Shear wall identification method based on recombination segmentation graph

Abstract

The invention relates to the technical field of shear wall recognition, in particular to a shear wall recognition method based on a recombined segmented graph, and the method comprises the steps: obtaining an indoor wall image from different point locations through image collection equipment, and synchronously obtaining distance measurement data and positioning data; an image is processed through a convolutional neural network, a wall joint is automatically calibrated and segmented, the height and length of a single wall are calculated in combination with distance measurement data, and positioning and orientation data during wall image acquisition are mapped into a two-dimensional matrix. And through geometric morphology analysis and traversal of the shear wall structure library on the two-dimensional matrix, identifying candidate single walls of the shear wall. And carrying out clustering analysis on the spatial positions and the connection relations of the adjacent wall bodies, determining candidate areas of the shear wall, and generating a recombination graph. The semantic segmentation algorithm analyzes the recombined graph, and judges whether the recombined graph accords with the shear wall features through the BP neural network, thereby reducing misjudgment and improving the recognition precision. And finally, an identification result is displayed through three-dimensional visualization, and analysis by engineers is facilitated.

Description

Technical Field

[0001] The present invention relates to the technical field of shear wall identification, and particularly to a shear wall identification method based on recombined segmented graphics. Background Art

[0002] Shear walls are key load-bearing components in buildings. Especially in high-rise buildings and seismic design, they are responsible for bearing vertical loads and resisting horizontal seismic forces to ensure the stability of the building structure. The design and layout of shear walls are not only directly related to the overall structural safety of the building but also play a crucial role in seismic performance. In actual engineering, the correct identification of shear walls is of great significance for the design, analysis, and construction safety of building structures.

[0003] In the scenarios of indoor renovation and secondary decoration, the identification and retention of shear walls become particularly important. As load-bearing walls, shear walls have the property that they cannot be randomly demolished. If a shear wall is accidentally demolished or its existence is ignored without reasonable retention during the renovation process, the building structure may be easily damaged. Therefore, how to accurately and quickly identify and reasonably retain shear walls in indoor renovation is an important link to ensure the safety and sustainability of buildings. However, traditional shear wall identification methods mostly rely on manual marking, and the position of shear walls is determined by manually inspecting building drawings and the site. Such methods highly rely on manual experience, are time-consuming, and extremely error-prone. Especially in complex indoor environments, the features of shear walls often resemble those of other walls, making it difficult to quickly and accurately distinguish them. Summary of the Invention

[0004] To solve the above problems, the present invention provides a shear wall identification method based on recombined segmented graphics, which automatically identifies shear walls by combining image acquisition and the positioning and orientation data during image acquisition with a deep learning algorithm, effectively realizing the automation of shear wall identification and improving the accuracy of shear wall identification, avoiding reliance on manual experience, and improving the identification efficiency.

[0005] To achieve the above object, the technical solution adopted by the present invention is:

[0006] A shear wall identification method based on recombined segmented graphics, comprising the following steps:

[0007] S1. Obtain a plurality of indoor wall images and ranging data from different positions or orientations through an image acquisition device, and simultaneously collect the positioning and orientation data during the acquisition of the indoor wall images;

[0008] S2. Calibrate the wall connection points in the indoor image through a convolutional neural network, segment a plurality of single-wall graphics based on the wall connection points, and calculate the height and length of the single-wall based on the ranging data;

[0009] S3. Initialize a two-dimensional matrix, and based on the height and length of the single wall and the positioning and orientation data when the wall image is acquired, map several single walls onto the two-dimensional matrix to obtain an indoor plane matrix;

[0010] S4. Based on the indoor plane matrix, identify the shear wall candidate areas through geometric shape analysis, select single wall graphics according to the shear wall candidate areas for recombination to obtain a recombined graphic, parse the semantics of the recombined graphic based on the semantic segmentation algorithm and use it as the input of the BP neural network to output the shear wall recognition result;

[0011] S5. Perform three-dimensional visualization according to the recognition result of the shear wall.

[0012] Furthermore, the calibration of the wall connection in the indoor image by the convolutional neural network includes the following steps:

[0013] Input the acquired indoor wall image into the trained convolutional neural network model;

[0014] Extract the edge feature points in the wall image through the convolutional layer, and perform downsampling through the pooling layer;

[0015] Based on the fully connected layer in the convolutional neural network, calibrate the extracted wall connection features.

[0016] Furthermore, the training of the convolutional neural network includes the following steps:

[0017] Collect several indoor wall images with uncalibrated wall connections, manually label the wall connections among them to generate corresponding calibration images, and form a paired training sample set;

[0018] According to the visual features of the wall connections, customize the network structures of the convolutional layer, pooling layer and fully connected layer, and initialize the network weight parameters;

[0019] Input the uncalibrated indoor wall image into the convolutional neural network model, use the corresponding calibration image as the supervision signal, and use the backpropagation algorithm and the loss function to iteratively update the network parameters.

