Point cloud technology processing method for defect positioning of complex building scene

Abstract

The invention provides a complex building scene defect positioning-oriented point cloud technology processing method, which comprises the steps of firstly preprocessing collected point cloud data, then dividing a complex building scene through a semantic segmentation model, obtaining independent sub-point cloud units, and directly performing feature extraction and defect positioning on a small-scale structure, so that the defect positioning accuracy of the small-scale structure is improved. A large-scale structure is subjected to multi-stage farthest point sampling and strategy pruning, the large-scale structure is segmented into sub-point cloud blocks, then feature extraction and defect positioning are carried out, local and global feature extraction is carried out by adopting a PointConv neural network, structural unbiasedness and position robustness of feature expression are ensured through relative coordinate normalization, and the feature extraction accuracy is improved. Further introducing an uncertainty factor to enhance feature expression, performing interpolation up-sampling and feature splicing, fusing upper and lower layer features and confidence level by level to realize adaptive multi-level defect positioning, and finally performing global defect fusion and point cloud reconstruction. Through the scheme of the invention, rapid positioning and efficient identification of complex building scene defects are realized.

Description

Technical Field

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

[0002] Integrated construction defect detection in complex building scenes can not only avoid safety hazards during construction, but also effectively avoid potential safety hazards. At the same time, the integrated detection technology significantly improves the efficiency of quality monitoring, transforming traditional manual inspections towards intelligence and automation.

[0003] Early building defect detection mainly relied on manual inspections, suffering from problems such as low efficiency, strong subjectivity, and the inability to detect concealed defects. With the development of computer vision technology, building defect detection has gradually moved towards automation and intelligence. The initial automated detection methods used traditional image processing techniques. Researchers proposed methods based on edge detection and texture analysis, analyzing building surface images by designing artificial feature extraction operators to achieve preliminary identification of surface defects such as cracks. However, these methods rely heavily on manually designed features, with unstable defect feature expressions, poor robustness, and difficulty in dealing with complex actual scenes. At the same time, the high cost of collecting and processing the massive data required to obtain complete texture maps, as well as the significant increase in the difficulty of equipment layout and data synchronization.

[0004] With the rise of deep learning, detection methods based on convolutional neural networks have become mainstream. Researchers used the improved Region with Convolutional Neural Network Features (Faster R-CNN) network structure to process 2D images of building facades, significantly improving the accuracy of surface defect detection through end-to-end feature learning. However, these methods are still limited to the surface information obtained from visible light images and cannot effectively detect internal building structure defects. To obtain richer three-dimensional information, researchers began to attempt to use 3D laser scanning technology. The point cloud processing method based on PointNet++ achieved precise detection of building components through hierarchical sampling and local feature extraction. However, when dealing with large-scale complex building scenes, this method faces problems such as low computational efficiency and susceptibility to occlusion.

[0005] Current point cloud defect detection still remains at traditional lidar, millimeter-wave radar, and synthetic aperture radar technologies. In such a working background with low sampling points and simple scenes, the following main problems exist in the defect recognition and location of complex building scenes:

[0006] Insufficient processing capacity: For traditional point cloud scenarios, the scope of data processing is mainly limited by sampling devices. Currently, the accuracy of sampling devices such as lidar, millimeter-wave radar, and synthetic aperture radar cannot meet the need to capture the global point cloud scenario at one time, and traditional point cloud technologies are also difficult to extract and process point cloud scenarios with diverse defects in complex scenarios. These multiple limitations make it difficult to implement integrated construction defect detection;

[0007] Low processing efficiency: Traditional point cloud defect processing methods generally rely on the single-sampling and single-measurement method. For an entire building, it is necessary to collect the surfaces to be detected one by one. On the one hand, the defect location accuracy is limited by the sampling coverage rate, and on the other hand, the defect location efficiency is limited by the sampling rate. Moreover, in the face of the collected complete data, traditional methods also lack the ability to locate defects uniformly in complex scenarios.

[0008] Lack of comprehensive adaptation to the complexity and diversity of building structures: The current integrated construction defect detection system has limited adaptability to different types of building structures (such as curved surfaces and irregular geometric shapes). For example, in special scenarios such as curved surface curtain walls, long-span bridges, or prefabricated buildings, the generality of existing algorithms is insufficient, making it difficult to accurately model and detect, resulting in a large number of customized adjustments being required in actual projects, increasing the deployment difficulty and cost. Summary of the Invention

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

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

[0011] Step S1, perform preprocessing such as normalization and denoising on the original 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, use the efficient farthest point sampling (FPS) to reduce the data dimension, and automatically divide the overall scene into multiple structurally independent and granularity-adjustable sub-point cloud units through a pre-trained semantic segmentation model;

[0012] Step S2: For each sub-point cloud unit obtained after semantic segmentation, different processing strategies are set according to its scale. For small-scale point cloud structures, they are directly used as independent point cloud blocks for hierarchical feature extraction and defect localization; for large-scale point cloud structures, based on farthest point sampling, backbone points are generated, and point cloud cells are constructed with the backbone points as the core. Combining 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;

[0013] Step S3: For the independent point cloud blocks of small structures and the split sub-point cloud blocks, hierarchical feature extraction and defect localization are performed. The PointConv neural network is used for local and global feature extraction. By relative coordinate normalization, the structural unbiasedness and position robustness of feature expression are ensured. Further, an uncertainty factor is introduced to enhance the feature expression. Through interpolation upsampling and feature stitching, each resolution layer is restored in turn to achieve adaptive multi-level defect localization, and integrated spatial defect features and point-level confidence are obtained;

[0014] Step S4: For large-scale point cloud structures, based on the detection results of each sub-point cloud block, multi-strategy pruning fusion is introduced to cover and integrate the defect results under different pruning segmentation paths. Through overlapping stitching, confidence-weighted, and uncertainty-driven credibility evaluation, the bias caused by single segmentation is effectively eliminated to achieve unbiased fusion of defects in the global range. Finally, the defect information of all sub-point cloud blocks is integrated, and through interpolation and upsampling strategies, the defect points and their confidence are mapped back to the original point cloud resolution to achieve point cloud reconstruction with complete structure and detailed defect annotation.

[0015] Step S1 specifically includes:

[0016] First, for the overly dense sampling data, a hierarchical farthest point sampling strategy is adopted for downsampling and simplifying to an appropriate order of magnitude, and the overall topological structure of the building is ensured to be unified. Due to the locality of defects, the pre-trained model KPConv is used for semantic segmentation of complex large-scale point cloud scenes to obtain each relatively independent sub-point cloud unit. The specific process is as follows:

[0017] ;

[0018] ;

[0019] Among them, is the original point cloud scene data directly collected by the terahertz radar, is the point cloud scene data after being downsampled to an appropriate order of magnitude by FPS, represents each column of sub-point cloud units after semantic segmentation, such as platforms, scaffolds, etc.

[0020] Step S2 specifically includes:

[0021] For each sub - point cloud unit obtained after semantic segmentation, according to the scale of each semantic structure, a segmentation strategy for point cloud structures of different orders of magnitude is generated: for small - scale point cloud structures, they can be directly regarded as independent point cloud blocks for hierarchical division; for large - scale point cloud structures, further segmentation is carried out to obtain smaller second - level sub - point cloud blocks. 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 samplings until the point cloud set presents a line - level topological backbone structure. Subsequently, a multi - group cross - point segmentation strategy with coverage is carried out on this line structure to obtain each group of parallel second - level sub - point cloud blocks.

