Water accumulation detection method, electronic equipment and storage medium for underground pipe gallery

Abstract

The present application provides a method, electronic device and storage medium for detecting water accumulation in underground pipe corridors, including obtaining image samples of water accumulation in underground pipe corridors; generating a heat map of candidate areas of positive samples according to a two-dimensional Gaussian kernel function and image samples; adding pixel points whose distance from the center point of a true bounding box is lower than a preset distance to a final sample set; obtaining a positive sample quantity indicator value and a regression loss according to the true bounding box and the final sample set; adding pixel points whose positive sample quantity indicator value has the lowest regression loss to the final sample set; processing the final sample set according to a basic feature extraction backbone network and an attention correction module to obtain input features of water accumulation in underground pipe corridors; processing input features of water accumulation in underground pipe corridors according to bounding box regression and heat map prediction, respectively, to obtain candidate areas; and inputting the candidate areas into a classification network to effectively identify water accumulation in underground pipe corridors.

Description

Technical Field

[0001] The present application relates to the technical field of water accumulation identification in underground pipe corridors, and in particular to a water accumulation detection method, electronic equipment and storage medium for underground pipe corridors. Background Art

[0002] In the infrastructure of modern cities, underground pipe corridor systems undertake key tasks such as transmitting electricity, communications and water. Maintaining the safety and normal operation of underground pipe corridors is crucial to the daily life and economic activities of cities. However, the problem of water accumulation in underground pipe corridors frequently troubles operation and maintenance personnel, especially in the rainy season, when water accumulation can easily cause short circuits in power equipment and even cause safety accidents such as fires.

[0003] At present, there are the following problems when the target detection algorithm is directly applied to the detection of water accumulation in underground pipe corridors: the light in the underground pipe corridor is dim and the water surface has serious reflections, which affects the quality of the monitoring video. The boundary between the water accumulation area and the surrounding ground is blurred. The size and shape of the water accumulation are also different. The area of ​​the water accumulation area may differ by several to dozens of times, making it difficult for the detection model to process targets of different sizes at the same time. Summary of the invention

[0004] In view of this, it is necessary to provide a water accumulation detection method, electronic equipment and storage medium for underground pipe corridors, which can overcome at least one of the above defects.

[0005] In a first aspect, an embodiment of the present application provides a method for detecting water accumulation in an underground pipe gallery, the method comprising:

[0006] Acquire an image sample of water accumulation in an underground pipe gallery, wherein the image sample includes a real bounding box;

[0007] A positive sample candidate region heat map is generated according to the two-dimensional Gaussian kernel function and the image sample. The heat map generation formula is:

[0008]

[0009] Among them, (x, y) is the center point of the heat map, (x c ,y c ) is the center point of the real bounding box, σ x =w / 6,σ y =h / 6, w is the width of the true bounding box, h is the height of the true bounding box;

[0010] Adding pixel points whose distance from the center point of the true bounding box is less than a preset distance to the final sample set;

[0011] Obtaining a positive sample quantity indicator value and a regression loss according to the true bounding box and the final sample set;

[0012] Adding the pixel point of the positive sample quantity indicator value with the lowest regression loss to the final sample set;

[0013] The final sample set is processed according to the basic feature extraction backbone network and the attention correction module to obtain the input features of the underground pipe gallery water accumulation. The attention correction module includes a variable convolution and a dynamic linear correction. The variable convolution is used to improve the spatial attention, and the dynamic linear correction is used to improve the channel attention. The formula of the variable convolution is:

[0014]

[0015] Among them, p represents a certain position on the feature map, f(p) represents the calculation result of the deformable convolution kernel at the feature map position p, and o k represents the adaptive offset of the deformable convolution kernel relative to position p, w k is the convolution kernel parameter, Δm k Indicates that the shielding factor ranges from [0,1], p k is the offset of the convolution kernel sampling relative to position p, and x(·) represents the response value of the corresponding position on the feature map;

[0016] The formula for the dynamic linear correction is:

[0017]

[0018] Among them, x is the input feature map tensor, y i Represents the activation value calculated for the i-th channel of the feature map x. The linear parameter θ(x) is adaptively generated based on the input feature map x, including the dynamic linear activation function f θ(x) Slope of (x) and the intercept K represents the number of groups of linear parameters, the default value is 2;

[0019] The input features of the underground pipe gallery water accumulation are processed according to bounding box regression and heat map prediction respectively to obtain candidate areas;

[0020] The candidate region is input into a classification network, and the classification network comprises the following steps:

[0021] Predicting the classification confidence of the candidate region to obtain a preliminary classification result;

[0022] Performing an IoU-aware classification label adjustment on the preliminary classification result to generate an IoU-aware classification confidence;

[0023] Combining the classification confidence with the foreground confidence to generate a final classification confidence;

[0024] Based on the final classification confidence, non-maximum suppression is performed, overlapping areas are removed, and the final detection results are output.

[0025] In one embodiment, the method further comprises:

[0026] The cross entropy loss function is applied to supervise the heat map training, and the expression of the cross entropy loss function is:

[0027]

[0028] Among them, x×y represents the number of pixels in the heat map, Y x,y and Represents the value of the real heat map and the predicted heat map at the position (x, y), Y x,y =1 indicates a positive sample point, and α and β are hyperparameters.

[0029] In one embodiment, the method further comprises:

[0030] Gaussian blur processing is performed on the heat map to smooth the heat map. The formula for the Gaussian blur processing is:

[0031]

[0032] Among them, H(x,y) is the thermal map value after fuzzy processing, σ 2 is the variance.

[0033] In one embodiment, the method further comprises:

[0034] The input features are batch normalized to accelerate convergence and improve detection accuracy. The batch normalization formula is:

[0035]

[0036] Among them, μ is the mean of the data, σ 2 is the variance of the data, γ and β are trainable parameters, and ∈ is a minimum value.

[0037] In one embodiment, the method further comprises:

[0038] The feature map is processed by applying an attention gate, and different weights are applied to different regions in the feature map to improve the feature response of the region of interest and suppress invalid feature regions. The expression of the attention gate is:

[0039]

[0040] α i =σ2(q att (x i ,g i ; Θatt ))

[0041] Among them, x i is a pixel of the input feature map, g i is the gate signal vector, used to indicate pixel x i The attention focus area, σ1 is the ReLU activation function, σ2 is the sigmoid activation function, α i For x i The corresponding attention weight, Θ att are the learnable parameters involved in the attention module, including the coefficient W of the linear transformation 1×1 convolution x , W g and ψ, and the bias b g and b ψ .