[0020] Furthermore, the calculation of the height and length of the single wall based on the ranging data includes the following steps:

[0021] Extract the horizontal distance between the single wall and the image acquisition device, the distances between the vertical endpoints of the single wall and the image acquisition device respectively, and the distances between the parallel endpoints of the single wall and the image acquisition device respectively from the ranging data;

[0022] Calculate the height of the single wall through the horizontal distance between the single wall and the image acquisition device and the distances between the vertical endpoints of the single wall and the image acquisition device respectively;

[0023] Calculate the length of the single wall based on the horizontal distance between the single wall and the image acquisition device and the distances between the parallel endpoints of the single wall and the image acquisition device respectively.

[0024] Further, S3 includes the following steps:

[0025] Perform size normalization on each single wall based on the height and length data of the single wall;

[0026] Determine the sub-two-dimensional matrix of the two-dimensional projection for mapping according to the orientation of the image acquisition device and the spatial orientation of the single wall, and determine the matrix coordinates covered by the single wall according to the length data of the normalized single wall;

[0027] Map the height of the normalized single wall to the matrix element value in the two-dimensional matrix according to the matrix coordinates covered by the single wall to obtain the indoor plane matrix.

[0028] Further, the step of identifying the shear wall candidate area through geometric shape analysis includes the following steps:

[0029] Calculate the geometric shape parameters of each single wall based on the single wall data in the indoor plane matrix, including the aspect ratio, area, and spatial occupancy ratio;

[0030] Traverse the indoor plane matrix through a pre-configured shear wall structure library, and extract the indoor structure part that conforms to the shear wall structure;

[0031] Screen the single walls whose geometric shape parameters meet the standards through a preset geometric standard parameter threshold in the indoor structure part that conforms to the shear wall structure as the shear wall candidate single walls;

[0032] Perform cluster analysis on the selected candidate single walls according to their spatial positions to identify the shear wall candidate areas composed of several single walls.

[0033] Further, the selection of single wall graphics for recombination according to the shear wall candidate area includes:

[0034] Based on the indoor plane matrix, determine the adjacent single walls that meet the shear wall characteristics through spatial positions and geometric shape parameters in the shear wall candidate area;

[0035] Combine according to the spatial connection relationship between the adjacent single walls that meet the shear wall characteristics to obtain several combinations of adjacent single walls that meet the shear wall characteristics;

[0036] Select single wall graphics according to the combination result for recombination splicing of the graphic edges.

[0037] Further, the step of parsing the semantics of the recombined graphics based on the semantic segmentation algorithm and using it as the input of the BP neural network to output the shear wall recognition result includes the following steps:

[0038] Input the recombined graph into the semantic segmentation model to parse the semantic information contained in the recombined graph, including walls, windows, balcony openings, and door openings;

[0039] Import the semantic information into the BP neural network model to currently determine whether the recombined graph meets the standard features of shear walls.

[0040] Furthermore, the formula of the BP neural network is as follows:

[0041] y = f(W·x + b);

[0042] Among them, y is the output result, including whether the current recombined graph is a shear wall; W is the weight matrix; b is the bias term; f is the activation function; x is the feature vector of the input layer, that is, the combination of semantic information.

[0043] The beneficial effects of the present invention are as follows: The present invention obtains indoor wall images from different positions or orientations through an image acquisition device, combines ranging data and positioning data, ensures the multi-dimensionality and accuracy of data acquisition, and avoids the subjective errors of manual measurement and annotation. Compared with traditional methods, automated data acquisition can cover more details of the indoor environment, laying a more reliable foundation for subsequent analysis. Further, the collected images are processed by a convolutional neural network to automatically calibrate the joints of the walls and segment the wall images based on this. The calculation of the height and length of individual walls based on ranging data ensures the matching of the physical dimensions of the wall structure in the image with the actual environment. This step greatly improves the accuracy of recognition, especially in the case where shear walls are similar in appearance to other walls, and can more accurately distinguish the characteristics of shear walls. By combining geometric shape analysis with a shear wall structure library, individual walls in the two-dimensional matrix are identified. Different from traditional geometric rule analysis, this solution traverses the individual wall data using a pre-configured shear wall structure library, and combines geometric standard parameters to screen out candidate individual walls that meet the characteristics of shear walls in the indoor structure part that conforms to the shear wall structure. By clustering and analyzing the spatial positions and connection relationships of adjacent individual walls, the candidate areas of shear walls are determined, solving the problem of difficult extraction of shear wall characteristics in complex indoor environments. The semantic segmentation algorithm parses the semantic information of the recombined graph and combines the BP neural network to judge whether the recombined graph meets the standard characteristics of shear walls, effectively reducing the possibility of misjudgment and optimizing the recognition accuracy of shear walls. Finally, three-dimensional visualization is used to intuitively display the recognition results of shear walls to engineering personnel. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is the flowchart of step 1 of a method for identifying shear walls based on recombined and segmented graphs in the present invention.

[0045] Figure 2 It is the flowchart of step S3 in the present invention. Detailed implementation manners

[0046] Please refer to Figure 1-2 As shown, the present invention relates to a shear wall recognition method based on recombined segmented graphics, including the following steps:

[0047] S1. Obtain a plurality of indoor wall images and ranging data from different positions or orientations through an image acquisition device, and synchronously collect the positioning and orientation data when the indoor wall images are obtained;

[0048] S2. Calibrate the wall joints in the indoor image through a convolutional neural network, segment several single-wall graphics at the wall joints, and calculate the height and length of the single-wall based on the ranging data;

[0049] S3. Initialize a two-dimensional matrix and map several single-walls into the two-dimensional matrix based on the height and length of the single-wall and the positioning and orientation data when the wall image is obtained, to obtain an indoor plane matrix;

[0050] S4. Based on the indoor plane matrix, identify the shear wall candidate areas through geometric shape analysis, select single-wall graphics for recombination according to the shear wall candidate areas to obtain a recombined graphic, analyze the semantics of the recombined graphic based on a semantic segmentation algorithm and use it as the input of a BP neural network, and output the shear wall recognition result;

[0051] S5. Perform three-dimensional visualization according to the shear wall recognition result.