[0022] Step S3 specifically includes:

[0023] Step S31: Based on each segmented sub - point cloud block, with the backbone points obtained by farthest - point sampling as the center, a point cloud cell is constructed by aggregating around each backbone point, that is, the k - nearest - neighbor region of the backbone point. For each point cloud cell, a PointConv convolutional layer is used to perform local feature aggregation and learning to extract node spatial features. 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 spatially - normalized features are concatenated with the original features to obtain a set of point cloud feature representations with rich multi - level and spatial - structure information. Through multi - level recursive extraction, multi - scale expression of point cloud features is achieved. The features extracted at each level are integrated by concatenation 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 defect localization and uncertainty estimation.

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

[0025] Step S33: After obtaining the coarse - grained defect localization result, through up - sampling and interpolation mapping, the coarse - grained defect localization and uncertainty estimation results are mapped back to a higher - density point cloud level and concatenated and integrated with finer - grained local features. Subsequently, they are input into a linear layer for feature linear transformation to obtain the final defect localization results of each point cloud block, and the uncertainty estimation is corrected by combining the true - label information. Asynchronous feedback optimization is achieved through a loss function.

[0026] The core idea of the hierarchical feature extraction in step S3 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 designed in a hierarchical and progressive manner. In each iteration, the point set is downsampled, and the features in the local space are aggregated and enhanced, enabling the network to gradually enrich its understanding of the spatial structure from global to local and from coarse to fine.

[0027] In the hierarchical feature extraction, the specific process of each level of feature extraction is as follows: , represents the point set of the point cloud sub-structure obtained by the i-th downsampling of the point cloud block m, represents the point set of the point cloud sub-structure obtained by the (i + 1)-th downsampling of the point cloud block m. At the same time, can be regarded as the backbone node of. Next, according to we get the point cloud cells centered on the backbone points generated based on the backbone points in , which is denoted as . After such segmentation, the point cloud has strong spatial hierarchy, facilitating the next step of feature acquisition. Among them, represents the point cloud cell of the r-th backbone node in the point cloud sub-structure obtained by the i-th downsampling, represents the top k neighbor points in that are closest to in terms of Euclidean distance. For each of the obtained point cloud cells, the PointConv layer is used to perform aggregated feature learning on it to extract its node spatial features , where r represents the r-th backbone node, represents the richer spatial features of node r after aggregation, represents the module layer. To avoid feature noise caused by spatial offset, taking each point cloud cell as a unit, the local features of the neighbor points of each point cloud cell, that is, the spatial features, are referenced with the center point as the reference system, and the difference from the center point is used as the unbiased structural feature, that is, , represents the point set of the point cloud cell of the backbone point r the spatial features standardized with respect to the center point, that is, the unbiased structural features, represents the set of neighbor nodes, Then, after concatenating with we get more global spatial features.

[0028] After obtaining the local features layer by layer (usually five - layer feature extraction is performed, and it will increase if the number of points is too large to ensure that the pixel level of the last layer meets the actual requirements), the deepest dense features are used for coarse - grained defect localization and uncertainty prediction. Coarse - grained defect localization is carried out through non - linear transformation combined with an activation function, which is expressed as:

[0029] ;

[0030] Among them, is the activation function, are trainable parameters, d is the dimension of the feature channel, is the BN normalization layer, represents the deepest spatial feature 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. The first dimension represents the uncertainty probability and the second dimension represents the defect localization result .

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

[0032] Step S331, upsampling and interpolation mapping, restore (upsample) the features of the th layer ( ) and the coarse - grained defect localization result to the point cloud distribution of the previous layer , which is expressed as:

[0033] ;

[0034] Among them, is the interpolation mapping function, whose role is to map / interpolate the features and defect localization results of the th layer (coarse - grained layer) from to the corresponding point cloud positions of according to the spatial position, realizing the restoration from coarse to fine. Methods such as nearest - neighbor interpolation and inverse - distance weighting can be used. The three - dimensional position coordinate set of the point cloud of its th layer, is the three - dimensional position coordinate set of the point cloud of the previous layer, that is, the th layer (finer - grained), is the The coarse-grained defect localization results and uncertainty probabilities for all points in the layer is the number of points, and 2 represents the two components of the predicted output, namely the uncertainty probability and the defect localization result represents the global spatial features of all points in the layer, where d is the dimension of the feature channels represents the upsampled defect localization result for each point in the layer after interpolation is the upsampled feature result after interpolation alignment

[0035] Step S332, fusion splicing and recursive update. Concatenate and fuse the features and defect localization results obtained by upsampling with the features inherent in the current layer, and input them into a linear layer for feature linear transformation to achieve cross-layer information integration and prepare for subsequent layers. The fused feature combination is:

[0036] ;

[0037] where represents the fused feature combination, including the features of this layer, upsampled features, and localization information represents the feature splicing operation, which concatenates each part along the feature channels one by one to form a new input

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

[0039] ;

[0040] where is the linear transformation layer, which performs weighted summation on the concatenated feature combination 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 reduction, finally restore to the original point cloud resolution, and obtain the final predicted output by passing the linearly transformed feature combination through the defect head:

[0042] ;

[0043] where represents the number of points in the original point cloud is the predicted uncertainty of the th point is the defect determination probability of the th point

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

[0045] ;

[0046] where is the true label (defect / non-defect) of the -th point, is the defect localization result of the -th point, is the predicted uncertainty probability of the -th point. 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 to prevent the model from overextending the 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, a layer-by-layer fusion method is adopted to integrate the defect information from different sub-point cloud structures, different levels, and different pruning strategies. First, through the parallel detection of heterogeneous pruning strategies, the candidate defect points of each sub-point cloud block are collected, and their uncertainty defect probability scores, i.e., confidence levels, and the defect localization result groups are calculated. Pruning is to further segment the point cloud block. Subsequently, for the detection results output by different pruning strategies, through a support degree statistics and uncertainty weighted fusion mechanism, the local bias caused by a single segmentation method is effectively eliminated, and the credibility and consistency of the global results are improved. During the fusion process, the high-level structure can contain and accumulate the defect information of the lower-level granularity. Through feature stitching, residual accumulation, and confidence progression, the organic integration of multi-scale information is realized. Finally, the global credible defect point set and its point-level confidence distribution after debiasing superposition and confidence weighted are output, significantly enhancing the robustness and accuracy of large-scale point cloud defect localization under complex structures;

[0049] Step S42, point cloud reconstruction with defect information. The sparse and integrated defect localization results are mapped back to the original high-density point cloud to achieve explicit defect annotation and structure 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 according to the confidence level threshold, and denoising filtering and structure recovery are performed on the defect area to improve the boundary smoothness and spatial continuity. Finally, the defect labels and confidence levels are embedded into the original point cloud to obtain the reconstructed point cloud data with both structural integrity and accurate defect annotation, providing high-quality and seamless basic support for subsequent application scenarios such as analysis, visualization display, and automated modeling.

