A method and apparatus for landslide crack identification based on neighborhood-wide perception hierarchical convolution.

Abstract

This invention discloses a method and apparatus for identifying landslide cracks based on neighborhood-wide perceptual hierarchical convolution. First, the invention calculates the deep spatial features of each point based on the 3D data of the landslide point cloud. Then, these features and the point cloud data are used as input to the network. Next, the network is used to find the descriptive points with the best detailed structural representation within the neighborhood of the landslide crack and divides them into sub-neighborhoods. Then, by using the sub-neighborhoods corresponding to the determined descriptive points and combining them with the spatial features of the point cloud data in each sub-neighborhood, the fine local features of the landslide cracks are extracted. Finally, based on the extracted fine local features, the landslide cracks can be identified. Therefore, this invention, through spatial features and sub-neighborhood division, extracts local features, focusing on the overall structural morphology of the landslide cracks. This reduces misjudgments and omissions caused by complex terrain variations, thereby improving the accuracy of landslide crack identification.

Description

Technical Field

[0001] This invention belongs to the field of landslide crack recognition technology, specifically relating to a landslide crack recognition method and device based on neighborhood full-perception hierarchical convolution. Background Technology

[0002] Due to the complexity of landslide surface structures and the indistinctness of boundary contours, landslide identification based on point cloud data is difficult to achieve directly through surface morphology. Therefore, micro-topography analysis has become an effective way to improve the accuracy of landslide identification. However, current micro-topography detection methods based on point cloud data have the following shortcomings: the feature descriptions and expressions in existing micro-topography detection methods lack specificity for micro-topography, resulting in insufficient extraction of fine local features of micro-topography, thus leading to inaccurate identification of landslide cracks. Therefore, how to provide an identification method that can accurately identify landslide cracks has become an urgent problem to be solved. Summary of the Invention

[0003] The purpose of this invention is to provide a method and apparatus for identifying landslide cracks based on neighborhood-wide perception hierarchical convolution, in order to solve the problem of inaccurate landslide crack identification in the prior art.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] Firstly, a landslide crack identification method based on neighborhood-wide perception hierarchical convolution is provided, including:

[0006] Obtain a point cloud dataset of the target landslide area, and based on the point cloud dataset, determine the spatial features of each point cloud data in the point cloud dataset, wherein the spatial features of any point cloud data include the slope aspect and slope corresponding to that point cloud data.

[0007] A landslide crack identification model is obtained, wherein the landslide crack identification model is trained by taking the sample point cloud dataset of multiple sample landslide areas and the sample spatial features of each sample point cloud data in the corresponding sample point cloud dataset as input, and the crack identification results of each sample landslide area as output. The landslide crack identification model is a trained neighborhood full-perception hierarchical convolutional neural network model. During training, after the sample point cloud dataset of any sample landslide area and the sample spatial features of each corresponding sample point cloud data are input into the neighborhood full-perception hierarchical convolutional neural network model, the neighborhood full-perception hierarchical convolutional neural network model determines the crack center point of any sample landslide area, the sub-neighborhood points of each crack center point, and the sub-neighborhood point cloud set corresponding to each sub-neighborhood point from the sample point cloud dataset corresponding to the sample landslide area. Based on the sample spatial features of each sub-neighborhood point cloud data in the sub-neighborhood point cloud set corresponding to each sub-neighborhood point of each crack center point, the local features of each crack center point are extracted, and the crack identification result of any sample landslide area is obtained based on the local features.

[0008] The point cloud dataset of the target landslide area and the spatial features of each point cloud data in the point cloud dataset are input into the landslide crack identification model to obtain the crack identification result of the target landslide area.

[0009] Based on the above-disclosed content, this invention pre-trains a landslide crack identification model. This model is trained using sample point cloud datasets of multiple sample landslide areas and the sample spatial features of each sample point cloud data point in the corresponding datasets as input, and the crack identification results of each sample landslide area as output. After the sample point cloud dataset and corresponding sample spatial features of any sample landslide area are input into the model, the crack center point, the center point's sub-neighborhood points, and the corresponding sub-neighborhood point clouds of each sub-neighborhood point are extracted. Thus, it is equivalent to finding the cracks within the neighborhood of the landslide crack. The model uses descriptive points with the best ability to represent detailed structures and divides them into sub-neighborhoods. Then, by using the sub-neighborhoods corresponding to the descriptive points and combining the sample spatial features of each point cloud data in the sub-neighborhoods, the fine local features of each center point can be extracted. Finally, the crack identification result can be obtained based on the extracted fine local features. Based on this, in practical applications, the spatial features of each point cloud data in the point cloud dataset of the target landslide area can be determined first. Then, the point cloud data and the corresponding spatial features can be input into the aforementioned trained model to obtain the crack identification result of the target landslide area.

[0010] Through the above design, this invention first calculates the deep spatial features of each point based on the three-dimensional data of landslide point clouds, and then uses these features and point cloud data as input to the network. Next, the network is used to find the descriptive points with the best ability to represent the detailed structure within the neighborhood of the landslide crack and divides them into sub-neighborhoods. Then, by using the sub-neighborhoods corresponding to the determined descriptive points and combining the spatial features of the point cloud data in each sub-neighborhood, the fine local features of the landslide cracks can be extracted. Finally, based on the extracted fine local features, the identification of landslide cracks can be completed. Thus, this invention uses spatial features and sub-neighborhood division to extract local features, which can focus on the overall structural morphology of landslide cracks. Therefore, it can reduce misjudgments and omissions caused by complex terrain changes, thereby improving the accuracy of landslide crack identification. Based on this, this invention is very suitable for large-scale application and promotion.

[0011] In one possible design, based on the point cloud dataset, the spatial characteristics of each point cloud data point in the point cloud dataset are determined, including:

[0012] The nearest neighbor search algorithm is used to determine the neighboring point cloud data of each point cloud data in the point cloud dataset;

[0013] Based on the neighboring point cloud data of each point cloud data, and using the singular value decomposition algorithm, the surface normal of each point cloud data is obtained;

[0014] The spatial characteristics of each point cloud data are determined based on the surface normal of each point cloud data.

[0015] In one possible design, the crack center point of any landslide area is determined from the sample point cloud dataset corresponding to any sample landslide area based on the neighborhood-wide perception hierarchical convolutional neural network model, including:

[0016] In the neighborhood full-perception hierarchical convolutional neural network model, the farthest distance sampling method is used to downsample the target point cloud dataset, and each sampling point obtained by downsampling is used as the crack center point of any sample landslide area, wherein the target point cloud dataset is the sample point cloud dataset corresponding to any sample landslide area.

