A point cloud technology processing method for defect location in complex building scenes

Abstract

The present invention provides a point cloud technology processing method for defect location in complex building scenes. First, the collected point cloud data is preprocessed. Then, the complex building scene is divided using a semantic segmentation model to obtain independent sub-point cloud units. Feature extraction and defect location are directly performed on small-scale structures. Multi-level farthest point sampling and strategic pruning are used for large-scale structures to segment them into sub-point cloud blocks for feature extraction and defect location. A PointConv neural network is used for local and global feature extraction. Relative coordinate normalization is used to ensure the structural unbiasedness and positional robustness of the feature expression. Uncertainty factors are further introduced to enhance the feature expression. Through interpolation upsampling and feature splicing, the upper and lower layer features and confidence levels are fused step by step to achieve adaptive multi-level defect location. Finally, global defect fusion and point cloud reconstruction are performed. The solution of the present invention achieves rapid location and efficient identification of defects in complex building scenes.

Description

Technical Field

[0001] The present invention relates to the field of neural network technology in machine learning, and in particular to a point cloud technology processing method for defect positioning in complex building scenes. Background Art

[0002] Integrated construction defect detection in complex building scenarios can not only avoid safety hazards during construction, but also effectively circumvent potential safety hazards. At the same time, integrated detection technology significantly improves the efficiency of quality monitoring and enables traditional manual inspections to shift towards intelligent and automated directions.

[0003] Early building defect detection relied primarily on manual inspections, which were characterized by low efficiency, strong subjectivity, and an inability to detect hidden defects. With the development of computer vision technology, building defect detection has gradually become automated and intelligent. Initial automated detection methods used traditional image processing techniques. Researchers proposed methods based on edge detection and texture analysis. By designing artificial feature extraction operators to analyze building surface images, they achieved preliminary identification of surface defects such as cracks. However, these methods relied heavily on manually designed features, resulting in unstable and poorly robust representation of defect features, making them difficult to handle in complex real-world scenarios. Furthermore, the cost of collecting and processing the massive amount of data required to obtain a complete map was high, significantly increasing the difficulty of equipment deployment and data synchronization.

[0004] With the rise of deep learning, detection methods based on convolutional neural networks have become mainstream. Researchers use an improved fast regional convolutional neural network, or Faster R-CNN, to process 2D images of building facades. This significantly improves the accuracy of surface defect detection through end-to-end feature learning. However, this type of method is still limited to surface information obtained from visible light images and cannot effectively detect internal structural defects in buildings. To obtain richer three-dimensional information, researchers have begun to experiment with 3D laser scanning technology. The point cloud processing method based on PointNet++ achieves accurate detection of building components through layered sampling and local feature extraction. However, when dealing with large-scale and complex building scenes, this method faces problems such as low computational efficiency and susceptibility to occlusion.

[0005] Current point cloud defect detection still relies on traditional LiDAR, millimeter-wave radar, and synthetic aperture radar technologies. With such low sampling points and simple scenes, defect identification and location in complex building scenes present the following challenges:

[0006] Insufficient processing power: For traditional point cloud scenarios, the data processing scope is primarily limited by the sampling equipment. Current sampling equipment such as lidar, millimeter-wave radar, and synthetic aperture radar cannot achieve the accuracy required to capture a global point cloud scene in one go. Furthermore, traditional point cloud technology struggles to extract and process point clouds with diverse defects in complex scenarios. These limitations hinder the implementation of integrated construction defect detection.

[0007] Low processing efficiency: Traditional point cloud defect processing methods are generally based on single sampling and single measurement. For an entire building, it is necessary to collect the surfaces to be inspected one by one. On the one hand, the defect positioning accuracy is limited by the sampling coverage, and on the other hand, the defect positioning efficiency is limited by the sampling rate. In addition, when faced with complete collected data, traditional methods lack the ability to consistently locate defects in complex scenarios.

[0008] Lack of comprehensive adaptation to the complexity and diversity of building structures: Current integrated construction defect detection systems have limited adaptability to different types of building structures (such as curved surfaces and irregular geometric shapes). For example, for special scenarios such as curved curtain walls, long-span bridges, or prefabricated buildings, the existing algorithms are not versatile enough and difficult to accurately model and detect. This leads to the need for a large number of customized adjustments in actual projects, increasing the difficulty and cost of deployment. Summary of the Invention

[0009] To address the above problems, the present invention proposes a point cloud technology processing method for defect location in complex architectural scenes. This method adopts spatial structure recognition based on semantic segmentation, step-by-step adaptive hierarchical feature learning, and a multi-strategy fusion mechanism based on uncertainty to achieve hierarchical high-precision defect detection and reliable reconstruction for point cloud scenes with complex structures and huge data volumes.

[0010] A point cloud technology processing method for defect location in complex building scenes includes the following steps:

[0011] In step S1, the raw point cloud data collected by the terahertz radar is preprocessed by normalization and denoising. Based on multi-source interpolation, the point cloud structure is completed to obtain a high-fidelity macro point cloud input. The data dimension is then reduced using efficient farthest point sampling (FPS). A pre-trained semantic segmentation model is then used to automatically divide the entire scene into multiple structurally independent, granularly adjustable sub-point cloud units.

[0012] Step S2: For each sub-point cloud unit obtained after semantic segmentation, different processing strategies are set according to its size. For small point cloud structures, they are directly treated as independent point cloud blocks for step-by-step feature extraction and defect location. For large point cloud structures, backbone points are generated based on the farthest point sampling, and point cloud cells are constructed with the backbone points as the core. Combined with multiple groups of cross-pruning strategies, the large point cloud is adaptively split into multi-layer, overlapping and structurally complete sub-point cloud blocks, realizing multi-scale spatial decomposition of the point cloud.

[0013] In step S3, feature extraction and defect location are performed on the independent point cloud blocks and the split sub-point cloud blocks of the small structure. The PointConv neural network is used for local and global feature extraction. Relative coordinate normalization is used to ensure the structural unbiasedness and position robustness of the feature expression. Uncertainty factors are further introduced to enhance the feature expression. After interpolation upsampling and feature splicing, each resolution layer is restored in turn to achieve adaptive multi-level defect location, and obtain integrated spatial defect features and point-level confidence.

[0014] In step S4, for large-scale point cloud structures, based on the detection results of each sub-point cloud block, multi-strategy pruning and fusion are introduced to integrate the defect results under different pruning segmentation paths. Through overlapping splicing, confidence weighting and uncertainty-driven credibility evaluation, the deviation caused by single segmentation is effectively eliminated, and unbiased fusion of defects in a global scope is achieved. Finally, the defect information of all sub-point cloud blocks is integrated, and the defect points and their confidence are mapped back to the original point cloud resolution through interpolation and upsampling strategies, realizing point cloud reconstruction with complete structure and detailed defect annotations.

[0015] Step S1 specifically includes:

[0016] First, for overly densely sampled data, a step-by-step farthest point sampling strategy is used to downsample to an appropriate order of magnitude while ensuring the uniformity of the building's overall topological structure. Due to the local nature of defects, a pre-trained KPConv model is used to perform semantic segmentation on complex point cloud scenes to obtain relatively independent sub-point cloud units. The specific process is as follows:

[0017] ;

[0018] ;

[0019] in, It is the original point cloud scene data directly collected by the terahertz radar. It is the point cloud scene data after FPS downsampling to a suitable order of magnitude. Represents the columns of sub-point cloud units after semantic segmentation, such as platforms, scaffolding, etc.

[0020] Step S2 specifically includes:

[0021] For each sub-point cloud unit obtained after semantic segmentation, segmentation strategies for point cloud structures of different orders of magnitude are generated according to the scale of each semantic structure: small point cloud structures can be directly regarded as independent point cloud blocks for hierarchical division; and large-scale point cloud structures are further segmented to obtain smaller secondary sub-point cloud blocks. In order to avoid the cutting point noise introduced by segmentation, a multi-group cross-pruning strategy is adopted. The logic of the pruning strategy is to use multiple farthest point sampling until the point cloud set presents a line-level topological backbone structure, and then the line structure is subjected to a multi-group intersection point segmentation strategy with coverage to obtain parallel secondary sub-point cloud block groups.