[0052] In some embodiments, in step S1, a handheld mobile device integrated with a high-resolution camera and a laser rangefinder is used to collect indoor wall images from multiple points and different angles. The handheld device simultaneously obtains the distance data between the wall and the device through the laser ranging module, and obtains the positioning and orientation information during image acquisition through the built-in inertial measurement unit (IMU). These information ensure the accurate correspondence between the image and the physical space. The device continuously records its position changes in space during the acquisition process, and through real-time data synchronization, ensures that each image and ranging data can reflect the current device orientation and distance. In step S2, the collected images and ranging data are transmitted to the computing unit, and a convolutional neural network is used to calibrate the wall joints in the images. Through the multi-layer convolution and feature extraction of the network on the images, the system can automatically identify the connection points of the walls. Based on the depth information in the images and the ranging data provided by the handheld device for each connection point, the system further calculates the height and length of each individual wall to ensure that the geometric dimensions of the individual walls match the actual physical dimensions. Next, in step S3, according to the height, length of the individual walls already obtained, as well as the device positioning and orientation data, the system maps them into a two-dimensional matrix to generate an indoor plane matrix. The spatial position information of the handheld device ensures the accurate position of all individual walls in the two-dimensional plane is consistent with the actual indoor environment through coordinate transformation. In step S4, the present invention identifies the shear wall candidate areas through geometric shape analysis. First, based on the individual wall data in the indoor plane matrix, the geometric shape parameters of each individual wall are calculated, including the length-to-height ratio, area, and spatial occupancy ratio in the entire plane matrix. These parameters can provide a basic basis for subsequent screening to ensure the accurate extraction of shear wall features. Then, the system traverses the indoor plane matrix through a pre-configured shear wall structure library. The shear wall structure library contains various common shear wall structure patterns and their geometric features. The system compares each individual wall in the two-dimensional matrix one by one according to these structure patterns, and extracts the parts that conform to the shear wall structure. In the indoor structure parts that conform to the shear wall structure, the system further screens the individual walls whose geometric shape parameters meet the standard through preset geometric standard parameter thresholds (such as length-to-height ratio, area range, etc.), and marks them as shear wall candidate individual walls. After the screening of the candidate individual walls is completed, the system performs clustering analysis based on the spatial position relationship of these individual walls. The purpose of clustering analysis is to identify the spatial connection situation of multiple individual walls and determine whether they jointly form a continuous shear wall structure area. Through this analysis, the system can identify the shear wall candidate areas composed of several adjacent individual walls, laying a foundation for subsequent shear wall recombination and further analysis. Finally, in step S5, the system displays the identification results of the shear walls through a three-dimensional visualization module. The three-dimensional visualization system visually presents the specific position, size of the shear walls and their spatial relationship in the indoor environment based on the data collected by the handheld device and the results of shear wall identification.This process can be integrated with a Building Information Modeling (BIM) system to provide accurate shear wall location information for architects and construction workers, ensuring the reasonable retention and use of shear walls during interior renovation, and avoiding potential structural safety hazards caused by accidental demolition of shear walls.

[0053] Furthermore, calibrating the wall connection in the indoor image through the convolutional neural network includes the following steps:

[0054] Input the collected indoor wall image into the trained convolutional neural network model;

[0055] Extract edge feature points in the wall image through the convolutional layer and perform downsampling through the pooling layer;

[0056] Calibrate the extracted wall connection features based on the fully connected layer in the convolutional neural network.

[0057] In some embodiments, an indoor image is processed by a Convolutional Neural Network (CNN) to achieve precise calibration of wall joints. First, the indoor wall image collected by a handheld mobile device is input into a pre-trained Convolutional Neural Network model. This model is optimized with a large amount of labeled training data and can automatically identify the wall features in the image. The Convolutional Neural Network adopts a hierarchical structure, and each layer extracts features of different levels from the image data. The first layer of the Convolutional Neural Network is the convolutional layer, whose main function is to extract the edge feature points of the image. Specifically, the convolution operation detects the edge information in different directions by applying multiple convolutional kernels to the image. The convolutional kernel is a trained weight matrix that extracts local features, such as vertical, horizontal, or diagonal edges of the wall, when scanning the image. Through the convolution operation, the low-level features in the original image are gradually transformed into high-level features, forming a more abstract representation. In particular, the edge features at the wall joints are captured at this stage. After the convolutional layer processes, the extracted feature map is input into the pooling layer for downsampling. The main function of the pooling layer is to reduce the computational amount by reducing the size of the feature map and enhance the anti-noise ability of the model. The commonly used pooling method is max pooling, which can retain the most significant features in the local area and eliminate the redundant detailed information in the image. Downsampling not only reduces the data dimension but also makes the subsequent calculations more efficient. By shrinking the size of the feature map, it can highlight the core features at the wall joints and further improve the recognition accuracy. After the convolution and pooling operations are completed, the extracted feature map is fed into the fully connected layer. The fully connected layer is responsible for combining the local features extracted by the convolutional layer into global features and finally achieving the calibration of the wall joints. The fully connected layer linearly transforms all the input features through a weight matrix and then performs a non-linear transformation through an activation function, and finally outputs a calibration result. Specifically, through the calculation of the fully connected layer, the model can accurately identify the positions of each wall joint in the image based on the wall edge features extracted by the convolution and perform precise calibration on them.

