A method for quantitatively characterizing spatial form of bridge disease by fusing three-dimensional laser point cloud and two-dimensional image

Abstract

The application discloses a kind of bridge disease spatial form quantitative characterization method of three-dimensional laser point cloud and two-dimensional image fusion, comprising the following steps: synchronous acquisition bridge disease image data and laser point cloud data;Data and are preprocessed;PSAG-Net model is constructed, point cloud data input PSAG-Net model training;CM-FPN model is constructed, image data input CM-FPN model training;With PSAG-Net model, crack point cloud data is segmented, for the large-scale crack point cloud of complete segmentation, post-processing is carried out and the missing crack area is completed, the depth value of large crack is calculated, and the depth distribution is counted;For the small-scale crack point cloud of incomplete segmentation, convert small-scale crack point cloud into depth map and binarization, use CM-FPN model to carry out crack pixel level detection and segmentation, back projection is back to three-dimensional coordinate system, and depth value is calculated.The application realizes accurate quantitative characterization of depth information.

Description

Technical Field

[0001] This invention belongs to the field of bridge engineering inspection technology, and in particular relates to a method for quantitative characterization of the spatial morphology of bridge defects by fusing three-dimensional laser point clouds with two-dimensional images. Background Technology

[0002] In the operation and maintenance of bridges, timely and accurate detection and quantitative characterization of bridge defects are crucial for ensuring bridge safety and durability. Currently, routine detection indicators for concrete surface defects cover quantity, length, width, and area. Although depth has also been included in the detection scope, the effective acquisition of depth information still faces significant challenges in practical applications due to limitations in existing technologies. Depth, as a key indicator of the in-depth development trend of defects, is a core basis for judging the degree of internal development of defects and an indispensable reference for implementing a comprehensive assessment of bridge safety and durability. However, most current detection methods struggle to efficiently and accurately acquire depth information of defects, especially in complex scenarios, extreme weather conditions, or dynamic objects. Relying solely on two-dimensional image processing is insufficient for accurate identification, and images themselves only provide two-dimensional projection information, failing to directly reflect depth and distance in three-dimensional space. Complex algorithms are needed for estimation and reconstruction, making accuracy difficult to guarantee. Furthermore, while three-dimensional laser scanning technology can provide rich spatial information, its data processing is complex and it has shortcomings in extracting detailed features. Summary of the Invention

[0003] Purpose of the invention: The purpose of this invention is to provide a method for quantitatively representing the spatial morphology of bridge defects by fusing three-dimensional laser point clouds with two-dimensional images, so as to achieve accurate quantitative representation of depth information.

[0004] Technical Solution: To achieve the above objectives, this invention discloses a method for quantitative characterization of the spatial morphology of bridge defects by fusing three-dimensional laser point clouds with two-dimensional images, comprising the following steps: S1. Define a dual-modal temporal synchronous acquisition mechanism to simultaneously acquire two-dimensional image data and three-dimensional laser point cloud data of bridge defects; S2. Preprocess the point cloud data and image data to obtain point cloud datasets and image datasets with spatiotemporal consistency; S3. Construct a multi-scale spatial attention network (PSAG-Net) model, input the point cloud dataset into the PSAG-Net model for training, and achieve segmentation of large cracks; construct a lightweight feature pyramid (CM-FPN) model, input the image dataset into the CM-FPN model for training, and achieve segmentation of small cracks. S4. Use the trained PSAG-Net model to segment the crack point cloud data and output the classification results. The complete segmentation results are large-scale crack point clouds, indicating that the crack point clouds are large in scale and have obvious features; the incomplete segmentation results are small-scale crack point clouds, indicating that the crack point clouds are small in scale and have insignificant features. S5. For the large-scale crack point cloud that is completely segmented by the PSAG-Net model, the 3D-3D registration algorithm is used for post-processing to fill in the crack regions that are missing due to downsampling during the segmentation process. After filling in the cracks, the depth values ​​of the large cracks are calculated and the depth distribution is statistically analyzed. S6. For small-scale crack point clouds that are not fully segmented by the PSAG-Net model, the small-scale crack point cloud is converted into a depth map and binarized by a 3D-2D cross-modal fusion algorithm. Then, the CM-FPN model is used for crack pixel-level detection and segmentation. Finally, the crack is back-projected back to the three-dimensional coordinate system and the depth value is calculated in combination with the reference plane.

[0005] Optionally, step S1 includes the following specific steps: S101. The camera resolution is no less than 2 million pixels, the scanning accuracy of the Leica P50 laser scanner is no less than 5mm, and the frame rate is set to 25Hz. S102. The relative positions and angles of the camera and laser scanner are kept fixed. Synchronization between the two is achieved through external synchronization triggering to ensure the consistency of the acquired two-dimensional images and three-dimensional laser point cloud data in time and space.

[0006] Optionally, step S2 includes the following specific steps: S201. Filter and denoise the point cloud data to remove outliers and noise points. Use a voxel filtering algorithm to divide the space into equal-length voxels, retaining the point with the most points within each voxel, to obtain the denoised point cloud data. ; Step S201 specifically includes the following steps: S2011, First, analyze the original point cloud. Constructing local geometric feature fields using KD-Tree acceleration Nearest neighbor search , definition point neighborhood Calculation points neighborhood covariance matrix The calculation formula is: , in, For the neighborhood centroid, For the first point cloud One point, For the neighborhood The points in the diagram are T, where T is the transpose; then, the points are extracted through eigenvalue decomposition. curvature eigenvalues curvature eigenvalues The region is a high-curvature region, which is marked as a potential crack edge. The curvature characteristic value... The calculation formula is: , in, Represents the covariance matrix eigenvalues, and satisfying , This indicates the degree of dispersion of neighborhood points along the first principal direction; This indicates the degree of dispersion of neighboring points along the second principal direction; This indicates the degree of dispersion of the neighborhood points along the third principal direction; S2012. Perform adaptive voxel partitioning on the point cloud data, and define the formula for calculating the voxel resolution adjustment amount for each point as follows: , For point The amount of voxel resolution adjustment, The baseline resolution is the baseline resolution for the crack-sensitive area. =0.5cm, α and β are the control parameters determined by cross-validation, α=0.8, β=10; then voxel sampling is used, at each voxel The point with the maximum curvature is preserved within the inner region; the expression is: , in, Indicated in voxels The point with the largest curvature retained within the data is used to obtain the denoised point cloud data. ; S202. Perform geometric correction and feature enhancement processing on the image data to generate an enhanced response map with brighter crack areas and darker background areas, and output the enhanced image data. Step S202 includes the following specific steps: S2021. Construct a geometric correction model driven by optical flow field, and obtain the displacement field by minimizing the energy function. The expression for minimizing the energy function is: , in, For the original image at position pixel values, Indicates passing through the displacement field The transformed image at position pixel values, Weights for smoothing terms; and Indicates the image in and Displacement in the direction, Let be the smoothing term of the displacement field, and let represent the square of the Frobenius norm of the displacement field gradient. The optimal displacement field is obtained by iteratively solving using the multi-scale Gauss-Newton method. And generate a geometrically corrected image, the original image at the location Corrected pixel values The expression is: , in, Indicates passing through the optimal displacement field The transformed image at position Pixel values; S2022, In the corrected pixel values Image features formed on Point features are obtained by combining point-side cues or projections. A feature mapping function is constructed to improve the signal-to-noise ratio in the crack region, and an enhanced response map is output. Enhanced response map The expression for the feature mapping function is: , in, This represents an activation function, such as the sigmoid function; and The weight matrix of the feature map. and These are image features and point features.