[0050] Step S41 specifically is as follows: For a large-scale point cloud structure, according to the defect recognition and confidence prediction results of sub-point cloud blocks, defect information is fused layer by layer to generate global defect localization and comprehensive confidence evaluation 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 a large-scale point cloud structure, the 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 localization results of each group, by designing multiple groups of pruning strategies, the coverage overlap between each cutting seam is realized. According to the uncertainty probability, the results of each group are weighted and fused, continuously purifying the global representation of the defect, and finally obtaining a complete and unbiased global defect distribution. Among them, reliable and unbiased defect fusion needs to eliminate the bias that may be brought by a single segmentation strategy. By fusing the defect point sets generated by multiple parallel segmentation strategies, the reliability and unbiasedness are ensured. Through cross-validation and parallel information superposition, finally, a reliable defect point set and confidence distribution are output, that is: , where represents the array of uncertainty defect probability scores of point cloud block k, = [{ , ,..., }}, { , ,..., }}, ……, { , ,..., }], Each element in represents the uncertainty defect probability corresponding to the point cloud block after pruning segmentation, represents the defect localization result group of point cloud block k, , , , Each element in represents the defect localization result corresponding to the point cloud block after pruning segmentation, taking values of 1 or 0. 1 represents a defect point, and 0 represents not a defect point. g1... gt represent pruning strategies, N represents the number of point cloud blocks segmented by each pruning strategy, and the function TrustFuse is the fusion module for the multi-strategy pruning results. The fusion module realizes the following functions:

[0052] Step S411, multi-pruning strategy fusion and consensus evaluation, for different pruning strategies , first, merge the defect point sets detected by all pruning strategies to obtain the combined overall candidate defect point set: , where t represents the total number of pruning strategies, Denote the set of defect points output by the j-th pruning strategy, which represents the set of all candidate defect points obtained by merging all strategies;

[0053] Subsequently, count each point for the number of times it is identified as a defect point in each strategy:

[0054] ;

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

[0056] Step S412, uncertainty weighted fusion, fuse the confidence predictions of multiple strategies for point p, that is, the uncertainty probability, to construct the final credible confidence distribution:

[0057] ;

[0058] Among them, is the confidence after global fusion 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] Among them, is the support label of point p in the j-th strategy, taking values of 1 or 0; is the uncertainty probability predicted by the j-th strategy for point p; is the numerical stability term to prevent division by zero;

[0061] , is the normalization factor to ensure the normalization of the fusion result;

[0062] This mechanism fuses and considers two important factors: one is the consensus support degree of multiple strategies , and the other is the confidence reliability of the predictions of each strategy (the smaller the uncertainty, the greater the weight).

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

[0064] The final output is: , that is, the credible global defect point set and point-level confidence distribution obtained by fusing the detection results of multiple strategies and debiasing through uncertainty weighting and cross-validation.

[0065] Step S42 is specifically as follows: Corresponding the fused defect localization result with the original point cloud structure, restoring the defect information to the original point cloud density through linear interpolation, and denoising the defects according to the original structure, finally obtaining a reconstruction result with the same order of magnitude as the original point cloud and accurate defect annotation. This reconstruction method not only maintains the integrity of the original structure but also realizes the precise localization and visualization of defects. It is implemented by the defect-annotated point cloud reconstruction module. The defect-annotated point cloud reconstruction module performs interpolation propagation and structure restoration, which is expressed as: where the function f represents the process of embedding the final credible defect information into the original macro point cloud and completing the explicit defect reconstruction, including the following steps:

[0066] S421, Defect information position alignment and interpolation propagation. Since the defect detection process is carried out in the relatively sparse structure after downsampling or pruning, it is necessary to map the detection result back to the original high-density point cloud and adopt a simple distance-weighted interpolation propagation strategy:

[0067] ;

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

[0069] S422, Defect determination and mask generation. Generate a point-level mask , , used to distinguish the defect area from the normal area:

[0070] ;

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

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

[0073] S424, Defect annotation and result output. Finally, the defect mask and confidence information are embedded 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 determination label , and the corresponding defect confidence . It 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 localization in complex building scenes, which realizes efficient splitting of the point cloud scene, 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 localization module, and a global multi-strategy defect fusion and reconstruction module. The complex scene adaptive segmentation module performs preprocessing such as normalization and denoising on the original point cloud data collected by the terahertz radar, and based on multi-source interpolation, complements the point cloud structure to obtain a high-fidelity macro point cloud input. Then, it uses the efficient farthest point sampling (FPS) to reduce the data dimension, and automatically divides the overall scene into multiple structurally independent and adjustable granularity sub-point cloud units through a pre-trained semantic segmentation model. According to the scale of the sub-point cloud units, it generates processing strategies for point cloud structures of different orders of magnitude. For small-scale point cloud structures, they are directly used as independent point cloud blocks for step-by-step feature extraction and defect localization; 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. Combining spatial layering and 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;

[0077] The hierarchical feature learning and defect localization module performs step-by-step feature extraction and defect localization. It uses neural networks such as PointConv for local and global feature extraction, ensures the structural unbiasedness and position robustness of the feature expression through relative coordinate normalization, further introduces an uncertainty factor to enhance the feature expression, and through residual update and multi-step loop, gradually fuses the features and confidences of the upper and lower layers to achieve adaptive multi-level defect localization and obtain integrated spatial defect features and point-level confidences;

[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 and fusion, covers and integrates the defect results under different pruning / splitting paths, and effectively eliminates the bias caused by single splitting through overlapping splicing, confidence weighting, and uncertainty-driven credibility evaluation, realizing unbiased fusion of defects in the global scope. Finally, it integrates the defect information of all sub-point cloud blocks, and through interpolation and upsampling strategies, maps the defect points and their confidence levels back to the original point cloud resolution, realizing point cloud reconstruction with complete structure and detailed defect annotation.

[0079] A point cloud technology processing method for defect location in complex building scenes provided by the present invention realizes precise defect location in complex building scenes using terahertz radar technology through adaptive hierarchical processing based on semantic segmentation, sub-structure defect location and upsampling restoration based on uncertainty, defect location methods integrating multiple levels and multiple strategies, and an integrated point cloud reconstruction mechanism with defect information.

[0080] For the adaptive hierarchical processing based on semantic segmentation, first, semantic segmentation is performed on the complex building scene 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 divided into hierarchical levels as independent point cloud blocks, while for large structures, farthest point sampling is used to generate backbone points, point cloud cells are generated based on the backbone points, and then different hierarchical sub-point cloud blocks are generated through multiple groups of breakpoint pruning strategies. This hierarchical and adaptive processing method ensures the efficient processing and structural integrity of complex scenes.

[0081] For the sub-structure defect location and upsampling restoration based on uncertainty, for each level of point cloud structure obtained by hierarchical processing, first, feature encoding is realized within the point cloud cell based on PointConv, and the relative coordinate system is adopted to improve the translation robustness of features. Combining with PointConvD, hierarchical downsampling and multi-scale feature aggregation are completed, and the discrimination ability is enhanced through local and global information fusion. An uncertainty branch is introduced as a feature channel. In the upsampling stage, interpolation and feature splicing are used to restore the deep coarse-grained prediction to a high resolution, and the uncertainty learning is continuously guided based on the true label, and it is dynamically used as a weighting factor for feature update, adaptively enhancing the sensitivity of the model to the boundary and abnormal regions, realizing fine-grained defect location and uncertainty estimation.