[0022] Step S3 specifically includes:

[0023] Step S31: Based on each segmented sub-point cloud block, the backbone point obtained by sampling the farthest point is used as the center, and a point cloud cell is constructed around each backbone point, that is, the k-nearest neighbor area of the backbone point. For each point cloud cell, the PointConv convolution layer is used to aggregate and learn local features, and the node spatial features are extracted. In order to reduce the feature noise caused by spatial offset, the local coordinate system of each point cloud cell is normalized with the backbone point as the center. Finally, the standardized spatial features are spliced with the original features to obtain a set of multi-level, spatial structure information-rich point cloud feature representations. Through multi-level recursive extraction, the multi-scale expression of point cloud features is realized. The features extracted at each level are spliced and integrated to form a multi-scale point cloud spatial structure feature with both global topology and local geometric details, providing high-quality feature support for the next step of defect location and uncertainty estimation.

[0024] Step S32: Based on the local structural features extracted from the deepest layer, nonlinear transformation and activation function are used to perform coarse-grained defect location and uncertainty prediction for each point cloud block. This step outputs the defect location result and uncertainty estimation value corresponding to each point cloud block, laying the foundation for subsequent defect refinement and criterion fusion.

[0025] In step S33, after obtaining the coarse-grained defect localization results, the coarse-grained defect localization and uncertainty estimation results are mapped back to a higher-density point cloud level through upsampling and interpolation mapping, and spliced and integrated with finer-grained local features. Subsequently, the results are input into the linear layer for feature linear transformation to obtain the final defect localization results of each point cloud block, and the uncertainty estimation is corrected in combination with the real label information, and asynchronous feedback optimization is achieved through the loss function.

[0026] The core idea of the step-by-step feature extraction in step S3 is to capture the structural features of the point cloud at different spatial scales through the effective coupling of the farthest point sampling FPS and the PointConv aggregation layer. The entire feature extraction process is designed in a hierarchical and progressive manner. Each round of iteration will downsample the point set and aggregate and enhance the features in the local space, enabling the network to gradually enrich its understanding of the spatial structure from global to local and from coarse to fine.

[0027] In the step-by-step feature extraction, the feature extraction process at each level is as follows: , represents the point set of the point cloud substructure obtained by downsampling the point cloud block m for the i-th time, represents the point set of the point cloud substructure obtained by downsampling the point cloud block m for the i+1th time, and at the same time, Can be regarded as The backbone nodes, next, according to Get in The point cloud cells centered on the backbone points generated based on the backbone points are expressed as ,The segmented point cloud has a strong spatial hierarchy, which is convenient for the next step of feature collection. Represents the point cloud cell of the rth backbone node in the point cloud substructure obtained by the i-th downsampling, Represents the Zhongyu The top k neighbor points with the closest Euclidean distance are selected. For each point cloud cell obtained, the PointConv layer is used to perform aggregate feature learning and extract its node spatial features. , where r represents backbone node number r, represents the richer spatial features of node r after aggregation, express At the module level, in order to avoid feature noise caused by spatial offset, the local features of the neighboring points of each point cloud cell, i.e., the spatial features, are taken as the reference system with the center point as the reference system, and the difference with respect to the center point is taken as the unbiased structural feature, i.e. , Represents the point cloud cell set of backbone point r The spatial features after standardization of the center point, that is, the unbiased structural features, represents the set of neighbor nodes, Then with After splicing, more global spatial features are obtained.

[0028] After obtaining local features layer by layer (usually five layers of feature extraction are performed. If there are too many points, the number will increase to ensure that the pixel level of the last layer meets the actual needs), the deepest dense features are used for coarse-grained defect localization and uncertainty prediction. Coarse-grained defect localization is performed through nonlinear transformation combined with activation function, which is expressed as:

[0029] ;

[0030] in, is the activation function, is a trainable parameter, d is the dimension of the feature channel, is the BN normalization layer, Represents the deepest spatial features of the secondary sub-point cloud block, Represents the coarse-grained defect localization result of the point cloud block, which is a two-dimensional output result, where the first dimension represents the uncertainty probability , the second dimension represents the defect location result .

[0031] After obtaining the coarse-grained defect localization results, we use layer-by-layer upsampling and restoration to restore the defect localization results and uncertainty probabilities to the input point cloud scale. During the upsampling and restoration process, we continuously splice local features with finer-grained features and input them into the linear layer for feature linear transformation to obtain the final defect localization results for each point cloud block. In this process, we calculate the loss of the uncertainty output results based on the true label, specifically including:

[0032] Step S331, upsampling and interpolation mapping, layer( ) features and coarse-grained defect localization results are restored (upsampled) to the previous layer The point cloud distribution is expressed as:

[0033] ;

[0034] in, is the interpolation mapping function, which is used to convert the Layer (coarse-grained layer) characteristics and defect location results, according to the spatial position, from Mapping / interpolation to The corresponding point cloud position can be restored from coarse to fine by using methods such as nearest neighbor interpolation and inverse distance weighting. The first The three-dimensional position coordinate set of the layer point cloud, For the previous layer, A collection of 3D position coordinates of a layer (finer-grained) point cloud, For the Coarse-grained defect localization results and uncertainty probabilities for all points in the layer, is the number of points, 2 represents the two components of the prediction output, namely the uncertainty probability and the defect location result, Indicates the The global spatial features of all points in the layer, d is the feature channel dimension, After interpolation, for the The upsampled defect localization results of each point in the layer, is the up-sampled feature result after interpolation alignment.

[0035] Step S332: fusion splicing and recursive update, the up-sampled features and defect location results are spliced and fused with the features of the current layer, and input into the linear layer for feature linear transformation to achieve cross-layer information integration and prepare for subsequent layers. The fused feature combination is:

[0036] ;

[0037] in, Represents the fused feature combination, including the current layer features, up-sampled features and positioning information, Represents the feature splicing operation, which splices each part along the feature channel one by one to form a new input;

[0038] Then perform a linear transformation on the feature combination:

[0039] ;

[0040] in, It is a linear transformation layer that performs weighted summation on the concatenated feature combinations to achieve linear integration of cross-layer information; is the feature after linear transformation, is the number of output feature channels.

[0041] Step S333: Multi-scale fusion output and loss function definition. Through multi-scale linear transformation and layer-by-layer upsampling, the original point cloud resolution is finally restored. The linearly transformed features are combined through the defect head to obtain the final prediction output:

[0042] ;

[0043] in, represents the number of points in the original point cloud, For the The prediction uncertainty of each point, For the The probability of defect determination at each point;

[0044] The maximum likelihood estimation is used as the loss function to jointly optimize uncertainty and prediction error, which can be expressed as:

[0045] ;

[0046] in, For the The true label of each point (defect / non-defect), For the The defect location results of each point, For the The first term of the loss function measures the prediction error and is dynamically weighted by the uncertainty. The second term is a regularization constraint that prevents the model from over-expanding uncertainty.

[0047] Step S4 specifically includes:

[0048] Step S41, multi-strategy defect fusion and reconstruction, based on the defect localization results of the sub-point cloud blocks, adopts a layer-by-layer fusion method to integrate defect information from different sub-point cloud structures, different levels, and different pruning strategies. First, through parallel detection of heterogeneous pruning strategies, candidate defect points of each sub-point cloud block are collected, and their uncertainty defect probability scores, i.e., confidence scores and defect localization result groups, are calculated. Pruning is to further segment the point cloud blocks. Subsequently, for the detection results output by different pruning strategies, the support statistics and uncertainty weighted fusion mechanism are used to effectively eliminate the local deviation caused by a single segmentation method, thereby improving the credibility and consistency of the global results. During the fusion process, the high-level structure can contain and accumulate defect information of the lower-level granularity. Through feature splicing, residual accumulation and confidence progression, the organic integration of multi-scale information is achieved. Finally, the global credible defect point set and its point-level confidence distribution after debiased superposition and confidence weighting are output, which significantly enhances the robustness and accuracy of large-scale point cloud defect localization under complex structures.

[0049] Step S42, reconstructing the point cloud with defect information, mapping the sparse, fused and integrated defect positioning results back to the original high-density point cloud to achieve explicit defect labeling and structural restoration. First, the credible defect points and their confidence levels are propagated and restored to the original point cloud density through methods such as distance weighted interpolation. Subsequently, a point-level defect mask is generated based on the confidence threshold, and the defect area is denoised and filtered and the structure is restored to improve the boundary smoothness and spatial continuity. Finally, the defect label and confidence level are embedded in the original point cloud to obtain reconstructed point cloud data with both structural integrity and accurate defect annotations, providing high-quality, seamless basic support for subsequent analysis, visualization, and automated modeling.