[0007] Optionally, step S3 includes the following specific steps: S301. Construct a multi-scale spatial attention network (PSAG-Net) model. The multi-scale spatial attention network includes a point cloud hierarchical feature extraction module and a semantic segmentation module. The point cloud hierarchical feature extraction module includes a sampling grouping module and a spatial attention module. The input data of the point cloud hierarchical feature extraction module is a point cloud dataset. The point cloud data is input into the sampling grouping module for upsampling and grouping operations to capture the first-layer geometric features. The first-layer geometric features are input into the spatial attention module for fusion processing to obtain the first-layer optimized features. The first-layer optimized features are input into the sampling grouping module for upsampling and grouping operations to capture the second-layer geometric features. The second-layer geometric features are input into the spatial attention module for fusion processing to obtain the second-layer optimized features. The semantic segmentation module includes an upsampling layer, a grouping layer, and a spatial attention module. The second-layer optimized features, after being processed by the upsampling layer, are input together with the second-layer geometric features into the grouping layer for grouping processing to obtain a local fusion feature tensor. This tensor is then input into the spatial attention module for processing to obtain attention-enhanced semantic context features. The semantic context features and the first-layer geometric features are input together into the grouping layer for grouping processing, and the segmented point cloud and point-level classification results are output. The spatial attention module includes convolutional layers, global pooling layers, fully connected (FC) layers, ReLU activation functions, Sigmoid activation functions, concatenation layers, and normalization layers. In the feature input spatial attention module, multi-level convolutional transformations are first performed to obtain feature maps. Then, global pooling and FC fully connected layers are used to extract global descriptions. Attention weights are generated by combining ReLU activation functions, FC fully connected layers, and Sigmoid functions. The features are then weighted and fused point by point. After fusion, the features are concatenated and normalized to output the features that have undergone attention weighting and multi-scale fusion. S302. Construct a lightweight feature pyramid CM-FPN model. The lightweight feature pyramid includes a feature extraction module, a feature fusion module, and an output module. The feature extraction module includes a convolutional layer, a first C3k2 module, a second C3k2 module, a spatial pyramid pooling layer, and a C2PSA spatial attention module. The image dataset is input into the feature extraction module and sequentially passes through two levels of convolutional layers, the first C3k2 module, a convolutional layer, and the second C3k2 module to extract the first feature. The first feature is then extracted through a convolutional layer and the first C3k2 module to obtain the second feature. The second feature is then processed through a convolutional layer, the second C3k2 module, the spatial pyramid pooling layer, and the C2PSA spatial attention module to obtain the third feature. The first C3k2 module includes a convolutional layer, a segmentation layer, and a C3k2 module. Initial features are input into the first C3k2 module, and after processing by the convolutional layer, convolutional features are obtained. The convolutional features are processed by the segmentation layer to obtain segmentation features. The segmentation features are processed by two C3k2 modules and then input together with the convolutional features and segmentation features into two convolutional layers for processing to output the final features. The second C3k2 module includes convolutional layers, segmentation layers, and a bottleneck structure. Initial features are input into the second C3k2 module, processed by the convolutional layers to obtain convolutional features, which are then processed by the segmentation layers to obtain segmentation features. The segmentation features are processed by two bottleneck structures and then input together with the convolutional features and segmentation features into two convolutional layers for further processing to output the final features. The bottleneck structure includes two convolutional layers. Initial features are input into the bottleneck structure, processed by the two convolutional layers, and then summed with the initial features to obtain the output features. The feature fusion module includes an upsampling layer, a stitching layer, a C3k2 module, and a convolutional layer. The third and second features, after passing through the upsampling layer, are input together with the first feature into the stitching layer. After stitching, they are sequentially input into the C3k2 module to form the first fused feature. The first fused feature, after passing through the upsampling layer, is stitched together with the first feature to form the first stitched feature. The first fused feature, after passing through the stitching layer and the convolutional layer, is input together with the first stitched feature into the C3k2 module to form the second fused feature. The first fused feature, after passing through the stitching layer and the C3k2 module, forms the third fused feature. The third fused feature, after passing through the convolutional layer, is stitched together with the second feature to form the second stitched feature. The second fused feature, the second stitched feature, and the third fused feature are input together into the output module for segmentation, generating a crack segmentation mask. S303. Train the PSAG-Net model using the point cloud dataset of the cracks, and train the CM-FPN model using the image dataset of the cracks. After training, evaluate the models and save the best PSAG-Net and CM-FPN models.

[0008] Optionally, step S5 includes the following specific steps: S501. For large-scale crack point clouds with complete segmentation results, use a registration algorithm to calculate the rotation matrix R and eigenvector t, and then transfer the target point cloud... Compared with the original point cloud Perform feature alignment to obtain complete geometric information of the crack; S502. Calculate the depth value of each crack point, statistically analyze the depth distribution of the crack point cloud, and obtain the mean, maximum and distribution histogram of the depth to form the overall depth distribution characteristics of the crack. Calculate the depth of the crack point using the formula for the distance from a point to a plane. For any crack point... Its fit plane The expression for the distance d is: , in, This is the distance from the point to the plane, i.e., the depth value.

