the point cloud data. During the training process, if only the query vector from the superpoint is used, there may be a phenomenon of losing some feature information. Introducing instance residuals can also ensure the integrity of the information and effectively prevent the gradient disappearance phenomenon.

[0082] The decoder is used to convert the high-dimensional query vector into a low-dimensional feature representation. This module consists of six sequential converter decoder layers, which accept the query vector from the instance module and the keys and values ​​from the super-point features for cross-attention input and convert them into feature convolution kernels. These kernels are then convolved with the super-point features, and the instance mask and residual mask are generated through the convolution operation. The instance mask is used to predict different instances in the point cloud, and the residual mask is used to capture the difference between the predicted results and the actual target. By analyzing the difference between the predicted mask and the true mask, the model can adjust its parameters more finely to better fit the training data, which helps to accelerate convergence and improve the generalization ability of the model.

[0083] The prediction head is used to convert the output mask feature perception information and instance masks into final prediction results and evaluate the quality of each prediction. This module bypasses the traditional bipartite graph matching method based on the Hungarian algorithm through decoupled matching. Based on the mask feature perception information, it directly establishes the correspondence between the instance mask and the real instance, significantly improving the efficiency of the matching process. After matching, the instance loss can be finally calculated. Classification errors are penalized using the cross-entropy loss.

[0084] The training hyperparameter module configuration of the deep learning model described in step S4 includes the following training hyperparameters: number of threads num_workers, learning rate lr, training batch size batch_size, number of iterations iteration, weight decay weight_decay, optimizer optimizer, and learning rate scheduler param_scheduler.

[0085] Furthermore, the post-processing optimization in step S5 uses a combination of statistical filtering algorithm and cluster analysis method to process the instance information point cloud data to obtain optimized instance information point cloud data, aiming to not only eliminate outliers but also accurately identify and eliminate abnormal instance labels.

[0086] Furthermore, the specific process of obtaining the size, spatial position and other information of the building components in step S6 is as follows:

[0087] S6-1: First, the building components are divided into three-dimensional components (such as columns and beams) and two-dimensional components (such as walls, panels, doors, and windows) according to their point cloud scanning features.

[0088] S6-2: For 3D components with multiple planes, a multi-plane fitting method based on random consensus (RANSAC) is used to fit each plane and calculate the distance between parallel planes. Single-plane RANSAC fitting is inefficient and ineffective when used with point cloud data with uneven density acquired by laser scanners. Multi-plane RANSAC fitting, on the other hand, uses multiple random sampling and model estimations to find the optimal set of planes, fitting multiple planes simultaneously, improving fitting accuracy.

[0089] S6-3: For planar components with only a single plane, the oriented bounding box (OBB) method can more accurately represent its boundary. The size information of the planar component can be obtained by measuring the size and center coordinates of the oriented bounding box.

[0090] Example 2:

[0091] This embodiment introduces a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the OneFormer3D-based lightweight building component detection method as described in any one of the embodiments 1 is implemented.

[0092] Example 3:

[0093] A computer device comprising:

[0094] Memory, used to store instructions.

[0095] The processor is configured to execute the instructions so that the computer device performs an operation according to any one of the OneFormer3D building component lightweight detection methods described in Example 1.

[0096] Example 4:

[0097] This embodiment introduces the application process of the OneFormer3D building component lightweight detection method according to the present invention, which specifically includes the following steps:

[0098] S1: Point cloud data collection: using Figure 2 The terrestrial laser scanner shown collects point cloud data in multiple construction scenes, including concrete structure and decoration and renovation stages, such as Figure 3 As shown, the scanning resolution is 8,000 points per meter (Pt / m). Future plans call for expansion to more construction sites.

[0099] S2: De-noise and simplify the point cloud data obtained in step S1. The processed point cloud data is divided into different areas, and each area does not exceed 5 million point clouds, such as Figure 4 shown.

[0100] S3: The point cloud data processed in step S2 is labeled as wall, column, beam, plate, door, window and other instances according to the building component category, and a point cloud dataset that can be used for quality inspection is obtained (the total number of point clouds is about 350 million). Figure 5 As shown in Figures (a) and (b), the point cloud dataset in this implementation is divided into a ratio of 7:2:1 and the dataset is saved in the S3DIS format.