[0050] Step S41 specifically includes: for large-scale point cloud structures, based on the defect identification and confidence prediction results of the sub-point cloud blocks, defect information is fused layer by layer to generate global defect location and comprehensive confidence assessment results. The defect points in the higher layer will contain the information of the lower layer, and the confidence of the defect points is weighted and adjusted during the fusion process, gradually accumulating the upper and lower layer information. The specific fusion process is as follows:

[0051] For large-scale point cloud structures, point cloud branch groups generated by multiple groups of heterogeneous pruning strategies are processed in parallel. Pruning is to further segment the point cloud blocks. After obtaining the defect positioning results of each group, multiple groups of pruning strategies are designed to achieve overlap between each cut. The results of each group are weighted and fused according to the uncertainty probability, and the global representation of the defects is continuously purified to finally obtain a complete and unbiased global defect distribution. Among them, the credible and unbiased defect fusion needs to eliminate the deviation that may be caused by a single segmentation strategy. By fusing the defect point sets generated by multiple parallel segmentation strategies, credibility and unbiasedness are ensured. Through cross-validation and parallel information superposition, a credible defect point set and confidence distribution are finally output, namely: ,in, Represents the uncertainty defect probability score array of point cloud block k, =[{ , ,..., }, { , ,..., }, ..., { , ,..., }], Each element in represents the uncertainty defect probability corresponding to the point cloud block after pruning and segmentation. represents the defect location result group of point cloud block k, , , ], Each element in represents the defect localization result corresponding to the point cloud block after pruning and segmentation, and takes a value of 1 or 0, where 1 represents a defect point and 0 represents a non-defect point. g1…gt represents the pruning strategy, and N represents the number of point cloud blocks after segmentation by each pruning strategy. The TrustFuse function is a fusion module for multi-strategy pruning results. The fusion module implements the following functions:

[0052] Step S411: Multiple pruning strategies fusion and consensus evaluation, for different pruning strategies First, merge the defect point sets detected by all pruning strategies to obtain the joint set of all candidate defect points: , where t represents the total number of pruning strategies, represents the set of defect points output by the j-th pruning strategy, Represents the set of all candidate defect points obtained by combining all strategies;

[0053] Then, count each point Number of supports identified as weak points in each strategy:

[0054] ;

[0055] Among them, c(p) represents the number of point p that is simultaneously identified as defective points in multiple pruning strategies. This value is used to measure the degree of policy consensus of point p and serves as part of the subsequent fusion weight.

[0056] Step S412: uncertainty weighted fusion, which fuses the confidence predictions of multiple strategies for point p, i.e., uncertainty probabilities, to construct the final credible confidence distribution:

[0057] ;

[0058] in, is the global fusion confidence of point p, is the confidence estimate of the j-th strategy for point p, is the fusion weight of the j-th strategy at point p, defined as:

[0059] ;

[0060] in, is the support mark of point p in the j-th strategy, which takes a value of 1 or 0; is the uncertainty probability of the j-th strategy predicting point p; It is a numerical stability term to prevent division by zero;

[0061] , is the normalization factor, which ensures the normalization of the fusion result;

[0062] This mechanism integrates two important factors: first, the consensus support of multiple strategies. , and the second is the confidence reliability of each strategy’s prediction (the smaller the uncertainty, the greater the weight).

[0063] Step S413: Extract credible defect points and output the results, setting the confidence threshold , filter out the final set of credible defect points: ;

[0064] The final output is: , which is the credible global defect point set and point-level confidence distribution obtained by integrating the multi-strategy detection results and removing the bias through uncertainty weighting and cross-validation.

[0065] Step S42 specifically includes: matching the fused defect positioning result with the original point cloud structure, restoring the defect information to the original point cloud density through linear interpolation, and performing defect denoising based on the original structure, and finally obtaining a reconstruction result with the same magnitude as the original point cloud and accurate defect annotation. This reconstruction method not only maintains the integrity of the original structure but also achieves accurate positioning and visualization of defects. It is implemented by the defect annotation point cloud reconstruction module, which performs interpolation propagation and structure restoration, and is expressed as: , where function f represents the process of embedding the final credible defect information into the original macro point cloud and completing the defect explicit reconstruction, which includes the following steps:

[0066] S421, defect information position alignment and interpolation propagation. Since the defect detection process is performed in a relatively sparse structure after downsampling or pruning, the detection results need to be mapped back to the original high-density point cloud. , using a simple distance-weighted interpolation propagation strategy:

[0067] ;

[0068] in, represents the i-th credible defect point, , M is the total number of credible defect points, for The corresponding confidence level, is the interpolation kernel function, such as Gaussian kernel, represents the Euclidean distance between points, Represents the defect confidence of point p estimated in the macro point cloud.

[0069] S422, defect determination and mask generation, generate point-level masks based on the fusion confidence results , , used to distinguish defective areas from normal areas:

[0070] ;

[0071] in, It is a preset confidence threshold (such as 0.5) used to control the strictness of defect judgment.

[0072] S423, defect area denoising and structure restoration, for the sub-point cloud marked as the defect area in the mask ( ), filtering or reconstruction processing: , BilateralFilter is a bilateral filter. This step can improve the smoothness and visual quality of the defect edge and enhance the structural continuity. This step is optional.

[0073] S424, defect annotation and result output, finally embeds the defect mask and confidence information into the original point cloud to generate a reconstructed point cloud result with explicit annotations:

[0074] ;

[0075] Among them, each reconstructed point contains: the original geometric position of point p (i.e. coordinates x, y, z), the defect judgment label , corresponding to the defect confidence , which not only retains the point cloud data of the original structure, but also embeds defect annotation information (mask and confidence) into it for visualization, analysis or subsequent reconstruction.

[0076] The present invention also provides a point cloud technology processing system for defect positioning in complex building scenes, which realizes efficient segmentation of point cloud scenes, layer-by-layer feature enhancement and global defect consistent fusion. The system includes a complex scene adaptive segmentation module, a hierarchical feature learning and defect positioning module, and a global multi-strategy defect fusion and reconstruction module. The complex scene adaptive segmentation module performs pre-processing such as normalization and denoising on the original point cloud data collected by the terahertz radar, and based on multi-source interpolation, completes the point cloud structure to obtain high-fidelity macro point cloud input, and then uses efficient farthest point sampling FPS to reduce the data dimension, and through The pre-trained semantic segmentation model automatically divides the entire scene into multiple structurally independent and granularly adjustable sub-point cloud units. Based on the size of the sub-point cloud units, processing strategies are generated for point cloud structures of different orders of magnitude. For small point cloud structures, they are directly treated as independent point cloud blocks for step-by-step feature extraction and defect location. For large-scale point cloud structures, backbone points are generated based on farthest point sampling, and point cloud cells are constructed with backbone points as the core. Combined with spatial stratification and multi-group cross-pruning strategies, large point clouds are adaptively split into multi-layer, overlapping, and structurally complete sub-point cloud blocks, achieving multi-scale spatial decomposition of point clouds.

[0077] The hierarchical feature learning and defect localization module performs step-by-step feature extraction and defect localization, using neural networks such as PointConv for local and global feature extraction. Relative coordinate normalization is used to ensure the structural unbiasedness and positional robustness of feature expression. Uncertainty factors are further introduced to enhance feature expression. Through residual updates and multi-step loops, upper and lower layer features and confidence levels are gradually integrated to achieve adaptive multi-level defect localization, obtaining integrated spatial defect features and point-level confidence levels.

[0078] The global multi-strategy defect fusion and reconstruction module is used for large-scale point cloud structures. Based on the detection results of each sub-point cloud block, it introduces multi-strategy pruning fusion, covers and integrates the defect results under different pruning / segmentation paths, and effectively eliminates the deviation caused by single segmentation through overlapping splicing, confidence weighting and uncertainty-driven credibility evaluation, realizing unbiased fusion of defects in a global scope. Finally, the defect information of all sub-point cloud blocks is integrated, and the defect points and their confidence are mapped back to the original point cloud resolution through interpolation and upsampling strategies, realizing point cloud reconstruction with complete structure and detailed defect annotations.

[0079] The present invention provides a point cloud technology processing method for defect location in complex building scenes. Through adaptive layered processing based on semantic segmentation, uncertainty-based substructure defect location and upsampling restoration, a multi-level and multi-strategy fusion defect location method, and an integrated point cloud reconstruction mechanism with defect information, it achieves accurate defect location in complex building scenes using terahertz radar technology.