[0009] Optionally, step S501 includes the following specific steps: S5011, using RANSAC in the target point cloud A plane is fitted over the crack neighborhood, and the fitted plane is: The equivalent representation is: , in, For target point cloud The coefficients of the plane equation, Represents the coordinates of point i; The fitted plane serves as the reference crack plane, thus establishing a unified reference surface for subsequent alignment and depth measurement. S5012, via target point cloud The plane equation coefficients are used to calculate the target point cloud. Normal vector of the fitted plane In the original point cloud The original point cloud is obtained by fitting the plane. The plane equation coefficients are used to calculate the original point cloud. normal vector To ensure that the normals of the two clouds are aligned, a rotation axis is constructed: , , in, Let be the rotation axis vector. Let be the rotation angle; and give the rotation matrix using Rodriguez's formula, the expression is: , Obtain the target point cloud Rotate the normal to the original point cloud Consistent minimum angular rotation R eliminates the attitude difference between the two clouds, ensuring that subsequent translation and depth calculations are performed in the same normal coordinate system; S5013, Calculate the centroid of two clouds and Calculate the translation vector based on the centroid difference: , Obtain the translation vector for rigid registration Under the premise of consistent normals, position alignment is completed so that the two clouds coincide in the same global coordinate system; S5014, Finally, based on the translation vector and rotation matrix Align the original point cloud and target point cloud : , Obtain the target point cloud Aligned point set The aligned crack point cloud set is obtained to provide consistent coordinates and a unified reference surface for subsequent depth estimation and statistical calculation.

[0010] Optionally, step S6 includes the following specific steps: S601, through a 3D-2D cross-modal fusion algorithm, integrates point clouds... Mapped to a structured depth map The coordinates of the third dimension are mapped to the pixel value range of 0-255, so that the depth map contains the third dimension information, and the depth map of the crack is binarized. S602. Use the CM-FPN model to perform pixel-level prediction on the binarized depth map to obtain the crack pixel set, and then back-project it back to the three-dimensional coordinate system to obtain the crack point cloud set. S603. Normalize the depth value corresponding to the pixel, calculate the z-coordinate of the point cloud, and reconstruct the three-dimensional coordinates of the crack point cloud. S604. Calculate the depth corresponding to each pixel point based on the three-dimensional coordinates, and collect the depth information of the micro-cracks.

[0011] Optionally, step S601 includes the following specific steps: S6011. First, the point cloud data is corrected according to the method in step S501 to complete the spatial transformation of the coordinate system. After correction, the Y-axis in the point cloud coordinate system is used to represent the depth direction, while the X-axis and Z-axis correspond to the plane where the crack is located. S6012. Then, normalize the Y coordinate values ​​of the point cloud and map them between 0 and 255 pixel values ​​to map the point cloud into a two-dimensional depth map. , in, The raw point cloud data, For depth map, Represents depth information, This represents a combination of depth map and point cloud data; S6013. Subsequently, threshold segmentation is performed to obtain a binarized depth map: , in, This represents a binarized depth map, retaining pixel values ​​greater than or equal to a threshold. Pixels.

[0012] Optionally, step S602 includes the following specific steps: S6021. Use the CM-FPN model to perform crack detection and segmentation on the depth map. The expression is: , in, This is the set of pixels after the crack is segmented. This indicates that the CM-FPN model is used for crack detection and segmentation on the binary depth map; S6022, Set the segmented crack pixels Back-projecting back to the 3D coordinate system yields the point cloud set of the crack. Point cloud set of cracks The k-th point cloud coordinates for: , Where K represents the point cloud set The total number of midpoints These are the coordinates of the pixels in the segmented crack.

[0013] Optionally, step S604 includes the following specific steps: According to the fitted plane For crack point clouds Each point in Calculate its orthogonal distance to the plane: , in, This is the distance from the point to the line, i.e., the depth value; Calculate the depth value of each point Describe the distribution; , in, This represents the set of depth values ​​for each point obtained from the final statistics. Represents a set of point clouds A function to calculate the depth value for each point in the array.

[0014] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: By simultaneously acquiring high-resolution images and high-precision point cloud data, and employing a deep learning model for targeted semantic segmentation, combined with cross-modal data fusion technology, this invention achieves accurate quantitative characterization of the spatial morphology of bridge defects; it solves the shortcomings of traditional methods in depth estimation accuracy and the problem that a single data source cannot comprehensively characterize defect features; it significantly improves the accuracy and efficiency of defect detection, providing a reliable basis for timely bridge maintenance and safety assessment; this invention uses differentiated segmentation models for cracks of different sizes, further optimizing the precision of defect identification and ensuring that all types of defect features can be accurately detected and quantitatively analyzed. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the process in this invention; Figure 2 This is a schematic diagram of the dual-channel data acquisition and point cloud coordinate system in this invention; Figure 3 This is a schematic diagram of point cloud preprocessing in this invention; Figure 4 This is a schematic diagram of the processed point cloud and image data in this invention; Figure 5 This is a structural diagram of the PSAG-Net model in this invention; Figure 6 This is a structural diagram of the spatial attention module in this invention; Figure 7 This is a structural diagram of the CM-FPN model in this invention; Figure 8 This is a schematic diagram of the point cloud registration algorithm in this invention. Detailed Implementation

[0016] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0017] like Figure 1 As shown, this invention discloses a method for quantitatively characterizing the spatial morphology of bridge defects by fusing three-dimensional laser point clouds with two-dimensional images, comprising the following steps: S1. Define a dual-modal temporal synchronous acquisition mechanism, using a camera and a laser scanner to simultaneously acquire two-dimensional image data and three-dimensional laser point cloud data of bridge defects; like Figure 2 As shown, step S1 includes the following specific steps: S101. The camera resolution is no less than 2 million pixels, the scanning accuracy of the Leica P50 laser scanner is no less than 5mm, and the frame rate is set to 25Hz. S102. The relative positions and angles of the camera and laser scanner are kept fixed. Precise synchronization between the two is achieved through external synchronization triggering, ensuring the consistency of the acquired two-dimensional images and three-dimensional laser point cloud data in time and space.

[0018] S2. Preprocess the point cloud data and image data to obtain point cloud datasets and image datasets with spatiotemporal consistency; like Figure 3 and Figure 4 As shown, step S2 includes the following specific steps: S201. Filter and denoise the point cloud data to remove outliers and noise points. A voxel filtering algorithm is used to divide the space into equal-length voxels, retaining the point with the most points within each voxel to reduce data volume and smooth the point cloud, resulting in denoised point cloud data. ; Step S201 specifically includes the following steps: S2011, First, analyze the original point cloud. Constructing local geometric feature fields using KD-Tree acceleration Nearest neighbor search , definition point neighborhood Calculation points neighborhood covariance matrix The calculation formula is: , in, For the neighborhood centroid, For the first point cloud One point, For the neighborhood The points in the diagram are T, where T is the transpose; then, the points are extracted through eigenvalue decomposition. curvature eigenvalues curvature eigenvalues The region is a high-curvature region, which is marked as a potential crack edge. The curvature characteristic value... The calculation formula is: , in, Represents the covariance matrix eigenvalues, and satisfying , This indicates the degree of dispersion of neighborhood points along the first principal direction; This indicates the degree of dispersion of neighboring points along the second principal direction; Indicates the degree of dispersion of neighborhood points along the third principal direction; curvature eigenvalue A larger value indicates that the point is located in a region of high curvature and is more likely to be a potential crack edge point. The purpose of this step is to identify crack-related edge regions in the point cloud through geometric feature extraction.