[0042] In one embodiment, the formula for the final classification confidence is:

[0043]

[0044] Among them, s p is the foreground confidence, The classification confidence for each category.

[0045] In one embodiment, the image sample includes temporal information and spatial information, and the method further includes:

[0046] A three-dimensional convolutional neural network is applied to obtain the spatiotemporal features in the image sample. The formula of the three-dimensional convolutional neural network is:

[0047]

[0048] Among them, X is the input three-dimensional data, H is the height of the three-dimensional tensor, W is the width of the three-dimensional tensor, C in is the number of input channels, K is the weight of the three-dimensional convolution, K H is the height of the convolution kernel, K W is the width of the convolution kernel, D is the depth of the time dimension, C in is the number of input channels, t is the position index of the time dimension, i is the position index of the height dimension, j is the position index of the width dimension, c is the index of the output channel, b c is a scalar representing the bias of the cth output channel, Y is the output tensor, m is the offset in the height dimension, n is the offset in the width dimension, and d is the depth offset in the time dimension.

[0049] In one embodiment, the image sample includes temporal information and spatial information, and the method further includes:

[0050] Applying graph convolutional networks and the spatiotemporal features to obtain the changing trend of waterlogging;

[0051] The graph convolutional network formula is:

[0052]

[0053] in, is the adjacency matrix with self-loops added, I is the self-loop, A is the adjacency matrix, D is the degree matrix, W is the learnable weight matrix, and σ is the activation function;

[0054] The graph convolutional network is used to detect water accumulation according to the spatiotemporal features.

[0055] In a second aspect, an embodiment of the present application provides an electronic device, comprising: a processor; a memory for storing instructions executable by the processor; wherein the processor is configured to implement the water accumulation detection method for an underground utility tunnel described in the first aspect when executing the instructions.

[0056] In a third aspect, an embodiment of the present application provides a computer-readable storage medium, comprising instructions, wherein the instructions instruct a device to execute the water accumulation detection method for an underground utility tunnel as described in the first aspect.

[0057] The water accumulation detection method, electronic device and storage medium for underground utility tunnels provided in the embodiments of the present application can dynamically select more high-quality positive samples during the training process by introducing an adaptive positive sample selection algorithm based on a Gaussian kernel function, significantly improving the training effect of the model and the final detection accuracy. In addition, the intersection-over-union-perceived category label combined with the progressive bounding box correction network can effectively alleviate the problem of inaccurate bounding box quality assessment, reduce the probability of erroneous filtering, and further improve the positioning accuracy and classification confidence. At the same time, the feature extraction backbone network based on the attention mechanism and path aggregation structure enhances the algorithm's detection ability for multi-scale targets by transmitting and fusing feature maps of different levels and highlighting the features of the area of ​​interest. It effectively solves the challenges of poor lighting environment and unclear boundaries of waterlogged areas in underground utility tunnels, so that the water accumulation detection method for underground utility tunnels provided in the present application can still work stably and accurately in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 A schematic flow chart of a method for detecting water accumulation in an underground pipe gallery provided in one embodiment of the present application.

[0059] Figure 2 A schematic diagram of a water accumulation detection model provided in an embodiment of the present application.

[0060] Figure 3 This is a diagram of an example of Gaussian kernel generation provided in one embodiment of the present application.

[0061] Figure 4 A schematic diagram of a classified network process provided for an embodiment of the present application.

[0062] Figure 5 A schematic diagram of an electronic device provided in one embodiment of the present application.

[0063] Main component symbols

[0064] Water Detection Model 10

[0065] Data acquisition module 11

[0066] Regional candidate network 12

[0067] Multi-scale feature extraction backbone network13

[0068] Progressive Bounding Box Refinement Network14

[0069] Electronic devices 20

[0070] Processor 21

[0071] Memory 22

[0072] Method steps S100-S800; S810-S840 DETAILED DESCRIPTION

[0073] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments.

[0074] It should be noted that, in the embodiments of the present application, "at least one" refers to one or more, and "more" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art in the present application. The terms used in the specification of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application.

[0075] It should be noted that, in the embodiments of the present application, words such as "first" and "second" are only used for the purpose of distinguishing descriptions, and cannot be understood as indicating or implying relative importance, nor can they be understood as indicating or implying order. Features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the embodiments of the present application, words such as "exemplary" or "for example" are used to indicate examples, illustrations or explanations. Any embodiment or design described as "exemplary" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present related concepts in a concrete way.

[0076] Based on the implementations in this application, all other implementations obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0077] In the infrastructure of modern cities, underground pipe corridor systems undertake key tasks such as transmitting electricity, communications and water. Maintaining the safety and normal operation of underground pipe corridors is crucial to the daily life and economic activities of cities. However, the problem of water accumulation in underground pipe corridors frequently troubles operation and maintenance personnel, especially in the rainy season, when water accumulation can easily cause short circuits in power equipment and even cause safety accidents such as fires.

[0078] At present, the direct application of target detection algorithms to underground pipe gallery water accumulation detection has the following problems: the light in the underground pipe gallery is dim, the water surface reflects light seriously, which affects the quality of the surveillance video, the boundary between the water accumulation area and the surrounding ground is blurred, the size and shape of the water accumulation are also different, and the area of ​​the water accumulation area may vary from several times to dozens of times, making it difficult for the detection model to process targets of different sizes at the same time

[0079] Therefore, the embodiments of the present application provide a method, electronic device and storage medium for detecting water accumulation in underground pipe corridors. By introducing an adaptive positive sample selection algorithm based on a Gaussian kernel function, more high-quality positive samples can be dynamically selected during the training process, significantly improving the training effect of the model and the final detection accuracy. In addition, the intersection-over-union-perceived category label combined with the progressive bounding box correction network can effectively alleviate the problem of inaccurate bounding box quality assessment, reduce the probability of erroneous filtering, and further improve the positioning accuracy and classification confidence. At the same time, the feature extraction backbone network based on the attention mechanism and path aggregation structure enhances the algorithm's ability to detect multi-scale targets by transmitting and fusing feature maps of different levels and highlighting the features of the area of ​​interest. It effectively solves the challenges of poor lighting environment and unclear boundaries of water accumulation areas in underground pipe corridors, so that the water accumulation detection method for underground pipe corridors provided by the present application can still work stably and accurately in complex environments.