[0080] Adaptive hierarchical processing based on semantic segmentation first performs semantic segmentation on complex building scenes to obtain relatively independent sub-point cloud units. Different processing strategies are adopted according to the scale characteristics of the sub-point cloud units. Small structures are directly hierarchically divided into independent point cloud blocks. For large structures, backbone points are generated by sampling the farthest points, and point cloud cells are generated based on the backbone points. Then, multiple groups of breakpoint pruning strategies are used to generate sub-point cloud blocks of different levels. This hierarchical adaptive processing method ensures efficient processing of complex scenes and structural integrity.

[0081] Uncertainty-based substructure defect localization and upsampling restoration: For the hierarchical point cloud structures at all levels, feature encoding is first implemented within the point cloud cells based on PointConv. A relative coordinate system is used to improve the translation robustness of the features. This is then combined with PointConvD to complete hierarchical downsampling and multi-scale feature aggregation. The discrimination capability is enhanced by fusing local and global information. Uncertainty branches are introduced as feature channels. In the upsampling stage, interpolation and feature splicing are used to restore deep coarse-grained predictions to high resolution. Uncertainty learning is continuously guided based on true labels, and their dynamics are used as a weighting factor for feature updates. This adaptively improves the model's sensitivity to boundaries and abnormal areas, achieving fine-grained defect localization and uncertainty estimation.

[0082] A multi-level and multi-strategy fusion defect localization method is designed to target substructures of different scales and spatial distributions through a multi-level fusion and multi-strategy pruning parallel processing mechanism. Specifically, this includes precise local defect localization within micro-point cloud blocks, cross-layer fusion of defect points and confidence levels at each level, and cross-coverage of results from different pruning strategies. This method maximizes the extraction and fusion of multi-source, redundant, and heterogeneous information, effectively improving the accuracy and robustness of global defect localization, ensuring unbiased and credible fusion output, and providing reliable support for defect localization in a wide range of scenarios.

[0083] The integrated point cloud reconstruction mechanism with defect information, after completing the fusion of reliable defect points, maps the defect information back to the original point cloud density structure through linear interpolation, and performs denoising processing in combination with the original geometric information, ultimately achieving point cloud reconstruction with complete structure and accurate defect annotation, taking into account both reconstruction accuracy and defect visualization effect.

[0084] The method of the present invention has the following beneficial technical effects:

[0085] Improved processing capabilities: This invention uses independent structure recognition and backbone point parallel segmentation technology to efficiently process large-scale point cloud data in complex architectural scenes. The macro point cloud multi-view segmenter divides the huge point cloud set into independent structures, providing a high-quality data foundation for subsequent processing, effectively reducing data redundancy, and improving processing capabilities.

[0086] Improved processing efficiency: This invention significantly reduces the order of magnitude of point cloud data by optimizing the farthest point sampling and multi-view segmentation scheme, enabling complex scenes to be processed at a lower computational cost. Furthermore, micro-point cloud defect localization combines multi-scale feature learning and global information fusion to achieve rapid defect location and efficient identification, improving the overall system processing efficiency.

[0087] Improve positioning accuracy. Through multi-scale structural feature extraction and global information fusion, micro-point cloud defect positioning can accurately identify local and global defects. At the same time, parallel micro-point cloud reconstruction and fusion combined with uncertainty assessment ensure high-precision defect positioning. The entire method effectively eliminates noise and errors, achieving accurate identification and positioning of defects.

[0088] Enhanced credibility: This invention introduces an uncertainty assessment model and adopts a weighted fusion strategy to eliminate the errors caused by different segmentation schemes, ensuring the credibility of the defect location results. Through the defect denoising and defect restoration processes, the final defect distribution map is highly consistent, providing reliable support for subsequent applications.

[0089] The solution of the present invention not only effectively solves the computational bottleneck problem of large-scale point cloud data, but also significantly improves the efficiency and accuracy of defect positioning through semantic understanding and multi-scale feature learning. It can efficiently adapt to terahertz point cloud data of various scales and types, and provides feasibility and reliability for the practical application of terahertz radar technology in complex building scenes. At the same time, the semantic-based adaptive segmentation strategy and uncertainty-guided feature learning method proposed in the present invention take into account both global and local feature expressions, effectively improve the defect positioning accuracy in scenes such as complex buildings, expand the technical boundaries of point cloud data processing, and provide new ideas for intelligent manufacturing, lossless defect positioning and other fields requiring high-precision three-dimensional data processing. It has important engineering application value and broad market prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0091] Figure 1 This is a flow chart of a point cloud technology processing method for defect location in complex building scenes provided by an embodiment of the present invention;

[0092] Figure 2 It is a schematic diagram of defect segmentation, positioning and fusion of a large-scale point cloud structure provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0093] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0094] This paper proposes a point cloud technology processing method for defect location in complex building scenes. The method mainly includes three parts:

[0095] Semantically guided hierarchical processing strategy: After appropriate farthest point downsampling, complex scenes are semantically segmented using a pre-trained model to obtain independent sub-point cloud structures. These structures are then flexibly graded based on point cloud size. Small-scale structures are directly output to the feature extraction module, while large-scale structures are pruned using multi-level farthest point sampling and a fixed strategy, ensuring processing adaptability and efficiency.

[0096] Uncertainty-based substructure defect localization and upsampling restoration: Within each independent substructure, the deep feature extraction module first performs high-dimensional encoding of its spatial features and obtains coarse-grained defect localization and model uncertainty through nonlinear transformations involving probability and uncertainty branches. Then, from the bottom up, interpolation mapping and feature splicing methods are used to upsample the deep localization layer by layer to a higher resolution. Linear transformations of features are then performed through linear layers to integrate cross-layer information. The final output is fine-grained defect recognition and uncertainty estimation with the same resolution as the original point cloud. This fusion and progressive process enables the model to achieve "coarse-to-fine" discrimination and adaptively adjust feature weights based on uncertainty, improving the reliability and discrimination ability of overall defect localization in an unsupervised manner.

[0097] Defect reconstruction using multi-strategy fusion: A multi-strategy fusion mechanism is designed for the defect detection results of all sub-point cloud structures. This includes uncertainty-guided candidate aggregation, regional merging based on spatial position and semantic consistency, and probability-weighted global reconstruction. This achieves unified defect fusion and restoration for regions of different scales and complexities, eliminates redundancy and noise, and generates a seamless final defect distribution map that meets actual engineering requirements. This multi-strategy fusion step further enhances the capabilities of fine positioning and global reconstruction, achieving high-confidence reconstruction of defects in complex scenarios.

[0098] like Figure 1 As shown, the present invention provides a point cloud technology processing method for defect location in complex building scenes, comprising the following steps:

[0099] In step S1, the raw point cloud data collected by the terahertz radar is preprocessed by normalization and denoising. Based on multi-source interpolation, the point cloud structure is completed to obtain a high-fidelity macro point cloud input. The data dimension is then reduced using efficient farthest point sampling (FPS). A pre-trained semantic segmentation model is then used to automatically divide the entire scene into multiple structurally independent, granularly adjustable sub-point cloud units.

[0100] In this step, the high-density point cloud data collected by the terahertz radar is first downsampled using the progressive farthest point sampling (FPS) algorithm, which not only reduces the data size but also maintains the consistency of the overall topological structure of the building. On this basis, the pre-trained KPConv model is used to perform semantic segmentation on the downsampled point cloud to automatically extract relatively independent sub-point cloud units and achieve precise positioning of the defect area.

[0101] Step S2: For each sub-point cloud unit obtained after semantic segmentation, processing strategies for point cloud structures of different orders of magnitude are generated according to their size. For small point cloud structures, they are directly treated as independent point cloud blocks for step-by-step feature extraction and defect location. For large point cloud structures, backbone points are generated based on farthest point sampling, and point cloud cells are constructed with backbone points as the core. Combined with spatial stratification and multi-group cross pruning strategies, the large point cloud is adaptively split into multi-layer, overlapping, and structurally complete sub-point cloud blocks, realizing multi-scale spatial decomposition of the point cloud.

[0102] Point cloud structures with points ranging from a few thousand to tens of thousands are classified as small point cloud structures. Small point cloud structures are directly processed as independent point cloud blocks without further subdivision. Dense point cloud structures with points ranging from hundreds of thousands to tens of millions are classified as large point cloud structures. Large point cloud structures continue to use hierarchical farthest point sampling and combine multiple groups of random pruning strategies to recursively segment and optimize large point clouds, ensuring that all substructures can be processed efficiently and avoiding artificial division noise.