[0019] S2012. Perform adaptive voxel partitioning on the point cloud data, and define the formula for calculating the voxel resolution adjustment amount for each point as follows: , For point The amount of voxel resolution adjustment, The baseline resolution is the baseline resolution for the crack-sensitive area. =0.5cm, α and β are the control parameters determined by cross-validation, α=0.8, β=10; then voxel sampling is used, at each voxel The point with the maximum curvature is preserved within the inner region; the expression is: , in, Indicated in voxels The point with the largest curvature retained within the data is used to obtain the denoised point cloud data. The purpose of this step is to improve sampling accuracy in crack-sensitive areas while reducing the number of points in non-critical areas, achieving data compression of approximately 40% and a crack edge point retention rate of over 95%.

[0020] S202. Perform geometric correction and feature enhancement processing on the image data to generate an enhanced response map with brighter crack areas and darker background areas, and output the enhanced image data. Step S202 includes the following specific steps: S2021. Construct a geometric correction model driven by optical flow field, and obtain the displacement field by minimizing the energy function. The expression for minimizing the energy function is: , in, For the original image at position pixel values, Indicates passing through the displacement field The transformed image at position pixel values, Weights for smoothing terms; and Indicates the image in and Displacement in the direction, Let be the smoothing term of the displacement field, and let represent the square of the Frobenius norm of the displacement field gradient. The optimal displacement field is obtained by iteratively solving using the multi-scale Gauss-Newton method. And generate a geometrically corrected image, the original image at the location Corrected pixel values The expression is: , in, Indicates passing through the optimal displacement field The transformed image at position The pixel values; the purpose of this step is to eliminate the spatiotemporal inconsistencies caused by motion / distortion, and to provide an alignment baseline for subsequent feature enhancement and dynamic range extension.

[0021] S2022, In the corrected pixel values Image features formed on Point features are obtained by combining point-side cues or projections. A feature mapping function is constructed to improve the signal-to-noise ratio in the crack region, and an enhanced response map is output. Enhanced response map The expression for the feature mapping function is: , in, This represents an activation function, such as the sigmoid function; and The weight matrix of the feature map. and The functions of this step are image features and point features; the purpose of this step is to output the enhanced response map. The response map has a higher value at the crack, which is used to significantly improve the contrast between the crack and the background.

[0022] S3. Construct a multi-scale spatial attention network PSAG-Net, and train it with point cloud datasets to achieve high-precision segmentation of large cracks; construct a lightweight feature pyramid CM-FPN, and train it with image datasets to achieve fine segmentation of small cracks. Step S3 includes the following specific steps: S301. Construct the multi-scale spatial attention network PSAG-Net model (Point Cloud Spatial Attention Guided Network), such as... Figure 5 As shown, the multi-scale spatial attention network includes a point cloud hierarchical feature extraction module and a semantic segmentation module. The point cloud hierarchical feature extraction module includes a sampling grouping module and a spatial attention module. The input data of the point cloud hierarchical feature extraction module is a point cloud dataset, represented as a matrix (N, d+C), where N is the number of points, d is the three-dimensional coordinates (x, y, z) of each point, and C represents additional features, such as color or normal information. The point cloud data is input into the sampling grouping module for upsampling and grouping operations to capture the first layer of geometric features. The first layer of geometric features are input into the spatial attention module for fusion processing to obtain the first layer of optimized features. The first layer of optimized features are input into the sampling grouping module for upsampling and grouping operations to capture the second layer of geometric features. The second layer of geometric features are input into the spatial attention module for fusion processing to obtain the second layer of optimized features. The semantic segmentation module includes an upsampling layer, a grouping layer, and a spatial attention module. The optimized features of the second layer are processed by the upsampling layer and then input together with the geometric features of the second layer into the grouping layer for grouping processing to obtain a local fusion feature tensor. This tensor is then input into the spatial attention module for processing to obtain attention-enhanced semantic context features. The semantic context features and the geometric features of the first layer are input together into the grouping layer for grouping processing, and the output is the segmented point cloud and point-level classification results. Crack points are marked as 1, and background points are marked as 0. like Figure 6As shown, the spatial attention module includes convolutional layers, global pooling layers, fully connected (FC) layers, ReLU activation function, Sigmoid activation function, concatenation layer, and normalization layer. In the feature input spatial attention module, multi-level convolutional transformations are first performed to obtain feature maps. Then, global pooling and FC fully connected layers are used to extract global descriptions. Attention weights are generated by combining ReLU activation function, FC fully connected layer, and Sigmoid function. The features are then weighted and fused point by point. After fusion, the features are concatenated and normalized to output the features after attention weighting and multi-scale fusion. The backbone of the multi-scale spatial attention network uses the PointNet++ architecture for hierarchical feature extraction, capturing local geometric features at each scale through upsampling and grouping operations. A spatial attention module is then introduced, obtaining feature maps through convolutional transformations, followed by global pooling and fully connected layers to extract global descriptions. ReLU activation and the sigmoid function are combined to generate attention weights, dynamically adjusting the importance of features in different regions and highlighting crack areas. Subsequently, skip connections are used to concatenate and fuse shallow geometric features with upsampled deep semantic features, allowing the network to preserve edge details while utilizing high-level abstract information for crack discrimination. The fused features enter the semantic segmentation module, where upsampling restores the original resolution. Further local grouping and optional spatial attention module enhancements refine crack boundaries. Finally, 1×1 convolutions map the crack to both crack and background channels, and softmax normalization yields point-level classification results. In the network output, crack points are labeled as 1, and background points as 0.