[0080] In conjunction with the accompanying drawings, some embodiments of the application are described in detail below. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0081] Figure 1 FIG. 1 is a flow chart of a method for detecting water accumulation in an underground pipe gallery provided in an embodiment of the present application. Figure 1 As shown, at least the following steps are included: S100: obtaining an image sample of water accumulation in an underground pipe gallery; S200: generating a heat map of a positive sample candidate area according to a two-dimensional Gaussian kernel function and image samples; S300: adding pixels whose distance from the center point of the true bounding box is lower than a preset distance to the final sample set; S400: obtaining a positive sample quantity indicator value and a regression loss according to the true bounding box and the final sample set; S500: adding pixels whose positive sample quantity indicator value has the lowest regression loss to the final sample set; S600: processing the final sample set according to the basic feature extraction backbone network and the attention correction module to obtain input features of water accumulation in the underground pipe gallery; S700: processing the input features of water accumulation in the underground pipe gallery according to the bounding box regression and the heat map prediction, respectively, to obtain candidate areas; S800: inputting the candidate areas into the classification network.

[0082] S100: Obtain an image sample of water accumulation in the underground pipe gallery.

[0083] In an embodiment of the present application, the method for detecting water accumulation in an underground pipe gallery includes obtaining an image sample of water accumulation in the underground pipe gallery in step S100. The image sample includes a real bounding box. It is understandable that the real bounding box marks the location and range of the water accumulation area.

[0084] Please also read Figure 2 , Figure 2 Schematic diagram of a water accumulation detection model provided by an embodiment of the present application. Figure 2 As shown, the water accumulation detection model 10 includes at least the following parts: a data acquisition module 11, a region candidate network 12, a multi-scale feature extraction backbone network 13 and a progressive bounding box correction network.

[0085] In the embodiment of the present application, the data acquisition module 11 is used to obtain water accumulation image samples of the underground pipe gallery. This module is responsible for collecting monitoring video or image data and transmitting these data to the subsequent processing unit for further analysis. The data acquisition module 11 ensures the quality and reliability of the input data and provides basic data support for the entire detection process.

[0086] In the embodiment of the present application, the region candidate network 12 is used to generate candidate regions to locate potential waterlogging areas. The region candidate network 12 generates multiple candidate regions through a series of convolution operations based on the input image data, and selects high-quality positive samples through a specific algorithm (such as an adaptive positive sample selection algorithm based on a Gaussian kernel function).

[0087] It can be understood that the region candidate network 12 includes two branches, a bounding box regression branch and a heat map prediction branch. The bounding box regression branch is used to densely predict the distances (l, r, b, t) from all points to the four sides of the bounding box. Compared with the center point representation (x center ,y center ,w,h), using (l,r,b,t) representation can add more positive samples to participate in training. The heat map prediction branch is used to generate a heat map. The peak point of the heat map indicates the location of the key point, and the peak point response value gives the confidence of whether the target exists in the candidate area, that is, the foreground confidence.

[0088] In addition, the region proposal network 12 takes the output of the basic feature extraction backbone network 13 as input, and there are 5 feature maps of different scales, p3, p4, p5, p6, and p7. The hierarchical allocation of the feature pyramid network, the regression range is set to [0,80], [64,160], [128,320], [256,640], and [512,INF]. Since convolutional neural networks are very sensitive to scale, the practice of each layer only processing bounding boxes of one scale range will reduce the recall rate. Therefore, for all positive sample points in the feature map of each layer, samples whose half of the diagonal length of the corresponding true bounding box is not within the regression size range will be ignored by the layer. This allows the same positive sample to be regressed by feature maps of multiple layers, rather than feature maps of different layers being responsible for regressing bounding boxes of different size ranges.

[0089] In an embodiment of the present application, the multi-scale feature extraction backbone network 13 is used to extract and fuse feature maps of different scales to enhance the model's ability to detect multi-scale targets. The multi-scale feature extraction backbone network 13 adopts a design based on an attention mechanism and a path aggregation structure, which can transfer and fuse feature maps of different levels and highlight the features of the area of ​​interest. In this way, the model can better handle the diversity of waterlogged areas and improve the accuracy and robustness of detection.

[0090] In the embodiment of the present application, the progressive bounding box correction network 14 is used to further optimize the position and size of the bounding box of the candidate area. The progressive bounding box correction network 14 gradually adjusts and corrects the bounding box by combining the intersection-over-union ratio perception category label with the progressive bounding box correction method, so that it more accurately surrounds the waterlogged area. Through iterative correction, the network can effectively reduce the error of the bounding box quality assessment and improve the accuracy of the final detection result.

[0091] S200: Generate a positive sample candidate region heat map according to the two-dimensional Gaussian kernel function and the image samples.

[0092] In the embodiment of the present application, the method for detecting water accumulation in underground pipe corridors includes generating a positive sample candidate area heat map according to a two-dimensional Gaussian kernel function and image samples in step S200. The heat map generation formula is:

[0093]

[0094] Among them, (x, y) is the center point of the heat map, (x c ,y c ) is the center point of the real bounding box, σ x =w / 6,σ y =h / 6, w is the width of the true bounding box, and h is the height of the true bounding box.

[0095] It is understandable that for special detection targets such as water accumulation, the center point of the true rectangular bounding box and its small neighborhood in a large number of annotated images often fall on background pixels. In this case, if the candidate region bounding box is regressed from the background center point, it is very likely that a predicted box with a high intersection over union (IoU) ratio with the true bounding box will be regressed during training, but it does not actually contain enough target foreground pixels. When using such low-quality candidate regions for region of interest (RoI) pooling, the quality of the cropped features is also very poor, which directly affects the bounding box regression and target classification process in the subsequent stages.

[0096] If the bounding box is regressed from the foreground point, it can be guaranteed that there are more foreground pixels in the prediction box, thereby giving the subsequent network richer and more effective foreground features. For the case where the true center point falls on the background pixel, it is more reasonable to regress the bounding box from the foreground position than to force the network to regress the bounding box from the background position. In addition, in addition to the center point and the small center neighborhood, there are many pixels in the true bounding box that have the potential to regress to a high-quality predicted bounding box. To this end, the present application proposes an adaptive positive sample selection algorithm based on a Gaussian kernel function, in order to improve the detection accuracy by improving the quality and quantity of positive samples involved in training.

