Wall flatness detection method and system based on three-dimensional scanning technology

Abstract

The invention relates to the technical field of wall surface detection, in particular to a wall surface flatness detection method and system based on a three-dimensional scanning technology, and provides the following scheme: obtaining global point cloud data through a laser scanner, calculating macroscopic flatness and marking a problem area; performing local scanning on the problem area by using a structured light scanner to obtain local point cloud data; identifying defect types by applying a deep learning algorithm, and calculating local flatness; registering the global point cloud data and the local point cloud data, and calculating a correction coefficient based on an optimal registration result; and weighting the local flatness by using the correction coefficient, and fusing the local flatness with the macroscopic flatness to generate a wall flatness report. The report includes overall flatness evaluation, local defect details, and a flatness profile. The invention provides a high-precision global and local flatness fusion analysis method, and the accuracy and comprehensiveness of wall flatness detection are effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of wall surface detection, and in particular to a wall surface flatness detection method and system based on three-dimensional scanning technology. Background Art

[0002] Wall flatness is an important parameter for building construction quality control, which directly affects the aesthetics, functionality and service life of the wall. Traditional wall flatness inspection mainly relies on manual methods, such as using tools such as level, ruler and laser marking instrument for measurement. Although these methods are simple and easy to use, the measurement accuracy of manual operation is limited, and the inspection efficiency is low, which makes it difficult to meet the accuracy and real-time requirements of large-area wall flatness inspection. In addition, manual methods have great limitations in identifying subtle defects and cannot effectively meet the needs of high-standard building quality control.

[0003] In recent years, 3D scanning technology has been widely used in the field of building quality inspection. 3D scanning technology based on laser scanners, structured light scanners and other equipment can generate high-precision wall point cloud data, which can fully obtain wall surface information and provide new possibilities for flatness detection. Current 3D scanning technology has been used for wall flatness detection, usually by laser scanning the entire wall and calculating the flatness of the wall by processing and analyzing the point cloud data. However, relying solely on the analysis of global point cloud data, it is difficult to perform fine detection of local minor defects, which may lead to the loss of detail information, thereby affecting the accuracy of flatness assessment.

[0004] For example, a Chinese patent with publication number CN117760344A discloses a wall flatness detection system and method for civil engineering projects, which relates to the field of wall flatness detection technology and includes a drone and a cloud server. The drone is equipped with a depth camera for acquiring depth images, and the acquired depth images are uploaded to the cloud server. The cloud server is provided with a building model construction module, a path planning module, a path navigation module and a flatness detection module. The present invention constructs a three-dimensional model of a building and then extracts a number of visual reference parts from the three-dimensional model as the flight route of the drone. At the same time, the walls in the rooms of the building are automatically identified, and wall flatness check points are formulated according to the parameters of each wall, which are connected with the flight route to form a path route. During the detection process, the drone can complete the flatness detection of all the walls in the room by flying according to the path route. Not only is the detection speed fast but also the detection accuracy is high, which effectively reduces the use of manual labor and improves the detection efficiency.

[0005] The above patents all have the problem raised by this background technology: the existing wall flatness detection method is difficult to simultaneously take into account the fine detection of global flatness and local defects, resulting in insufficient accuracy and completeness of the detection results. In order to solve the above problems, this application designs a wall flatness detection method and system based on three-dimensional scanning technology. Summary of the invention

[0006] The technical problem to be solved by the present invention is to address the deficiencies of the prior art, and provide a wall flatness detection method and system based on three-dimensional scanning technology. The global point cloud data is obtained by a laser scanner, the macroscopic flatness is calculated and the problem area is marked; the problem area is locally scanned using a structured light scanner to obtain local point cloud data; a deep learning algorithm is applied to identify the defect type and calculate the local flatness; the global and local point cloud data are aligned, and the correction coefficient is calculated based on the optimal alignment result; the local flatness is weighted using the correction coefficient and merged with the macroscopic flatness to generate a wall flatness report. The report includes an overall flatness assessment, local defect details and a flatness distribution map. The present invention provides a high-precision global and local flatness fusion analysis method, which effectively improves the accuracy and comprehensiveness of wall flatness detection.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A wall flatness detection method based on three-dimensional scanning technology, the method comprising:

[0009] Use a laser scanner to globally scan the wall surface and obtain global point cloud data;

[0010] Processing the global point cloud data, calculating the macroscopic flatness of the wall through multi-scale analysis, and marking the problem areas;

[0011] Use a structured light scanner to locally scan the marked problem area to obtain local point cloud data;

[0012] Processing the local point cloud data, identifying the defect type of the problem area through a deep learning algorithm, and calculating the local flatness according to the defect type;

[0013] Performing point cloud registration on the global point cloud data and the local point cloud data, and calculating a correction coefficient according to the registration result;

[0014] The macroscopic flatness is corrected according to the local flatness and the correction coefficient, and a wall flatness report is generated.

[0015] The macroscopic flatness of the wall is calculated by multi-scale analysis, including:

[0016] The global point cloud data is converted into a world coordinate system, the world coordinate system is divided into grids at equal intervals, the grids are clustered according to a clustering algorithm, the grids in the same central cluster are merged, and the geometric features of the point cloud are extracted;

[0017] Decomposing the global point cloud data into scale-level regions according to the point cloud geometric features, and calculating the flatness in each scale-level region, wherein the scale-level region includes a macro-scale region, a local-scale region, and a micro-scale region;

[0018] The flatness of all areas is weighted and fused to calculate the macroscopic flatness.