[0082] Defect localization method integrating multiple levels and multiple strategies. For substructures with different scales and different spatial distributions, a parallel processing mechanism of multi-level fusion and multi-strategy pruning is designed, which specifically includes accurate localization of local defects within micro point cloud blocks, cross-layer fusion of hierarchical defect points and confidence levels, and cross-coverage of results from different pruning strategies. This method can extract and fuse multi-source, redundant, and heterogeneous information to the greatest extent, effectively improve the accuracy and robustness of global defect localization, ensure that the fusion output is unbiased and credible, and provide reliable support for defect localization in large-scale scenarios;

[0083] Integrated point cloud reconstruction mechanism with defect information. After completing the fusion of credible defect points, the defect information is mapped back to the original point cloud density structure through linear interpolation and denoised in combination with the original geometric information, finally realizing 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] Improve processing capacity. Through the independent structure recognition and parallel segmentation technology of backbone points, the present invention can efficiently process large-scale point cloud data in complex building scenes. The macro point cloud multi-view splitter cuts the huge point cloud set into independent structures, provides a high-quality data basis for subsequent processing, effectively reduces data redundancy, and improves processing capacity;

[0086] Improve processing efficiency. Through the optimization of the farthest point sampling and multi-view segmentation scheme, the present invention significantly reduces the order of magnitude of point cloud data, enabling complex scenes to be processed at a lower computational cost. At the same time, the micro point cloud defect localization combines multi-scale feature learning and global information fusion to achieve fast localization and efficient recognition of defects, overall improving the processing efficiency of the system;

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

[0088] Enhance credibility. The present invention introduces an uncertainty evaluation model and adopts a weighted fusion strategy to eliminate errors caused by different segmentation schemes, ensuring the credibility of the defect localization results. Through the defect denoising and defect restoration processes, the finally generated defect distribution map has high consistency, 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 localization through semantic understanding and multi-scale feature learning. It can efficiently adapt to various scales and types of terahertz point cloud data, providing feasibility and reliability for the practical application of terahertz radar technology in complex building scenarios. At the same time, the semantic-based adaptive segmentation strategy and uncertainty-guided feature learning method proposed in the present invention take into account the global and local feature expressions, effectively improving the defect localization accuracy in scenarios such as complex buildings, expanding the technical boundaries of point cloud data processing, and providing new ideas for intelligent manufacturing, non-destructive defect localization, and other fields that require high-precision three-dimensional data processing, having important engineering application value and broad market prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0091] Figure 1 It is a schematic flowchart of a point cloud technology processing method for defect localization in complex building scenarios provided by an embodiment of the present invention;

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

[0093] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0094] The present invention proposes a point cloud technology processing method for defect localization in complex building scenarios, which mainly includes three parts:

[0095] Semantic-guided hierarchical processing strategy: After appropriate farthest point downsampling, semantic segmentation of the complex scene is performed through a pre-trained model to obtain independent sub-point cloud structures, and then elastic grading is performed according to the point cloud scale. The small-scale structures are directly output to the feature extraction module, while the large-scale structures adopt multi-level farthest point sampling and fixed strategy pruning to ensure the adaptability and efficiency of processing;

[0096] Substructure Defect Localization and Upsampling Restoration Based on Uncertainty: Within each independent substructure, the deep feature extraction module first performs high-dimensional encoding on its spatial features, and obtains coarse-grained defect localization and model uncertainty through a non-linear transformation that includes probability and uncertainty branches. Then, from bottom to top, using the methods of interpolation mapping and feature splicing, the deep localization is upsampled layer by layer to a higher resolution, and feature linear transformation is performed through a linear layer to achieve the integration of cross-layer information. Finally, fine-grained defect recognition and uncertainty estimation with the same resolution as the original point cloud are output. This fusion and progressive process enables the model to achieve "coarse to fine" discrimination, adaptively adjust feature weights according to uncertainty, and improve the reliability and discrimination ability of overall defect localization without unsupervised assistance;

[0097] Defect Reconstruction with Multi-Strategy Fusion: For the defect detection results of all sub-point cloud structures, a multi-strategy fusion mechanism is designed, including methods such as uncertainty-guided candidate aggregation, region merging with spatial position and semantic consistency, and probability-weighted global reconstruction, etc., to achieve unified defect fusion and restoration for regions of different scales and different complexities, eliminate redundancy and noise, and generate a final defect distribution map that seamlessly connects and meets the actual engineering requirements. Through this multi-strategy fusion step, the fine localization and global reconstruction capabilities are further enhanced, and high-confidence reconstruction of defects in complex scenarios is achieved.

[0098] As Figure 1 shown, a point cloud technology processing method for defect localization in complex building scenarios provided by the present invention includes the following steps:

[0099] Step S1, perform preprocessing such as normalization and denoising on the original 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, use the efficient farthest point sampling (FPS) to reduce the data dimension, and automatically divide the overall scene into multiple sub-point cloud units with independent structures and adjustable granularity through a pre-trained semantic segmentation model;

[0100] In this step, for the high-density point cloud data collected by the terahertz radar, first perform downsampling through the hierarchical farthest point sampling (FPS) algorithm, which not only reduces the data scale but also maintains the consistency of the overall topology of the building. On this basis, use the pre-trained KPConv model to perform semantic segmentation on the downsampled point cloud to automatically extract relatively independent sub-point cloud units and achieve precise localization of the defect area.

[0101] Step S2: For each sub - point - cloud unit obtained after semantic segmentation, according to its scale, generate processing strategies for point - cloud structures of different orders of magnitude. For small - scale point - cloud structures, directly use them as independent point - cloud blocks for hierarchical feature extraction and defect localization; for large - scale point - cloud structures, generate backbone points based on farthest - point sampling, construct point - cloud cells with the backbone points as the core, and combine spatial stratification and multiple groups of cross - pruning strategies to adaptively split the large point - cloud into multiple layers of overlapping and structurally complete sub - point - cloud blocks, realizing the multi - scale spatial decomposition of the point - cloud.

[0102] Point - cloud structures with the number of points in the range of several thousand to several tens of thousands are small - scale point - cloud structures. Small - scale point - cloud structures are directly processed as independent point - cloud blocks without further subdivision. Dense point - cloud structures with the number of points in the range of several hundred thousand to tens of millions are large - scale point - cloud structures. Large - scale point - cloud structures continue to use hierarchical farthest - point sampling and combine multiple groups of random pruning strategies to recursively segment and optimize the large - volume point - cloud, ensuring that each level of sub - structure can be efficiently processed and avoiding artificial division noise.

[0103] In this step, for point - cloud structures of different volumes, adopt the strategies of hierarchical refinement and multiple groups of cross - pruning: for small - scale point - cloud structures, directly process them as independent point - cloud blocks; for large - scale point - cloud structures, recursively use multi - level farthest - point sampling and multiple groups of pruning strategies to further divide them into several sub - structure blocks, and perform defect detection and fusion processing on each one, effectively improving the accuracy and robustness of defect localization, and being particularly suitable for the automatic segmentation and defect recognition of large - volume point - cloud data.