[0097] It is understandable that σ x =w / 6 and σ y The setting method of =h / 6 is to limit the range of the elliptical Gaussian kernel function area to about one third of the width and height of the corresponding bounding box, so as to improve the detection accuracy of the water accumulation detection method for underground pipe corridors.

[0098] It can be understood that the heat map generation process is to weight the center point of each true bounding box so that the area closer to the center point has a higher weight, thereby generating a heat map that highlights the candidate areas of positive samples. This method can more accurately reflect the distribution characteristics of the waterlogged area and provide a basis for the subsequent positive sample selection.

[0099] S300: adding pixel points whose distance from the center point of the true bounding box is less than a preset distance to the final sample set.

[0100] In an embodiment of the present application, the method for detecting water accumulation in an underground utility tunnel includes, in step S300, adding pixel points whose distance from the center point of the real bounding box is less than a preset distance to the final sample set.

[0101] It can be understood that by calculating the distance between each pixel and the center point of the true bounding box, the pixels below the preset distance are screened out and added to the final sample set. This step ensures that the selected sample points have high relevance and improves the training effect of the model.

[0102] It can be understood that in the embodiment of the present application, the preset distance is 3, that is, all points within a 3×3 neighborhood from the center of the true bounding box are added to the final positive sample set. In addition, for all other samples in the positive sample candidate area, except for the points that have been selected as positive samples by other true bounding boxes, the remaining points are added to the candidate positive sample set.

[0103] In an embodiment of the present application, the final sample set is processed according to the basic feature extraction backbone network and the attention correction module to obtain the input features of the underground pipe gallery water accumulation. The attention correction module includes variable convolution and dynamic linear correction. The variable convolution is used to improve spatial attention, and the dynamic linear correction is used to improve channel attention. The formula of variable convolution is:

[0104]

[0105] Among them, p represents a certain position on the feature map, f(p) represents the calculation result of the deformable convolution kernel at the feature map position p, and o k represents the adaptive offset of the deformable convolution kernel relative to position p, w k is the convolution kernel parameter, Δm k Indicates that the shielding factor ranges from [0,1], p k is the offset of the convolution kernel sampling relative to the position p, and x(·) represents the response value of the corresponding position on the feature map.

[0106] It is understandable that deformable convolution is used to improve spatial attention. When performing convolution operations, the sampling points of the convolution kernel are fixed, while deformable convolution can dynamically adjust the sampling point positions of the convolution kernel according to the characteristics of the input image, thereby better capturing the geometric deformation of the target object.

[0107] Understandably, p k ∈{(-1,-1),(-1,0),(-1,1),(0,-1),(0,0),(0,1),(1,-1),(1,0),(1,1)} is the offset of the ordinary 3*3 convolution kernel relative to the position p.

[0108] The formula for dynamic linear correction is:

[0109]

[0110]

[0111] Among them, x is the input feature map tensor, y i Represents the activation value calculated for the i-th channel of the feature map x. The linear parameter θ(x) is adaptively generated based on the input feature map x, including the dynamic linear activation function f θ(x) Slope of (x) and the intercept K represents the number of groups of linear parameters, and the default value is 2.

[0112] It can be understood that by dynamically adjusting the activation function of each channel, dynamic linear correction can better highlight important channel features and suppress irrelevant or redundant information, thereby enhancing the expressive power of the model.

[0113] S400: Obtaining a positive sample quantity indicator value and a regression loss according to the true bounding box and the final sample set.

[0114] In an embodiment of the present application, the water accumulation detection method for underground pipe corridors includes obtaining a positive sample quantity indicator value and a regression loss according to a true bounding box and a final sample set in step S400. Specifically, by analyzing the samples in the final sample set, the regression loss value of each sample point is calculated, and the quantity indicator value of the positive samples is counted. Among them, the regression loss is mainly used to measure the deviation between the sample point and the true bounding box, and the quantity indicator value is used to evaluate the effect of sample selection.

[0115] Specifically, the intersection-over-union ratio of the true bounding box and the predicted box regressed by all samples in the candidate positive sample set is calculated, and the N values ​​with the highest intersection-over-union ratio are summed and rounded to obtain the positive sample number indicator value (K). Among them, N is fixed to 20, and the number of additional positive samples selected is limited to no more than 20 to avoid obtaining too many noise samples.

[0116] In the embodiment of the present application, the distance intersection over union (D-IoU) is used to calculate the regression loss. The calculation formula of the distance intersection over union is:

[0117]

[0118] Among them, ρ(·) represents the Euclidean distance, and c represent the center points of the predicted box and the true box, respectively, and l represents the diagonal length of the minimum circumscribed rectangle of the predicted box and the true box.

[0119] It is understandable that the distance intersection and union ratio is used to calculate the regression loss. The distance intersection and union ratio improves the positioning accuracy by comprehensively considering the overlapping area of ​​the bounding box and the center point distance. By introducing the Euclidean distance between the center points of the predicted box and the true box, the distance intersection and union ratio not only considers the overlapping area of ​​the bounding box, but also introduces the position relationship of the bounding box, so that the position and size of the bounding box can be optimized more accurately during the regression process.

[0120] S500: Add the pixel point with the lowest positive sample quantity indicator value of regression loss to the final sample set.

[0121] In an embodiment of the present application, the method for detecting water accumulation in an underground pipe gallery includes, in step S500, adding the pixel points with the lowest positive sample quantity indicator value of the regression loss to the final sample set. Specifically, by sorting the regression loss of each sample point, a part of the sample points with the lowest regression loss is selected, and these sample points are added to the final sample set. This step can effectively reduce errors, improve the quality of samples, and thus improve the training accuracy of the model.

[0122] It can be understood that the number of pixels equal to the positive sample number indicator value (K) is selected to be added to the final positive sample set. For points with the same regression loss value, the larger value in the Gaussian kernel function heat map is preferentially selected.

[0123] Please also read Figure 3 , Figure 3 : is a Gaussian kernel generation example diagram provided by an embodiment of the present application. Figure 3 As shown, the water accumulation detection method for underground pipe corridors provided by the embodiment of the present application can effectively generate a heat map of the positive sample candidate area. In the generation process, a heat map is first generated based on a two-dimensional Gaussian kernel function and an image sample. After the heat map is generated, the pixel points that are less than a preset distance from the center point of the true bounding box are added to the final sample set. In this way, it is possible to ensure that the bounding box is regressed from the foreground point, ensuring that there are more foreground pixels in the prediction box, thereby improving the quality of feature extraction.