[0101] S4: A deep learning model based on Onefomer3D network structure optimization was used to train point cloud datasets. The backbone was pre-trained on the S3DIS, Structured3D, and ScanNet datasets. On this basis, a downsampling module and a residual connection module were added to reduce computational overhead, ensure the integrity of the deep neural network, keep the model lightweight, and improve training efficiency and accuracy. Figure 7 (a) shows the input point cloud original dataset, Figure 7 (b) is a schematic diagram of the instance segmentation performed by the deep network. By setting different training hyperparameter modules and performing multiple training runs, the optimal hyperparameter module is determined. This hyperparameter module configuration is then used for training, and the training parameters are saved. The model training results for the point cloud training set selected in this example are shown in Table 1. The corresponding instance information can be obtained from the unlabeled point cloud after being processed by this deep learning model, as shown in Table 1.

[0102] Table 1 is a diagram of specific data for model training under an example.

[0103]

[0104] The detailed steps for configuring the hyperparameter module of the deep learning model based on Onefomer3D network structure optimization are as follows:

[0105] We used LightGBM to tune hyperparameters, using Bayesian optimization to automatically find the optimal configuration. We generated multiple hyperparameter combinations by setting the maximum depth from 5 to 50 and sampling with a step size of 5. Each combination was evaluated using k-fold cross-validation to ensure good generalization performance.

[0106] The hyperparameters that need to be adjusted include: number of threads num_workers, learning rate lr, training batch size batch_size, number of iterations iteration, weight decay weight_decay, optimizer optimizer, and learning rate scheduler param_scheduler.

[0107] In this embodiment, the training is performed on a server with an Intel(R) Xeon(R) Platinum 8336C CPU @ 2.30GHz, a NVIDIA GeForce RTX 4090 (24G) GPU, and an Ubuntu 20.04 operating system. Figure 8 As shown in .

[0108] S5: To optimize the point cloud data derived from S4 that contains some incorrect instance labels, we use advanced data processing techniques that combine statistical filtering and cluster analysis for post-processing. This not only effectively removes outliers but also accurately identifies and removes abnormal instance labels, further improving the reliability and accuracy of the data.

[0109] S6: The optimized data is divided into three-dimensional components (such as columns and beams) and planar components (such as walls, panels, doors and windows). For three-dimensional components with multiple planes, the multi-plane fitting technology based on random consensus (RANSAC) is used to fit each plane and calculate the parallel plane distance, such as Figure 9 As shown. Traditional single-plane RANSAC fitting has strong assumptions about data distribution and can only fit a single model. It is not effective and efficient when facing point cloud data with uneven density obtained by laser scanners. Multi-plane RANSAC fitting finds the optimal plane set through multiple random sampling and model estimation, and fits multiple planes at the same time, which improves the fitting accuracy and thus obtains information such as the size, spatial position and flatness of the three-dimensional component. For planar components that usually only have a single plane, using an directed bounding box (OBB) can more accurately represent its boundary. By measuring the size and center coordinates of the envelope box, the size information of the planar component can be obtained, such as Figure 10 shown.

[0110] S7: Combine the position information of all components in S6 with the specifications to achieve lightweight component detection.

[0111] This paper proposes a lightweight building component inspection method based on OneFormer3D. This method combines point cloud processing technology with a deep neural network model and can be effectively applied to building component inspection in concrete and decoration projects. The method includes collecting point cloud data, downsampling, denoising, and annotating the data to create a training dataset, optimizing a deep neural network model, and training the model using this dataset. The point cloud data to be predicted is fed into the deep learning model to obtain instance information about the building interior. The results are further processed and optimized. Using a multi-plane fitting technique based on randomized consensus (RANSAC) and an directed bounding box (OBB) algorithm, component dimensions, spatial position, and other information are obtained. This information is then compared with relevant specifications to complete lightweight inspection. This method enables efficient and accurate inspection of building components and has broad application prospects.