[0017] Accordingly, based on the neighborhood-wide perception hierarchical convolutional neural network model, the sub-neighborhood points of each crack center point and the sub-neighborhood point cloud set corresponding to each sub-neighborhood point are determined from the sample point cloud dataset corresponding to any sample landslide area. This includes:

[0018] For any crack center point among several crack center points corresponding to any sample landslide area, a neighborhood point cloud set of any crack center point is determined from the target point cloud dataset;

[0019] Calculate the distance between each neighboring point cloud in the neighborhood point cloud set and any crack center point, and take the neighboring point cloud with the smallest distance as the i-th child neighbor point of any crack center point, where the initial value of i is 1;

[0020] The first target neighborhood point set is obtained by removing the i-th sub-neighbor point from the neighborhood point cloud set;

[0021] Calculate the distance between each first target neighborhood point cloud in the first target neighborhood point cloud set and the i-th sub-neighbor point, and sort each first target neighborhood point cloud in ascending order of distance to obtain a sorted point cloud sequence;

[0022] Using the first J first target neighborhood point clouds in the sorted point cloud sequence, a sub-neighbor point cloud set of the i-th sub-neighbor point is formed, where J is a positive integer greater than 1;

[0023] The second target neighborhood point set is obtained by removing the sub-neighbor point cloud data of the i-th sub-neighbor point from the first target neighborhood point cloud set.

[0024] The neighborhood point cloud is updated to the second target neighborhood point cloud and i is incremented by 1. The distance between each neighborhood point cloud in the neighborhood point cloud and any crack center point is recalculated until all neighborhood point clouds in the neighborhood point cloud are divided, so as to obtain the sub-neighbor points corresponding to any crack center point and the sub-neighbor point cloud corresponding to each sub-neighbor point.

[0025] In one possible design, any crack center point corresponds to at least one sub-neighborhood point. Based on the sample spatial features of the sub-neighborhood point cloud data corresponding to the sub-neighborhood points of each crack center point, local features of each crack center point are extracted, including:

[0026] For any crack center point in any sample landslide area, the spatial features of each sub-neighbor point cloud data in the sub-neighbor point cloud set corresponding to each sub-neighbor point of the crack center point are convolved using the neighborhood full-perception hierarchical convolutional neural network model to obtain the local features of each sub-neighbor point corresponding to the crack center point.

[0027] The local features of each sub-neighborhood point corresponding to any crack center point are updated to obtain the updated local features of each sub-neighborhood point.

[0028] Based on the neighborhood-wide perception hierarchical convolutional neural network model, the local features of each updated sub-neighborhood point are aggregated to obtain the local features corresponding to any crack center point.

[0029] In one possible design, the neighborhood-wide perception hierarchical convolutional neural network model is used to perform convolution processing on the spatial features of each sub-neighbor point cloud data in the sub-neighbor point cloud set corresponding to each sub-neighbor point of any crack center point, to obtain the local features of each sub-neighbor point corresponding to any crack center point, including:

[0030] The local features of each sub-neighborhood point corresponding to any crack center point are extracted using the following formula (1);

[0031] f i =ReLU(conv j×1 (H ij ,f ij ),j=1,2,...,J (1)

[0032] In the above formula (1), f i conv represents the local feature of the i-th sub-neighborhood point corresponding to any crack center point. j×1 H represents a convolution operation with a kernel size of j×1. ij f represents the convolution kernel weight of the j-th sub-neighbor point cloud data in the sub-neighbor point cloud set corresponding to the i-th sub-neighbor point. ij Let J represent the spatial features of the j-th sub-neighbor point cloud data in the sub-neighbor point cloud set corresponding to the i-th sub-neighbor point, J represent the total number of sub-neighbor point cloud data in the sub-neighbor point cloud set corresponding to the i-th sub-neighbor point, and ReLU represent the activation function, where i = 1, 2, 3, ..., n, and n represents the total number of sub-neighbor points corresponding to any crack center point.

[0033] In one possible design, the local features of each sub-neighborhood point corresponding to any crack center point are updated to obtain the updated local features of each sub-neighborhood point, including:

[0034] The following formula (2) is used to update the local features of each sub-neighborhood point corresponding to any crack center point, so as to obtain the updated local features of each sub-neighborhood point.

[0035]

[0036] In the above formula (2), f i f represents the local feature of the i-th sub-neighborhood point corresponding to any crack center point. i' represents the updated local feature of the i-th sub-neighborhood point corresponding to any crack center point, x i Let pi represent the i-th sub-neighborhood point, p0 represent the center point of any crack, and dist(x) represent the i-th sub-neighborhood point. i (p0) represents the distance between the i-th sub-neighborhood point and the center point of any of the cracks, W i Let X represent the weight matrix of the i-th sub-neighborhood point, X represent the set of all sub-neighborhood points corresponding to any crack center point, and n represent the total number of sub-neighborhood points corresponding to any crack center point.

[0037] In one possible design, based on the neighborhood-wide perceptual hierarchical convolutional neural network model, the local features of each updated sub-neighborhood point are aggregated to obtain the local features corresponding to any crack center point, including:

[0038] The local features of each updated sub-neighborhood point are aggregated using the following formula (3) to obtain the local features corresponding to any crack center point.

[0039] f0 = ReLU(conv) i×1 (H i ,f i ′)), i=1,2,3,...,n (3)

[0040] In the above formula (3), f0 represents the local feature corresponding to any crack center point, f i ' represents the updated local feature of the i-th sub-neighborhood point corresponding to any crack center point, conv i×1 ReLU represents a convolution operation with a kernel size of i×1, H represents the activation function. i represents the convolution kernel weight of the i-th sub-neighborhood point, and n represents the total number of sub-neighborhood points corresponding to any crack center point.

[0041] Secondly, a landslide crack recognition device based on neighborhood-wide perception hierarchical convolution is provided, comprising:

[0042] The data acquisition unit is used to acquire a point cloud dataset of the target landslide area and, based on the point cloud dataset, determine the spatial characteristics of each point cloud data in the point cloud dataset, wherein the spatial characteristics of any point cloud data include the slope aspect and slope corresponding to that point cloud data.

[0043] A model acquisition unit is used to acquire a landslide crack identification model. This model is trained using a sample point cloud dataset of multiple sample landslide areas and the sample spatial features of each sample point cloud data point in the corresponding dataset as input, with the crack identification results of each sample landslide area as output. The landslide crack identification model is a trained neighborhood-wide perception hierarchical convolutional neural network model. During training, the sample point cloud dataset of any sample landslide area and the sample spatial features of each corresponding sample point cloud data are input into the neighborhood-wide perception hierarchical convolutional neural network model. Based on this model, the crack center point, the sub-neighborhood points of each crack center point, and the sub-neighborhood point cloud set corresponding to each sub-neighborhood point are determined from the sample point cloud dataset corresponding to the landslide area. Furthermore, based on the sample spatial features of each sub-neighborhood point cloud set corresponding to the sub-neighborhood points of each crack center point, local features of each crack center point are extracted. Based on these local features, the crack identification result of the landslide area is obtained.