[0124] It is understandable that the final performance of target detection is seriously affected by the quality of training samples. By adopting the method provided in the embodiment of the present application, more high-quality positive samples can be adaptively selected during model training, thereby significantly improving the performance of the model. In addition, the method provided in the embodiment of the present application only increases the computational overhead during the training phase and does not affect the number of parameters and reasoning time of the model. Based on the method provided in the embodiment of the present application, for different input images, the positive samples used in training are all adaptively selected, and the number of positive samples selected each time will also be dynamically adjusted with the iterative process.

[0125] S600: Process the final sample set according to the basic feature extraction backbone network and the attention correction module to obtain the input features of the underground pipe gallery water accumulation.

[0126] In an embodiment of the present application, the method for detecting water accumulation in underground pipe corridors includes, in step S600, processing the final sample set according to the basic feature extraction backbone network and the attention correction module to obtain the input features of water accumulation in the underground pipe corridor. Specifically, step S600 analyzes the samples in the final sample set, calculates the regression loss value of each sample point, and counts the number indicator value of positive samples. The regression loss is mainly used to measure the deviation between the sample point and the true bounding box, while the number indicator value is used to evaluate the effect of sample selection.

[0127] S700: Processing input features of water accumulation in the underground pipe gallery according to bounding box regression and heat map prediction to obtain candidate areas.

[0128] In an embodiment of the present application, the method for detecting water accumulation in underground pipe corridors includes, in step S700, processing input features of water accumulation in underground pipe corridors according to bounding box regression and heat map prediction, respectively, to obtain candidate areas. Specifically, the input features are processed by bounding box regression prediction and heat map prediction to generate candidate areas for water accumulation areas. This process combines the accuracy of regression prediction and the wide coverage of heat map prediction to ensure the reliability of candidate areas.

[0129] S800: Input the candidate region into the classification network.

[0130] In an embodiment of the present application, the method for detecting water accumulation in an underground pipe gallery includes inputting a candidate area into a classification network in step S800. Specifically, the generated candidate area is input into a pre-trained classification network, and each candidate area is classified and judged to determine whether it is a water accumulation area. The classification network extracts features and classifies the candidate area, and outputs the final detection result. This process combines the two steps of area generation and classification recognition, which improves the accuracy and robustness of water accumulation detection.

[0131] Please also read Figure 4 , Figure 4 is a schematic diagram of a classification network flow provided by an embodiment of the present application, such as Figure 4 As shown, the classification network includes at least the following steps: S810: predicting the classification confidence of the candidate area to obtain a preliminary classification result; S820: adjusting the classification label of the preliminary classification result based on the intersection-over-union ratio perception to generate the classification confidence based on the intersection-over-union ratio perception; S830: combining the classification confidence with the foreground confidence to generate the final classification confidence; S840: performing non-maximum suppression based on the final classification confidence, removing the overlapping areas, and outputting the final detection result.

[0132] S810: Predict the classification confidence of the candidate region to obtain a preliminary classification result.

[0133] In an embodiment of the present application, the method for detecting water accumulation in an underground pipe gallery performs classification confidence prediction on the candidate area in step S810 to obtain a preliminary classification result. Specifically, a confidence prediction is performed on each candidate area through a classification network to determine whether it belongs to a water accumulation area. The preliminary classification result includes a classification confidence score for each candidate area.

[0134] S820: Perform an IoU-aware classification label adjustment on the preliminary classification result to generate an IoU-aware classification confidence.

[0135] In the embodiment of the present application, the water accumulation detection method for underground pipe corridors, in step S820, performs an IoU-aware classification label adjustment on the preliminary classification result to generate an IoU-aware classification confidence. Specifically, by calculating the IoU of the candidate area and the true bounding box, the preliminary classification result is adjusted and the classification label is corrected. This can more accurately reflect whether the candidate area belongs to a water accumulation area and improve the classification accuracy.

[0136] S830: Combine the classification confidence and the foreground confidence to generate a final classification confidence.

[0137] In an embodiment of the present application, in the water accumulation detection method for underground pipe corridors, in step S820, the classification confidence and the foreground confidence are combined to generate a final classification confidence. Specifically, the classification confidence and the foreground confidence are weighted averaged or otherwise fused to obtain the final classification confidence of each candidate area. This step combines the output of the classification network and the foreground confidence predicted by the heat map to improve the reliability of the classification result.

[0138] S840: Perform non-maximum suppression based on the final classification confidence, remove overlapping areas, and output the final detection result.

[0139] In an embodiment of the present application, in step S820, the method for detecting water accumulation in underground pipe corridors performs non-maximum suppression based on the final classification confidence, removes overlapping areas, and outputs the final detection result. Specifically, a non-maximum suppression (NMS) algorithm is used to filter out overlapping candidate areas according to the final classification confidence, and only the optimal detection result is retained. In this way, repeated detection can be removed and the accuracy and precision of the detection results can be improved.

[0140] In an embodiment of the present application, a method for detecting water accumulation in an underground pipe gallery is provided: a cross entropy loss function is applied to supervise the heat map training, and the expression of the cross entropy loss function is:

[0141]

[0142] Among them, x×y represents the number of pixels in the heat map, Y x,y and Represents the value of the real heat map and the predicted heat map at the position (x, y), Y x,y =1 indicates a positive sample point, and α and β are hyperparameters.

[0143] It is understandable that α mainly adjusts the balance between positive and negative samples, and β is mainly used to balance the impact of difficult samples and easy samples. By introducing α and β parameters, it is possible to better cope with sample imbalance and difficult samples, so that the model can pay more attention to the key positive sample area during training.

[0144] In an embodiment of the present application, the method for detecting water accumulation in an underground pipe gallery further includes: performing Gaussian blur processing on the thermal map to smooth the thermal map, and the formula for Gaussian blur processing is:

[0145]

[0146] Among them, H(x,y) is the thermal map value after fuzzy processing, σ 2 is the variance. σ is used to control the degree of Gaussian blur. The larger σ is, the more obvious the blur effect is.

[0147] It can be understood that by performing Gaussian blurring on the heat map, the heat map can be smoothed and the impact of noise on the detection results can be reduced, thereby improving the detection accuracy and stability of the model.