[0019] The performing cluster analysis on the grid according to the clustering algorithm comprises:

[0020] Performing preliminary clustering on the global point cloud data in the grid according to the DBSCAN clustering algorithm to obtain core points, and determining the number of cluster center clusters according to the number of core points;

[0021] The core points after DBSCAN clustering are used as the initial center clusters of the K-means clustering algorithm, and the Euclidean distances from the global point cloud data in the grid to each of the initial center clusters are calculated;

[0022] Clustering is performed according to the Euclidean distance, and the average clustering error of the initial central cluster is calculated. The central cluster is updated according to the average clustering error, and it is determined whether the central cluster meets the convergence condition after the update. If the convergence condition is not met, the central cluster is continued to be updated. If the convergence condition is met, the clustering result is output.

[0023] The defect types identified in the problem area through deep learning algorithms include:

[0024] Randomly select three non-collinear points in the local point cloud data, calculate the point cloud plane equation, iterate the point cloud plane equation by calculating the point cloud standard deviation, and calculate the best fitting plane equation of the local point cloud data;

[0025] Filter out the point cloud data whose plane normal vector in the best fitting plane equation is equal to zero in the y-axis, mark the point cloud data as vertical point cloud, calculate the vertical average distance from the vertical point cloud to the corresponding best fitting plane equation, and calculate the wall point cloud features according to the vertical average distance. The calculation formula of the wall point cloud features is:

[0026] ,

[0027] in, Represents the wall point cloud feature, i represents a single vertical point cloud, N represents the total number of vertical point clouds, represents the gradient of the i-th vertical point cloud, represents the vertical distance of the i-th vertical point cloud, d represents the vertical average distance, represents the second-order curvature of the i-th vertical point cloud, represents the normal vector of the i-th vertical point cloud, Represents the component of the i-th vertical point cloud normal vector on the y-axis, Represents the point cloud density of the i-th vertical point cloud;

[0028] A defect recognition network is constructed, the wall point cloud features are used as input parameters of the defect recognition network, the input parameters are processed by the defect recognition network, and the defect type is output.

[0029] The defect recognition network comprises:

[0030] An input layer, which establishes a feature sequence according to input parameters and calculates the fitting value of the feature sequence;

[0031] A hidden layer converts the fitting value of the feature sequence of the input layer into a high-dimensional space, and calculates the output value of the hidden layer in the high-dimensional space;

[0032] The output layer is used to output the defect recognition result.

[0033] The processing of the input parameters by the defect recognition network includes:

[0034] Initializing the defect recognition network according to the number of nodes in the input layer, the number of nodes in the hidden layer, and the number of nodes in the output layer of the defect recognition network, and determining the connection weights between the input layer and the hidden layer and the node thresholds between the hidden layer and the output layer;

[0035] The input parameters are biased by the connection weights of the input layer and the hidden layer to obtain the output of the hidden layer, and the output of the hidden layer is updated by the node threshold to obtain the output layer input;

[0036] Calculating the error of the output layer input according to the error function of the output layer, judging whether the error of the output layer input meets the convergence requirement of the output layer, and if not, updating the connection weight and the node threshold;

[0037] If the convergence requirement is met, the defect identification result is output.

[0038] The point cloud registration comprises:

[0039] Extracting feature points from the global point cloud data and the local point cloud data respectively, and calculating a rotation matrix and a translation matrix according to the feature points;

[0040] The local point cloud data is transformed according to the rotation matrix and the translation matrix to obtain the iterative point cloud, and the average distance between the corresponding position of the iterative point cloud in the global point cloud data and the global point cloud data is calculated;

[0041] If the average distance is less than the point cloud distance threshold, stop the iterative calculation; if the average distance is greater than or equal to the point cloud distance threshold, use the iterative point cloud as the new local point cloud data to recalculate the minimum average distance, rotation matrix and translation matrix;

[0042] According to the minimum average distance, rotation matrix and translation matrix after the iteration, the correction coefficient is calculated. The calculation formula of the correction coefficient is:

[0043] ,

[0044] in, represents the correction coefficient, j represents a single point cloud to be registered, M represents the total number of point clouds to be registered, represents the density difference of the jth point cloud to be registered, represents the minimum average distance, represents the morphological complexity of the jth point cloud to be registered, represents the consistency value of the normal vector of the jth point cloud to be registered, R represents the rotation matrix, T represents the translation matrix, r(•) represents the rank calculation function, which is used to solve the rank of the matrix. represents the jth point cloud to be registered in the local point cloud data, and w represents the jth point cloud to be registered in the global point cloud data.

[0045] The generating of the wall flatness report includes:

[0046] Performing weighted processing on the local flatness according to the correction coefficient corresponding to the problem area;

[0047] The weighted local roughness set is merged with the macro-roughness to generate the corrected macro-roughness;

[0048] A wall surface flatness report is generated based on the corrected macro-flatness, the report including an overall flatness assessment, local defect details, and a flatness distribution map.

[0049] A wall flatness detection system based on three-dimensional scanning technology, the system comprising a data acquisition module, a data processing module, a point cloud correction module and a report generation module;

[0050] The data acquisition module is used to collect all and part of the three-dimensional point cloud data of the wall;

[0051] The data processing module is configured with a flatness analysis strategy, and the flatness analysis strategy is used to process the collected point cloud data and calculate the flatness information;

[0052] The point cloud correction module is used to align the global and local point cloud data, calculate the correction coefficient, and correct the macroscopic flatness according to the correction coefficient;

[0053] The report generation module is used to generate a wall surface flatness inspection report based on the corrected macroscopic flatness, and output the overall evaluation and local defect details.