[0058] Furthermore, the training of the Convolutional Neural Network includes the following steps:

[0059] Collect several indoor wall images with uncalibrated wall joints, manually label the wall joints among them, generate corresponding calibrated images, and form a paired training sample set;

[0060] According to the visual features of the wall joints, customize the network structures of the convolutional layer, pooling layer, and fully connected layer, and initialize the network weight parameters;

[0061] Input the uncalibrated indoor wall image into the Convolutional Neural Network model, use the corresponding calibrated image as the supervision signal, and iteratively update the network parameters using the backpropagation algorithm and the loss function.

[0062] In some embodiments, first, a large number of uncalibrated indoor wall images need to be collected. These images are obtained from different positions and angles using a handheld mobile device, covering a variety of indoor scenarios and wall features. After obtaining these images, annotators will label the wall joints in each image. This process is achieved through software tools. The annotators will precisely circle the wall connection points in each image to generate a calibrated image corresponding to the original image. Each pair of training sample sets includes an original image and its corresponding calibrated image, forming the basic data set required for supervised learning. The number and depth of convolutional kernels need to be tuned through multiple experiments. An increase in the number of convolutional kernels enables the neural network to extract more diverse features in different regions of the image, covering various possible connection forms such as vertical, horizontal, and inclined edges. When designing the network depth, a shallower network may be insufficient to extract complex features. Therefore, the network design usually includes multiple convolutional layers. As the layers deepen, the feature representation gradually transitions from low-level pixel information to more high-level abstract features. Considering the feature complexity of the wall joints, the depth and width (i.e., the number of convolutional kernels in each layer) of the convolutional layer should be adjusted in the experiment to ensure that the network's feature extraction ability is strong enough to adapt to various complex indoor environments. After the convolutional layer, a pooling layer is introduced to reduce the dimension of the feature map. The core purpose of the pooling operation is to reduce data redundancy through downsampling while retaining the most important feature information. In this solution, Max Pooling is commonly used for extracting the features of wall joints. Max Pooling retains the largest feature value in each local area. This method can effectively enhance the robustness of the network and resist noise or interference in the indoor environment. Usually, the size of the pooling kernel is 2x2, that is, the size of the feature map is reduced by half each time downsampling is performed. The pooling operation ensures that the convolutional neural network can extract the most significant connection features from a large-scale input image without losing the important geometric forms of the wall joints. After the pooling layer, the network enters the fully connected layer, which is responsible for integrating the extracted features into a global feature vector. In the fully connected layer, all the feature maps generated by the previous convolutional and pooling processes are unfolded into a vector, and each element represents a specific feature in the image. Through linear transformation, all features are weighted and combined to form the output of the model. To increase the nonlinear expression ability of the model, the fully connected layer usually also adds nonlinear activation functions such as ReLU (Rectified Linear Unit). For the calibration task of the wall joints, the output of the fully connected layer is the specific coordinates of the wall joints in the image or the calibration result of a specific area.

[0063] Network initialization is crucial for the success of training. The weight parameters of the network are usually set through Gaussian random initialization or the He initialization method. Gaussian initialization allows the weight parameters to start with small random values, ensuring stable gradient flow in the initial stage of training and avoiding problems such as vanishing or exploding gradients. He initialization dynamically adjusts the initialization range of the weights according to the number of input units in the network layer. This method performs excellently in the training of deep networks. Especially when dealing with complex features at the wall joints, it can ensure the sensitivity of the convolutional layer to edges and corners. During the training process, with each iteration, the convolutional neural network gradually learns the significant features of the wall joints from the training data. To ensure the generalization ability of the model, data augmentation techniques such as rotation, scaling, flipping, etc. are usually adopted to randomly transform the original images to increase the diversity of training samples. This helps the model adapt to different environmental changes and enhances its robustness in identifying wall joints in different indoor scenarios.

[0064] Furthermore, calculating the height and length of a single wall based on the ranging data includes the following steps:

[0065] Extract the horizontal distance between the single wall and the image acquisition device, the distances between the vertical endpoints of the single wall and the image acquisition device, and the distances between the parallel endpoints of the single wall and the image acquisition device from the ranging data;

[0066] Calculate the height of the single wall through the horizontal distance between the single wall and the image acquisition device and the distances between the vertical endpoints of the single wall and the image acquisition device;

[0067] Calculate the length of the single wall through the horizontal distance between the single wall and the image acquisition device and the distances between the parallel endpoints of the single wall and the image acquisition device.