[0044] The crack identification unit is used to input the point cloud dataset of the target landslide area and the spatial features of each point cloud data in the point cloud dataset into the landslide crack identification model to obtain the crack identification result of the target landslide area.

[0045] Thirdly, another landslide crack identification device based on neighborhood-wide perception hierarchical convolution is provided. Taking the device as an electronic device as an example, it includes a memory, a processor, and a transceiver that are connected in sequence. The memory is used to store a computer program, the transceiver is used to send and receive messages, and the processor is used to read the computer program and execute the landslide crack identification method based on neighborhood-wide perception hierarchical convolution as described in the first aspect or any possible design in the first aspect.

[0046] Fourthly, a storage medium is provided, on which instructions are stored, which, when executed on a computer, perform the landslide crack identification method based on neighborhood full-perception hierarchical convolution as described in the first aspect or any possible design of the first aspect.

[0047] Fifthly, a computer program product containing instructions is provided, which, when executed on a computer, causes the computer to perform the landslide crack identification method based on neighborhood full-perception hierarchical convolution as described in the first aspect or any possible design of the first aspect.

[0048] Beneficial effects:

[0049] (1) This invention first calculates the deep spatial features of each point based on the three-dimensional data of landslide point cloud, and then uses these features and point cloud data as input to the network. Next, the network is used to find the descriptive points with the best detailed structural representation ability within the neighborhood of the landslide crack and divides them into sub-neighborhoods. Then, the fine local features of the landslide crack can be extracted by using the sub-neighborhoods corresponding to the determined descriptive points and combining the spatial features of each point cloud data in the sub-neighborhoods. Finally, the landslide crack identification can be completed based on the extracted fine local features. Thus, this invention extracts local features through spatial features and sub-neighborhood division, which can focus on the overall structural morphology of the landslide crack. Therefore, it can reduce misjudgments and omissions caused by complex changes in terrain, thereby improving the accuracy of landslide crack identification. Based on this, this invention is very suitable for large-scale application and promotion. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the landslide crack identification method based on neighborhood full-perception hierarchical convolution provided in an embodiment of the present invention.

[0051] Figure 2 This is a schematic diagram of the structure of the neighborhood-wide perceptual hierarchical convolutional neural network model provided in an embodiment of the present invention;

[0052] Figure 3 A schematic diagram illustrating the processing flow of the neighborhood-wide perceptual hierarchical convolutional neural network model provided in an embodiment of the present invention;

[0053] Figure 4 This is a schematic diagram illustrating the effect of landslide crack identification using the method provided in this embodiment of the invention.

[0054] Figure 5 ROC curve for landslide crack identification using the method provided in this embodiment of the invention;

[0055] Figure 6 This is a schematic diagram of the structure of the project result analysis device provided in an embodiment of the present invention;

[0056] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be briefly introduced below in conjunction with the accompanying drawings and descriptions of the embodiments or the prior art. Obviously, the following description of the structure of the accompanying drawings is only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0058] It should be understood that although the terms first, second, etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are only used to distinguish one unit from another. For example, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit, without departing from the scope of the exemplary embodiments of the invention.

[0059] It should be understood that the term "and / or" that may appear in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" that may appear in this document describes another relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " that may appear in this document generally indicates that the related objects before and after it are in an "or" relationship.

[0060] Example:

[0061] See Figure 1 As shown, the landslide crack identification method based on neighborhood full-perception hierarchical convolution provided in this embodiment extracts local features by utilizing the spatial characteristics of each point cloud data of the landslide crack and combining it with a sub-neighborhood partitioning method. Compared with traditional techniques, it can focus on the overall structural morphology of the landslide crack. Therefore, it can reduce misjudgments and omissions caused by complex terrain changes, thereby improving the accuracy of landslide crack identification. Based on this, this method is very suitable for large-scale application and promotion in the field of landslide crack identification. For example, this method can be run on the landslide crack identification end. Optionally, the landslide crack identification end can be, but is not limited to, a personal computer (PC). It is understood that the aforementioned execution subject does not constitute a limitation on the embodiments of this application. Accordingly, the operation steps of this method can be, but are not limited to, the steps S1 to S3 below.

[0062] S1. Obtain the point cloud dataset of the target landslide area, and based on the point cloud dataset, determine the spatial features of each point cloud data in the point cloud dataset. The spatial features of any point cloud data include the slope aspect and slope corresponding to that point cloud data. In this embodiment, the spatial features of each point cloud data can be determined by calculating the surface normal of each point cloud data. Specifically, the nearest neighbor search algorithm is first used to determine the neighboring point cloud data of each point cloud data in the point cloud dataset. Then, based on the neighboring point cloud data of each point cloud data, the singular value decomposition algorithm is used to obtain the surface normal of each point cloud data. Finally, the spatial features of each point cloud data can be determined according to the surface normal of each point cloud data. Of course, using neighboring point clouds and singular value decomposition to determine the surface normal of point cloud data, and obtaining spatial features such as slope and slope aspect based on the surface normal, are common techniques for extracting spatial features of point cloud data. The principle of these techniques will not be elaborated here.

[0063] By using the aforementioned design and incorporating the spatial features as additional inputs into the network, the learning direction of the network model can be controlled, enabling the network to learn the overall structural morphology of landslide cracks. Therefore, the accuracy of model recognition can be improved. Furthermore, the model processing procedure is described in detail in step S2 below.

[0064] After extracting the spatial features of each point cloud data corresponding to the target landslide area, a pre-trained landslide crack identification model can be obtained, so as to identify cracks in the target landslide area based on the landslide crack identification model; wherein, the model acquisition process can be, but is not limited to, as shown in step S2 below.

[0065] S2. Obtain a landslide crack identification model, wherein the landslide crack identification model is trained by taking the sample point cloud dataset of multiple sample landslide areas and the sample spatial features of each sample point cloud data in the corresponding sample point cloud dataset as input, and the crack identification results of each sample landslide area as output; in this embodiment, the landslide crack identification model may be, but is not limited to, a trained neighborhood full-perception hierarchical convolutional neural network model.

[0066] Meanwhile, the following describes the processing of the sample point cloud dataset of any sample landslide area and the corresponding sample spatial features of each sample point cloud data in the aforementioned model.