[0054] The data processing module comprises:

[0055] A preprocessing unit, used for preprocessing the collected global point cloud data and local point cloud data, wherein the preprocessing includes noise filtering, outlier removal and data format conversion;

[0056] A multi-scale analysis unit is used to divide the pre-processed global point cloud data into hierarchical regions of different scales and calculate the corresponding flatness value in each scale region;

[0057] The defect recognition unit is used to identify and classify the defect in the flatness abnormal area in the local point cloud data, and extract the defect type and the flatness of the defect area;

[0058] The flatness analysis strategy includes global analysis logic and local analysis logic;

[0059] The global analysis logic is configured in the multi-scale analysis unit, and is used to calculate the macroscopic smoothness of the wall surface through multi-scale analysis, and mark the problem areas;

[0060] The local analysis logic is configured in the defect recognition unit, and is used to identify the defect type of the problem area through a deep learning algorithm, and calculate the local flatness according to the defect type.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] 1. The present invention combines laser scanning and structured light scanning technology to achieve high-precision fusion of global and local point cloud data, making the detection results more accurate, especially suitable for the detection of complex wall structures;

[0063] 2. The present invention combines multi-scale analysis with the process of automatically generating a flatness report, greatly improving the detection efficiency, and provides the overall flatness of the wall, details of local defects and a flatness distribution map, providing a clear basis for subsequent construction and repair. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0065] Figure 1 This is a schematic diagram of the flow of a wall flatness detection method based on three-dimensional scanning technology according to Embodiment 1 of the present invention;

[0066] Figure 2 This is a flowchart of point cloud geometric feature extraction in Example 1 of the present invention;

[0067] Figure 3 This is a module diagram of a wall flatness detection system based on three-dimensional scanning technology according to Example 2 of the present invention. DETAILED DESCRIPTION

[0068] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0069] Embodiment 1:

[0070] See also Figure 1 The present invention provides an embodiment: a wall flatness detection method based on three-dimensional scanning technology, wherein the specific steps of the method are as follows:

[0071] S1: Use a laser scanner to globally scan the wall surface and obtain global point cloud data;

[0072] In this step, the global point cloud data contains the 3D coordinate information and positional relationship of each point on the wall, which can fully record the overall shape, unevenness, inclination and other characteristics of the wall, and provide basic data for subsequent flatness analysis and local defect identification. The high precision and large-scale scanning capabilities of the laser scanner ensure the comprehensiveness and accuracy of the global point cloud data.

[0073] S2: Process the global point cloud data, calculate the macroscopic flatness of the wall through multi-scale analysis, and mark the problem areas;

[0074] In this step, multi-scale analysis involves dividing the global point cloud data into different levels of scale and calculating the flatness at each scale. Through this hierarchical flatness analysis, areas on the wall that may have flatness problems can be identified and marked, and possible defects on the wall can be preliminarily located, providing a reference for subsequent detailed scanning and defect identification. The multi-scale analysis method allows flatness inspection to have both an overall perspective and focus on details.

[0075] S3: Use a structured light scanner to perform local scanning on the marked problem area to obtain local point cloud data;

[0076] In this step, the structured light scanner has higher accuracy than the laser scanner and can capture tiny surface features. Through this refined scanning, detailed 3D data of the problem area can be obtained to identify the specific type of wall defects and evaluate local flatness. This step ensures the accuracy and detail expression of flatness data in key areas.

[0077] S4: Process the local point cloud data, identify the defect type of the problem area through the deep learning algorithm, and calculate the local flatness according to the defect type;

[0078] S5: performing point cloud registration on the global point cloud data and the local point cloud data, and calculating the correction coefficient according to the registration result;

[0079] In this step, the local point cloud is accurately superimposed on the global point cloud by calculating the rotation matrix and the translation matrix. After the registration is completed, the correction coefficient is calculated based on the registration result. The correction coefficient reflects the weight of the influence of the local defect area on the overall flatness. The accuracy of the registration ensures the alignment consistency of the local data and the global data, and the correction coefficient provides a quantitative basis for the subsequent steps to adjust the macro flatness value.

[0080] S6: Correct the macroscopic flatness according to the local flatness and correction coefficient, and generate a wall flatness report.

[0081] The specific steps of S2 are as follows:

[0082] S2.1: converting the global point cloud data into a world coordinate system, dividing the world coordinate system into grids at equal intervals, performing cluster analysis on the grids according to a clustering algorithm, merging grids in the same central cluster, and extracting point cloud geometric features;

[0083] The acquired global point cloud data is converted from the laser scanner's coordinate system to the world coordinate system to ensure the consistency and accuracy of subsequent processing. Specifically, according to the installation position, scanning angle and distance parameters of the laser scanner, the matrix transformation method is used to obtain the three-dimensional plane, and the point cloud data is transformed to map it to the global world coordinate system. The goal of clustering is to classify grids with similar geometric features on the wall into the same cluster and extract the geometric features of each cluster.

[0084] The point cloud geometric features reflect the geometric shape and spatial relationship of each point on the wall, including normal vector, curvature and point cloud density. The normal vector represents the surface direction of each point, which is used to describe the orientation consistency of the points on the wall. It is used to detect overall tilt on a macro scale and to identify local subtle unevenness, such as cracks and bumps, on a micro scale. The curvature describes the degree of curvature of the surface around the point. The point cloud density is the number of points per unit volume, which is used to describe the density of the point cloud. Density changes can indicate irregular areas on the surface. For example, the density is usually low in cracks and holes, and more uniform in smoother areas.