[0103] In this step, a hierarchical refinement and multi-group cross-pruning strategy is adopted for point cloud structures of different sizes: for small point cloud structures, they are directly processed as independent point cloud blocks; for large-scale point cloud structures, multi-level farthest point sampling and multi-group pruning strategies are recursively adopted to further divide them into several sub-structure blocks, and defect detection and fusion processing are performed one by one, which effectively improves the accuracy and robustness of defect positioning, and is particularly suitable for the automated segmentation and defect identification of large-volume point cloud data.

[0104] In step S3, feature extraction and defect location are performed on the independent point cloud blocks and the split sub-point cloud blocks of the small structure. Neural networks such as PointConv are used for local and global feature extraction. Relative coordinate normalization is used to ensure the structural unbiasedness and position robustness of the feature expression. Uncertainty factors are further introduced to enhance the feature expression. After interpolation upsampling and feature splicing, each resolution layer is restored in turn to achieve adaptive multi-level defect location, obtaining integrated spatial defect features and point-level confidence.

[0105] In step S4, for large-scale point cloud structures, based on the detection results of each sub-point cloud block, multi-strategy pruning and fusion are introduced to integrate the defect results under different pruning segmentation paths. Through overlapping splicing, confidence weighting and uncertainty-driven credibility evaluation, the deviation caused by single segmentation is effectively eliminated, and unbiased fusion of defects in a global scope is achieved. Finally, the defect information of all sub-point cloud blocks is integrated, and the defect points and their confidence are mapped back to the original point cloud resolution through interpolation and upsampling strategies, realizing point cloud reconstruction with complete structure and detailed defect annotations.

[0106] In this step, global defect fusion and point cloud reconstruction are required. After completing the defect detection of the sub-point cloud block, multi-strategy pruning and fusion technology is used for larger point cloud structures to integrate defect information under different granularities and strategies, and finally achieve global defect visualization and point cloud reconstruction with defect annotations. The entire stage focuses on eliminating deviations under different segmentation strategies or granularities to improve the reliability and global consistency of defect identification.

[0107] The present invention can efficiently adapt to terahertz point cloud data of various sizes and types, taking into account both global and local feature expression, effectively improving the defect detection accuracy in scenes such as complex buildings, as well as the reliability and consistency of defect point cloud visualization and subsequent engineering applications, providing an innovative and efficient technical solution for large-scale applications of high-density point clouds.

[0108] Step S1 specifically includes:

[0109] First, for overly densely sampled data, a step-by-step farthest point sampling strategy is used to downsample to an appropriate order of magnitude while ensuring the uniformity of the building's overall topological structure. Due to the local nature of defects, a pre-trained KPConv model is used to perform semantic segmentation on complex point cloud scenes to obtain relatively independent sub-point cloud units. The specific process is as follows:

[0110] ;

[0111] ;

[0112] in, It is the original point cloud scene data directly collected by the terahertz radar. It is the point cloud scene data after FPS downsampling to a suitable order of magnitude. Represents the columns of sub-point cloud units after semantic segmentation, such as platforms, scaffolding, etc.

[0113] Step S2 specifically includes:

[0114] For each sub-point cloud unit obtained after semantic segmentation, segmentation strategies for point cloud structures of different orders of magnitude are generated according to the scale of each semantic structure: small point cloud structures can be directly regarded as independent point cloud blocks for hierarchical division; and large-scale point cloud structures are further segmented to obtain smaller secondary sub-point cloud blocks. In order to avoid the cutting point noise introduced by segmentation, a multi-group cross-pruning strategy is adopted. The logic of the pruning strategy is to use multiple farthest point sampling until the point cloud set presents a line-level topological backbone structure, and then the line structure is subjected to a multi-group intersection point segmentation strategy with coverage to obtain parallel secondary sub-point cloud block groups.

[0115] Step S3 specifically includes:

[0116] Step S31, step-by-step feature extraction, the core idea is to realize the structural feature capture of point cloud at different spatial scales through the effective coupling of the farthest point sampling FPS and the PointConv aggregation layer. The whole feature extraction process is designed in a hierarchical and progressive manner. Each round of iteration will downsample the point set, aggregate and enhance the features in the local space, so that the network can gradually enrich the understanding of the spatial structure from global to local, from coarse to fine. Each feature extraction includes constructing a point cloud cell around each backbone point based on each segmented sub-point cloud block, with the backbone point obtained by the farthest point sampling as the center, i.e., the k-nearest neighbor area of the backbone point. For each point cloud cell, sampling The PointConv convolution layer is used to aggregate and learn local features and extract node spatial features. To reduce feature noise caused by spatial offset, the local coordinate system of each point cloud cell is normalized with the backbone point as the center. Finally, the standardized spatial features are spliced with the original features to obtain a set of multi-level point cloud feature representations with rich spatial structure information. Through multi-level recursive extraction, multi-scale expression of point cloud features is achieved. The features extracted at each level are spliced and integrated to form multi-scale point cloud spatial structure features that combine global topology and local geometric details, providing high-quality feature support for the next step of defect location and uncertainty estimation.

[0117] Step S32: Based on the local structural features extracted from the deepest layer, nonlinear transformation and activation function are used to perform coarse-grained defect location and uncertainty prediction for each point cloud block. This step outputs the defect location result and uncertainty estimation value corresponding to each point cloud block, laying the foundation for subsequent defect refinement and criterion fusion.

[0118] In step S33, after obtaining the coarse-grained defect localization results, the coarse-grained defect localization and uncertainty estimation results are mapped back to a higher-density point cloud level through upsampling and interpolation mapping, and spliced and integrated with finer-grained local features. Subsequently, the results are input into the linear layer for feature linear transformation to obtain the final defect localization results of each point cloud block, and the uncertainty estimation is corrected in combination with the real label information, and asynchronous feedback optimization is achieved through the loss function.

[0119] In the step-by-step feature extraction, the feature extraction process at each level is as follows: , represents the point set of the point cloud substructure obtained by downsampling the point cloud block m for the i-th time, represents the point set of the point cloud substructure obtained by downsampling the point cloud block m for the i+1th time, and at the same time, Can be regarded as The backbone nodes, next, according to Get in The point cloud cells centered on the backbone points generated based on the backbone points are expressed as ,The segmented point cloud has a strong spatial hierarchy, which is convenient for the next step of feature collection. Represents the point cloud cell of the rth backbone node in the point cloud substructure obtained by the i-th downsampling, Represents the Zhongyu The top k neighbor points with the closest Euclidean distance are selected. For each point cloud cell obtained, the PointConv layer is used to perform aggregate feature learning and extract its node spatial features. , where r represents backbone node number r, represents the richer spatial features of node r after aggregation, express At the module level, in order to avoid feature noise caused by spatial offset, the local features of the neighboring points of each point cloud cell, i.e., the spatial features, are taken as the reference system with the center point as the reference system, and the difference with respect to the center point is taken as the unbiased structural feature, i.e. , Represents the point cloud cell set of backbone point r The spatial features after standardization of the center point, that is, the unbiased structural features, represents the set of neighbor nodes, Then with After splicing, more global spatial features are obtained.

[0120] After obtaining the local features layer by layer, the deepest dense features are used for coarse-grained defect localization and uncertainty prediction. Coarse-grained defect localization is performed through nonlinear transformation combined with activation function, which is expressed as:

[0121] ;

[0122] in, is the activation function, is a trainable parameter, d is the dimension of the feature channel, is the BN normalization layer, Represents the deepest spatial features of the secondary sub-point cloud block, Represents the coarse-grained defect localization result of the point cloud block, which is a two-dimensional output result, where the first dimension represents the uncertainty probability , the second dimension represents the defect location result , the value is 1 or 0, 1 means it is a defect point, 0 means it is not a defect point.