[0067] During training, the sample point cloud dataset of any landslide area and the sample spatial features of each corresponding sample point cloud dataset are input into the neighborhood fully perceptual hierarchical convolutional neural network model. First, based on the neighborhood fully perceptual hierarchical convolutional neural network model, the crack center point, the sub-neighborhood points of each crack center point, and the sub-neighborhood point cloud corresponding to each sub-neighborhood point are determined from the sample point cloud dataset corresponding to each landslide area. Then, after dividing the sub-neighborhoods of each crack center point into sub-neighborhoods, the local features of each crack center point are extracted based on the sample spatial features of each sub-neighborhood point cloud dataset corresponding to each crack center point. Finally, the crack identification result of any landslide area can be obtained based on these local features.

[0068] Optionally, for example but not limited to, in the neighborhood-wide perceptual hierarchical convolutional neural network model, the furthest distance sampling method can be used to downsample the target point cloud dataset (the target point cloud dataset is the sample point cloud dataset corresponding to any sample landslide area), and each sampled point obtained by downsampling can be used as the crack center point of any sample landslide area; in this way, based on the aforementioned method, the crack center point of any sample landslide area can be determined; then, based on the crack center point, the descriptive points with the best detailed structure representation ability can be found (i.e., the sub-neighborhood points of each crack center point), and the division of their sub-neighborhood sets can be performed; wherein, taking any crack center point as an example, the aforementioned process can be specifically illustrated as shown in steps A to G below.

[0069] A. For any crack center point among several crack center points corresponding to any sample landslide area, determine the neighborhood point cloud of any crack center point from the target point cloud dataset; in specific implementation, for example, but not limited to, the KNN nearest neighbor algorithm can be used to determine the neighborhood point cloud of any crack center point from the target point cloud dataset; of course, the KNN nearest neighbor algorithm is a commonly used method of neighborhood search, and its principle will not be elaborated here.

[0070] After obtaining the neighborhood point cloud of any of the aforementioned crack center points, the sub-neighborhood description points and corresponding sub-neighborhood point clouds of any crack center point can be extracted, as shown in steps B to G below.

[0071] B. Calculate the distance between each neighboring point cloud in the neighborhood point cloud set and any crack center point, and take the neighboring point cloud with the smallest distance as the i-th sub-neighbor point of any crack center point, where the initial value of i is 1; In this embodiment, the distance between each neighboring point cloud and any crack center point can be calculated using the three-dimensional coordinates of each neighboring point cloud and the three-dimensional coordinates of any crack center point; Then, the neighboring point cloud with the smallest distance can be taken as the i-th sub-neighbor point of any crack center point (i.e., as a sub-neighbor description point); After the extraction of the i-th sub-neighbor point is completed, the i-th sub-neighbor point can be deleted from the original point cloud set (i.e., the neighboring point cloud set in step A above); Then, the distance between the point clouds is used to determine the sub-neighbor point cloud set of the i-th sub-neighbor point; Optionally, the extraction process of the sub-neighbor point cloud set of the i-th sub-neighbor point can be, but is not limited to, as shown in steps C to E below.

[0072] C. Remove the i-th sub-neighborhood point from the neighborhood point cloud to obtain the first target neighborhood point cloud.

[0073] D. Calculate the distance between each first target neighborhood point cloud in the first target neighborhood point cloud set and the i-th sub-neighbor point, and sort each first target neighborhood point cloud in ascending order of distance to obtain a sorted point cloud sequence.

[0074] E. Using the first J first target neighborhood point clouds in the sorted point cloud sequence, form the sub-neighbor point cloud set of the i-th sub-neighbor point, where J is a positive integer greater than 1.

[0075] Thus, through the aforementioned steps C to E, it is equivalent to selecting the J first target neighborhood point clouds that are closest to the i-th sub-neighbor point from the first target neighborhood point cloud set; then, the selected J first target neighborhood point clouds can be used to form the sub-neighbor point cloud set of the i-th sub-neighbor point; optionally, for example, J can be, but is not limited to, 3. Of course, the value of J can be specifically set according to actual use, and is not limited to the aforementioned example.

[0076] After extracting the first sub-neighbor point and its corresponding sub-neighbor point cloud, the point cloud data of the i-th sub-neighbor point and its corresponding sub-neighbor point cloud can be deleted from the original point cloud. Then, the above steps are repeated, and this process is repeated until all point cloud data in the original point cloud is divided. At this point, the sub-neighbor points corresponding to any crack center point and the sub-neighbor point cloud corresponding to each sub-neighbor point can be obtained. The process of updating the original point cloud data and extracting the sub-neighbor points in a loop can be, but is not limited to, the steps F and G below.

[0077] F. Remove the sub-neighbor point cloud data of the i-th sub-neighbor point from the first target neighborhood point cloud set to obtain the second target neighborhood point cloud set.

[0078] G. Update the neighborhood point cloud to the second target neighborhood point cloud and increment i by 1, and recalculate the distance between each neighborhood point cloud in the neighborhood point cloud and any crack center point, until all neighborhood point clouds in the neighborhood point cloud are divided, so as to obtain the sub-neighbor points corresponding to any crack center point and the sub-neighbor point cloud corresponding to each sub-neighbor point.

[0079] Therefore, through the aforementioned steps A to G, this embodiment sequentially searches for sub-neighborhood descriptor points based on their distance to the center point. Each time a sub-neighborhood descriptor point is determined, a k-NN search is used to divide the sub-neighborhood, and the already divided sub-neighborhood points do not participate in subsequent sub-neighborhood divisions. By repeatedly performing the above operations, the division of all sub-neighborhoods can be completed. Based on this, since the sub-neighborhood descriptor points come from the original point cloud data, they can more accurately describe the true detailed structural features of the point cloud.

[0080] After extracting the sub-neighborhood description points of each crack center point and dividing the corresponding sub-neighborhood point cloud based on the aforementioned steps A to G, the spatial features of each point cloud in the sub-neighborhood point cloud can be used to extract the local features of each crack center point. Here, we take any crack center point as an example to illustrate the process of extracting local features, which can be, but is not limited to, the steps shown below.

[0081] Step 1: For any crack center point in any sample landslide area, the spatial features of each sub-neighbor point cloud data in the sub-neighbor point cloud set corresponding to each sub-neighbor point of the crack center point are convolved using the neighborhood full-perception hierarchical convolutional neural network model to obtain the local features of each sub-neighbor point corresponding to the crack center point. In specific implementation, for example, but not limited to, the following formula (1) can be used to extract the local features of each sub-neighbor point corresponding to the crack center point.