[0085] S2.2: decomposing the global point cloud data into scale-level regions according to the point cloud geometric features, and calculating the flatness in each scale-level region, wherein the scale-level region includes a macro-scale region, a local-scale region, and a micro-scale region;

[0086] The key to decomposing into scale-level regions is to use the geometric features in the point cloud data to perform hierarchical processing on the wall at different analysis scales. Different scale-level regions have different characteristics and can capture the flatness information of the wall at the macro, local and micro levels. The macro-scale region is used to capture the overall flatness trend of the wall and evaluate the inclination of the overall wall or large-scale volatility. The local-scale region is used to detect flatness changes in the medium range and identify smaller but relatively concentrated concave and convex or uneven areas on the wall. The micro-scale region is used to capture subtle defects on the wall, such as cracks, small-scale convexities or concave depressions, which are the details that need to be paid attention to in the refined flatness analysis.

[0087] The specific decomposition rules are as follows:

[0088] The normal vector in the point cloud is selected as the main geometric feature at the macro scale. Specifically, by analyzing the consistency of the normal vector in a large range, the mean of the normal vector direction of each region is calculated. When the normal vector consistency of a region is high and the normal vector direction is consistent with the overall wall direction, the region is considered to be a region with good flatness at the macro scale; if the consistency is low, it may be a tilted or fluctuating region at the macro scale.

[0089] At the local scale, curvature and point cloud density are mainly used as geometric features. For each region, the mean and variance of the point cloud density are calculated. Regions with uniform density and low curvature indicate high flatness, while regions with uneven density distribution or high curvature may be locally uneven regions. The local scale region division is set to a medium size so that smaller ranges of unevenness can be captured without losing overall features.

[0090] At the microscopic scale, we mainly focus on curvature, normal vector consistency and surface compactness. For each small area, we calculate the mean and change of curvature. Areas with larger curvature usually indicate more dramatic surface changes, and there may be small bumps or depressions. Normal vector consistency is used to detect normal vector changes between adjacent points. If the consistency is low, it means that the surface changes significantly. Further combined with the surface compactness, low compactness areas may indicate cracks or holes on the surface.

[0091] S2.3: weighted fusion of the flatness of all regions to calculate the macroscopic flatness;

[0092] After obtaining the global point cloud data and calculating the overall flatness, first calculate the flatness deviation of each point. Flatness deviation refers to the distance between the point cloud surface and the ideal plane, which can be used to evaluate the flatness of each point on the wall. After calculating the macro flatness deviation, set a flatness threshold based on experience to determine whether there is a flatness problem. By comparing with the flatness threshold, filter out areas that deviate from normal flatness and mark areas on the wall that may have problems;

[0093] See also Figure 2 , the point cloud geometric feature extraction flow chart of the embodiment of the present invention, the specific steps of S2.1 are as follows:

[0094] S2.1.1: Find the three-dimensional plane that best fits the global point cloud data through plane fitting, perform coordinate transformation on the global point cloud data according to the three-dimensional plane, and transform the global point cloud data from the radar coordinate system to the world coordinate system;

[0095] An optimal plane is fitted by the least squares method to maximize the fit of the point cloud data of the entire wall. After the fitting is completed, the reference plane is used to transform the global point cloud data, and the point cloud data is transformed from the scanner's radar coordinate system to the world coordinate system to unify the data processing coordinate system. Through plane fitting and coordinate transformation, it can ensure that the point cloud data is processed in a unified world coordinate system, effectively reducing the errors caused by the scanner installation position and angle, and facilitating subsequent grid division and cluster analysis.

[0096] S2.1.2: Divide the world coordinate system into grids at equal intervals, voxelize the global point cloud data, and project it into the corresponding grids according to the voxel index of the global point cloud data;

[0097] The voxelization process discretizes continuous 3D point cloud data to facilitate subsequent indexing and calculations, divides the space into 3D voxels, and sets the size of each voxel according to the data accuracy and wall scale. Each point cloud point is mapped to the corresponding grid through the voxel index. Each grid records the number of points and geometric features contained. Voxelization effectively reduces the computational complexity and converts point cloud data from irregular coordinates to regular grid indexes, which facilitates subsequent clustering and geometric analysis. At the same time, voxelization also facilitates data storage and compression, improving processing efficiency;

[0098] S2.1.3: performing preliminary clustering on the global point cloud data in the grid according to the DBSCAN clustering algorithm to obtain core points, and determining the number of cluster center clusters according to the number of core points;

[0099] In this step, the density clustering algorithm DBSCAN is used to perform preliminary clustering on the voxelized point cloud data. DBSCAN is suitable for irregularly shaped point cloud data and can automatically identify the number of clusters. The core points of each cluster are obtained through DBSCAN, and the number of core points is counted to determine the initial number of central clusters for subsequent K-means clustering. Through core point extraction, low-density noise points can be filtered out and the main flatness feature areas can be retained. The number of core points determines the number of high-density areas, providing the initial number of central points for K-means.

[0100] S2.1.4: using the core points after DBSCAN clustering as the initial center clusters of the K-means clustering algorithm, and calculating the Euclidean distance from the global point cloud data in the grid to each of the initial center clusters;

[0101] The accuracy and convergence speed of K-means clustering are greatly affected by the selection of the initial center. Using DBSCAN core points as the initial center cluster can make the clustering more stable and avoid the local minimum problem. At the same time, the Euclidean distance calculation provides a division standard.