[0023] S302. Construct a lightweight feature pyramid CM-FPN model (Cross-Modal Feature Pyramid Network), such as Figure 7 As shown, the lightweight feature pyramid includes a feature extraction module, a feature fusion module, and an output module. The feature extraction module includes a convolutional layer, a first C3k2 module, a second C3k2 module, a spatial pyramid pooling layer, and a C2PSA spatial attention module. The image dataset is input into the feature extraction module and sequentially passes through two levels of convolutional layers, the first C3k2 module, another convolutional layer, and the second C3k2 module to extract the first feature. The first feature is then extracted through a convolutional layer and the first C3k2 module to obtain the second feature. The second feature is then processed through a convolutional layer, the second C3k2 module, a spatial pyramid pooling layer, and the C2PSA spatial attention module to obtain the third feature. The first C3k2 module includes a convolutional layer, a segmentation layer, and a C3k2 module. Initial features are input into the first C3k2 module, and after processing by the convolutional layer, convolutional features are obtained. The convolutional features are processed by the segmentation layer to obtain segmentation features. The segmentation features are processed by two C3k2 modules and then input together with the convolutional features and segmentation features into two convolutional layers for processing to output the final features. The second C3k2 module includes convolutional layers, segmentation layers, and a bottleneck structure. Initial features are input into the second C3k2 module, processed by the convolutional layers to obtain convolutional features, which are then processed by the segmentation layers to obtain segmentation features. The segmentation features are processed by two bottleneck structures and then input together with the convolutional features and segmentation features into two convolutional layers for further processing to output the final features. The bottleneck structure includes two convolutional layers. Initial features are input into the bottleneck structure, processed by the two convolutional layers, and then summed with the initial features to obtain the output features. The feature fusion module includes an upsampling layer, a stitching layer, a C3k2 module, and a convolutional layer. The third and second features, after passing through the upsampling layer, are input together with the first feature into the stitching layer. After stitching, they are sequentially input into the C3k2 module to form the first fused feature. The first fused feature, after passing through the upsampling layer, is stitched together with the first feature to form the first stitched feature. The first fused feature, after passing through the stitching layer and the convolutional layer, is input together with the first stitched feature into the C3k2 module to form the second fused feature. The first fused feature, after passing through the stitching layer and the C3k2 module, forms the third fused feature. The third fused feature, after passing through the convolutional layer, is stitched together with the second feature to form the second stitched feature. The second fused feature, the second stitched feature, and the third fused feature are input together into the output module for segmentation, generating a crack segmentation mask, where crack pixels are marked as 1 and background pixels are marked as 0. The lightweight feature pyramid takes a high-resolution image of the bridge surface as input, with the input tensor containing spatial texture features of cracks. The backbone network of the lightweight feature pyramid employs a C3k2 module and a Spatial Pyramid Pooling (SPPF) module. The C3k2 module extracts basic semantic features of the image through a standard convolutional structure. The first C3k2 module (c3k=0) focuses more on capturing edge details, while the second C3k2 module (c3k=1) combines bottleneck structures to enhance channel representation and reduce redundancy. After stacking C3k2 modules, the SPPF module extracts multi-scale contextual information through parallel operations of pooling kernels at different scales, enabling the network to robustly perceive cracks of different sizes. Subsequently, a C2PSA spatial attention module is introduced in the feature fusion layer to adaptively adjust the importance of features at different levels and highlight the response of crack regions. Based on this, a top-down feature pyramid network (FPN) is constructed, which restores spatial resolution from high-semantic low-resolution features through progressive upsampling and concatenates them with shallow high-resolution features from the backbone network. The fused features are then further refined by the C3k2 module and convolutional layers. In the output layer, the fused features are mapped to the crack / background channels and pixel-level classification and normalization are performed through the Softmax function to generate a crack segmentation mask, in which crack pixels are marked as 1 and background pixels are marked as 0, thereby achieving accurate detection and segmentation of cracks on complex bridge surfaces.

[0024] S303. Train the PSAG-Net model using the point cloud dataset of the cracks, and train the CM-FPN model using the image dataset of the cracks. After training, evaluate the models and save the best PSAG-Net model and CM-FPN model. Step S303 includes the following sub-steps: The preprocessed crack point cloud dataset is divided into a training set and a validation set in an 8:2 ratio. The PSAG-Net model is trained using the training set, and the PSAG-Net model with the best performance on the validation set is saved based on the evaluation results of the validation set. The preprocessed crack image dataset was divided into a training set and a validation set in an 8:2 ratio. The CM-FPN model was trained using the training set, and the CM-FPN model with the best performance on the validation set was saved based on the evaluation results of the validation set. S4. Use the trained PSAG-Net model to segment the crack point cloud data and output the classification results. The complete segmentation results are large-scale crack point clouds, indicating that the crack point clouds are large in scale and have obvious features; the incomplete segmentation results are small-scale crack point clouds, indicating that the crack point clouds are small in scale and have insignificant features. S5. For the large-scale crack point cloud that is completely segmented by the PSAG-Net model, the 3D-3D registration algorithm is used for post-processing to fill in the crack regions that are missing due to downsampling during the segmentation process. After filling in the cracks, the depth values ​​of the large cracks are calculated and the depth distribution is statistically analyzed. like Figure 8 As shown, step S5 includes the following sub-steps: S501. For large-scale crack point clouds with complete segmentation results, the output crack point cloud becomes sparse due to the hierarchical feature extraction and downsampling operations of the PSAG-Net model. A registration algorithm is used to calculate the rotation matrix R and the feature vector t, and then... Compared with the original point cloud Perform feature alignment to obtain complete geometric information of the crack; Step S501 includes the following specific steps: S5011, using RANSAC in the target point cloud A plane is fitted over the crack neighborhood, and the fitted plane is: The equivalent representation is: , in, For target point cloud The coefficients of the plane equation, Represents the coordinates of point i; The fitted plane serves as the reference crack plane, thus establishing a unified reference surface for subsequent alignment and depth measurement. S5012, via target point cloud The plane equation coefficients are used to calculate the target point cloud. Normal vector of the fitted plane In the original point cloud The original point cloud is obtained by fitting the plane. The plane equation coefficients are used to calculate the original point cloud. normal vector To ensure that the normals of the two clouds are aligned, a rotation axis is constructed: , , in, Let be the rotation axis vector. Let be the rotation angle; and give the rotation matrix using Rodriguez's formula, the expression is: , Obtain the target point cloud Rotate the normal to the original point cloud Consistent minimum angular rotation R eliminates the attitude difference between the two clouds, ensuring that subsequent translation and depth calculations are performed in the same normal coordinate system; S5013, Calculate the centroid of two clouds and Calculate the translation vector based on the centroid difference: , Obtain the translation vector for rigid registration Position alignment is achieved under the premise of consistent normal direction, so that the two clouds coincide in the same global coordinate system.

[0025] S5014, Finally, based on the translation vector and rotation matrix Align the original point cloud and target point cloud : , Obtain the target point cloud Aligned point set The aligned crack point cloud set is obtained to provide consistent coordinates and a unified reference surface for subsequent depth estimation and statistical calculation. S502. Calculate the depth value of each crack point, statistically analyze the depth distribution of the crack point cloud, and obtain the mean, maximum and distribution histogram of the depth to form the overall depth distribution characteristics of the crack. Calculate the depth of the crack point using the formula for the distance from a point to a plane. For any crack point... Its fit plane The expression for the distance d is: , in, This is the distance from the point to the plane, i.e., the depth value.