[0123] After obtaining the coarse-grained defect localization results (usually five layers of feature extraction are performed. If the number of points is too large, the number will increase to ensure that the pixel level of the last layer meets the actual needs), layer-by-layer upsampling restoration is used to restore the defect localization results and uncertainty probabilities to the input point cloud scale. During the upsampling restoration process, local features are continuously spliced with finer-grained features and input into the linear layer for feature linear transformation to obtain the final defect localization results for each point cloud block. In this process, the uncertainty output results are calculated based on the true label. Specifically, the following steps are performed:

[0124] Step S331, upsampling and interpolation mapping, layer( ) features and coarse-grained defect localization results are restored (upsampled) to the previous layer The point cloud distribution is expressed as:

[0125] ;

[0126] in, is the interpolation mapping function, which is used to convert the Layer (coarse-grained layer) characteristics and defect location results, according to the spatial position, from Mapping / interpolation to The corresponding point cloud position can be restored from coarse to fine by using methods such as nearest neighbor interpolation and inverse distance weighting. The first The three-dimensional position coordinate set of the layer point cloud, For the previous layer, A collection of 3D position coordinates of a layer (finer-grained) point cloud, For the Coarse-grained defect localization results and uncertainty probabilities for all points in the layer, is the number of points, 2 represents the two components of the prediction output, namely the uncertainty probability and the defect location result, Indicates the The global spatial features of all points in the layer, d is the feature channel dimension, After interpolation, for the The upsampled defect localization results of each point in the layer, is the up-sampled feature result after interpolation alignment.

[0127] Step S332: fusion splicing and recursive update, the up-sampled features and defect location results are spliced and fused with the features of the current layer, and input into the linear layer for feature linear transformation to achieve cross-layer information integration and prepare for subsequent layers. The fused feature combination is:

[0128] ;

[0129] in, Represents the fused feature combination, including the current layer features, up-sampled features and positioning information, Represents the feature splicing operation, which splices each part along the feature channel one by one to form a new input;

[0130] Then perform a linear transformation on the feature combination:

[0131] ;

[0132] in, It is a linear transformation layer that performs weighted summation on the concatenated feature combinations to achieve linear integration of cross-layer information; is the feature after linear transformation, is the number of output feature channels.

[0133] Step S333: Multi-scale fusion output and loss function definition. Through multi-scale linear transformation and layer-by-layer upsampling, the original point cloud resolution is finally restored. The linearly transformed features are combined through the defect head to obtain the final prediction output:

[0134] ;

[0135] in, represents the number of points in the original point cloud, For the The prediction uncertainty of each point, For the The probability of defect determination at each point;

[0136] The maximum likelihood estimation is used as the loss function to jointly optimize uncertainty and prediction error, which can be expressed as:

[0137] ;

[0138] in, For the The true label of each point (defect / non-defect), For the The defect location results of each point, For the The first term of the loss function measures the prediction error and is dynamically weighted by the uncertainty. The second term is a regularization constraint that prevents the model from over-expanding uncertainty.

[0139] Step S4 specifically includes:

[0140] Step S41, multi-strategy defect fusion and reconstruction, based on the defect localization results of the sub-point cloud blocks, adopts a layer-by-layer fusion approach to integrate defect information from different sub-point cloud structures, different levels, and different pruning strategies. First, through parallel detection of heterogeneous pruning strategies, candidate defect points of each sub-point cloud block are collected, and their uncertainty defect probability scores, i.e., confidence scores and defect localization result groups, are calculated. Subsequently, for the detection results output by different pruning strategies, a fusion mechanism of support statistics and uncertainty weighting is used to effectively eliminate the local deviations caused by a single segmentation method, thereby improving the credibility and consistency of the global results. During the fusion process, high-level structures can contain and accumulate defect information of lower-level granularity. Through feature splicing, residual accumulation, and confidence progression, the organic integration of multi-scale information is achieved. Finally, the global credible defect point set and its point-level confidence distribution after debiased superposition and confidence weighting are output, which significantly enhances the robustness and accuracy of large-scale point cloud defect localization under complex structures.

[0141] Step S42, reconstructing the point cloud with defect information, mapping the sparse, fused and integrated defect positioning results back to the original high-density point cloud to achieve explicit defect labeling and structural restoration. First, the credible defect points and their confidence levels are propagated and restored to the original point cloud density through methods such as distance weighted interpolation. Subsequently, a point-level defect mask is generated based on the confidence threshold, and the defect area is denoised and filtered and the structure is restored to improve the boundary smoothness and spatial continuity. Finally, the defect label and confidence level are embedded in the original point cloud to obtain reconstructed point cloud data with both structural integrity and accurate defect annotations, providing high-quality, seamless basic support for subsequent analysis, visualization, and automated modeling.

[0142] Figure 2 The defect segmentation, positioning and fusion of large-scale point cloud structures are demonstrated. For structures with too large point cloud volumes, the spatial resources of the computing device are unable to meet the needs of extracting high-dimensional spatial features of each node. Therefore, segmentation is performed according to multiple groups of pruning strategies, and defect detection is performed on each sub-point cloud block obtained. After obtaining the defect detection results of multiple sub-point cloud blocks, the defect positioning results of multiple groups of sub-point cloud blocks are fused to avoid inaccurate defect positioning caused by cutting noise.

[0143] Step S41 specifically includes: for large-scale point cloud structures, based on the defect identification and confidence prediction results of the sub-point cloud blocks, defect information is fused layer by layer to generate global defect location and comprehensive confidence assessment results. The defect points in the higher layer will contain the information of the lower layer, and the confidence of the defect points is weighted and adjusted during the fusion process, gradually accumulating the upper and lower layer information. The specific fusion process is as follows:

[0144] For large-scale point cloud structures, point cloud branch groups generated by multiple groups of heterogeneous pruning strategies are processed in parallel. Pruning is to further segment the point cloud blocks. After obtaining the defect positioning results of each group, multiple groups of pruning strategies are designed to achieve overlap between each cut. The results of each group are weighted and fused according to the uncertainty probability, and the global representation of the defects is continuously purified to finally obtain a complete and unbiased global defect distribution. Among them, the credible and unbiased defect fusion needs to eliminate the deviation that may be caused by a single segmentation strategy. By fusing the defect point sets generated by multiple parallel segmentation strategies, credibility and unbiasedness are ensured. Through cross-validation and parallel information superposition, a credible defect point set and confidence distribution are finally output, which can be expressed as: ,in, Represents the uncertainty defect probability score array of point cloud block k, =[{ , ,..., }, { , ,..., }, ..., { , ,..., }], Each element in represents the uncertainty defect probability corresponding to the point cloud block after pruning and segmentation. represents the defect location result group of point cloud block k, , , …… ， { P k_gt_1 def , P k_gt_2 def ,..., P k_gt_N def } ] , Each element in represents the defect localization result corresponding to the point cloud block after pruning and segmentation, and takes a value of 1 or 0, where 1 represents a defect point and 0 represents a non-defect point. g1…gt represents the pruning strategy, and N represents the number of point cloud blocks after segmentation by each pruning strategy. The TrustFuse function is a fusion module for multi-strategy pruning results. The fusion module implements the following functions:

[0145] Step S411: Multiple pruning strategies fusion and consensus evaluation, for different pruning strategies First, merge the defect point sets detected by all pruning strategies to obtain the joint set of all candidate defect points: , where t represents the total number of pruning strategies, represents the set of defect points output by the j-th pruning strategy, Represents the set of all candidate defect points obtained by combining all strategies;

[0146] Then, count each point Number of supports identified as weak points in each strategy:

[0147] ;

[0148] Among them, c(p) represents the number of point p that is simultaneously identified as defective points in multiple pruning strategies. This value is used to measure the degree of policy consensus of point p and serves as part of the subsequent fusion weight.

[0149] Step S412: uncertainty weighted fusion, which fuses the confidence predictions of multiple strategies for point p, i.e., uncertainty probabilities, to construct the final credible confidence distribution:

[0150] ;

[0151] in, is the global fusion confidence of point p, is the confidence estimate of the j-th strategy for point p, is the fusion weight of the j-th strategy at point p, defined as:

[0152] ;

[0153] in, is the support mark of point p in the j-th strategy, which takes a value of 1 or 0; is the uncertainty probability of the j-th strategy predicting point p; It is a numerical stability term to prevent division by zero;

[0154] , is the normalization factor, which ensures the normalization of the fusion result;

[0155] This mechanism integrates two important factors: first, the consensus support of multiple strategies. , and the second is the confidence reliability of each strategy’s prediction (the smaller the uncertainty, the greater the weight).

[0156] Step S413: Extract credible defect points and output the results, setting the confidence threshold , filter out the final set of credible defect points: ;

[0157] The final output is: , which is the credible global defect point set and point-level confidence distribution obtained by integrating the multi-strategy detection results and removing the bias through uncertainty weighting and cross-validation.