[0102] S2.1.5: Perform clustering according to the Euclidean distance, calculate the average clustering error of the initial central cluster, update the central cluster according to the average clustering error, and determine whether the central cluster meets the convergence condition after the update. If the convergence condition is not met, continue to update the central cluster. If the convergence condition is met, output the clustering result;

[0103] By calculating the average error and iteratively updating it, the stability and accuracy of the clustering results can be ensured, making the divided areas more consistent with the actual wall features. The setting of the convergence condition avoids unnecessary calculations and improves the clustering efficiency. When the clustering average error is less than the predetermined threshold, the algorithm is considered to have converged;

[0104] The specific steps of S4 are as follows:

[0105] S4.1: randomly selecting three non-collinear points in the local point cloud data, calculating the point cloud plane equation, iterating the point cloud plane equation by calculating the point cloud standard deviation, and calculating the best fitting plane equation of the local point cloud data;

[0106] This method can accurately fit the best plane of the local point cloud area, providing a stable reference plane for subsequent vertical distance calculation. At the same time, it reduces the uncertainty in the calculation process of local defect features and ensures the accuracy of the feature value.

[0107] S4.2: Filter out the point cloud data whose plane normal vector in the best fitting plane equation has a y-axis component equal to zero, mark the point cloud data as a vertical point cloud, calculate the vertical average distance from the vertical point cloud to the corresponding best fitting plane equation, and calculate the wall point cloud feature according to the vertical average distance. The calculation formula of the wall point cloud feature is:

[0108] ,

[0109] in, Represents the wall point cloud feature, i represents a single vertical point cloud, N represents the total number of vertical point clouds, represents the gradient of the i-th vertical point cloud, represents the vertical distance of the i-th vertical point cloud, d represents the vertical average distance, represents the second-order curvature of the i-th vertical point cloud, represents the normal vector of the i-th vertical point cloud, Represents the component of the i-th vertical point cloud normal vector on the y-axis, Represents the point cloud density of the i-th vertical point cloud;

[0110] After finding the best fit plane, the vertical distance from each point in the local point cloud data to the plane is calculated. Then, based on the distance and other geometric features of the point cloud, the wall features are calculated for further analysis of the flatness and defects of the wall. These geometric features include the rate of change of the surface, point cloud density, curvature, and normal vector;

[0111] The surface change rate reflects the degree of surface change at each point on the wall, and is used to identify the tilted or fluctuating areas on the wall. The surface flatness is judged by observing the vertical deviation of the normal vector. A consistent normal vector usually indicates a relatively flat surface. Through the combination of multiple geometric features, the wall feature value can accurately characterize the wall flatness and defects. Especially for subtle changes in the wall, these features can effectively capture cracks, bumps and other problems on the wall.

[0112] S4.3: Input the extracted wall point cloud features into the defect recognition network, and use the deep learning algorithm to analyze the wall features, so as to automatically identify and classify the wall defect types. The defect recognition network is trained with a large amount of wall data, learns and extracts the correlation between features, and can identify different feature patterns as different defect types, including depressions, protrusions and cracks;

[0113] The introduction of deep learning networks makes the defect recognition process automated and intelligent, and can accurately distinguish various defects on the wall, reducing the errors of manual judgment. This automatic recognition method greatly improves the efficiency of wall flatness detection;

[0114] The defect recognition network consists of three layers, namely, the input layer, the hidden layer and the output layer. Each layer plays a different role in the recognition and classification of wall defect features. The defect recognition network includes:

[0115] The input layer receives the wall point cloud features as input parameters. These features include geometric information such as surface change rate, point cloud density, surface curvature, and normal vector offset. The input layer converts these features into feature sequences and normalizes the sequences to ensure the stability of the network during training. In addition, the input layer also calculates the fitting value of the feature sequence to preliminarily eliminate noise and abnormal data.

[0116] By normalizing and fitting the features, we can ensure that the data received by the network is stable and balanced, thereby improving the accuracy of the defect recognition network. At the same time, normalization can avoid the deviation caused by the difference in the range of feature values ​​and enhance the generalization ability of the network.

[0117] Hidden layer, used to map input features into high-dimensional space, so as to extract complex features of wall defects. The hidden layer can be composed of multiple neurons, each neuron uses a nonlinear activation function to enable the network to express nonlinear feature relationships, and calculate the output value of the hidden layer in the high-dimensional space. The hidden layer uses connection weights and biases to enable interactions between different input features, so as to capture complex patterns of wall features;

[0118] High-dimensional mapping and nonlinear feature extraction enhance the network’s sensitivity to complex defect features, enabling it to identify tiny cracks, bumps and other details in the wall. In addition, the nonlinear characteristics of the hidden layer enable the network to express complex geometric relationships, thereby improving the accuracy of defect classification.