[0026] S6. For small-scale crack point clouds that are not fully segmented using the PSAG-Net model, the small-size crack point cloud is converted into a depth map and binarized using a 3D-2D cross-modal fusion algorithm. Then, the CM-FPN model is used to realize pixel-level crack detection and segmentation. Finally, the crack is back-projected back to the three-dimensional coordinate system and the depth value is calculated in conjunction with the reference plane.

[0027] Step S6 includes the following specific steps: S601, through a 3D-2D cross-modal fusion algorithm, integrates point clouds... Mapped to a structured depth map The coordinates of the third dimension are mapped to a pixel value range of 0-255, so that the depth map contains the third dimension information, and the depth map of the crack is binarized.

[0028] Step S601 includes the following specific steps: S6011. First, the point cloud data is corrected according to the method in step S501 to complete the spatial transformation of the coordinate system. After correction, the Y-axis in the point cloud coordinate system is used to represent the depth direction, while the X-axis and Z-axis correspond to the plane where the crack is located. S6012. Then, normalize the Y coordinate values ​​of the point cloud and map them between 0 and 255 pixel values ​​to map the point cloud into a two-dimensional depth map. , in, The raw point cloud data, For depth map, Represents depth information, This represents a combination of depth map and point cloud data; S6013. Subsequently, threshold segmentation is performed to obtain a binarized depth map: , in, This represents a binarized depth map, retaining pixel values ​​greater than or equal to a threshold. Pixels; S602. Use the CM-FPN model to perform pixel-level prediction on the binarized depth map to obtain the crack pixel set. Step S602 includes the following specific steps: S6021. Use the CM-FPN model to perform crack detection and segmentation on the depth map. The expression is: , in, This is the set of pixels after the crack is segmented. This step involves using the CM-FPN model to detect and segment cracks in a binary depth map. The purpose of this step is to achieve accurate detection and pixel-level segmentation of small cracks, and output the distribution area of ​​cracks in the two-dimensional depth map.

[0029] S6022, Set the segmented crack pixels Back-projecting back to the 3D coordinate system yields the point cloud set of the crack. Point cloud set of cracks The k-th point cloud coordinates for: , Where K represents the point cloud set The total number of midpoints These are the coordinates of the segmented crack pixels; S603. Inversely normalize the depth values ​​corresponding to the pixels to calculate the z-coordinate of the point cloud and reconstruct the three-dimensional coordinates of the crack point cloud. The purpose of this step is to revert the two-dimensional segmentation result to a three-dimensional crack point cloud and restore its true spatial position.

[0030] S604. Calculate the depth corresponding to each pixel point based on the three-dimensional coordinates, and collect the depth information of the micro-cracks.

[0031] According to the fitted plane in step S402 For crack point clouds Each point in Calculate its orthogonal distance to the plane: , in, This is the distance from the point to the line, i.e., the depth value; Calculate the depth value of each point , describes the distribution.

[0032] , in, This represents the set of depth values ​​for each point obtained from the final statistics. Represents a set of point clouds A function to calculate the depth value for each point in the array.

Claims

1. A method for quantitatively characterizing the spatial morphology of bridge defects by fusing three-dimensional laser point clouds with two-dimensional images, characterized in that, Includes the following steps: S1. Define a dual-modal temporal synchronous acquisition mechanism to simultaneously acquire two-dimensional image data and three-dimensional laser point cloud data of bridge defects; S2. Preprocess the point cloud data and image data to obtain point cloud datasets and image datasets with spatiotemporal consistency; S3. Construct a multi-scale spatial attention network (PSAG-Net) model, input the point cloud dataset into the PSAG-Net model for training, and achieve segmentation of large cracks; construct a lightweight feature pyramid (CM-FPN) model, input the image dataset into the CM-FPN model for training, and achieve segmentation of small cracks. The backbone of the multi-scale spatial attention network uses the PointNet++ architecture for hierarchical feature extraction, capturing local geometric features at each scale through upsampling and grouping operations; it also introduces... The spatial attention module obtains feature maps through convolutional transformations, then extracts global descriptions through global pooling and fully connected layers, and generates attention weights by combining ReLU activation and the Sigmoid function to weight the original features point by point; subsequently, a skip connection method is used to concatenate and fuse shallow geometric features with upsampled deep semantic features. The fused features enter the semantic segmentation module, are upsampled to restore the original resolution, and then refined by combining local grouping again and the enhancement operation of the optional spatial attention module. Finally, they are convolved and mapped to the crack / background channels, and then normalized to obtain point-level classification results. The backbone network of the lightweight feature pyramid employs a first C3k2 module, a second C3k2 module, and a spatial pyramid pooling SPPF module. The first and second C3k2 modules extract basic semantic features of the image through standard convolutional structures, while the spatial pyramid pooling SPPF module extracts multi-scale contextual information through parallel operations of pooling kernels at different scales. The feature fusion layer introduces a C2PSA spatial attention module to adaptively adjust the importance of features at different levels and highlight the response of crack regions. A top-down feature pyramid network FPN is then constructed, which restores spatial resolution from high semantic low-resolution features through stepwise upsampling and concatenates and fuses them with shallow high-resolution features from the backbone network. The fused features are then further refined by the C3k2 module and convolutional layers. S4. Use the trained PSAG-Net model to segment the crack point cloud data and output the classification results. The point cloud with complete segmentation results is the large-scale crack point cloud, and the point cloud with incomplete segmentation results is the small-scale crack point cloud. S5. For the large-scale crack point cloud that is completely segmented by the PSAG-Net model, the 3D-3D registration algorithm is used for post-processing to fill in the crack regions that are missing due to downsampling during the segmentation process. After filling in the cracks, the depth values ​​of the large cracks are calculated and the depth distribution is statistically analyzed. S6. For small-scale crack point clouds that are not fully segmented by the PSAG-Net model, the small-scale crack point cloud is converted into a depth map and binarized by a 3D-2D cross-modal fusion algorithm. Then, the CM-FPN model is used for crack pixel-level detection and segmentation. Finally, the crack is back-projected back to the three-dimensional coordinate system and the depth value is calculated in combination with the reference plane.