[0068] In some embodiments, the height and length of a single wall are accurately calculated based on ranging data to ensure that the mapping of the wall structure in the two-dimensional plane matrix is consistent with the physical environment. First, while the handheld mobile device captures wall images at different positions and orientations, the ranging data between each single wall and the device is recorded through the integrated laser rangefinder. Specifically, the ranging data includes the horizontal distance between the single wall and the device, the distance between the vertical endpoint and the device, and the distance between the parallel endpoints and the device. These measurement points provide the core parameters for subsequent height and length calculations. For the height calculation of a single wall, the system first extracts the horizontal distance between the single wall and the image acquisition device and the distance between the vertical endpoint and the device. The horizontal distance refers to the distance between the wall baseline and the position of the device, while the vertical endpoint distance represents the vertical ranging between a point at the top or bottom of the wall and the device. With these two key parameters and combined with the inclination angle of the wall or the elevation angle of the image acquisition device, the system can calculate the height of the single wall using trigonometric formulas (such as sine or cosine functions). Specifically, when the image acquisition device is in a horizontal or non-horizontal position, the vertical distance can be converted into the height information of the wall through the actual measurement values obtained by the ranging device and the inclination angle of the device. The length calculation of a single wall is completed by the horizontal distance between the single wall and the image acquisition device and the distances between the parallel endpoints of the single wall and the device respectively. The distance measurement of the parallel endpoints usually adopts the horizontal scanning mode of the laser rangefinder, which can accurately measure the length of the wall in the horizontal direction. By measuring the distances between the two parallel endpoints of the single wall and the image acquisition device and combining with the horizontal distance, the system can further calculate the actual length of the wall. In this process, the system also uses trigonometric formulas to convert these distance data into the length value of the actual wall according to the relative angles between the measurement points and the device.

[0069] Further, step S3 includes the following steps:

[0070] Based on the height and length data of the single wall, perform size normalization processing on each single wall;

[0071] According to the orientation of the image acquisition device and the spatial orientation of the single wall, determine the sub-two-dimensional matrix of the two-dimensional projection for mapping, and determine the matrix coordinates covered by the single wall according to the normalized length data of the single wall;

[0072] According to the matrix coordinates covered by the single wall, map the height of the normalized single wall to the matrix element value in the two-dimensional matrix to obtain the indoor plane matrix.

[0073] In some embodiments, first, based on the measured physical height and length data of each individual wall, size normalization is performed on each individual wall. The purpose of normalization is to convert wall data of different scales into a unified ratio, enabling all individual walls to be represented in the same two-dimensional matrix and ensuring that its ratio is consistent with the actual space. The normalization process determines a reference scale (such as the height or length of the largest or smallest wall) as the standard and scales the actual dimensions of each individual wall according to this standard. The normalized data is represented in standard units, ensuring that all individual walls can be accurately calculated and processed in the same matrix even if there are significant differences in the sizes of different walls. After completing the normalization process, the system determines a sub-two-dimensional matrix of the two-dimensional projection for mapping based on the orientation of the image acquisition device and the spatial orientation of the individual wall. To ensure the accuracy of the projection, the system calculates the projection angle and position of each individual wall in the two-dimensional space based on the positioning and orientation information of the device. This process relies on the orientation data of the image acquisition device, such as the pitch angle, horizontal rotation angle, etc., and converts the individual wall data in the three-dimensional space into a projection matrix on the two-dimensional plane through a geometric transformation algorithm. The sub-two-dimensional matrix is a subset of the main two-dimensional matrix, specifically used to represent the spatial range and geometric features of a specific individual wall. To ensure the accuracy of the projection, the system performs real-time correction on the coordinates of each individual wall, taking into account the orientation changes and position errors of the image acquisition device. Next, the system determines the covered coordinates of the individual wall in the sub-two-dimensional matrix based on the normalized length data of the individual wall. The normalized length of each individual wall determines the horizontal range it occupies on the two-dimensional plane, that is, its horizontal coordinate range. The system determines the specific starting and ending coordinates of the individual wall in the matrix by matching the normalized length with the scale factor of the two-dimensional plane matrix. This process ensures that the horizontal occupancy of the individual wall is accurately reflected in the two-dimensional matrix, avoiding coordinate misalignment problems caused by data deviation. Finally, the system maps the normalized height of the individual wall to the matrix element value in the two-dimensional matrix based on the matrix coordinates covered by the individual wall. The value of each matrix element represents the normalized height of the individual wall at that position. By calibrating the longitudinal position of the matrix, the system can accurately reflect the height distribution of the individual wall on the two-dimensional plane. In this way, an indoor plane matrix is generated, where the height and length of each individual wall are clearly represented in the two-dimensional space, ensuring a one-to-one correspondence between the physical structure and the matrix structure. The finally obtained indoor plane matrix not only contains the geometric information of the wall but can also be used for further processing operations such as shear wall identification and three-dimensional reconstruction in the future.

[0074] Further, the step of identifying the shear wall candidate area through geometric shape analysis includes the following steps:

[0075] Based on the individual wall data in the indoor plane matrix, calculate the geometric shape parameters of each individual wall, including the aspect ratio, area, and spatial occupancy ratio;

[0076] Traverse the indoor plane matrix through a pre-configured shear wall structure library, and extract the indoor structural parts that conform to the shear wall structure;

[0077] In the indoor structural parts that conform to the shear wall structure, screen out the single walls whose geometric shape parameters meet the standards through the preset geometric standard parameter thresholds as the candidate single walls for shear walls;

[0078] Perform clustering analysis on the selected candidate single walls according to their spatial positions to identify the candidate areas for shear walls composed of several single walls.

