Automatic accounting method for desilting project amount in road construction

Abstract

The invention discloses an automatic accounting method for a desilting project amount in road construction, and the method comprises the following steps: S1, carrying out the aerial survey of a road construction region before and after desilting through an unmanned plane, collecting original data, and collecting a range file for limiting the boundary of the desilting region; s2, preprocessing the original data and executing range file verification; s3, a triangulation network ground model is constructed according to the range file and the ground point cloud data, and a residual neural network model is used for desilting area identification; s4, a spider population algorithm is adopted, and each spider individual corresponds to a group of residual network structure parameters; s5, the positions of the spider individuals are updated according to the optimal parameters of the spider individuals and the optimal parameters of the adjacent spider individuals; s6, pixel-level segmentation is carried out on the dredging area based on the residual neural network structure parameters; and S7, generating an achievement preview file based on the accounting result. According to the method, the residual neural network model, the spider gregarious algorithm and the three-dimensional volume estimation technology are combined, and automatic accounting of the desilting project amount in highway construction is achieved.

Description

Technical Field

[0001] The present invention relates to the field of highway construction, and particularly to an automatic calculation method for the amount of dredging work in highway construction. Background Art

[0002] In highway construction projects, the calculation of project quantities is not only the basis for project cost estimation, but also an important basis for project management, progress control, and quality assessment. Among them, the dredging project, as an important part of the basic construction stage, the accurate calculation of its project quantity is directly related to the rationality of the overall project budget and the scientific nature of construction scheduling. However, in the existing technology, the acquisition and calculation methods of the dredging project quantity still mainly rely on traditional manual measurement methods, facing problems in terms of efficiency, accuracy, safety, and data integrity, seriously restricting the development of the informatization and automation level of modern highway engineering construction.

[0003] The traditional method for calculating the amount of dredging work mainly relies on on-site manual measurement. Usually, GPS-RTK equipment is used to collect points in the field, and the topographic information of the area to be dredged is obtained through point layout measurement. Then, combined with manual calculation or CAD modeling software for cross-section estimation, the volume calculation is finally completed. There are significant bottlenecks in this operation process. First of all, the data collection efficiency is extremely low. In large linear engineering scenarios, such as highway, mountain tunnel entrances and exits, or dredging areas around culverts, it not only takes a long time and has sparse coverage to complete the global point layout manually, but also the sampling density cannot meet the requirements of refined measurement. To improve the accuracy of project quantity calculation, a large amount of point information is required, but it is almost impossible to complete such high-density sampling under the traditional method, and the project progress is thus seriously affected.

[0004] Secondly, there are obvious safety hazards in the traditional method. The dredging construction site is often accompanied by soft soil, accumulated water, and silt deposits. Especially in scenarios such as river reconstruction and ditch pond excavation, the on-site environment is complex. Manual measurement is extremely easy to get stuck in the mud, which not only threatens the personal safety of the operators, but also increases the equipment failure rate. In addition, due to the drastic changes in the terrain of the dredging area, it is difficult to place and fix the instruments, and it is easily interfered by the mud, resulting in large deviations in the measurement data and frequent repeated operations, making it difficult to guarantee the measurement accuracy.

[0005] Thirdly, the data quality under the traditional measurement mode is difficult to systematically guarantee, and there are problems such as a large space for human intervention and untraceable errors. The recording, transmission, and processing of measurement results often rely on manual transcription and later manual modeling. The data integrity is poor and the logical association is weak. It is extremely easy to cause information loss or error accumulation due to human negligence, thus affecting the accuracy and credibility of the entire dredging volume calculation. In the context of the increasing requirements for digital engineering construction and the whole-process supervision of construction, relying on the traditional manual method for dredging volume calculation obviously cannot meet the accuracy, efficiency, and intelligence requirements of modern projects.

[0006] Based on this, in the existing technology, exploration paths for obtaining three-dimensional data by introducing means such as unmanned aerial vehicle (UAV) aerial survey, lidar scanning, and point cloud modeling have gradually emerged. However, most methods still remain at the data acquisition stage and have not yet formed an automated engineering quantity accounting system integrating "acquisition - processing - recognition - calculation - upload". Especially in the aspect of identifying dredging engineering quantities, the following deficiencies still exist: First, the identification of the dredging area boundary relies on manual annotation and cannot adapt to complex environments such as different landforms, materials, and occlusion situations. Second, there is a lack of an intelligent structural parameter optimization mechanism in the modeling link after the fusion of point cloud and image data. The generalization ability of the recognition model is weak, and it is easy to misjudge or miss the dredging range in complex terrains. Third, the calculation of the dredging volume fails to effectively combine factors such as recognition accuracy, boundary fitting, and point density. The volume calculation still relies on static rules or empirical models and lacks intelligent feedback capabilities.

[0007] Therefore, how to provide an automatic accounting method for dredging engineering quantities in highway construction is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0008] An object of the present invention is to propose an automatic accounting method for dredging engineering quantities in highway construction. The present invention combines a residual neural network model, a spider swarm algorithm, and three-dimensional volume estimation technology. By constructing a multi-layer convolutional structure to extract the spatial boundary features of the dredging area and using skip connections to enhance the edge information retention ability, high-precision recognition of the dredging area under complex terrains is achieved. At the same time, a swarm intelligence optimization algorithm that simulates the vibration behavior of social spiders is introduced to dynamically optimize the structural parameters of the residual network, enabling the model to adapt to the recognition requirements in different dredging scenarios. The system establishes a spatial voxel structure based on three-dimensional point cloud data, calculates the volume based on the recognition results, and realizes automatic accounting and result preview. This method has the characteristics of high recognition accuracy, strong parameter self-optimization ability, rigorous volume calculation, and full-process automation, effectively improving the efficiency and reliability of engineering quantity accounting.