[0104] Step S3: For the independent point - cloud blocks of small structures and the split sub - point - cloud blocks, perform hierarchical feature extraction and defect localization. Use neural networks such as PointConv for local and global feature extraction, ensure the structural unbiasedness and position robustness of feature expression through relative - coordinate normalization, further introduce uncertainty factors to enhance the feature expression, and through interpolation up - sampling and feature stitching, restore each resolution layer in turn, realizing adaptive multi - level defect localization, and obtaining integrated spatial defect features and point - level confidence.

[0105] Step S4: For large - scale point - cloud structures, based on the detection results of each sub - point - cloud block, introduce multi - strategy pruning fusion, perform coverage integration on the defect results under different pruning and segmentation paths, and through overlapping stitching, confidence - weighted and uncertainty - driven credibility evaluation, effectively eliminate the bias caused by single segmentation, realize unbiased fusion of defects in the global range. Finally, integrate the defect information of all sub - point - cloud blocks, and through interpolation and up - sampling strategies, map the defect points and their confidence back to the original point - cloud resolution, realizing point - cloud reconstruction with complete structure and detailed defect annotation.

[0106] In this step, global defect fusion and point cloud reconstruction are required. After defect detection of sub-point cloud blocks, for larger point cloud structures, multi-strategy pruning and fusion techniques are used to integrate defect information under different granularities and strategies, and finally global defect visualization and point cloud reconstruction with defect annotations are achieved. This entire stage focuses on eliminating biases under different segmentation strategies or granularities, and improving the reliability and global consistency of defect recognition.

[0107] The present invention can efficiently adapt to terahertz point cloud data of various scales and types, take into account global and local feature expressions, effectively improve the defect detection accuracy in scenarios 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 dense sampling data, a progressive farthest point sampling strategy is adopted for downsampling to an appropriate order of magnitude, and the overall topological structure of the building is ensured to be unified. Since defects are local, a pre-trained model KPConv is used for semantic segmentation of complex large-scale point cloud scenes to obtain relatively independent sub-point cloud units. The specific process is as follows:

[0110] ;

[0111] ;

[0112] Among them, is the original point cloud scene data directly collected by the terahertz radar, is the point cloud scene data after downsampling to an appropriate order of magnitude by FPS, represents each column 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, according to the scale of each semantic structure, a segmentation strategy for point cloud structures of different orders of magnitude is generated: for small-scale point cloud structures, they can be directly regarded as independent point cloud blocks for hierarchical division; for large-scale point cloud structures, further segmentation is carried out to obtain smaller second-level 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 samplings until the point cloud set presents a line-level topological backbone structure, and then a covered multi-group cross-point segmentation strategy is carried out on this line structure to obtain each parallel second-level sub-point cloud block group.

[0115] Step S3 specifically includes:

[0116] Step S31: Hierarchical feature extraction. The core idea 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 designed in a hierarchical and progressive manner. In each iteration, the point set is downsampled, and the features in the local space are aggregated and enhanced, enabling the network to gradually enrich its understanding of the spatial structure from global to local and from coarse to fine. Each feature extraction includes, based on each segmented sub-point cloud block, taking the backbone points obtained by the farthest point sampling as the center, and aggregating to construct a point cloud cell around each backbone point, that is, the k-nearest neighbor region of the backbone point. For each point cloud cell, the PointConv convolutional layer is used to perform local feature aggregation and learning to extract the node spatial features. 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 spatially normalized features are concatenated with the original features to obtain a set of point cloud feature representations rich in multi-level and spatial structure information. Through multi-level recursive extraction, multi-scale expression of the point cloud features is achieved. The features extracted at each level are integrated by concatenation to form multi-scale point cloud spatial structure features with both global topology and local geometric details, providing high-quality feature support for the next defect localization and uncertainty estimation.

[0117] Step S32: Based on the local structural features extracted at the deepest layer, use non-linear transformation and activation functions to perform coarse-grained defect localization and uncertainty prediction for each point cloud block. This step outputs the defect localization results and uncertainty estimation values corresponding to each point cloud block, laying a foundation for subsequent defect refinement and criterion fusion.

[0118] Step S33: After obtaining the coarse-grained defect localization results, through upsampling and interpolation mapping, map the coarse-grained defect localization and uncertainty estimation results back to a higher-density point cloud level, and concatenate and integrate them with finer-grained local features. Subsequently, input them into a linear layer for feature linear transformation to obtain the final defect localization results of each point cloud block, and correct the uncertainty estimation in combination with the true label information. Asynchronous feedback optimization is achieved through the loss function.

[0119] In the hierarchical feature extraction, the specific process of each level of feature extraction is as follows: , represents the point set of the point cloud sub-structure obtained by the i-th downsampling of the point cloud block m. represents the point set of the point cloud sub-structure obtained by the (i + 1)-th downsampling of the point cloud block m. Meanwhile, can be regarded as 's backbone nodes. Next, according to obtain the point cloud cells centered on the backbone points generated based on the backbone points in, denoted as , the segmented point cloud has strong spatial hierarchy, facilitating the next-step feature acquisition. Among them, represents the point cloud cell of the r-th backbone node in the point cloud substructure obtained by the i-th downsampling. represents the in the The top k neighbor points with the closest Euclidean distance. For each obtained point cloud cell, the PointConv layer is used to perform aggregated feature learning on it to extract its node spatial features. , where r represents the r-th backbone node. represents the richer spatial features of node r after aggregation. represents module layer. To avoid feature noise caused by spatial offset, taking each point cloud cell as a unit, the local features of the neighbor points of each point cloud cell, that is, the spatial features, are referenced with the center point as the reference system, and the difference from the center point is used as the unbiased structural feature, that is, . represents the point cloud cell point set of the backbone point r. The spatial features standardized for the center point, that is, the unbiased structural features. represents the neighbor node set. Then are concatenated to obtain more global spatial features.

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

[0121] ;

[0122] Among them, is the activation function. are trainable parameters, d is the feature channel dimension. 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. The first dimension represents the uncertainty probability. , and the second dimension represents the defect localization result. , taking values of 1 or 0. 1 indicates a defective point, and 0 indicates a non-defective point.

[0123] After obtaining the coarse-grained defect localization results (usually by performing five-layer feature extraction, which will increase if the number of points is too large, ensuring that the pixel level of the last layer meets the actual requirements), use upsampling layer by layer to restore the defect localization results and uncertainty probabilities to the input point cloud scale. During the upsampling restoration process, continuously splice local features with finer-grained features and input them into a linear layer for feature linear transformation to obtain the final defect localization results output for each point cloud block. And in this process, calculate the loss of the uncertainty output results according to the true labels, specifically including:

[0124] Step S331, upsampling and interpolation mapping, to restore (upsample) the features of the -th layer ( ) and the coarse-grained defect localization results to the point cloud distribution of the previous layer , expressed as:

[0125] ;

[0126] Among them, is the interpolation mapping function, whose role is to map / interpolate the features and defect localization results of the -th layer (coarse-grained layer) from to the corresponding point cloud positions of according to the spatial positions, realizing the restoration from coarse to fine. Methods such as nearest neighbor interpolation and inverse distance weighting can be used. Its three-dimensional position coordinate set of the point cloud of the -th layer, is the three-dimensional position coordinate set of the point cloud of the previous layer, i.e., the -th layer (finer-grained), is the coarse-grained defect localization results and uncertainty probabilities of all points of the -th layer, is the number of points, 2 represents the two components of the predicted output, i.e., the uncertainty probability and the defect localization result, represents the global spatial features of all points of the -th layer, d is the dimension of the feature channel, represents the upsampled defect localization results for each point of the -th layer after interpolation, is the upsampled feature result after interpolation alignment.