[0119] The output layer is used to receive the high-dimensional feature output of the hidden layer and classify it to identify the specific defect type on the wall. The output result of the output layer will be mapped to the probability value of each defect type, and the maximum probability value will be taken as the final classification of the defect. The output layer uses the cross entropy loss function to calculate the error to guide the learning of the network;

[0120] Through the error feedback mechanism of the output layer, the network can gradually learn the characteristic patterns of various wall defects, so as to accurately identify the defect type in the test phase. At the same time, the probability output of the output layer enables the network to provide the confidence level of each defect type, providing more intuitive identification results;

[0121] The processing of the input parameters by the defect recognition network includes:

[0122] Initializing the defect recognition network according to the number of nodes in the input layer, the number of nodes in the hidden layer, and the number of nodes in the output layer of the defect recognition network, and determining the connection weights between the input layer and the hidden layer and the node thresholds between the hidden layer and the output layer;

[0123] The input parameters are biased by the connection weights of the input layer and the hidden layer to obtain the output of the hidden layer, and the output of the hidden layer is updated by the node threshold to obtain the output layer input;

[0124] Calculating the error of the output layer input according to the error function of the output layer, judging whether the error of the output layer input meets the convergence requirement of the output layer, and if not, updating the connection weight and the node threshold;

[0125] If the convergence requirement is met, output the defect identification result;

[0126] Specifically, first, the defect recognition network is initialized according to the preset number of nodes in the input layer, hidden layer, and output layer. The network is given an initial state by randomly initializing the connection weights and node thresholds for subsequent training iterations. The purpose of network initialization is to avoid weights being zero, thereby ensuring that the network can learn complex patterns. During the training process, the features are biased by the connection weights of the input layer and the hidden layer to generate the output of the hidden layer. The purpose of the bias is to introduce additional flexibility so that the network can be adjusted to non-zero outputs. The output of the hidden layer is processed by the activation function and passed to the next layer. At the output layer, the error between the predicted result and the true label is calculated by the loss function, and the error is passed back to the hidden layer and the input layer. Based on the size of the error, the connection weights and node thresholds are updated to gradually approach the target output.

[0127] The output layer uses the softmax function to generate probability distribution, which can intuitively display the confidence of each defect type and improve the interpretability and reliability of the recognition results.

[0128] The defect recognition network has output the defect type of each area. Then the geometric feature data of these defect areas are extracted, including:

[0129] The area of ​​the defect reflects the size of the defect area. Defects with larger areas have a more significant impact on flatness. The depth or height of the defect describes the vertical offset of the defect. The larger the value, the greater the impact of the defect on flatness. The density distribution of defects describes the density of defects per unit volume or unit area, which helps to determine the degree of defect aggregation. The distribution of defect types records the specific location distribution of each defect type in order to perform segmented weighting of local flatness.

[0130] Determine the weight value according to the defect type. For each defect type, calculate the local flatness deviation value of the defect area according to its area, depth (or height) and density parameters, and calculate the local flatness in combination with the weight value.

[0131] When registering a local point cloud with a global point cloud, there is often a deviation in the matching of the local point cloud and the global point cloud due to errors caused by different data sources (such as laser scanners and structured light scanners), slight differences in coordinate systems, and the complexity of the scanning environment. The correction coefficient is used to compensate for the slight errors in the registration process between the local point cloud and the global point cloud, ensuring that the final registered point cloud data remains highly consistent in spatial position. This coefficient can adjust the registered point cloud so that the local features are seamlessly connected to the global features, thereby accurately evaluating the overall flatness and defect locations, and using the correction coefficient to make a weighted adjustment to the local flatness so that the local flatness is more in line with the actual situation. The specific steps of S5 are as follows:

[0132] S5.1: extracting feature points from the global point cloud data and the local point cloud data respectively, and calculating a rotation matrix and a translation matrix according to the feature points;

[0133] Feature points are points on the wall with significant geometric features, such as corner points, edge points, and density mutation points. These points are relatively stable and can be kept consistent in point cloud data of different scales;

[0134] Using curvature analysis, the curvature value of each point on the wall is calculated. Points with significant curvature mutations are marked as corner points. Based on the normal vector difference, points with obvious normal vector changes between adjacent points are extracted and defined as edge points. In the local and global point clouds, the point cloud density changes in the local area are analyzed, and points with obvious density mutations are extracted as feature points.

[0135] Based on the extracted feature point set, the least square method is used to optimize the rotation matrix and translation matrix between the two sets of point clouds, so as to preliminarily align the local point cloud data with the global point cloud data;

[0136] S5.2: transform the local point cloud data according to the rotation matrix and the translation matrix, obtain the iterative point cloud, and calculate the average distance between the corresponding position of the iterative point cloud in the global point cloud data and the global point cloud data;

[0137] The generation of iterative point clouds and the calculation of average distance provide a preliminary evaluation of the effect of registration. If the average distance is small, it means that the preliminary alignment between the local point cloud and the global point cloud is good; otherwise, further optimization is required.

[0138] S5.3: If the average distance is less than the point cloud distance threshold, stop the iterative calculation; if the average distance is greater than or equal to the point cloud distance threshold, use the iterative point cloud as the new local point cloud data to recalculate the minimum average distance, rotation matrix and translation matrix;

[0139] S5.4: Calculate the correction coefficient based on the minimum average distance, rotation matrix and translation matrix after the iteration. The calculation formula of the correction coefficient is:

[0140] ,

[0141] in, represents the correction coefficient, j represents a single point cloud to be registered, M represents the total number of point clouds to be registered, Represents the density difference of the jth point cloud to be registered, represents the minimum average distance, represents the morphological complexity of the jth point cloud to be registered, represents the consistency value of the normal vector of the jth point cloud to be registered, R represents the rotation matrix, T represents the translation matrix, r(•) represents the rank calculation function, which is used to solve the rank of the matrix. represents the jth point cloud to be registered in the local point cloud data, and w represents the jth point cloud to be registered in the global point cloud data;

[0142] Morphological complexity reflects the geometric complexity of a local area. Regions with high complexity usually have more concave and convex features, so they require higher registration accuracy. Regions with high complexity will also be given higher weights in the correction coefficient.