[0082] f i =ReLU(conv j×1 (H ij ,f ij ),j=1,2,...,J (1)

[0083] In the above formula (1), f i conv represents the local feature of the i-th sub-neighborhood point corresponding to any crack center point. j×1 H represents a convolution operation with a kernel size of j×1. ijf represents the convolution kernel weight of the j-th sub-neighbor point cloud data in the sub-neighbor point cloud set corresponding to the i-th sub-neighbor point. ij Let J represent the spatial features of the j-th sub-neighbor point cloud data in the sub-neighbor point cloud set corresponding to the i-th sub-neighbor point, J represent the total number of sub-neighbor point cloud data in the sub-neighbor point cloud set corresponding to the i-th sub-neighbor point, and ReLU represent the activation function, where i = 1, 2, 3, ..., n, and n represents the total number of sub-neighbor points corresponding to any crack center point.

[0084] After calculating the local features of each sub-neighborhood point corresponding to any crack center point based on the aforementioned formula (1), it is also necessary to update the features based on the distance from each sub-neighborhood point corresponding to any crack center point to the crack center point. The feature update process is as shown in the second step below.

[0085] Step 2: Perform feature update processing on the local features of each sub-neighborhood point corresponding to any crack center point to obtain the local features of each updated sub-neighborhood point; In specific applications, sub-neighborhood descriptor points are used to describe the detailed structure of each sub-neighborhood and extract fine local features, and the way the sub-neighborhood descriptor points are determined makes them present a distribution characteristic from far to near the center point; Therefore, in order to balance the contribution of the sub-neighborhood descriptor point features to the center point, after the local features of the sub-neighborhood points are aggregated into the sub-neighborhood descriptor points, it is necessary to balance the weights according to the distance between the sub-neighborhood descriptor points and the center point, so that the weights of the sub-neighborhood points far from the center point are small and the weights of the sub-neighborhood points close to the center point are large; For example, but not limited to, the following formula (2) can be used to perform feature update processing on the local features of each sub-neighborhood point corresponding to any crack center point to obtain the local features of each updated sub-neighborhood point.

[0086]

[0087] In the above formula (2), f i f represents the local feature of the i-th sub-neighborhood point corresponding to any crack center point. i ' represents the updated local feature of the i-th sub-neighborhood point corresponding to any crack center point, x i Let pi represent the i-th sub-neighborhood point, p0 represent the center point of any crack, and dist(x) represent the i-th sub-neighborhood point. i (p0) represents the distance between the i-th sub-neighborhood point and the center point of any of the cracks, W i Let X represent the weight matrix of the i-th sub-neighborhood point, X represent the set of all sub-neighborhood points corresponding to any crack center point, and n represent the total number of sub-neighborhood points corresponding to any crack center point. In this embodiment, the weight matrix of each sub-neighborhood point can be set within the landslide crack identification end.

[0088] After updating the local features of each sub-neighborhood point based on the aforementioned formula (2), the updated local features of each sub-neighborhood point can be aggregated using the model to obtain the fine local features corresponding to any crack center point; wherein, the aforementioned aggregation process can be, but is not limited to, the third step below.

[0089] Step 3: Based on the neighborhood full-perception hierarchical convolutional neural network model, the local features of each updated sub-neighborhood point are aggregated to obtain the local features corresponding to any crack center point; in a specific implementation, for example, but not limited to, the following formula (3) can be used to obtain the local features corresponding to any crack center point.

[0090] f0 = ReLU(conv) i×1 (H i ,f i ′)), i=1,2,3,...,n (3)

[0091] In the above formula (3), f0 represents the local feature corresponding to any crack center point, f i ' represents the updated local feature of the i-th sub-neighborhood point corresponding to any crack center point, conv i×1 ReLU represents a convolution operation with a kernel size of i×1, H represents the activation function. i represents the convolution kernel weight of the i-th sub-neighborhood point, and n represents the total number of sub-neighborhood points corresponding to any crack center point.

[0092] Based on the aforementioned steps describing the extraction process of local features at the center point of any crack, this embodiment, after grouping the sub-neighborhoods, aggregates the fine local features of the sub-neighborhood point set onto the sub-neighborhood descriptor point through convolution. Then, after updating the features of the sub-neighborhood descriptor point, the high-dimensional features of the sub-neighborhood descriptor point are aggregated onto the center point. In this way, fine feature extraction of the neighborhood can be achieved. Since the neighborhood is divided into multiple sub-neighborhoods, feature extraction on each sub-neighborhood can achieve a sufficiently fine granularity, thereby improving the precision of the extracted local features. At the same time, using convolution kernels of size i×1 and j×1 for feature extraction, compared with the traditional 1×1 convolution kernel, this embodiment can extract the contextual information between adjacent points, further improving the ability to extract fine local features of the neighborhood.

[0093] Thus, by using the aforementioned local feature extraction steps, the local features of the center point of each crack can be obtained. Then, by using the extracted local features, the crack identification of any sample landslide area can be completed. In this way, the model is continuously trained using the aforementioned method, and the model parameters are adjusted in reverse using the model's loss function value until the model converges, thus obtaining the landslide crack identification model.

[0094] Meanwhile, one of the network structures of the publicly disclosed neighborhood-wide perceptual hierarchical convolutional neural network model can be found in [link to relevant documentation]. Figure 2 As shown, it may include, but is not limited to, a feature extraction network, a first upsampling layer, a first feature concatenation layer, a one-dimensional convolutional structure layer 1, a second upsampling layer, a second feature concatenation layer, a one-dimensional convolutional structure layer 2, a third upsampling layer, a third feature concatenation layer, a one-dimensional convolutional structure layer 3, and an output layer. Meanwhile, the feature extraction network may include, but is not limited to, a first downsampling layer, a two-dimensional convolutional structure layer 1, a first feature update layer, a two-dimensional convolutional structure layer 2, a second downsampling layer, a two-dimensional convolutional structure layer 3, a second feature update layer, a two-dimensional convolutional structure layer 4, and a max pooling layer. The connection relationships between the aforementioned layers are described in [reference needed]. Figure 2 As shown, it will not be elaborated further here.

[0095] Furthermore, this embodiment, in conjunction with the network, discloses the aforementioned training process, the flowchart of which can be found here. Figure 3 As shown.

[0096] First, the deep spatial features (including the corresponding slope aspect and slope) calculated based on the 3D data of any sample landslide area (i.e., the corresponding sample point cloud dataset) and the 3D information of the point cloud data are used as input to the network, in the format of a matrix of size (B, N, C). Here, B is the number of point cloud samples input to the network in each batch, i.e., the batch size; N is the number of points in each point cloud sample. Before the samples are fed into the network, the number of points in each sample is unified for subsequent calculations; C is the number of features of each point in the point cloud sample, i.e., the number of 3D information features of the point cloud data plus the number of deep spatial features.