[0158] Step S42 specifically includes: matching the fused defect positioning result with the original point cloud structure, restoring the defect information to the original point cloud density through linear interpolation, and performing defect denoising based on the original structure, and finally obtaining a reconstruction result with the same magnitude as the original point cloud and accurate defect annotation. This reconstruction method not only maintains the integrity of the original structure but also achieves accurate positioning and visualization of defects. It is implemented by the defect annotation point cloud reconstruction module, which performs interpolation propagation and structure restoration, and is expressed as: , where function f represents the process of embedding the final credible defect information into the original macro point cloud and completing the defect explicit reconstruction, which includes the following steps:

[0159] S421, defect information position alignment and interpolation propagation. Since the defect detection process is performed in a relatively sparse structure after downsampling or pruning, the detection results need to be mapped back to the original high-density point cloud. , using a simple distance-weighted interpolation propagation strategy:

[0160] ;

[0161] in, represents the i-th credible defect point, , M is the total number of credible defect points, for The corresponding confidence level, is the interpolation kernel function, such as Gaussian kernel, represents the Euclidean distance between points, Represents the defect confidence of point p estimated in the macro point cloud.

[0162] S422, defect determination and mask generation, generate point-level masks based on the fusion confidence results , , used to distinguish defective areas from normal areas:

[0163] ;

[0164] in, It is a preset confidence threshold (such as 0.5) used to control the strictness of defect judgment.

[0165] S423, defect area denoising and structure restoration, for the sub-point cloud marked as the defect area in the mask ( ), filtering or reconstruction processing: , BilateralFilter is a bilateral filter. This step can improve the smoothness and visual quality of the defect edge and enhance the structural continuity. This step is optional.

[0166] S424, defect annotation and result output, finally embeds the defect mask and confidence information into the original point cloud to generate a reconstructed point cloud result with explicit annotations:

[0167] ;

[0168] Each reconstructed point contains: the original geometric position of point p (i.e., coordinates x, y, z), the defect judgment label , corresponding to the defect confidence , which not only retains the point cloud data of the original structure, but also embeds defect annotation information (mask and confidence) into it for visualization, analysis or subsequent reconstruction.

[0169] The present invention also provides a point cloud technology processing system for defect positioning in complex building scenes, which realizes efficient segmentation of point cloud scenes, layer-by-layer feature enhancement and global defect consistent fusion. The system includes a complex scene adaptive segmentation module, a hierarchical feature learning and defect positioning module, and a global multi-strategy defect fusion and reconstruction module. The complex scene adaptive segmentation module performs pre-processing such as normalization and denoising on the original point cloud data collected by the terahertz radar, and based on multi-source interpolation, completes the point cloud structure to obtain high-fidelity macro point cloud input, and then uses efficient farthest point sampling FPS to reduce the data dimension, and through The pre-trained semantic segmentation model automatically divides the entire scene into multiple structurally independent and granularly adjustable sub-point cloud units. Based on the size of the sub-point cloud units, processing strategies are generated for point cloud structures of different orders of magnitude. For small point cloud structures, they are directly treated as independent point cloud blocks for step-by-step feature extraction and defect location. For large-scale point cloud structures, backbone points are generated based on farthest point sampling, and point cloud cells are constructed with backbone points as the core. Combined with spatial stratification and multi-group cross-pruning strategies, large point clouds are adaptively split into multi-layer, overlapping, and structurally complete sub-point cloud blocks, achieving multi-scale spatial decomposition of point clouds.

[0170] The hierarchical feature learning and defect localization module performs step-by-step feature extraction and defect localization, using neural networks such as PointConv for local and global feature extraction. Relative coordinate normalization is used to ensure the structural unbiasedness and positional robustness of feature expression. Uncertainty factors are further introduced to enhance feature expression. Through residual updates and multi-step loops, upper and lower layer features and confidence levels are gradually integrated to achieve adaptive multi-level defect localization, obtaining integrated spatial defect features and point-level confidence levels.

[0171] The global multi-strategy defect fusion and reconstruction module is used for large-scale point cloud structures. Based on the detection results of each sub-point cloud block, it introduces multi-strategy pruning fusion, covers and integrates the defect results under different pruning / segmentation paths, and effectively eliminates the deviation caused by single segmentation through overlapping splicing, confidence weighting and uncertainty-driven credibility evaluation, realizing unbiased fusion of defects in a global scope. Finally, the defect information of all sub-point cloud blocks is integrated, and the defect points and their confidence are mapped back to the original point cloud resolution through interpolation and upsampling strategies, realizing point cloud reconstruction with complete structure and detailed defect annotations.

[0172] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A point cloud technology processing method for defect location in complex building scenes, characterized by: The method comprises: Step S1: Normalize and denoise the raw point cloud data collected by the terahertz radar, and complete the point cloud structure based on multi-source interpolation to obtain a high-fidelity macro point cloud input. Then, the farthest point sampling FPS is used to reduce the data dimension, and the pre-trained semantic segmentation model is used to divide the overall scene into multiple structurally independent and granularly adjustable sub-point cloud units. Step S2: For each sub-point cloud unit obtained after semantic segmentation, different processing strategies are set according to its size. For small point cloud structures, they are directly used as independent point cloud blocks for step-by-step feature extraction and defect location. For large-scale point cloud structures, backbone points are generated based on the farthest point sampling, and point cloud cells are constructed with the backbone points as the core. Combined with multiple groups of cross-pruning strategies, the large point cloud is adaptively split into multi-layer, overlapping and structurally complete sub-point cloud blocks to achieve multi-scale spatial decomposition of the point cloud. Among them, small point cloud structures are point cloud structures with a point count ranging from thousands to tens of thousands, and large-scale point cloud structures are dense point cloud structures with a point count ranging from hundreds of thousands to tens of millions. In step S3, feature extraction and defect location are performed on the independent point cloud blocks and the split sub-point cloud blocks. The PointConv neural network is used for local and global feature extraction. Relative coordinate normalization is performed to ensure the structural unbiasedness and position robustness of the feature expression. Uncertainty factors are further introduced to enhance the feature expression. After interpolation upsampling and feature splicing, each resolution layer is restored in turn to achieve adaptive multi-level defect location, and obtain integrated spatial defect features and point-level confidence. In step S4, for large-scale point cloud structures, based on the positioning results of each sub-point cloud block, multi-strategy pruning and fusion are introduced to integrate the defect results under different pruning segmentation paths. Through overlapping splicing, confidence weighting and uncertainty-driven credibility evaluation, the deviation caused by single segmentation is effectively eliminated, and unbiased fusion of defects in a global scope is achieved. Finally, the defect information of all sub-point cloud blocks is integrated, and the defect points and their confidence are mapped back to the original point cloud resolution through interpolation, realizing point cloud reconstruction with complete structure and detailed defect annotations.

2. The method according to claim 1, characterized in that The step S1 further comprises: The pre-trained model KPConv is used to perform semantic segmentation on complex point cloud scenes to obtain independent sub-point cloud units, which are expressed as: ； ； in, Represents the original point cloud scene data collected by the terahertz radar, is the point cloud scene data after FPS downsampling, Represents the columns of sub-point cloud units after semantic segmentation.

3. The method according to claim 1, characterized in that The logic of the pruning strategy in step S2 is to use multiple farthest point sampling until the point cloud set presents a line-level topological backbone structure, and then perform a multi-group intersection point segmentation strategy with covering on the line-level topological backbone structure to obtain parallel secondary sub-point cloud block groups.

4. The method according to claim 1, wherein The step S3 further comprises: Step S31: step-by-step feature extraction. The idea of step-by-step feature extraction is to effectively couple the farthest point sampling FPS with the PointConv aggregation layer to capture the structural features of the point cloud at different spatial scales. The entire feature extraction process is progressive in a hierarchical manner. Each round of iteration will downsample the point set and aggregate and enhance the features in the local space, so that the network can gradually enrich its understanding of the spatial structure from global to local, from coarse to fine; Step S32: After obtaining the local features layer by layer, based on the deepest spatial features, nonlinear transformation and activation function are used to perform coarse-grained defect location and uncertainty prediction for each point cloud block, and the defect location result and uncertainty estimation probability corresponding to each point cloud block are output; In step S33, after obtaining the coarse-grained defect localization results, the coarse-grained defect localization and uncertainty estimation results are mapped back to a higher-density point cloud level through upsampling and interpolation mapping, and spliced and integrated with finer-grained local features. Subsequently, the results are input into the linear layer for feature linear transformation to obtain the final defect localization results of each point cloud block, and the uncertainty estimation is corrected in combination with the real label information, and asynchronous feedback optimization is achieved through the loss function.