[0079] It should be noted that geometric morphology analysis is used to accurately identify the candidate areas of shear walls from the indoor plane matrix. First, the system calculates the geometric morphology parameters of each individual wall based on the individual wall data in the two-dimensional plane matrix. The geometric parameters include the aspect ratio, area, and spatial occupancy ratio. Specifically, the system calculates the aspect ratio of each individual wall through the length and height data of the wall, which can indicate the shape characteristics of the wall; determines the area of the individual wall by the number of grids it occupies in the matrix; and the spatial occupancy ratio reflects the proportion of the individual wall in the entire indoor structure. Next, the system traverses the indoor plane matrix through a pre-configured shear wall structure library. The shear wall structure library stores the geometric models and characteristic patterns of various standard shear walls, covering different types of buildings and structural forms. The system compares the geometric parameters of each individual wall with the models in the shear wall structure library. Specifically, the system traverses the matrix with a fixed step size by moving the structure window. Specifically, the structure window starts from a starting position in the plane matrix and moves row by row and column by column, moving a distance of one step each time to ensure that all individual walls in the entire plane matrix can be covered. During traversal, the geometric parameters of all individual walls within the window, including the aspect ratio, area, and spatial occupancy ratio, are extracted. For each structure window, the system compares the geometric parameters of the individual walls contained therein with the models in the shear wall structure library. The shear wall structure library stores a variety of standard geometric models of shear walls, which are predefined according to different architectural design and construction requirements. When comparing, the system first calculates the overall geometric characteristics of all individual walls within the window, including their spatial arrangement, connection relationships with each other, etc. Then, the system uses a matching algorithm to compare these characteristics with the shear wall models in the structure library one by one to find the sub-matrix that meets the shear wall standards. When the individual walls within the structure window match a certain model in the structure library, the system records the specific location of this sub-region and further screens the walls within the sub-region according to the preset geometric standard parameter thresholds. The geometric standard parameter thresholds include the upper and lower limits of the aspect ratio, the minimum and maximum values of the area, etc., to ensure that the system can exclude those individual walls that do not conform to the shear wall structural characteristics. This process is similar to classifying the wall characteristics. By excluding the individual walls that do not conform to the geometric morphology of the shear wall, the walls that meet the geometric standards of the shear wall are finally retained as candidate individual walls for the shear wall. To further accurately identify the shear wall, the system performs clustering analysis on the selected candidate individual walls. The system will further integrate and analyze these candidate individual walls through clustering analysis. Clustering analysis is based on the relative position relationships of these individual walls in space. By calculating the spatial distances, geometric morphology similarities, etc. between individual walls, it is determined which individual walls belong to the same shear wall. Specific clustering algorithms can use K-means clustering or hierarchical clustering, which classify adjacent individual walls into one category according to the geometric morphology and spatial distribution of the individual walls.For example, if two adjacent single - body walls are similar in aspect ratio and area and are close to each other, through cluster analysis, the system can identify the candidate shear - wall areas composed of multiple single - body walls.

[0080] Further, the selection of single - body wall graphics for recombination according to the shear - wall candidate areas includes:

[0081] Based on the indoor plane matrix, in the shear - wall candidate area, determine adjacent single - body walls that meet the shear - wall characteristics through spatial position and geometric shape parameters;

[0082] Combine according to the spatial connection relationship between adjacent single - body walls that meet the shear - wall characteristics to obtain several combinations of adjacent single - body walls that meet the shear - wall characteristics;

[0083] According to the combination result, select single - body wall graphics for recombination splicing of the graphic edges.

[0084] In some embodiments, first, within the candidate region of shear walls, based on the spatial positions and geometric shape parameters of each individual wall, the system determines which individual walls have geometric shape features that conform to the characteristics of shear walls, such as the aspect ratio, area, and degree of matching with the standard structure of shear walls. Specifically, the system calculates the distance of each individual wall relative to other individual walls through the position coordinates in a two-dimensional plane matrix, and combines its geometric features (such as length and width) to determine whether they may belong to the same shear wall structure. During this process, the system identifies which individual walls are adjacent in physical space and have similar geometric shape features through the calculation of spatial proximity. After determining the adjacent individual walls that conform to the characteristics of shear walls, the system combines them by calculating the spatial connection relationships between these individual walls. The spatial connection relationships mainly consider the boundary alignment, direction consistency, and geometric shape similarity of adjacent individual walls. The system uses the coordinate data in the matrix and determines the continuity of adjacent individual walls through adjacency matrix or topological structure analysis. For example, when the boundary lengths of two individual walls are close and their connection points are linearly arranged on the plane, the system determines that they are connected to each other and combines these adjacent individual walls. The combined individual walls can present a larger continuous wall surface, which conforms to the overall structural characteristics of shear walls. After completing the spatial connection analysis of adjacent individual walls, the system selects appropriate individual walls for the recombination and splicing of the graphic edges according to the results of these combinations. In this step, the system stitches together the adjacent wall graphics according to the boundaries of the previously combined individual walls to form a complete shear wall graphic. To ensure the accuracy of the splicing, the system smooths the edges of the individual walls to eliminate the misalignment problems caused by the slight deviations between different walls. During the splicing process, the system relocates these spliced walls in the two-dimensional plane matrix according to the normalized size of each individual wall and outputs the recombined shear wall graphic. In this way, the system can generate a recombination based on the initially identified candidate region of individual walls through recombination and splicing.