2. The method for quantitative characterization of the spatial morphology of bridge defects by fusing three-dimensional laser point clouds and two-dimensional images according to claim 1, characterized in that: Step S1 includes the following specific steps: S101. The camera resolution is no less than 2 million pixels, and the scanning accuracy of the Leica P50 laser scanner is no less than 5mm. S102. The relative positions and angles of the camera and laser scanner are kept fixed. Synchronization between the two is achieved through external synchronization triggering to ensure the consistency of the acquired two-dimensional images and three-dimensional laser point cloud data in time and space.

3. The method for quantitative characterization of the spatial morphology of bridge defects by fusing three-dimensional laser point clouds and two-dimensional images according to claim 1, characterized in that: Step S2 includes the following specific steps: S201. Filter and denoise the point cloud data to remove outliers and noise points. Use a voxel filtering algorithm to divide the space into equal-length voxels, retaining the point with the most points within each voxel, to obtain the denoised point cloud data. ; Step S201 specifically includes the following steps: S2011, First, analyze the original point cloud. Constructing local geometric feature fields using KD-Tree acceleration Nearest neighbor search , definition point neighborhood Calculation points neighborhood covariance matrix The calculation formula is: ， in, For the neighborhood centroid, For the first point cloud One point, For the neighborhood The points in the diagram are T, where T is the transpose; then, the points are extracted through eigenvalue decomposition. curvature eigenvalues curvature eigenvalues The region is a high-curvature region, which is marked as a potential crack edge. The curvature characteristic value... The calculation formula is: ， in, Represents the covariance matrix eigenvalues, and satisfying , This indicates the degree of dispersion of neighborhood points along the first principal direction; This indicates the degree of dispersion of neighboring points along the second principal direction; This indicates the degree of dispersion of the neighborhood points along the third principal direction; S2012. Perform adaptive voxel partitioning on the point cloud data, and define the formula for calculating the voxel resolution adjustment amount for each point as follows: ， For point The amount of voxel resolution adjustment, As the baseline resolution, α and β are control parameters determined through cross-validation; then voxel sampling is used, at each voxel... The point with the maximum curvature is preserved within the inner region; the expression is: ， in, Indicated in voxels The point with the largest curvature retained within the data is used to obtain the denoised point cloud data. ; S202. Perform geometric correction and feature enhancement processing on the image data to generate an enhanced response map with brighter crack areas and darker background areas, and output the enhanced image data. Step S202 includes the following specific steps: S2021. Construct a geometric correction model driven by optical flow field, and obtain the displacement field by minimizing the energy function. The expression for minimizing the energy function is: ， in, For the original image at position pixel values, Indicates passing through the displacement field The transformed image at position pixel values, Weights for smoothing terms; and Indicates the image in and Displacement in the direction, Let be the smoothing term of the displacement field, and let represent the square of the Frobenius norm of the displacement field gradient. The optimal displacement field is obtained by iteratively solving using the multi-scale Gauss-Newton method. And generate a geometrically corrected image, the original image at the location Corrected pixel values The expression is: ， in, Indicates passing through the optimal displacement field The transformed image at position Pixel values; S2022, In the corrected pixel values Image features formed on Point features are obtained by combining point-side cues or projections. A feature mapping function is constructed to improve the signal-to-noise ratio in the crack region, and an enhanced response map is output. Enhanced response map The expression for the feature mapping function is: ， in, Indicates the activation function; and The weight matrix of the feature map. and These are image features and point features.

4. The method for quantitative characterization of the spatial morphology of bridge defects by fusing three-dimensional laser point clouds and two-dimensional images according to claim 1, characterized in that: Step S3 includes the following specific steps: S301. Construct a multi-scale spatial attention network model, PSAG-Net. The multi-scale spatial attention network includes a point cloud hierarchical feature extraction module and a semantic segmentation module. The point cloud hierarchical feature extraction module includes a sampling grouping module and a spatial attention module. The input data of the point cloud hierarchical feature extraction module is a point cloud dataset. The point cloud data is input into the sampling grouping module for upsampling and grouping operations to capture the first layer of geometric features. The first layer of geometric features are input into the spatial attention module for fusion processing to obtain the first layer of optimized features; The first-layer optimized feature input sampling and grouping module performs upsampling and grouping operations to capture the second-layer geometric features; The second-layer geometric features are input into the spatial attention module for fusion processing to obtain the second-layer optimized features; The semantic segmentation module includes an upsampling layer, a grouping layer, and a spatial attention module. The optimized features of the second layer are processed by the upsampling layer and then input together with the geometric features of the second layer into the grouping layer for grouping processing to obtain a local fusion feature tensor. This tensor is then input into the spatial attention module for processing to obtain attention-enhanced semantic context features. The semantic context features and the geometric features of the first layer are input together into the grouping layer for grouping processing, and the segmented point cloud and point-level classification results are output. The spatial attention module includes convolutional layers, global pooling layers, fully connected (FC) layers, ReLU activation function, sigmoid activation function, concatenation layers, and normalization layers. In the feature input space attention module, multi-level convolution transformation is first performed to obtain feature mapping, then global pooling and FC fully connected layer are used to extract global description, and attention weights are generated by combining ReLU activation function, FC fully connected layer and Sigmoid function. The features are then weighted and fused point by point. After fusion, the features are concatenated and normalized to output the features after attention weighting and multi-scale fusion. S302. Construct a lightweight feature pyramid CM-FPN model. The lightweight feature pyramid includes a feature extraction module, a feature fusion module, and an output module. The feature extraction module includes a convolutional layer, a first C3k2 module, a second C3k2 module, a spatial pyramid pooling layer, and a C2PSA spatial attention module. The image dataset is input into the feature extraction module and sequentially passes through two levels of convolutional layers, the first C3k2 module, a convolutional layer, and the second C3k2 module to extract the first feature. The first feature is then extracted through a convolutional layer and the first C3k2 module to obtain the second feature. The second feature is then processed through a convolutional layer, the second C3k2 module, the spatial pyramid pooling layer, and the C2PSA spatial attention module to obtain the third feature. The first C3k2 module includes a convolutional layer, a segmentation layer, and a C3k2 module. Initial features are input into the first C3k2 module, and after processing by the convolutional layer, convolutional features are obtained. The convolutional features are processed by the segmentation layer to obtain segmentation features. The segmentation features are processed by two C3k2 modules and then input together with the convolutional features and segmentation features into two convolutional layers for processing to output the final features. The second C3k2 module includes convolutional layers, segmentation layers, and a bottleneck structure. Initial features are input into the second C3k2 module, processed by the convolutional layers to obtain convolutional features, which are then processed by the segmentation layers to obtain segmentation features. The segmentation features are processed by two bottleneck structures and then input together with the convolutional features and segmentation features into two convolutional layers for further processing to output the final features. The bottleneck structure includes two convolutional layers. Initial features are input into the bottleneck structure, processed by the two convolutional layers, and then summed with the initial features to obtain the output features. The feature fusion module includes an upsampling layer, a stitching layer, a C3k2 module, and a convolutional layer. The third and second features, after passing through the upsampling layer, are input together with the first feature into the stitching layer. After stitching, they are sequentially input into the C3k2 module to form the first fused feature. The first fused feature, after passing through the upsampling layer, is stitched together with the first feature to form the first stitched feature. The first fused feature, after passing through the stitching layer and the convolutional layer, is input together with the first stitched feature into the C3k2 module to form the second fused feature. The first fused feature, after passing through the stitching layer and the C3k2 module, forms the third fused feature. The third fused feature, after passing through the convolutional layer, is stitched together with the second feature to form the second stitched feature. The second fused feature, the second stitched feature, and the third fused feature are input together into the output module for segmentation, generating a crack segmentation mask. S303. Train the PSAG-Net model using the point cloud dataset of the cracks, and train the CM-FPN model using the image dataset of the cracks. After training, evaluate the models and save the best PSAG-Net and CM-FPN models.