[0148] In an embodiment of the present application, the method for detecting water accumulation in an underground pipe gallery further includes: performing batch normalization processing on the input features to accelerate the convergence speed and improve the detection accuracy. The formula for batch normalization is:

[0149]

[0150] Among them, μ is the mean of the data, σ 2is the variance of the data, γ and β are trainable parameters, and ∈ is a minimum value.

[0151] It is understandable that batch normalization can make the distribution of each batch of data consistent, eliminate internal covariate shift, speed up the convergence of the model, and improve detection accuracy.

[0152] In an embodiment of the present application, the method for detecting water accumulation in an underground pipe gallery further includes: applying an attention gate to process the feature map, applying different weights to different regions in the feature map to improve the feature response of the region of interest and suppress invalid feature regions, and the expression of the attention gate is:

[0153]

[0154] α i =σ2(q att (x i ,g i ; Θ att ))

[0155] Among them, x i is a pixel of the input feature map, g i is the gate signal vector, used to indicate pixel x i The attention focus area, σ1 is the ReLU activation function, σ2 is the sigmoid activation function, α i For x i The corresponding attention weight, Θ att are the learnable parameters involved in the attention module, including the coefficient W of the linear transformation 1×1 convolution x , W g and ψ, and the bias b g and b ψ .

[0156] It can be understood that by introducing the attention mechanism, the water accumulation detection method for underground pipe corridors can pay more attention to the characteristics of the water accumulation area and improve the accuracy and robustness of detection. The attention gate module effectively enhances the feature response of the area of ​​interest and suppresses the invalid feature area, so that the model can more accurately detect the water accumulation area in the underground pipe corridor.

[0157] In the embodiment of the present application, the formula for the final classification confidence is:

[0158]

[0159] Among them, s p is the foreground confidence, The classification confidence for each category.

[0160] In the embodiment of the present application, the candidate region bounding box is defined as (xp ,y p ,w p ,h p ), the corrected candidate region bounding box is represented as (x, y, w, h), and the predicted bounding box position correction value is (Δx, Δy, Δw, Δh). The process formula of bounding box correction is:

[0161] x=Δx·w p +x P

[0162] w=w p ·e Δw

[0163] y=Δy·h p +y P

[0164] h=h p ·e Δ h

[0165] It is understandable that for network training, since the ratio of positive and negative samples is always controlled at about 1:3 when sampling candidate regions, the problem of imbalance between positive and negative samples generally does not occur, so the classification task uses the cross entropy loss function. Considering that the regression value of the progressive bounding box position correction has a small change range, the effect of the IoU series loss is not as good as the smooth L1 loss, so the bounding box position correction branch uses the smooth L1 loss function. The expression of the smooth L1 loss function is:

[0166]

[0167] in, represents the predicted value, and Y represents the true value.

[0168] It is understandable that the smooth L1 loss is not sensitive to outliers and does not cause gradient explosion. Therefore, the overall loss function formula of the final stage network is:

[0169]

[0170] Among them, L focal is the focal loss used for heatmap. L diou is the distance intersection-over-union loss used for regression loss. is the classification cross entropy loss at different stages. is the bounding box regression loss at different stages. Through multi-stage progressive correction and combining multiple loss functions, the detection method of the embodiment of the present application can effectively improve the detection accuracy and robustness.

[0171] In some embodiments of the present application, the image sample includes time series information and spatial information, and the method for detecting water accumulation in an underground pipe gallery further includes: applying a three-dimensional convolutional neural network to obtain the temporal and spatial features in the image sample, and the formula of the three-dimensional convolutional neural network is:

[0172]

[0173] Among them, X is the input three-dimensional data, H is the height of the three-dimensional tensor, W is the width of the three-dimensional tensor, C in is the number of input channels, K is the weight of the three-dimensional convolution, K H is the height of the convolution kernel, K W is the width of the convolution kernel, D is the depth of the time dimension, C in is the number of input channels, t is the position index of the time dimension, i is the position index of the height dimension, j is the position index of the width dimension, c is the index of the output channel, b c is a scalar representing the bias of the cth output channel, Y is the output tensor, m is the offset in the height dimension, n is the offset in the width dimension, and d is the depth offset in the time dimension.

[0174] It can be understood that the 3D convolutional neural network is used to process image data containing temporal and spatial information, thereby extracting rich spatiotemporal features. The 3D convolution can capture the characteristics of image changes over time and combine them with spatial features to more accurately detect changes and trends in waterlogged areas.

[0175] In some embodiments of the present application, the image sample includes time series information and spatial information, and the method for detecting waterlogging in underground pipe corridors further includes: applying a graph convolutional network and spatiotemporal features to obtain a changing trend of waterlogging. The graph convolutional network is used to detect waterlogging based on spatiotemporal features.

[0176] The graph convolutional network formula is:

[0177]

[0178] in, is the adjacency matrix with self-loops added, I is the self-loop, A is the adjacency matrix, D is the degree matrix, W is the learnable weight matrix, and σ is the activation function.

[0179] It can be understood that the graph convolutional network is used to process the spatiotemporal features extracted by the three-dimensional convolutional neural network in order to better understand and predict the changing trend of waterlogging.

[0180] It is understandable that in the detection of water accumulation in underground pipe corridors, due to the insufficient number and diversity of samples in the self-built dataset, it is impossible to fully cover the complex and changeable actual situation. In order to enhance the model's ability to cope with various potential changes, such as different lighting environments, clarity, contrast, shape and color of water accumulation areas, this application has designed a series of data augmentation operations to expand the dataset, thereby improving the accuracy and generalization ability of the model.

[0181] It is understandable that, in terms of geometric transformation, the water accumulation detection method for underground pipe gallery provided in the embodiment of the present application randomly performs perspective transformation, translation and rotation on the image. These operations can simulate different perspectives and position changes, and improve the model's ability to recognize water accumulation areas under various perspectives. At the same time, a null operation is randomly added to each operation to keep some images unchanged, thereby increasing the diversity and randomness of the samples.

[0182] It is understandable that the water accumulation detection method for underground pipe corridors provided in the embodiment of the present application adds a random horizontal flip operation to enhance the model's ability to process symmetrical images. In terms of color transformation, the brightness and contrast of the image are randomly adjusted, the image is converted to a grayscale image, or the RGB channel is translated to simulate different lighting conditions and environmental changes. These color transformations can help the model maintain stable detection performance under different lighting conditions.