[0085] Furthermore, for the graphic, the steps of parsing the semantics of the recombined graphic based on the semantic segmentation algorithm and using it as the input of the BP neural network to output the shear wall recognition result include the following:

[0086] Input the recombined graphic into the semantic segmentation model to parse the semantic information contained in the recombined graphic, including walls, windows, balcony openings, and door openings;

[0087] Import the semantic information into the BP neural network model to currently determine whether the recombined graphic conforms to the standard features of shear walls.

[0088] In some embodiments, first, the reconstructed graph is input into a semantic segmentation model for semantic parsing. The semantic segmentation model also uses a convolutional neural network (CNN) learning architecture and is capable of performing pixel-level classification on the input image. Specifically, the semantic segmentation model classifies each pixel in the reconstructed graph to determine the class to which the pixel belongs. Through this method, the system can parse out the key semantic information related to shear walls in the reconstructed graph, including structures such as walls, windows, balcony openings, doorways, etc. For example, by identifying the edge features and textures of the wall, the system can accurately locate the contour of the wall and distinguish surrounding building structures such as window sills and balconies at the same time, which is very important for distinguishing shear walls from non-load-bearing walls. During the semantic segmentation process, the model gradually learns and distinguishes different structures in the building according to the pre-trained data set. For example, elements such as windows and balcony openings usually have unique geometric morphological features in the image, and the semantic segmentation model combines this feature information to distinguish these areas from the wall. The output result after semantic segmentation is usually a label matrix, where the label of each pixel point represents the class to which the pixel belongs to a wall, window, balcony or other structure. Next, the system imports the parsed semantic information as input into a BP neural network model. The input of the BP neural network model is the structural feature vector extracted from the semantic segmentation model, and this feature vector includes the spatial distribution of the wall, the semantic classes of the surrounding structures, and the relative position relationship of these elements. In this way, the BP neural network can identify the shear walls in the reconstructed graph according to the input feature vector and in combination with the standard features of shear walls learned during the training process. After receiving the feature data after semantic segmentation, the BP neural network calculates the matching degree with the shear wall features through forward propagation. The weight and bias parameters of the BP neural network have been optimized according to a large number of labeled shear wall data during the training process. The system performs multi-layer processing on the input data through these parameters and gradually judges whether the reconstructed graph conforms to the standard features of shear walls. The BP neural network performs non-linear processing on the output of each layer through an activation function to ensure that the model can accurately identify complex shear wall structures. Finally, the system determines whether the reconstructed graph contains a structure that conforms to the shear wall standard according to the output result of the BP neural network. This result is usually output in the form of a classification probability, indicating the matching degree between the reconstructed graph and the shear wall features. In this way, the system can accurately distinguish shear walls from other walls or building structures and finally output the identification result of the shear wall.

[0089] Furthermore, the formula of the BP neural network is as follows:

[0090] y = f(W·x + b);

[0091] Among them, y is the output result, including whether the current recombined graphic is a shear wall; W is the weight matrix; b is the bias term; f is the activation function; x is the feature vector of the input layer, that is, the combination of semantic information.

[0092] Specifically, the BP neural network receives a set of feature vectors x from the input layer. These feature vectors are the results after parsing the recombined graphic through the semantic segmentation algorithm and contain the semantic information of wall bodies, window openings, balcony openings, and door openings. Each input feature represents the semantic features of different regions in the recombined graphic, describing their geometric forms, spatial positions, and categories. These input features will be linearly combined with the weight matrix W in the network. The weight matrix is a set of pre-trained parameters used to measure the importance of each input feature in the entire network. Each input feature will be multiplied by the corresponding weight value. The larger the weight, the greater the impact of the feature on the identification of the shear wall. The calculation of the weight matrix is optimized through the backpropagation algorithm to ensure that the final output can accurately judge the shear wall. Then, the system adds a bias term b to each node. This bias term is used to adjust the calculation result of the network to ensure a more accurate output. The bias term provides a basic activation level for each node, enabling the network to still output reasonable results even if all input features are zero. Subsequently, the network passes the results of these linear combinations to the activation function. The activation function is a non-linear transformation that enables the network to process more complex data patterns. In the process of shear wall identification, commonly used activation functions include ReLU (Rectified Linear Unit) and the Sigmoid function. These functions can transform the linear output of the network into non-linear probability values to determine whether the current recombined graphic conforms to the standard features of the shear wall. Finally, the output layer of the network calculates a result, which is y, indicating the network's judgment on whether the current recombined graphic is a shear wall.