[0112] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A lightweight detection method for building components based on OneFormer3D, characterized by: Specifically include: Step S1: Obtaining point cloud data inside the building; Step S2: De-noising and simplifying the point cloud data inside the building to obtain pre-processed point cloud data; Step S3: Classify the pre-processed point cloud data by region, and label the building component types in the point cloud data in each region, so as to obtain each building component in each region and the label of each building component as a point cloud dataset; Step S4: Using a deep learning model optimized based on the OneFormer3D network structure, the point cloud dataset is trained. Multiple training runs are performed by setting different training hyperparameter modules to find the best training hyperparameter module. The deep learning model is configured with the best training hyperparameter module for training, resulting in a trained deep learning model. The point cloud data to be predicted is input into the trained deep learning model to obtain instance point cloud data of the building interior. Step S5: performing post-processing optimization on the instance information point cloud data to obtain optimized instance information point cloud data; Step S6: Using a multi-plane fitting method based on random consistency and a directed envelope box algorithm based on principal component analysis on the optimized instance information point cloud data, obtain the size information and spatial position information of the building components corresponding to the optimized instance information point cloud data; Step S7: Compare the size information and position information of the building components with the design specifications to complete the inspection of the building components.

2. The OneFormer3D-based lightweight building component detection method according to claim 1, characterized in that: The point cloud data inside the building in step S1 meets the following requirements: the data density of the point cloud data contains at least 100 points per square meter, and shall not be less than 50 points per square meter under special circumstances or requirements; the spatial range covers the entire target area, including the area within 30 meters outside the edge guide line; the file storage format is E57 file format, the scanning resolution is 8000 points per meter, and the ranging error is ≤±1mm at 25 meters; and the point cloud data includes RGB color information.

3. The OneFormer3D-based lightweight building component detection method according to claim 1, characterized in that: The denoising process uses Gaussian filtering, and the simplification uses an Octree tree structure to perform 12-level downsampling simplification on the point cloud data.

4. The OneFormer3D-based lightweight building component detection method according to claim 1, characterized in that: The deep learning model specifically includes: an output end of the downsampling module is respectively connected to the second input end of the mask module, the second input end of the instance module, and the input end of the superpoint extraction module; the output end of the superpoint extraction module is respectively connected to the first input end of the mask module, the first input end of the decoder, and the first input end of the instance module; the output end of the instance module is connected to the second input end of the decoder, the output end of the mask module is connected to the first input end of the prediction head, and the output end of the decoder is connected to the second input end of the prediction head; The downsampling module includes a first fully connected layer with a dimension of 6x32, a farthest point sampling layer, a K nearest neighbor layer, a second fully connected layer with a dimension of 32x16, and a maximum pooling layer connected in sequence; The farthest point sampling layer is used to iteratively calculate the distance between each point in the current point set and the center of the existing point cloud model, and select the N points with the farthest distance as the new subset; The K nearest neighbor layer is used to capture the local structural information of the point set; The superpoint extraction module is used to voxelize the input point cloud, convert it into a sparse three-dimensional convolutional format, and extract point-by-point features; The mask module is used to integrate the residual mask feature vector output by the downsampling module and the feature vector output by the superpoint extraction module to perceive the global feature; The instance module is used to obtain a query vector with instance features from the downsampling module and the superpoint extraction module; The decoder is used to convert the high-dimensional query vector into a low-dimensional feature representation; The prediction head is used to convert the output mask feature perception information and instance mask into a final prediction result and evaluate the quality of each prediction.

5. The OneFormer3D-based lightweight building component detection method according to claim 1, characterized in that: The training hyperparameter module configuration includes: number of threads, learning rate, training batch size, number of iterations, weight decay, optimizer and learning rate scheduler.

6. The OneFormer3D-based lightweight building component detection method according to claim 1, characterized in that: The post-processing optimization adopts statistical filtering algorithm and cluster analysis method.

7. The OneFormer3D-based lightweight building component detection method according to claim 1, characterized in that: The annotated objects of the point cloud dataset also include on-site objects.

8. The OneFormer3D-based lightweight building component detection method according to claim 1, characterized in that: The step S6 specifically includes: S6-1: Building components are divided into three-dimensional components and planar components according to their point cloud scanning features; S6-2: For a three-dimensional component with multiple planes, a multi-plane fitting method based on random consistency is used to fit each plane and calculate the distance between parallel planes to obtain the dimensional information of the three-dimensional component; S6-3: For a planar component having only a single plane, obtain the size information of the planar component by measuring the size and center coordinates of the directed bounding box.

9. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed by a processor, the method for detecting lightweight building components based on OneFormer3D as claimed in any one of claims 1 to 8 is implemented.

10. A computer device, characterized in that: include: a memory for storing instructions; The processor is configured to execute the instructions so that the computer device performs the operation of the OneFormer3D building component lightweight detection method according to any one of claims 1 to 8.

Images