[0183] It is understandable that the water accumulation detection method for underground pipe corridors provided in the embodiment of the present application also simulates a variety of weather effects, such as fog, sunlight, rainy days, and blurring. These operations make the image show different visual characteristics under different weather conditions, thereby enhancing the robustness of the model under variable weather conditions. In terms of occlusion and erasure, part of the image is randomly occluded or erased to simulate partial occlusion, thereby enhancing the model's ability to recognize waterlogged areas under partial occlusion.

[0184] It is understandable that the method for detecting water accumulation in underground pipe corridors provided in the embodiment of the present application performs oversampling and mosaic enhancement on some images in order to improve the detection effect of small targets. The oversampling operation increases the number of samples containing small targets, while the mosaic enhancement generates new complex scene samples by splicing multiple images together, thereby improving the model's ability to detect multiple targets.

[0185] It can be understood that the method for detecting water accumulation in underground pipe corridors provided in the embodiments of the present application not only increases the number and diversity of samples in the data set through these data augmentation operations, but also significantly improves the recognition accuracy and reliability of the model in different environments. The augmented data set can better simulate the actual situation, making the model perform better in the task of detecting water accumulation in real underground pipe corridors.

[0186] It is understandable that the water accumulation detection method for underground pipe corridors provided in the embodiment of the present application can dynamically select more high-quality positive samples during the training process, thereby improving the training effect and final accuracy of the model. The Gaussian kernel function generates a heat map based on the center point of the true bounding box, so that the pixels close to the center point have a higher weight, thereby selecting a more representative positive sample.

[0187] Secondly, the water accumulation detection method for underground pipe corridors provided in the embodiment of the present application introduces the intersection-over-union-perceived category label and the progressive bounding box correction network. By combining the classification confidence and positioning accuracy, the detection effect is improved. A progressive bounding box correction network is used to gradually optimize the position of the bounding box, making the bounding box quality assessment more accurate. This method effectively alleviates the problem of inaccurate bounding box quality assessment, reduces the possibility of erroneous filtering, and thus improves the robustness of detection.

[0188] Third, the method for detecting water accumulation in underground pipe corridors provided in the embodiment of the present application proposes a feature extraction backbone network based on an attention mechanism and a path aggregation structure. By transmitting and fusing the semantic information and positioning information of feature maps at different levels, the features of the area of ​​interest are highlighted during the feature fusion process, and the algorithm's ability to detect multi-scale targets is enhanced. The attention mechanism improves the efficiency of feature extraction and improves the detection effect of water accumulation areas by focusing on important areas.

[0189] In addition, the water accumulation detection method for underground pipe corridors provided in the embodiment of the present application introduces an adaptive enhancement mechanism to adaptively adjust the model parameters according to the characteristics of the input image, thereby improving the model's adaptability to different lighting conditions and reflection situations. The multi-scale target detection capability enables the model to simultaneously detect large-scale water accumulation areas and small water accumulation areas, ensuring the comprehensiveness and accuracy of the detection.

[0190] In the feature extraction stage, the water accumulation detection method for underground pipe corridors combines deformable convolution and dynamic linear correction to improve the accuracy of feature extraction. Deformable convolution can adapt to water accumulation areas of different shapes, while dynamic linear correction adjusts the weights of each channel to highlight important features and suppress noise.

[0191] Finally, in view of the special environment of underground pipe corridors, the water accumulation detection method for underground pipe corridors provided in the embodiment of the present application has designed a comprehensive robustness evaluation and model verification scheme, and verified the stability and accuracy of the model under different conditions through a large amount of actual data testing. This evaluation method ensures that the model can maintain efficient detection capabilities in various complex environments.

[0192] Figure 5 This is an electronic device 20 provided in an embodiment of the present application. Figure 5As shown, the electronic device 20 includes at least the following parts: a processor 21 and a memory 22 .

[0193] In the embodiment of the present application, the memory 22 is used to store instructions executable by the processor 21, and the processor 21 is configured to implement the following when executing the instructions: Figure 1 The method shown.

[0194] In an embodiment of the present application, a computer-readable storage medium includes instructions, and the instructions instruct a device to execute the method of the first aspect. For example, the instructions instruct the device to execute Figure 1 The method shown in steps S100 to S900.

[0195] The program operating in the electronic device 20 involved in one embodiment of the present application may be a program (a program that enables a computer to function) that controls a central processing unit (CPU) and the like to realize the functions of the above-mentioned embodiment involved in one scheme of the present invention. Then, the information processed by these devices is temporarily stored in a random access memory (RAM) during its processing, and then stored in various ROMs such as a read-only memory (Flash ROM) and a hard disk drive (HDD), and is read, modified, and written by the CPU as needed.

[0196] It should be noted that part of the electronic device 20 of the above embodiment may be implemented by a computer. In this case, a program for implementing the control function may be recorded in a computer-readable recording medium and implemented by reading the program recorded in the recording medium into a computer system and executing it.

[0197] It should be noted that the "computer system" mentioned here refers to a computer system built into the electronic device 20, and adopts a computer system including hardware such as an OS and peripheral devices. In addition, "computer-readable recording medium" refers to removable media such as a floppy disk, a magneto-optical disk, a ROM, a CD-ROM, and a storage device such as a hard disk built into the computer system.

[0198] Furthermore, "computer-readable recording media" may include: media that dynamically store programs for a short period of time, such as communication lines when sending programs via networks such as the Internet or communication lines such as telephone lines; and media that store programs for a fixed period of time, such as volatile memories inside computer systems that serve as servers or clients in this case. In addition, the above-mentioned program may be a program for realizing a part of the above-mentioned functions, or a program that can realize the above-mentioned functions by combining with a program already recorded in a computer system.

[0199] In addition, the electronic device 20 in the above-mentioned embodiment can also be implemented as a collection (device group) composed of multiple devices. Each device constituting the device group can have a part or all of the functions or functional blocks of the electronic device 20 in the above-mentioned embodiment. As a device group, it is sufficient to have all the functions or functional blocks of the electronic device 20.