[0093] The above embodiments are only descriptions of the preferred embodiments of the present invention and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A shear wall identification method based on reorganized segmentation graphics, characterized in that: The following steps are involved: S1. Acquire a number of indoor wall images and ranging data from different points or directions through an image acquisition device, and simultaneously collect the positioning and orientation data when the indoor wall images are acquired; S2. Use a convolutional neural network to calibrate the wall joints in the indoor image, segment several single wall graphics based on the wall joints, and calculate the height and length of the single wall based on the distance measurement data; S3, initializing a two-dimensional matrix and mapping several single walls into the two-dimensional matrix based on the height and length of the single wall and the positioning and orientation data when the wall image is acquired, to obtain an indoor plane matrix; S4. Based on the indoor plane matrix, the shear wall candidate area is identified through geometric morphology analysis. According to the shear wall candidate area, the single wall graphics are selected for reorganization to obtain the reorganized graphics. The semantics of the reorganized graphics are analyzed based on the semantic segmentation algorithm and used as the input of the BP neural network to output the shear wall identification result. S5. Perform three-dimensional visualization based on the identification results of the shear wall.

2. A shear wall identification method based on reorganized segmentation graph according to claim 1, characterized in that: The method of calibrating wall connections in indoor images by using a convolutional neural network comprises the following steps: Input the collected indoor wall images into the trained convolutional neural network model; The edge feature points in the wall image are extracted through the convolution layer and downsampled through the pooling layer; Based on the fully connected layer in the convolutional neural network, the extracted wall connection features are calibrated.

3. A shear wall identification method based on reorganized segmentation graph according to claim 2, characterized in that: The training of the convolutional neural network includes the following steps: Collect a number of indoor wall images without calibrated wall joints, manually annotate the wall joints, generate corresponding calibrated images, and form a paired training sample set; According to the visual features of the wall joints, the network structures of the convolutional layer, pooling layer, and fully connected layer are customized, and the network weight parameters are initialized; The uncalibrated indoor wall image is input into the convolutional neural network model, and the corresponding calibrated image is used as the supervision signal. The network parameters are iteratively updated using the back propagation algorithm and loss function.

4. The shear wall identification method based on reorganized segmentation graph according to claim 1, characterized in that: The method of calculating the height and length of a single wall based on the distance measurement data comprises the following steps: Extract the horizontal distance between the single wall and the image acquisition device, the distance between the vertical end point of the single wall and the image acquisition device, and the distance between the parallel end point of the single wall and the image acquisition device from the distance measurement data; The height of the single wall is calculated by the horizontal distance between the single wall and the image acquisition device and the distance between the vertical end points of the single wall and the image acquisition device; The length of the single wall is calculated by the horizontal distance between the single wall and the image acquisition device and the distance between the parallel end points of the single wall and the image acquisition device.

5. The shear wall identification method based on reorganized segmentation graph according to claim 1, characterized in that: The S3 comprises the following steps: Based on the height and length data of the single wall, the size of each single wall is normalized; Determine a sub-two-dimensional matrix of a two-dimensional projection for mapping according to the orientation of the image acquisition device and the spatial orientation of the single wall, and determine the matrix coordinates covered by the single wall according to the normalized length data of the single wall; According to the matrix coordinates covered by the single wall, the normalized height of the single wall is mapped to the matrix element value in the two-dimensional matrix to obtain the indoor plane matrix.

6. The shear wall identification method based on reorganized segmentation graph according to claim 1, characterized in that: The method of identifying the candidate shear wall area by geometric analysis comprises the following steps: Based on the single wall data in the indoor plane matrix, calculate the geometric parameters of each single wall, including aspect ratio, area and space proportion; The indoor plane matrix is ​​traversed through the pre-configured shear wall structure library to extract the indoor structural parts that meet the shear wall structure; In the indoor structure part that meets the shear wall structure, a single wall whose geometric parameters meet the standard is selected by a preset geometric standard parameter threshold as a candidate single wall of the shear wall; The selected candidate single walls are clustered according to their spatial positions to identify the candidate shear wall area composed of several single walls.

7. A shear wall identification method based on reorganized segmentation graph according to claim 6, characterized in that: The selecting of a single wall graphic according to the shear wall candidate area for reorganization includes: Based on the indoor plane matrix, the adjacent single wall that meets the shear wall characteristics is determined in the shear wall candidate area through the spatial position and geometric shape parameters; Combining the adjacent single walls that meet the characteristics of shear walls according to the spatial connection relationship between the adjacent single walls, to obtain a combination of several adjacent single walls that meet the characteristics of shear walls; According to the combination results, select the single wall graphics to reorganize and splice the edges of the graphics.

8. The shear wall identification method based on reorganized segmentation graph according to claim 1 is characterized in that: The method of analyzing and reorganizing the graphic semantics based on the semantic segmentation algorithm and using it as the input of the BP neural network to output the shear wall recognition result includes the following steps: The reconstructed image is input into the semantic segmentation model to analyze the semantic information contained in the reconstructed image, including walls, windows, balcony openings and doorways; By importing semantic information into the BP neural network model, it is currently determined whether the reconstructed graphics meet the standard characteristics of the shear wall.

9. The shear wall identification method based on reorganized segmentation graph according to claim 1, characterized in that: The formula of the BP neural network is as follows: y=f(W·x+b); Among them, y is the output result, including whether the current reorganized figure is a shear wall; W is the weight matrix; b is the bias term; f is the activation function; x is the feature vector of the input layer, that is, the combination of semantic information.

Images