[0097] After input into the network, the network first determines the center points through downsampling, and then determines the sub-neighborhood descriptor points (i.e., sub-neighborhood points) and divides the network into sub-neighborhoods (i.e., the set of sub-neighborhood points corresponding to the sub-neighborhood points) based on the center points, resulting in a matrix of size (B, N1, I1, J1, C), where N1 is the number of center points after downsampling, I1 is the number of sub-neighborhood descriptor points, and J1 is the number of points in the sub-neighborhood point set. Then, a two-dimensional convolution operation is performed on the matrix using a convolution kernel of size J1×1, resulting in a matrix of size (B, N1, I1, 64), where 64 indicates that the original C-dimensional features are increased to 64 dimensions through convolution. This resulting matrix is ​​the feature matrix of the sub-neighborhood descriptor points. Next, this feature matrix is ​​updated to obtain a new feature matrix of size (B, N1, I1, C). 1, 64), and then perform a two-dimensional convolution operation on the matrix using a convolution kernel of size I1×1 to obtain a matrix of size (B, N1, 128). The matrix obtained at this time is the feature matrix of the center point.

[0098] Furthermore, a new matrix of size (B, N², I², J², 128) is obtained through downsampling and sub-neighborhood partitioning. A two-dimensional convolution operation is then performed on this matrix using a J²×1 convolution kernel to obtain another matrix of size (B, N², I², 128). After another feature update of the sub-neighborhood description points, a new matrix of size (B, N², I², 128) is obtained. Simultaneously, a two-dimensional convolution operation is performed on this matrix using an I²×1 convolution kernel to obtain a matrix of size (B, N², 256). This resulting matrix represents the feature matrix of the center point abstracted to a high dimension after the second downsampling. Then, a two-dimensional convolution operation is performed on this matrix using a 1×1 convolution kernel to obtain the feature matrix (B, N², 1024) of the entire sample. Finally, max pooling is performed to obtain the global feature matrix (B, 1, 1024) of the entire sample.

[0099] Then, the global features are copied N2 times and concatenated with the high-dimensional center point feature matrix (B, N2, 256) extracted after the second downsampling, resulting in a feature matrix (B, N2, 1024+256) containing both global features and local center point features after the second sampling. Next, a one-dimensional convolution is performed on the matrix using a 1×1 convolution kernel to obtain the feature matrix (B, N2, 256). Then, the number of center points N2 is restored to N1 by inverse distance interpolation upsampling, and concatenated with the high-dimensional center point feature matrix (B, N1, 128) extracted after the first downsampling, resulting in a feature matrix (B, N1, 256+128) containing both global features and local center point features after the first sampling.

[0100] Next, a one-dimensional convolution is performed on the matrix using a 1×1 kernel to obtain the feature matrix (B, N1, 128). Simultaneously, inverse distance interpolation upsampling is used to restore the number of center points N1 to N, and the features of each point in the initial input sample are concatenated to obtain the feature matrix (B, N, 128+C). After three more one-dimensional convolutions with 1×1 kernels, the feature matrix (B, N, 128) is obtained. Finally, a one-dimensional convolution operation with a 1×1 kernel is performed on the feature matrix, outputting a result with a feature dimension of 2 (B, N, 2). The significance of this result is that each batch of the output contains B samples, each sample contains N points, and each point has two recognition results (is it a crack or not a crack). Furthermore, after each convolution operation in the network, normalization is performed using a BN layer and activation using the ReLU function.

[0101] After training the landslide crack identification model, the point cloud data and corresponding spatial features from step S1 above can be directly input into the landslide crack identification model to obtain the crack identification results of the target landslide area; the model application process can be, but is not limited to, as shown in step S3 below.

[0102] S3. Input the point cloud dataset of the target landslide area and the spatial features of each point cloud data in the point cloud dataset into the landslide crack identification model to obtain the crack identification result of the target landslide area; in specific applications, the crack identification result is whether there is a crack or no crack in the target landslide area.

[0103] In addition, this embodiment also provides a schematic diagram of the landslide crack identification effect of the aforementioned identification method; see [link to related documentation]. Figure 4 As shown, Figure 4 The example demonstrates the effectiveness of identifying two local structures of landslide cracks, from... Figure 4 As can be seen, its prediction results match the actual labels; furthermore, this embodiment also provides the ROC curve of the landslide crack identification results, which can be seen in [reference]. Figure 5 As shown, from Figure 5 The results also show that the accuracy of the binary classification is good; thus, it can be proved that the method provided in this embodiment can accurately identify landslide cracks and is very suitable for large-scale application and promotion.

[0104] Therefore, the landslide crack identification method based on neighborhood full-perception hierarchical convolution described in detail in steps S1 to S3 above uses the spatial features of each point cloud data of landslide cracks and combines them with a sub-neighborhood division method to extract local features. Compared with traditional techniques, this method can focus on the overall structural morphology of landslide cracks. Therefore, it can reduce misjudgments and omissions caused by complex changes in terrain, thereby improving the accuracy of landslide crack identification. Based on this, this method is very suitable for large-scale application and promotion in the field of landslide crack identification.

[0105] like Figure 6 As shown, the second aspect of this embodiment provides a hardware device for implementing the landslide crack identification method based on neighborhood full-perception hierarchical convolution described in the first aspect of the embodiment, comprising:

[0106] The data acquisition unit is used to acquire a point cloud dataset of the target landslide area, and based on the point cloud dataset, determine the spatial characteristics of each point cloud data in the point cloud dataset, wherein the spatial characteristics of any point cloud data include the slope aspect and slope corresponding to that point cloud data.

[0107] The model acquisition unit is used to acquire a landslide crack identification model. This model is trained using a sample point cloud dataset of multiple sample landslide areas and the sample spatial features of each sample point cloud data point in the corresponding dataset as input, with the crack identification results of each sample landslide area as output. The landslide crack identification model is a trained neighborhood-wide perception hierarchical convolutional neural network model. During training, after the sample point cloud dataset of any sample landslide area and the sample spatial features of each corresponding sample point cloud data point are input into the neighborhood-wide perception hierarchical convolutional neural network model, the model determines the crack center point of any sample landslide area, the sub-neighborhood points of each crack center point, and the sub-neighborhood point cloud set corresponding to each sub-neighborhood point from the sample point cloud dataset corresponding to that sample landslide area. Based on the sample spatial features of each sub-neighborhood point cloud set corresponding to the sub-neighborhood points of each crack center point, the model extracts the local features of each crack center point, and based on these local features, the crack identification result of any sample landslide area is obtained.