[0143] The specific steps of S6 are as follows:

[0144] S6.1: performing weighted processing on the local flatness according to the correction coefficient corresponding to the problem area;

[0145] S6.2: Fusing the weighted local roughness set with the macro-roughness to generate a corrected macro-roughness;

[0146] S6.3: Generate a wall surface flatness report based on the corrected macro-flatness, the report including an overall flatness assessment, local defect details, and a flatness distribution map.

[0147] Embodiment 2:

[0148] See also Figure 3, the present invention provides an embodiment: a wall flatness detection system based on three-dimensional scanning technology, the system includes a data acquisition module, a data processing module, a point cloud correction module and a report generation module;

[0149] The data acquisition module is used to collect all and part of the three-dimensional point cloud data of the wall;

[0150] The data processing module is configured with a flatness analysis strategy, and the flatness analysis strategy is used to process the collected point cloud data and calculate the flatness information;

[0151] The point cloud correction module is used to align the global and local point cloud data, calculate the correction coefficient, and correct the macroscopic flatness according to the correction coefficient;

[0152] The report generation module is used to generate a wall surface flatness inspection report based on the corrected macroscopic flatness, and output the overall evaluation and local defect details.

[0153] The data processing module comprises:

[0154] A preprocessing unit, used for preprocessing the collected global point cloud data and local point cloud data, wherein the preprocessing includes noise filtering, outlier removal and data format conversion;

[0155] A multi-scale analysis unit is used to divide the pre-processed global point cloud data into hierarchical regions of different scales and calculate the corresponding flatness value in each scale region;

[0156] The defect recognition unit is used to identify and classify the defect in the flatness abnormal area in the local point cloud data, and extract the defect type and the flatness of the defect area;

[0157] The flatness analysis strategy includes global analysis logic and local analysis logic;

[0158] The global analysis logic is configured in the multi-scale analysis unit, and is used to calculate the macroscopic smoothness of the wall surface through multi-scale analysis, and mark the problem areas;

[0159] The local analysis logic is configured in the defect recognition unit, and is used to identify the defect type of the problem area through a deep learning algorithm, and calculate the local flatness according to the defect type;

[0160] The point cloud correction module comprises:

[0161] A matrix calculation unit, used to extract feature points from the global point cloud data and the local point cloud data, respectively, and calculate a rotation matrix and a translation matrix according to the feature points;

[0162] A parameter iteration unit, used to transform the local point cloud data according to the rotation matrix and the translation matrix, obtain the iterative point cloud, and iteratively optimize the rotation matrix and the translation matrix by calculating the average distance;

[0163] The coefficient calculation unit is used to calculate the correction coefficient according to the minimum average distance, the rotation matrix and the translation matrix after the iteration is completed.

[0164] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A wall flatness detection method based on three-dimensional scanning technology, characterized in that: The method comprises: Use a laser scanner to globally scan the wall surface and obtain global point cloud data; Processing the global point cloud data, calculating the macroscopic flatness of the wall through multi-scale analysis, and marking the problem areas; Use a structured light scanner to locally scan the marked problem area to obtain local point cloud data; Processing the local point cloud data, identifying the defect type of the problem area through a deep learning algorithm, and calculating the local flatness according to the defect type; Performing point cloud registration on the global point cloud data and the local point cloud data, and calculating a correction coefficient according to the registration result; The macroscopic flatness is corrected according to the local flatness and the correction coefficient, and a wall flatness report is generated.

2. The wall flatness detection method based on three-dimensional scanning technology according to claim 1 is characterized in that: The macroscopic flatness of the wall is calculated by multi-scale analysis, including: The global point cloud data is converted into a world coordinate system, the world coordinate system is divided into grids at equal intervals, the grids are clustered according to a clustering algorithm, the grids in the same central cluster are merged, and the geometric features of the point cloud are extracted; Decomposing the global point cloud data into scale-level regions according to the point cloud geometric features, and calculating the flatness in each scale-level region, wherein the scale-level region includes a macro-scale region, a local-scale region, and a micro-scale region; The flatness of all areas is weighted and fused to calculate the macroscopic flatness.

3. The wall flatness detection method based on three-dimensional scanning technology according to claim 2 is characterized in that: The performing cluster analysis on the grid according to the clustering algorithm comprises: Performing preliminary clustering on the global point cloud data in the grid according to the DBSCAN clustering algorithm to obtain core points, and determining the number of cluster center clusters according to the number of core points; The core points after DBSCAN clustering are used as the initial center clusters of the K-means clustering algorithm, and the Euclidean distances from the global point cloud data in the grid to each of the initial center clusters are calculated; Clustering is performed according to the Euclidean distance, and the average clustering error of the initial central cluster is calculated. The central cluster is updated according to the average clustering error, and it is determined whether the central cluster meets the convergence condition after the update. If the convergence condition is not met, the central cluster is continued to be updated. If the convergence condition is met, the clustering result is output.