[0127] Step S332, fusion splicing and recursive update, splice and fuse the features and defect localization results obtained by upsampling with the features inherent in the current layer, and input them into a linear layer for feature linear transformation to achieve the integration of cross-layer information and prepare for the subsequent layer. The fused feature combination is:

[0128] ;

[0129] Among them, represents the fused feature combination, including the features of this layer, upsampled features, and location information. represents the feature concatenation operation, which concatenates each part along the feature channels one by one to form a new input.

[0130] Then, a linear transformation is performed on the feature combination:

[0131] ;

[0132] Among them, is the linear transformation layer, which performs weighted summation on the concatenated feature combination 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 upsampling restoration layer by layer, it is finally restored to the original point cloud resolution. The feature combination after linear transformation passes through the defect head to obtain the final prediction output:

[0134] ;

[0135] Among them, represents the number of points in the original point cloud. is the prediction uncertainty of the th point. is the defect determination probability of the th point.

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

[0137] ;

[0138] Among them, is the true label (defect / non-defect) of the th point. is the defect location result of the th point. is the prediction uncertainty probability of the th point. 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 to prevent the model from overextending the uncertainty.

[0139] Step S4 specifically includes:

[0140] Step S41: Multi-strategy defect fusion and reconstruction. Based on the defect localization results of sub-point cloud blocks, adopt 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, collect candidate defect points of each sub-point cloud block, and calculate their uncertainty defect probability scores, i.e., confidence levels and defect localization result groups. Subsequently, for the detection results output by different pruning strategies, through a support statistics and uncertainty weighted fusion mechanism, effectively eliminate the local bias caused by a single segmentation method, and improve the credibility and consistency of the global results. During the fusion process, the high-level structure can contain and accumulate defect information at the lower-level granularity. Through feature splicing, residual accumulation, and confidence progression, achieve the organic integration of multi-scale information. Finally, output the globally credible defect point set and its point-level confidence distribution after debiasing superposition and confidence weighting, significantly enhancing the robustness and accuracy of large-scale point cloud defect localization under complex structures;

[0141] Step S42: Point cloud reconstruction with defect information. Map the sparse, integrated defect localization results back to the original high-density point cloud to achieve explicit defect annotation and structure restoration. First, spread and restore the credible defect points and their confidence levels to the original point cloud density through methods such as distance-weighted interpolation. Subsequently, generate a point-level defect mask according to the confidence threshold, perform denoising filtering and structure restoration on the defect area to improve the boundary smoothness and spatial continuity. Finally, embed the defect labels and confidence levels into the original point cloud to obtain the reconstructed point cloud data with both structural integrity and accurate defect annotation, providing high-quality and seamless basic support for subsequent application scenarios such as analysis, visualization, and automated modeling.

[0142] Figure 2 It shows the defect segmentation, localization and fusion of large-scale point cloud structures. For structures with an overly large point cloud volume, the spatial resources of the computing device are difficult to meet the extraction of high-dimensional spatial features of each node. Therefore, perform segmentation according to multiple sets of pruning strategies, conduct defect detection on each obtained sub-point cloud block. After obtaining the defect detection results of multiple sub-point cloud blocks, fuse the defect localization results of multiple sets of sub-point cloud blocks to avoid inaccurate defect localization caused by cutting noise.

[0143] Specifically, Step S41 is as follows: For large-scale point cloud structures, based on the defect recognition and confidence prediction results of sub-point cloud blocks, fuse defect information layer by layer to generate global defect localization and comprehensive confidence evaluation results. The defect points at the higher level will contain the information at the lower level, and during the fusion process, weight and adjust the confidence levels of the defect points, gradually accumulating the upper and lower level information. The specific fusion process is as follows:

[0144] For large-scale point cloud structures, the point cloud branch groups generated by multiple heterogeneous pruning strategies are processed in parallel. Pruning is to further segment the point cloud blocks. After obtaining the defect localization results of each group, by designing multiple pruning strategies, the coverage overlap between each cut seam is realized. According to the uncertainty probability, the results of each group are weighted and fused to continuously refine the global representation of the defect, and finally a complete and unbiased global defect distribution is obtained. Among them, reliable and unbiased defect fusion needs to eliminate the bias that may be brought by a single segmentation strategy. By fusing the defect point sets generated by multiple parallel segmentation strategies, reliability and unbiasedness are ensured. Through cross-validation and parallel information superposition, a reliable defect point set and confidence distribution are finally output, which is expressed as: , where 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 segmentation, represents the defect localization 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 segmentation, and the value is 1 or 0. 1 indicates a defect point, and 0 indicates not a defect point. g1... gt represent pruning strategies, N represents the number of point cloud blocks segmented by each pruning strategy, and the function TrustFuse is the fusion module of the multi-strategy pruning results. The fusion module realizes the following functions:

[0145] Step S411, multi-pruning strategy fusion and consensus evaluation. For different pruning strategies , first, the defect point sets detected by all pruning strategies are merged to obtain a combined set of all candidate defect points: , where t represents the total number of pruning strategies, represents the defect point set output by the jth pruning strategy, Denote the set of all candidate defect points obtained by merging all strategies;

[0146] Subsequently, count each point The number of times supported as a defect point in each strategy:

[0147] ;

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

[0149] Step S412, uncertainty weighted fusion, fuse the confidence predictions of multiple strategies for point p, that is, the uncertainty probability, to construct the final credible confidence distribution:

[0150] ;

[0151] Among them, is the confidence after global fusion 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] Among them, is the support label of point p in the j-th strategy, taking values of 1 or 0; is the uncertainty probability predicted by the j-th strategy for point p; is the numerical stability term to prevent division by zero;

[0154] , is the normalization factor to ensure the normalization of the fusion result;

[0155] This mechanism takes into account two important factors: one is the consensus support degree of multiple strategies , and the other is the confidence reliability of the predictions of each strategy (the smaller the uncertainty, the greater the weight).

[0156] Step S413, extract and output credible defect points, set the confidence threshold , and screen out the final credible defect point set: ;

[0157] The final output is: , that is, the credible global defect point set and point-level confidence distribution obtained by fusing the detection results of multiple strategies and debiasing through uncertainty weighting and cross-validation.