[0108] The crack identification unit is used to input the point cloud dataset of the target landslide area and the spatial features of each point cloud data in the point cloud dataset into the landslide crack identification model to obtain the crack identification result of the target landslide area.

[0109] The working process, working details and technical effects of the device provided in this embodiment can be found in the second aspect of the embodiment, and will not be repeated here.

[0110] like Figure 7 As shown, the third aspect of this embodiment provides another landslide crack identification device based on neighborhood-wide perception hierarchical convolution. Taking the device as an electronic device as an example, it includes: a memory, a processor, and a transceiver connected in sequence. The memory is used to store a computer program, the transceiver is used to send and receive messages, and the processor is used to read the computer program and execute the landslide crack identification method based on neighborhood-wide perception hierarchical convolution as described in the first aspect of the embodiment.

[0111] For specific examples, the memory may include, but is not limited to, random access memory (RAM), read-only memory (ROM), flash memory, first-in-first-out (FIFO) memory, and / or first-in-last-out (FILO) memory, etc.; specifically, the processor may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor may be implemented using at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), PLA (Programmable Logic Array). The processor may also include a main processor and a coprocessor. The main processor is used to process data in the wake-up state, also known as the CPU (Central Processing Unit); the coprocessor is a low-power processor used to process data in the standby state.

[0112] In some embodiments, the processor may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. For example, the processor may not be limited to microprocessors of the STM32F105 series, reduced instruction set computer (RISC) microprocessors, x86 architecture processors, or processors with integrated embedded neural network processing units (NPUs). The transceiver may be, but is not limited to, a Wi-Fi transceiver, a Bluetooth transceiver, a General Packet Radio Service (GPRS) transceiver, a ZigBee (a low-power LAN protocol based on the IEEE 802.15.4 standard) transceiver, a 3G transceiver, a 4G transceiver, and / or a 5G transceiver. Furthermore, the device may also include, but is not limited to, a power module, a display screen, and other necessary components.

[0113] The working process, working details and technical effects of the electronic device provided in this embodiment can be found in the first aspect of the embodiment, and will not be repeated here.

[0114] The fourth aspect of this embodiment provides a storage medium for storing instructions containing the landslide crack identification method based on neighborhood full-perception hierarchical convolution as described in the first aspect of the embodiment. That is, the storage medium stores instructions that, when executed on a computer, perform the landslide crack identification method based on neighborhood full-perception hierarchical convolution as described in the first aspect of the embodiment.

[0115] The storage medium refers to a carrier for storing data, which may include, but is not limited to, floppy disks, optical disks, hard disks, flash memory, USB flash drives, and / or memory sticks. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.

[0116] The working process, working details and technical effects of the storage medium provided in this embodiment can be found in the first aspect of the embodiment, and will not be repeated here.

[0117] The fifth aspect of this embodiment provides a computer program product containing instructions that, when executed on a computer, cause the computer to perform the landslide crack identification method based on neighborhood full-perception hierarchical convolution as described in the first aspect of the embodiment, wherein the computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device.

[0118] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A landslide crack identification method based on neighborhood-wide perception hierarchical convolution, characterized in that, include: Obtain a point cloud dataset of the target landslide area, and based on the point cloud dataset, determine the spatial features of each point cloud data in the point cloud dataset, wherein the spatial features of any point cloud data include the slope aspect and slope corresponding to that point cloud data. A landslide crack identification model is obtained, wherein the landslide crack identification model is trained by taking the sample point cloud dataset of multiple sample landslide areas and the sample spatial features of each sample point cloud data in the corresponding sample point cloud dataset as input, and the crack identification results of each sample landslide area as output. The landslide crack identification model is a trained neighborhood full-perception hierarchical convolutional neural network model. During training, after the sample point cloud dataset of any sample landslide area and the sample spatial features of each corresponding sample point cloud data are input into the neighborhood full-perception hierarchical convolutional neural network model, the model is used to downsample the sample point cloud dataset corresponding to any sample landslide area using the farthest distance sampling method, and each downsampled sampling point is used as the crack center point of any sample landslide area. For any crack center point among several crack center points corresponding to any sample landslide area, determine the neighborhood point cloud of any crack center point from the sample point cloud dataset corresponding to any sample landslide area; Calculate the distance between each neighboring point cloud in the neighborhood point cloud set and any crack center point, and take the neighboring point cloud with the smallest distance as the i-th child neighbor point of any crack center point, where the initial value of i is 1; The first target neighborhood point set is obtained by removing the i-th sub-neighbor point from the neighborhood point cloud set; Calculate the distance between each first target neighborhood point cloud in the first target neighborhood point cloud set and the i-th sub-neighbor point, and sort each first target neighborhood point cloud in ascending order of distance to obtain a sorted point cloud sequence; Using the first J first target neighborhood point clouds in the sorted point cloud sequence, a sub-neighbor point cloud set of the i-th sub-neighbor point is formed, where J is a positive integer greater than 1; The second target neighborhood point set is obtained by removing the sub-neighbor point cloud data of the i-th sub-neighbor point from the first target neighborhood point cloud set. The neighborhood point cloud is updated to the second target neighborhood point cloud and i is incremented by 1. The distance between each neighborhood point cloud in the neighborhood point cloud and any crack center point is recalculated until all neighborhood point clouds in the neighborhood point cloud are divided, so as to obtain the sub-neighbor points corresponding to any crack center point and the sub-neighbor point cloud corresponding to each sub-neighbor point. Based on the sample spatial features of each sub-neighbor point cloud data in the sub-neighbor point cloud set corresponding to each crack center point, the local features of each crack center point are extracted, and the crack identification result of any sample landslide area is obtained based on each local feature. The point cloud dataset of the target landslide area and the spatial features of each point cloud data in the point cloud dataset are input into the landslide crack identification model to obtain the crack identification result of the target landslide area.

2. The method according to claim 1, characterized in that, Based on the point cloud dataset, the spatial features of each point cloud data in the point cloud dataset are determined, including: The nearest neighbor search algorithm is used to determine the neighboring point cloud data of each point cloud data in the point cloud dataset; Based on the neighboring point cloud data of each point cloud data, and using the singular value decomposition algorithm, the surface normal of each point cloud data is obtained; The spatial characteristics of each point cloud data are determined based on the surface normal of each point cloud data.