[0200] It can be understood that the water accumulation detection method, electronic device 20 and storage medium for underground pipe corridors provided in the embodiment of the present application can dynamically select more high-quality positive samples during the training process by introducing an adaptive positive sample selection algorithm based on a Gaussian kernel function, significantly improving the training effect of the model and the final detection accuracy. In addition, the intersection-over-union-perceived category label combined with the progressive bounding box correction network can effectively alleviate the problem of inaccurate bounding box quality assessment, reduce the probability of erroneous filtering, and further improve the positioning accuracy and classification confidence. At the same time, the feature extraction backbone network based on the attention mechanism and path aggregation structure enhances the algorithm's detection ability for multi-scale targets by transmitting and fusing feature maps of different levels and highlighting the features of the area of ​​interest. It effectively solves the challenges of poor lighting environment and unclear boundaries of waterlogged areas in underground pipe corridors, so that the water accumulation detection method for underground pipe corridors provided in the present application can still work stably and accurately in complex environments.

[0201] Those skilled in the art should recognize that the above embodiments are only used to illustrate the present application and are not intended to be limiting of the present application. As long as they are within the spirit and scope of the present application, appropriate changes and modifications to the above embodiments are within the scope of protection claimed in the present application.

Claims

1. A method for detecting water accumulation in an underground pipe gallery, characterized in that: The method comprises: Acquire an image sample of water accumulation in an underground pipe gallery, wherein the image sample includes a real bounding box; A positive sample candidate region heat map is generated according to the two-dimensional Gaussian kernel function and the image sample. The heat map generation formula is: ； in, is the center point of the heat map, is the center point of the ground-truth bounding box, , , is the true bounding box width, is the real bounding box height; Adding pixel points whose distance from the center point of the true bounding box is less than a preset distance to the final sample set; Obtaining a positive sample quantity indicator value and a regression loss according to the true bounding box and the final sample set; Adding the pixel point of the positive sample quantity indicator value with the lowest regression loss to the final sample set; The final sample set is processed according to the basic feature extraction backbone network and the attention correction module to obtain the input features of the underground pipe gallery water accumulation. The attention correction module includes a variable convolution and a dynamic linear correction. The variable convolution is used to improve the spatial attention, and the dynamic linear correction is used to improve the channel attention. The formula of the variable convolution is: ； in, represents a position on the feature map, Indicates the deformable convolution kernel at the feature map position The calculation results on Represents the deformable convolution kernel relative to the position The adaptive offset of is the convolution kernel parameter, Indicates that the shielding factor ranges from between, Convolution kernel sampling relative to position The offset of Indicates the response value of the corresponding position on the feature map; The formula for the dynamic linear correction is: ； ； in, is the input feature map tensor, Representation feature map No. The activation value calculated by the channel; linear parameters According to the input feature map Adaptively generated, including dynamic linear activation function The slope and the intercept , Indicates the number of groups of linear parameters, the default value is 2; The input features of underground pipe gallery water accumulation are processed according to bounding box regression and heat map prediction to obtain candidate areas; The candidate region is input into a classification network, and the classification network comprises the following steps: Predicting the classification confidence of the candidate region to obtain a preliminary classification result; Performing an IoU-aware classification label adjustment on the preliminary classification result to generate an IoU-aware classification confidence; Combining the classification confidence with the foreground confidence to generate a final classification confidence; Based on the final classification confidence, non-maximum suppression is performed, overlapping areas are removed, and the final detection results are output.

2. The method for detecting water accumulation in an underground pipe gallery according to claim 1, characterized in that: The method further comprises: The cross entropy loss function is applied to supervise the heat map training, and the expression of the cross entropy loss function is: ； in, Indicates the number of pixels in the heat map, and Respectively represent the real heat map and the predicted heat map at the location The value at represents the positive sample point, and is a hyperparameter.

3. The method for detecting water accumulation in an underground pipe gallery according to claim 2, characterized in that: The method further comprises: Gaussian blur processing is performed on the heat map to smooth the heat map. The formula for the Gaussian blur processing is: ； in, is the thermal map value after fuzzy processing, is the variance.

4. The method for detecting water accumulation in an underground pipe gallery according to claim 1, characterized in that: The method further comprises: The input features are batch normalized to accelerate convergence and improve detection accuracy. The batch normalization formula is: ； in, is the mean of the data, is the variance of the data, and is a trainable parameter, is a minimum value.

5. The method for detecting water accumulation in an underground pipe gallery according to claim 1, characterized in that: The method further comprises: The feature map is processed by applying an attention gate, and different weights are applied to different regions in the feature map to improve the feature response of the region of interest and suppress invalid feature regions. The expression of the attention gate is: ； ； in, is a pixel of the input feature map, is the gate signal vector, used to indicate the pixel The area of ​​focus, is the ReLU activation function, is the sigmoid activation function, for The corresponding attention weights are, are the learnable parameters involved in the attention module, including linear transformation Convolution coefficients , and , and bias and .

6. The method for detecting water accumulation in an underground pipe gallery according to claim 5, characterized in that: The formula for the final classification confidence is: ； in, is the number of categories, is the foreground confidence, The classification confidence for each category.

7. The method for detecting water accumulation in an underground pipe gallery according to claim 1, characterized in that: The image sample includes time sequence information and spatial information, and the method further includes: A three-dimensional convolutional neural network is applied to obtain the spatiotemporal features in the image sample. The formula of the three-dimensional convolutional neural network is: ； Where X is the input three-dimensional data, is the height of the three-dimensional tensor, is the width of the three-dimensional tensor, is the number of input channels, K is the weight of the three-dimensional convolution, is the height of the convolution kernel, is the width of the convolution kernel, D is the depth of the time dimension, is the number of input channels, t is the position index of the time dimension, i is the position index of the height dimension, j is the position index of the width dimension, and c is the index of the output channel. is a scalar representing the bias of the cth output channel, Y is the output tensor, m is the offset in the height dimension, n is the offset in the width dimension, and d is the depth offset in the time dimension.

8. The method for detecting water accumulation in an underground pipe gallery according to claim 6, characterized in that: The image sample includes time sequence information and spatial information, and the method further includes: Apply graph convolutional networks and spatiotemporal features to obtain the changing trend of waterlogging; The graph convolutional network formula is: ； in, , , is the adjacency matrix with self-loops added, For self-loop, is the adjacency matrix, is the degree matrix, is the learnable weight matrix, is the activation function; The graph convolutional network is used to detect water accumulation based on spatiotemporal features.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; Wherein, the processor is configured to implement the water accumulation detection method for underground pipe corridors as described in any one of claims 1 to 8 when executing the instructions.

10. A computer-readable storage medium, characterized in that: It includes instructions, and the instructions instruct the device to execute the water accumulation detection method for underground pipe gallery as described in any one of claims 1 to 8.

Images