[0158] Step S42 specifically includes: Corresponding the fused defect localization result to the original point cloud structure, restoring the defect information to the original point cloud density through linear interpolation, and denoising the defects according to the original structure. Finally, a reconstruction result with the same order of magnitude as the original point cloud and accurate defect annotation is obtained. This reconstruction method not only maintains the integrity of the original structure but also realizes the precise localization and visualization of defects. It is implemented by the defect-annotated point cloud reconstruction module. The defect-annotated point cloud reconstruction module performs interpolation propagation and structure restoration, which is expressed as: , where the function f represents the process of embedding the final credible defect information into the original macro point cloud and completing the explicit defect reconstruction, including the following steps:

[0159] S421, Defect information position alignment and interpolation propagation. Since the defect detection process is carried out in a relatively sparse structure after downsampling or pruning, it is necessary to map the detection result back to the original high-density point cloud , and a simple distance-weighted interpolation propagation strategy is adopted:

[0160] ;

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

[0162] S422, Defect determination and mask generation. Generate a point-level mask , , used to distinguish the defect area from the normal area:

[0163] ;

[0164] Among them, is a preset confidence threshold (such as 0.5), used to control the strictness of defect determination.

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

[0166] S424, Defect annotation and result output. Finally, the defect mask and confidence information are embedded into the original point cloud to generate a reconstructed point cloud result with explicit annotations:

[0167] ;

[0168] Among them, each reconstructed point contains: the original geometric position of point p (i.e., coordinates x, y, z), the defect determination label , and the corresponding defect confidence . It 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 localization in complex building scenes, which realizes efficient splitting of the point cloud scene, 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 localization module, and a global multi-strategy defect fusion and reconstruction module. The complex scene adaptive segmentation module preprocesses the original point cloud data collected by the terahertz radar, such as normalization and denoising, and based on multi-source interpolation, complements the point cloud structure to obtain a high-fidelity macro point cloud input. Then, it uses the efficient farthest point sampling (FPS) to reduce the data dimension, and automatically divides the overall scene into multiple structurally independent and granularity-adjustable sub-point cloud units through a pre-trained semantic segmentation model. According to the scale of the sub-point cloud units, it generates processing strategies for point cloud structures of different orders of magnitude. For small point cloud structures, they are directly used as independent point cloud blocks for hierarchical feature extraction and defect localization; 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. Combining spatial stratification and 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;

[0170] The hierarchical feature learning and defect localization module performs hierarchical feature extraction and defect localization. It uses neural networks such as PointConv for local and global feature extraction, ensures the structural unbiasedness and position robustness of feature expressions through relative coordinate normalization, further introduces uncertainty factors to enhance the feature expressions, and through residual updates and multi-step loops, hierarchically fuses the upper and lower layer features and confidences to achieve adaptive multi-level defect localization, obtaining integrated spatial defect features and point-level confidences;

[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 to perform coverage integration on the defect results under different pruning / splitting paths. Through overlapping stitching, confidence-weighting, and uncertainty-driven credibility evaluation, it effectively eliminates the bias caused by single segmentation, realizes unbiased defect fusion in the global scope. Finally, it integrates the defect information of all sub-point cloud blocks, and through interpolation and upsampling strategies, maps the defect points and their confidence levels back to the original point cloud resolution to achieve point cloud reconstruction with complete structure and detailed defect annotation.

[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 them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions 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 scenarios, characterized in that The method includes: Step S1: Normalize and denoise the original 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, use the farthest point sampling (FPS) to reduce the data dimension, and divide the overall scene into multiple structurally independent and granularity-adjustable sub-point cloud units through a pre-trained semantic segmentation model. Step S2: For each sub-point cloud unit obtained after semantic segmentation, set different processing strategies according to its scale. For small-scale point cloud structures, directly perform hierarchical feature extraction and defect localization as independent point cloud blocks. For large-scale point cloud structures, generate backbone points based on the farthest point sampling, construct point cloud cells with the backbone points as the core, and combine multiple groups of cross-pruning strategies to adaptively split the large point cloud into multi-layer, overlapping, and structurally complete sub-point cloud blocks, realizing the multi-scale spatial decomposition of the point cloud. Among them, small-scale point cloud structures are point cloud structures with the number of points in the range of thousands to tens of thousands, and large-scale point cloud structures are dense point cloud structures with the number of points in the range of hundreds of thousands to tens of millions. Step S3: For independent point cloud blocks and the split sub-point cloud blocks, perform hierarchical feature extraction and defect localization. Use the PointConv neural network to extract local and global features, ensure the structural unbiasedness and position robustness of feature expressions through relative coordinate normalization, further introduce an uncertainty factor to enhance the feature expressions, and through interpolation upsampling and feature splicing, restore each resolution layer in turn to achieve adaptive multi-level defect localization, obtaining integrated spatial defect features and point-level confidence. Step S4: For large-scale point cloud structures, based on the localization results of each sub-point cloud block, introduce multi-strategy pruning fusion, cover and integrate the defect results under different pruning segmentation paths, and effectively eliminate the bias caused by single segmentation through overlapping splicing, confidence weighting, and uncertainty-driven credibility evaluation to achieve unbiased fusion of defects in the global range. Finally, integrate the defect information of all sub-point cloud blocks, and map the defect points and their confidence to the original point cloud resolution through interpolation to achieve point cloud reconstruction with a complete structure and detailed defect annotation.

2. The method according to claim 1, wherein The step S1 further includes: Use the pre-trained model KPConv to perform semantic segmentation on the complex point cloud large scene to obtain each independent sub-point cloud unit, denoted as: ； ； Among them, represents the original point cloud scene data collected by the terahertz radar, is the point cloud scene data after FPS downsampling, represents each sub-point cloud unit column after semantic segmentation.

3. The method according to claim 1, wherein The logic of the pruning strategy in the step S2 is to use multiple farthest point samplings until the point cloud set presents a line-level topological backbone structure, and then perform a covered multi-group cross-point segmentation strategy on this line structure to obtain each parallel secondary sub-point cloud block group.

4. The method according to claim 1, wherein The step S3 further includes: Step S31: Hierarchical feature extraction. The idea of hierarchical 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 progresses hierarchically. In each round of iteration, the point set is downsampled, and the features are aggregated and enhanced 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. Step S32: After obtaining the local features layer by layer, based on the spatial features of the deepest layer, use non-linear transformation and activation function to perform coarse-grained defect localization and uncertainty prediction on each point cloud block, and output the defect localization result and uncertainty estimation probability corresponding to each point cloud block; Step S33: After obtaining the coarse-grained defect localization result, through upsampling and interpolation mapping, map the coarse-grained defect localization and uncertainty estimation results back to a higher-density point cloud level, and splice and integrate them with finer-grained local features. Then input them into a linear layer for feature linear transformation to obtain the final defect localization results of each point cloud block, and correct the uncertainty estimation in combination with the true label information, and achieve asynchronous feedback optimization through the loss function.

5. The method according to claim 4, wherein The specific feature extraction at each level in Step S31 includes: According to each segmented sub-point cloud block, with the backbone points obtained by farthest point sampling as the center, aggregate and construct point cloud cells around each backbone point, that is, the k-nearest neighbor region of the backbone point. For each point cloud cell, use the PointConv convolutional layer to perform local feature aggregation and learning, extract the node spatial features, and normalize the local coordinate system of each point cloud cell with the backbone point as the center. Finally, splice the spatially normalized features with the original features to obtain a set of point cloud feature representations with rich multi-level and spatial structure information, which is specifically expressed as: , where represents the point set of the point cloud sub-structure obtained by the i-th downsampling of the point cloud block m, represents the point set of the point cloud sub-structure obtained by the (i + 1)-th downsampling of the point cloud block m. Regarding as 's backbone node, according to to obtain the point cloud cells centered on the backbone points generated by the backbone points in , denoted as , where represents the point cloud cell of the r-th backbone node in the point cloud sub-structure obtained by the i-th downsampling, represents in the top k neighbor points with the closest Euclidean distance to . For each obtained point cloud cell, use the PointConv layer to perform aggregated feature learning on it and extract its node spatial features , where represents the richer spatial features of node r after aggregation, represents module layer. Then, taking each point cloud cell as a unit, use the local features of the neighbor points of each point cloud cell, that is, the spatial features, with the center point as a reference and the difference from the center point as an unbiased structural feature, that is , represents the point cloud cell point set of the backbone point r the spatial features standardized for the center point, that is, the unbiased structural features, represents the neighbor node set, and then are concatenated to obtain more global spatial features.