3. The method according to claim 1, characterized in that, Each crack center point corresponds to at least one sub-neighborhood point. Based on the sample spatial features of the sub-neighborhood point cloud data corresponding to the sub-neighborhood points of each crack center point, local features of each crack center point are extracted, including: For any crack center point in any sample landslide area, the spatial features of each sub-neighbor point cloud data in the sub-neighbor point cloud set corresponding to each sub-neighbor point of the crack center point are convolved using the neighborhood full-perception hierarchical convolutional neural network model to obtain the local features of each sub-neighbor point corresponding to the crack center point. The local features of each sub-neighborhood point corresponding to any crack center point are updated to obtain the updated local features of each sub-neighborhood point. Based on the neighborhood-wide perception hierarchical convolutional neural network model, the local features of each updated sub-neighborhood point are aggregated to obtain the local features corresponding to any crack center point.

4. The method according to claim 3, characterized in that, Using the aforementioned neighborhood-wide perception hierarchical convolutional neural network model, the spatial features of each sub-neighbor point cloud data in the sub-neighbor point cloud set corresponding to each sub-neighbor point of any crack center point are convolved to obtain the local features of each sub-neighbor point corresponding to any crack center point, including: The local features of each sub-neighborhood point corresponding to any crack center point are extracted using the following formula (1); （1） In the above formula (1), This represents the local features of the i-th sub-neighborhood point corresponding to any crack center point. Indicates the kernel size as Convolution operation, This represents the convolution kernel weight of the j-th sub-neighbor point cloud data in the sub-neighbor point cloud set corresponding to the i-th sub-neighbor point. This represents the spatial characteristics of the j-th sub-neighbor point cloud data in the sub-neighbor point cloud set corresponding to the i-th sub-neighbor point. This represents the total number of sub-neighbor point cloud data in the sub-neighbor point cloud set corresponding to the i-th sub-neighbor point. Let represent the activation function, where i = 1, 2, 3, ..., n, and n represents the total number of sub-neighborhood points corresponding to any crack center point.

5. The method according to claim 3, characterized in that, The local features of each sub-neighborhood point corresponding to any crack center point are updated to obtain the updated local features of each sub-neighborhood point, including: The following formula (2) is used to update the local features of each sub-neighborhood point corresponding to any crack center point, so as to obtain the updated local features of each sub-neighborhood point. （2） In the above formula (2), This represents the local features of the i-th sub-neighborhood point corresponding to any crack center point. This represents the updated local feature of the i-th sub-neighborhood point corresponding to any crack center point. Represents the i-th sub-neighborhood point. Indicates the center point of any of the aforementioned cracks. This represents the distance between the i-th sub-neighbor point and the center point of any of the cracks. This represents the weight matrix of the i-th sub-neighborhood point. This represents the set of all sub-neighborhood points corresponding to the center point of any crack. This represents the total number of sub-neighborhood points corresponding to any crack center point.

6. The method according to claim 3, characterized in that, Based on the neighborhood-wide perception hierarchical convolutional neural network model, the local features of each updated sub-neighborhood point are aggregated to obtain the local features corresponding to any crack center point, including: The local features of each updated sub-neighborhood point are aggregated using the following formula (3) to obtain the local features corresponding to any crack center point; （3） In the above formula (3), This represents the local feature corresponding to the center point of any of the cracks. This represents the updated local feature of the i-th sub-neighborhood point corresponding to any crack center point. Indicates the kernel size as Convolution operation, This represents the activation function. This represents the kernel weight of the i-th sub-neighborhood point. This represents the total number of sub-neighborhood points corresponding to any crack center point.

7. A landslide crack identification device based on neighborhood-wide perception hierarchical convolution, characterized in that, include: The data acquisition unit is used to acquire a point cloud dataset of the target landslide area and, based on the point cloud dataset, determine the spatial characteristics of each point cloud data in the point cloud dataset, wherein the spatial characteristics of any point cloud data include the slope aspect and slope corresponding to that point cloud data. The model acquisition unit is used to acquire a landslide crack recognition model. This model is trained using a sample point cloud dataset of multiple sample landslide areas and the sample spatial features of each sample point cloud data point in the corresponding dataset as input, with the crack recognition results of each sample landslide area as output. The landslide crack recognition model is a trained neighborhood-wide perception hierarchical convolutional neural network model. During training, the sample point cloud dataset of any sample landslide area and the sample spatial features of each corresponding sample point cloud data are input into the neighborhood-wide perception hierarchical convolutional neural network model. Based on this model, the model then employs... The farthest-distance sampling method downsamples the point cloud dataset corresponding to any sample landslide area and uses each downsampled point as the crack center point of the landslide area. Then, for any crack center point among several crack center points corresponding to the landslide area, a neighborhood point cloud is determined from the point cloud dataset. Next, the distance between each neighboring point cloud and the crack center point is calculated, and the neighboring point cloud with the smallest distance is selected as the i-th sub-neighbor of the crack center point, where i is initially 1. Then, the i-th sub-neighbor point is removed from the neighborhood point cloud set to obtain the first target neighborhood point cloud set. The distance between each first target neighborhood point cloud set and the i-th sub-neighbor point is calculated, and the first target neighborhood point clouds are sorted in ascending order of distance to obtain a sorted point cloud sequence. Next, the first J first target neighborhood point clouds in the sorted point cloud sequence are used to form the sub-neighbor point cloud set of the i-th sub-neighbor point, where J is a positive integer greater than 1. The sub-neighbor point cloud data of the i-th sub-neighbor point set is removed from the first target neighborhood point cloud set to obtain the second target neighborhood point cloud set. The neighborhood point cloud is then used to update the neighborhood point cloud to the second target neighborhood point cloud and increment i by 1. The distance between each neighborhood point cloud in the neighborhood point cloud and any crack center point is recalculated until all neighborhood point clouds in the neighborhood point cloud are divided, so as to obtain the sub-neighbor points corresponding to any crack center point and the sub-neighbor point cloud corresponding to each sub-neighbor point. Among them, based on the sample space features of each sub-neighbor point cloud data in the sub-neighbor point cloud corresponding to each crack center point, the local features of each crack center point are extracted, so as to obtain the crack identification result of any sample landslide area based on each local feature. The crack identification unit is used to input the point cloud dataset of the target landslide area and the spatial features of each point cloud data in the point cloud dataset into the landslide crack identification model to obtain the crack identification result of the target landslide area.

8. An electronic device, characterized in that, include: A memory, a processor, and a transceiver are sequentially connected in communication, wherein the memory is used to store computer programs, the transceiver is used to send and receive messages, and the processor is used to read the computer programs and execute the landslide crack identification method based on neighborhood full-perception hierarchical convolution as described in any one of claims 1 to 6.

9. A storage medium, characterized in that, The storage medium stores instructions that, when executed on a computer, perform the landslide crack identification method based on neighborhood full-perception hierarchical convolution as described in any one of claims 1 to 6.

Images