5. The method for quantitative characterization of the spatial morphology of bridge defects by fusing three-dimensional laser point clouds and two-dimensional images according to claim 1, characterized in that: Step S5 includes the following specific steps: S501. For large-scale crack point clouds with complete segmentation results, use a registration algorithm to calculate the rotation matrix R and eigenvector t, and then transfer the target point cloud... Compared with the original point cloud Perform feature alignment to obtain complete geometric information of the crack; S502. Calculate the depth value of each crack point, statistically analyze the depth distribution of the crack point cloud, and obtain the mean, maximum and distribution histogram of the depth to form the overall depth distribution characteristics of the crack. Calculate the depth of the crack point using the formula for the distance from a point to a plane. For any crack point... Its fit plane The expression for the distance d is: ， in, This is the distance from the point to the plane, i.e., the depth value.

6. The method for quantitative characterization of the spatial morphology of bridge defects by fusing three-dimensional laser point clouds and two-dimensional images according to claim 5, characterized in that: Step S501 includes the following specific steps: S5011, using RANSAC in the target point cloud A plane is fitted over the crack neighborhood, and the fitted plane is: The equivalent representation is: ， in, For target point cloud The coefficients of the plane equation, Represents the coordinates of point i; The fitted plane serves as the reference crack plane, thus establishing a unified reference surface for subsequent alignment and depth measurement. S5012, via target point cloud The plane equation coefficients are used to calculate the target point cloud. Normal vector of the fitted plane In the original point cloud The original point cloud is obtained by fitting the plane. The plane equation coefficients are used to calculate the original point cloud. normal vector To ensure that the normals of the two clouds are aligned, a rotation axis is constructed: ， ， in, Let be the rotation axis vector. Let be the rotation angle; and give the rotation matrix using Rodriguez's formula, the expression is: ， Obtain the target point cloud Rotate the normal direction to the original point cloud Consistent minimum angular rotation R eliminates the attitude difference between the two clouds, ensuring that subsequent translation and depth calculations are performed in the same normal coordinate system; S5013, Calculate the centroid of two clouds and Calculate the translation vector based on the centroid difference: ， Obtain the translation vector for rigid registration Under the premise of consistent normals, position alignment is completed so that the two clouds coincide in the same global coordinate system; S5014, Finally, based on the translation vector and rotation matrix Align the original point cloud and target point cloud : ， Obtain the target point cloud Aligned point set The aligned crack point cloud set is obtained to provide consistent coordinates and a unified reference surface for subsequent depth estimation and statistical calculation.

7. The method for quantitative characterization of the spatial morphology of bridge defects by fusing three-dimensional laser point clouds and two-dimensional images according to claim 1, characterized in that: Step S6 includes the following specific steps: S601, through a 3D-2D cross-modal fusion algorithm, integrates point clouds... Mapped to a structured depth map The coordinates of the third dimension are mapped to the pixel value range of 0-255, so that the depth map contains the third dimension information, and the depth map of the crack is binarized. S602. Use the CM-FPN model to perform pixel-level prediction on the binarized depth map to obtain the crack pixel set, and then back-project it back to the three-dimensional coordinate system to obtain the crack point cloud set. S603. Normalize the depth value corresponding to the pixel, calculate the z-coordinate of the point cloud, and reconstruct the three-dimensional coordinates of the crack point cloud. S604. Calculate the depth corresponding to each pixel point based on the three-dimensional coordinates, and collect the depth information of the micro-cracks.

8. The method for quantitative characterization of the spatial morphology of bridge defects by fusing three-dimensional laser point clouds and two-dimensional images according to claim 7, characterized in that: Step S601 includes the following specific steps: S6011. First, the point cloud data is corrected according to the method in step S501 to complete the spatial transformation of the coordinate system. After correction, the Y-axis in the point cloud coordinate system is used to represent the depth direction, while the X-axis and Z-axis correspond to the plane where the crack is located. S6012. Then, normalize the Y coordinate values ​​of the point cloud and map them between 0 and 255 pixel values ​​to map the point cloud into a two-dimensional depth map. ， in, The raw point cloud data, For depth map, Represents depth information, This represents a combination of depth map and point cloud data; S6013. Subsequently, threshold segmentation is performed to obtain a binarized depth map: ， in, This represents a binarized depth map, retaining pixel values ​​greater than or equal to a threshold. Pixels.

9. The method for quantitative characterization of the spatial morphology of bridge defects by fusing three-dimensional laser point clouds and two-dimensional images according to claim 7, characterized in that: Step S602 includes the following specific steps: S6021. Use the CM-FPN model to perform crack detection and segmentation on the depth map. The expression is: ， in, This is the set of pixels after the crack is segmented. This indicates that the CM-FPN model is used for crack detection and segmentation on the binary depth map; S6022, Set the segmented crack pixels Back-projecting back to the 3D coordinate system yields the point cloud set of the crack. Point cloud set of cracks The k-th point cloud coordinates for: ， Where K represents the point cloud set The total number of midpoints These are the coordinates of the pixels in the segmented crack.

10. The method for quantitative characterization of the spatial morphology of bridge defects by fusing three-dimensional laser point clouds and two-dimensional images according to claim 7, characterized in that: Step S604 includes the following specific steps: According to the fitted plane For crack point clouds Each point in Calculate its orthogonal distance to the plane: ， in, This is the distance from the point to the line, i.e., the depth value; Calculate the depth value of each point Describe the distribution; ， in, This represents the set of depth values ​​for each point obtained from the final statistics. Represents a set of point clouds A function to calculate the depth value for each point in the array.

Images