6. The method according to claim 4, wherein The specific content of Step S32 includes: Coarse-grained defect localization is performed through non-linear transformation combined with an activation function, expressed as: , where is the activation function, are trainable parameters, d is the dimension of the feature channel, is the BN normalization layer, represents the spatial features of the deepest layer 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 , and the second dimension represents the defect localization result .

7. The method according to claim 4, wherein The further content of Step S33 includes: Step S331, upsampling and interpolation mapping, restore the features of the layer and the coarse-grained defect localization result, that is, upsample to the previous layer, namely layer's point cloud distribution, expressed as: ； Among them, , is an interpolation mapping function, whose role is to map the features and defect localization results of the th layer, i.e., the coarse-grained layer, according to the spatial position, from map / interpolate to the corresponding point cloud position, realizing the reduction from coarse to fine. is the set of three-dimensional position coordinates of the point cloud of the th layer. is the set of three-dimensional position coordinates of the previous layer, i.e., the point cloud of the th layer with finer granularity. is the coarse-grained defect localization result and uncertainty probability of all points in the th layer. is the number of points. 2 represents the two components of the predicted output, i.e., the uncertainty probability and the defect localization result. represents the global spatial features of all points in the th layer, and d is the dimension of the feature channel. represents the upsampled defect localization result for each point in the th layer after interpolation. is the upsampled feature result after interpolation alignment; Step S332, fusion splicing and recursive update: splice and fuse the features obtained by upsampling and the defect localization results with the features of the current layer itself, and input them into a linear layer for linear feature transformation. The fused feature combination is: , where represents the fused feature combination, including the features of this layer, upsampled features, and localization information. represents the feature splicing operation, which splices each part along the feature channels one by one to form a new input. The linear transformation of the feature combination is as follows: , where is the linear transformation layer, which performs weighted summation on the concatenated feature combination 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 restoration, it is finally restored to the original point cloud resolution. The feature combination after the linear transformation is passed through the defect head to obtain the final prediction output: , where represents the number of points in the original point cloud, is the th point's prediction uncertainty, is the th point's defect determination probability; And use the maximum likelihood estimation as the loss function to jointly optimize the uncertainty and prediction error, which is expressed as: ； Among them, is the true label (defect / non-defect) of the th point, is the defect location result of the th point, is the predicted uncertainty probability of the th point.

8. The method according to claim 4, wherein The further content of Step S4 includes: Step S41: Multi-strategy defect fusion and reconstruction. For large-scale point cloud structures, use multiple groups of heterogeneous pruning strategies to generate point cloud branch groups for parallel processing, collect the candidate defect points of each sub-point cloud block, and calculate their uncertainty defect probability scores, that is, confidence levels and defect localization result groups. Then, for the detection results output by different pruning strategies, through a support degree statistics and uncertainty weighted fusion mechanism, effectively eliminate the local bias brought by a single segmentation method, and finally obtain a complete and unbiased global defect distribution, which is expressed as: ； Among them, represents the array of uncertainty defect probability scores of the point cloud block k, =[{ , ,..., }, { , ,..., }, ……, { , ,..., }], Each element in it represents the uncertainty defect probability corresponding to the point cloud block after pruning and segmentation, represents the array of defect localization results of the point cloud block k, , , Each element in it represents the defect localization result corresponding to the point cloud block after pruning and segmentation, with the value being 1 or 0. 1 represents a defect point, and 0 represents a non-defect point. g1……gt represent pruning strategies, N represents the number of point cloud blocks after each pruning strategy segmentation, and the function TrustFuse is the fusion module for the multi-strategy pruning results; Step S42: Point cloud reconstruction with defect information. Correlate the fused defect localization 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 structure restoration on the defect area. Finally, embed the defect label and confidence into the original point cloud to obtain a reconstruction result with the same magnitude as the original point cloud and accurate defect annotation.

9. The method according to claim 8, characterized in that The function of the fusion module TrustFuse is: Step S411, multi-pruning strategy fusion and consensus evaluation, for different pruning strategies , first, merge the sets of defect points detected by all pruning strategies to obtain a combined 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 merging all strategies; Subsequently, count each point The number of supports recognized as defect points in each strategy: ； Among them, c(p) represents the number of times that point p is simultaneously identified as a defect point in multiple pruning strategies. This value is used to measure the degree of strategy consensus of point p and is used as part of the subsequent fusion weight; Step S412: Uncertainty weighted fusion. Fuse the confidence predictions of multiple strategies for point p, that is, the uncertain probability, to construct the final credible confidence distribution: ； Among them, is the confidence after global fusion of point p, is the confidence estimation of the j-th strategy for point p, is the fusion weight of the j-th strategy at point p, defined as: ； Among them, is the support flag of point p in the j-th strategy, taking values of 1 or 0; is the uncertainty probability predicted by the j-th strategy for point p; is a numerical stability term to prevent division by zero; , is a normalization factor to ensure the normalization of the fusion result; Step S413, extraction of believable defect points and result output, set the confidence threshold , and screen out the final set of believable defect points: ; The final output is as follows: , which is a set of credible global defect points and point-level confidence distributions obtained by fusing multi-strategy positioning results and debiasing through uncertainty weighting and cross-validation.

10. The method according to claim 8, wherein The further content of Step S42 includes: S421, Defect information position alignment and interpolation propagation, mapping the defect localization result back to the original high-density point cloud , adopting a simple distance-weighted interpolation propagation strategy, expressed as: ； Among them, represents the i-th credible defect point, , where M is the total number of credible defect points, is the corresponding confidence level, is an interpolation kernel function, such as a Gaussian kernel, represents the Euclidean distance between points, represents the defect confidence level estimated by point p in the macro point cloud; S422, Defect determination and mask generation, generating a point-level mask based on the fused confidence result , , used to distinguish the defect area from the normal area: ； Among them, is a preset confidence threshold used to control the strictness of defect determination; S423, Defect area denoising and structure restoration. For the sub-point cloud marked as the defect area in the mask, i.e., perform filtering or reconstruction processing: , where BilateralFilter is a bilateral filter; S424, Defect annotation and result output. Finally, the defect mask and confidence information are embedded into the original point cloud to generate a reconstructed point cloud result with explicit annotations: ； Among them, each reconstruction point contains: the original geometric position of point p, and a defect determination label , corresponding defect confidence , which not only retains the point cloud data of the original structure, but also embeds defect annotation information, namely masks and confidence, therein.

Images