4. The wall flatness detection method based on three-dimensional scanning technology according to claim 1 is characterized in that: The defect types identified in the problem area through deep learning algorithms include: Randomly select three non-collinear points in the local point cloud data, calculate the point cloud plane equation, iterate the point cloud plane equation by calculating the point cloud standard deviation, and calculate the best fitting plane equation of the local point cloud data; Filter out the point cloud data whose plane normal vector in the best fitting plane equation is equal to zero in the y-axis, mark the point cloud data as vertical point cloud, calculate the vertical average distance from the vertical point cloud to the corresponding best fitting plane equation, and calculate the wall point cloud features according to the vertical average distance. The calculation formula of the wall point cloud features is: ， in, Represents the wall point cloud feature, i represents a single vertical point cloud, N represents the total number of vertical point clouds, represents the gradient of the i-th vertical point cloud, represents the vertical distance of the i-th vertical point cloud, d represents the vertical average distance, represents the second-order curvature of the i-th vertical point cloud, represents the normal vector of the i-th vertical point cloud, Represents the component of the i-th vertical point cloud normal vector on the y-axis, Represents the point cloud density of the i-th vertical point cloud; A defect recognition network is constructed, the wall point cloud features are used as input parameters of the defect recognition network, the input parameters are processed by the defect recognition network, and the defect type is output.

5. The wall flatness detection method based on three-dimensional scanning technology according to claim 4 is characterized in that: The defect recognition network comprises: An input layer, which establishes a feature sequence according to input parameters and calculates the fitting value of the feature sequence; A hidden layer converts the fitting value of the feature sequence of the input layer into a high-dimensional space, and calculates the output value of the hidden layer in the high-dimensional space; The output layer is used to output the defect recognition result.

6. The wall flatness detection method based on three-dimensional scanning technology according to claim 5 is characterized in that: The processing of the input parameters by the defect recognition network includes: Initializing the defect recognition network according to the number of nodes in the input layer, the number of nodes in the hidden layer, and the number of nodes in the output layer of the defect recognition network, and determining the connection weights between the input layer and the hidden layer and the node thresholds between the hidden layer and the output layer; The input parameters are biased by the connection weights of the input layer and the hidden layer to obtain the output of the hidden layer, and the output of the hidden layer is updated by the node threshold to obtain the output layer input; Calculating the error of the output layer input according to the error function of the output layer, determining whether the error of the output layer input meets the convergence requirement of the output layer, and if not, updating the connection weight and the node threshold; If the convergence requirement is met, the defect identification result is output.

7. The wall flatness detection method based on three-dimensional scanning technology according to claim 1 is characterized in that: The point cloud registration comprises: Extracting feature points from the global point cloud data and the local point cloud data respectively, and calculating a rotation matrix and a translation matrix according to the feature points; The local point cloud data is transformed according to the rotation matrix and the translation matrix to obtain the iterative point cloud, and the average distance between the corresponding position of the iterative point cloud in the global point cloud data and the global point cloud data is calculated; If the average distance is less than the point cloud distance threshold, stop the iterative calculation; if the average distance is greater than or equal to the point cloud distance threshold, use the iterative point cloud as the new local point cloud data to recalculate the minimum average distance, rotation matrix and translation matrix; According to the minimum average distance, rotation matrix and translation matrix after the iteration, the correction coefficient is calculated. The calculation formula of the correction coefficient is: ， in, represents the correction coefficient, j represents a single point cloud to be registered, M represents the total number of point clouds to be registered, represents the density difference of the jth point cloud to be registered, represents the minimum average distance, represents the morphological complexity of the jth point cloud to be registered, represents the consistency value of the normal vector of the jth point cloud to be registered, R represents the rotation matrix, T represents the translation matrix, r(•) represents the rank calculation function, which is used to solve the rank of the matrix. represents the jth point cloud to be registered in the local point cloud data, and w represents the jth point cloud to be registered in the global point cloud data.

8. The wall flatness detection method based on three-dimensional scanning technology according to claim 7 is characterized in that: The generating of the wall flatness report includes: Performing weighted processing on the local flatness according to the correction coefficient corresponding to the problem area; The weighted local roughness set is merged with the macro-roughness to generate the corrected macro-roughness; A wall surface flatness report is generated based on the corrected macro-flatness, the report including an overall flatness assessment, local defect details, and a flatness distribution map.

9. A wall flatness detection system based on three-dimensional scanning technology, implemented based on the wall flatness detection method based on three-dimensional scanning technology as claimed in any one of claims 1 to 8, characterized in that: The system includes a data acquisition module, a data processing module, a point cloud correction module and a report generation module; The data acquisition module is used to collect all and part of the three-dimensional point cloud data of the wall; The data processing module is configured with a flatness analysis strategy, and the flatness analysis strategy is used to process the collected point cloud data and calculate the flatness information; The point cloud correction module is used to align the global and local point cloud data, calculate the correction coefficient, and correct the macroscopic flatness according to the correction coefficient; The report generation module is used to generate a wall surface flatness inspection report based on the corrected macroscopic flatness, and output the overall evaluation and local defect details.

10. The wall flatness detection system based on three-dimensional scanning technology according to claim 9, characterized in that: The data processing module comprises: A preprocessing unit, used for preprocessing the collected global point cloud data and local point cloud data, wherein the preprocessing includes noise filtering, outlier removal and data format conversion; A multi-scale analysis unit is used to divide the pre-processed global point cloud data into hierarchical regions of different scales and calculate the corresponding flatness value in each scale region; The defect recognition unit is used to identify and classify the defect in the flatness abnormal area in the local point cloud data, and extract the defect type and the flatness of the defect area; The flatness analysis strategy includes global analysis logic and local analysis logic; The global analysis logic is configured in the multi-scale analysis unit, and is used to calculate the macroscopic smoothness of the wall surface through multi-scale analysis, and mark the problem areas; The local analysis logic is configured in the defect recognition unit, and is used to identify the defect type of the problem area through a deep learning algorithm, and calculate the local flatness according to the defect type.

Images