5. The method according to claim 4, characterized in that The feature extraction at each level in step S31 specifically includes: based on each segmented sub-point cloud block, with the backbone point obtained by sampling the farthest point as the center, aggregating and constructing a point cloud cell around each backbone point, that is, the k-nearest neighbor area of the backbone point, for each point cloud cell, using the PointConv convolution layer to aggregate and learn local features, extract node spatial features, and normalize the local coordinate system of each point cloud cell with the backbone point as the center. Finally, the standardized spatial features are spliced with the original features to obtain a set of multi-level, spatial structure information-rich point cloud feature representations, which are specifically expressed as follows: ,in, represents the point set of the point cloud substructure obtained by downsampling the point cloud block m for the i-th time, Represents the point set of the point cloud substructure obtained by downsampling the point cloud block m for the i+1th time. regarded as The backbone nodes, according to Get in The point cloud cells centered on the backbone points generated based on the backbone points are expressed as ,in, Represents the point cloud cell of the r-th backbone node in the point cloud substructure obtained by the i-th downsampling, Indicates Zhongyu The top k neighbor points with the closest Euclidean distance are selected. For each point cloud cell obtained, the PointConv layer is used to perform aggregate feature learning and extract its node spatial features. ,in, represents the richer spatial features of node r after aggregation, express Module layer, then take each point cloud cell as a unit, take the local features of the neighboring points of each point cloud cell, that is, the spatial features, with the center point as a reference, and use the difference with the center point as the unbiased structural feature, that is , Represents the point cloud cell set of backbone point r The spatial features after standardization of the center point, that is, the unbiased structural features, represents the set of neighbor nodes, Then with After splicing, more global spatial features are obtained.

6. The method according to claim 4, characterized in that The step S32 specifically includes: Coarse-grained defect localization is performed by combining nonlinear transformation with activation function, which is expressed as: ,in, is the activation function, is a trainable parameter, d is the dimension of the feature channel, is the BN normalization layer, Represents the deepest spatial features of the secondary sub-point cloud block, Represents the coarse-grained defect localization result of the point cloud block, which is a two-dimensional output result, where the first dimension represents the uncertainty probability , the second dimension represents the defect location result .

7. The method according to claim 4, characterized in that The step S33 further comprises: Step S331, upsampling and interpolation mapping, The features of the layer and the coarse-grained defect localization results are restored, that is, upsampled to the previous layer. The point cloud distribution of the layer is expressed as: ； in, , is the interpolation mapping function, which is used to convert the Layer is the characteristics and defect location results of the coarse-grained layer, according to the spatial position, from Mapping / interpolation to The corresponding point cloud position is restored from coarse to fine. For the The three-dimensional position coordinate set of the layer point cloud, For the previous layer, Layer is a set of 3D position coordinates of a finer-grained point cloud. For the Coarse-grained defect localization results and uncertainty probabilities for all points in the layer, is the number of points, 2 represents the two components of the prediction output, namely the uncertainty probability and the defect location result, Indicates the The global spatial features of all points in the layer, d is the feature channel dimension, After interpolation, for the The upsampled defect localization results of each point in the layer, is the up-sampled feature result after interpolation alignment; Step S332: Fusion, splicing and recursive updating. The up-sampled features and defect location results are spliced and fused with the features of the current layer. The features are then input into the linear layer for linear transformation. The fused feature combination is: ,in, Represents the fused feature combination, including the current layer features, up-sampled features and positioning information, Represents the feature splicing operation, which splices each part along the feature channel one by one to form a new input; The linear transformation of the feature combination is: ,in, It is a linear transformation layer that performs weighted summation on the concatenated feature combinations to achieve linear integration of cross-layer information. is the feature after linear transformation, is the number of output feature channels; Step S333: Multi-scale fusion output and loss function definition. Through multi-scale linear transformation and layer-by-layer upsampling, the original point cloud resolution is finally restored. The linearly transformed features are combined through the defect head to obtain the final prediction output: ,in, represents the number of points in the original point cloud, For the The prediction uncertainty of each point, For the The probability of defect determination at each point; And the maximum likelihood estimation is used as the loss function to jointly optimize the uncertainty and prediction error, which can be expressed as: ； in, For the The true labels of the points include defects and non-defects. For the The defect location results of each point, For the The prediction uncertainty probability of each point.

8. The method according to claim 4, characterized in that The step S4 further comprises: Step S41, multi-strategy defect fusion and reconstruction. For large-scale point cloud structures, point cloud branches generated by multiple groups of heterogeneous pruning strategies are processed in parallel. Candidate defect points of each sub-point cloud block are collected, and their uncertainty defect probability scores, i.e., confidence scores and defect location result groups, are calculated. Subsequently, for the detection results output by different pruning strategies, the support statistics and uncertainty weighted fusion mechanism are used to effectively eliminate the local deviation caused by a single segmentation method, and finally a complete and unbiased global defect distribution is obtained, which is expressed as: ； in, Represents the uncertainty defect probability score array of point cloud block k, =[{ , ,..., }, { , ,..., }, ..., { , ,..., }], Each element in represents the uncertainty defect probability corresponding to the point cloud block after pruning and segmentation. represents the defect location result group of point cloud block k, , ], Each element in represents the defect localization result corresponding to the point cloud block after pruning and segmentation, and its value is 1 or 0, 1 represents a defect point, and 0 represents a non-defect point. g1…gt represents the pruning strategy, N represents the number of point cloud blocks after segmentation by each pruning strategy, and the function TrustFuse is a fusion module for multi-strategy pruning results. Step S42, reconstruct the point cloud with defect information, correspond the fused defect positioning results with the original point cloud structure, restore the defect information to the original point cloud density through linear interpolation, then generate a point-level defect mask according to the confidence threshold, perform denoising filtering and structural restoration on the defect area, and finally, embed the defect label and confidence into the original point cloud to obtain a reconstruction result of the same magnitude as the original point cloud and with accurate defect annotation.

9. The method according to claim 8, characterized in that The functions of the fusion module TrustFuse are: Step S411: Multiple pruning strategies fusion and consensus evaluation, for different pruning strategies First, merge the defect point sets detected by all pruning strategies to obtain the joint set of all candidate defect points: , where t represents the total number of pruning strategies, represents the set of defect points output by the j-th pruning strategy, Represents the set of all candidate defect points obtained by combining all strategies; Then, count each point Number of supports identified as weak points in each strategy: ； Among them, c(p) represents the number of point p that is simultaneously identified as defective points in multiple pruning strategies. This value is used to measure the degree of policy consensus of point p and serves as part of the subsequent fusion weight; Step S412: uncertainty weighted fusion, which fuses the confidence predictions of multiple strategies for point p, i.e., uncertainty probabilities, to construct the final credible confidence distribution: ； in, is the global fusion confidence of point p, is the confidence estimate of the j-th strategy for point p, is the fusion weight of the j-th strategy at point p, defined as: ； in, is the support mark of point p in the j-th strategy, which takes a value of 1 or 0; is the uncertainty probability of the j-th strategy predicting point p; It is a numerical stability term to prevent division by zero; , is the normalization factor, which ensures the normalization of the fusion result; Step S413: Extract credible defect points and output the results, setting the confidence threshold , filter out the final set of credible defect points: ; The final output is: , which is the credible global defect point set and point-level confidence distribution obtained by integrating the multi-strategy positioning results and removing the bias through uncertainty weighting and cross-validation.

10. The method according to claim 8, characterized in that The step S42 further includes: S421, defect information position alignment and interpolation propagation, mapping the defect positioning results back to the original high-density point cloud , a simple distance-weighted interpolation propagation strategy is adopted, which can be expressed as: ； in, represents the i-th credible defect point, , M is the total number of credible defect points, for The corresponding confidence level, is the interpolation kernel function, such as Gaussian kernel, represents the Euclidean distance between points, represents the defect confidence of point p estimated in the macro point cloud; S422, defect determination and mask generation, generate point-level masks based on the fusion confidence results , , used to distinguish defective areas from normal areas: ； in, It is a preset confidence threshold used to control the strictness of defect determination; S423, defect area denoising and structure restoration, the defect area marked in the mask is Sub-point cloud, filtering or reconstruction processing: , BilateralFilter is a bilateral filter; S424, defect annotation and result output, finally embeds the defect mask and confidence information into the original point cloud to generate a reconstructed point cloud result with explicit annotations: ； Each reconstructed point contains: the original geometric position of point p, the defect judgment label , corresponding to the defect confidence , which not only retains the original structure of the point cloud data, but also embeds the defect annotation information, namely the mask and confidence.

Images