[0009] An automatic accounting method for dredging engineering quantities in highway construction according to an embodiment of the present invention includes the following steps: S1. Conduct aerial surveys on the highway construction area before and after dredging using a UAV, collect raw data, and collect range files that define the boundary of the dredging area; S2. Preprocess the raw data and perform range file verification to generate ground point cloud data of the dredging area; S3. Construct a triangular network ground model based on the range file and the ground point cloud data, and input the triangular network ground model into the residual neural network model to output an initial mask map of the dredging area. Each residual module of the residual neural network model extracts the features of the dredging area through convolutional layers and retains the spatial boundary information through skip connection technology; S4. Adopt the spider swarm algorithm to guide the search direction of the optimal parameters according to the dredging area recognition accuracy and the initial mask map, where each spider individual in the spider swarm algorithm corresponds to a set of residual network structure parameters; S5. The spider individuals update their positions according to their own and neighboring spider individuals' optimal parameters until the population converges, and output the optimized residual neural network model structure parameters; S6. Based on the optimized residual neural network model, perform pixel-level segmentation on the dredging area, calculate the three-dimensional total volume of the dredging area in combination with the point cloud space coordinate information, and output the accounting result of the dredging project quantity; S7. Generate a result preview file based on the accounting result.

[0010] Optionally, the original data specifically includes RGB images, lidar point cloud data, ground check point data, spatial coordinates, inertial measurement units, flight path planning, and timestamp markings.

[0011] Optionally, the dredging area features specifically include height changes, area undulations, and dredging traces.

[0012] Optionally, S2 specifically includes: S21. Convert the format of the original data into a common.las ground point cloud data format; S22. Conduct quality inspection on the ground point cloud data, use the collected ground check point data for precise point cloud inspection and accounting, and remove noise points and abnormal data through filtering algorithms; S23. The construction unit, based on the scope file, checks whether the scope is qualified according to the set measurement criteria including area size, construction depth, and volume calculation standards.

[0013] Optionally, S3 specifically includes: S31. Read the specified area boundary data from the scope file to determine the ground area range to be processed; S32. Use the triangulation algorithm to triangulate the processed ground point cloud data, connect every three adjacent ground points into a triangle, and generate a triangular mesh ground model covering the ground; S33. Crop the triangular mesh ground model according to the boundary provided by the scope file, remove the part outside the boundary, and only retain the ground data within the dredging area; S34. For the cropped triangular mesh ground model, eliminate triangles with an area greater than 5 square meters and acute triangles with the smallest interior angle less than 20 degrees by adjusting the size and angle of the triangles; S35. Input the adjusted triangular mesh ground model into the residual neural network model and output the initial mask map of the dredging area.

[0014] Optionally, S35 specifically includes: S351. The triangular mesh ground model is used as input data and passed into the residual neural network model. The convolutional layer extracts the dredging area features from the input triangular mesh ground model using nodes. ; Among them, is the dredging area feature of node i in the l-th layer of the residual neural network model, is the set of neighbor nodes of node i, and i is a node in the triangular mesh ground model, is the layer's convolutional kernel weight matrix for weighting nodes i and j, is the degree of node i, is the degree of node j, is the layer's dredging area feature of node j, is the layer's bias term; S352. In the convolutional operation, the skip connection technique is adopted to retain the spatial boundary information of the dredging area. The skip connection allows the input data to be directly transmitted to the output data stream. S353. Multiple feature maps are generated through the convolutional operation. The feature maps contain different spatial features of the dredging area, including the morphology of the dredging area, the fineness of the boundary, and the depth change. S354 flattens the two-dimensional or three-dimensional feature maps generated by the convolutional operation, that is, converts the pixel values of each feature map into a one-dimensional vector in sequence, and then uses the one-dimensional vector as input and passes it to the fully connected layer. S355. The fully connected layer performs matrix multiplication operations on each received input node and the set weight matrix and generates the initial mask map of the dredging area.

[0015] Optionally, S4 specifically includes: S41. Each spider individual corresponds to a set of residual network structure parameters, including the convolutional kernel size, the number of residual modules, the number of feature channels, and the learning rate. S42. The fitness value of each spider individual is determined by comparing the ratio of the intersection and union between the generated initial mask map and the manually annotated dredging area. If the ratio is greater than 0.5, it indicates that the recognition result of the residual neural network model is close to the actual dredging area and the fitness value is high. S43. The spider population simulates the vibration signal propagation behavior of social spiders according to the fitness value of each spider individual, so that the spider individuals with high fitness values guide the spiders to move towards the areas with high fitness values. S44. The spider individual guides the search direction of the optimal parameters by guiding the spider to move towards the area with a high fitness value: ; Among them, is the propagation state of the vibration signal within the spider population, is the search guidance direction vector of the spider individual p in the propagation state , is the set of all nodes in the residual neural network model graph structure. i is a node in the triangular mesh ground model, and N(i) is the set of adjacent nodes of node i. is the dredging area feature of node i in the -th layer of the residual neural network model, is the dredging area feature of node j in the -th layer of the residual neural network model. j is a neighbor node adjacent to node i. is the guidance weight of node j to node i in the propagation state .

[0016] Optionally, the S5 specifically includes: S51. In each round of iteration, the spider individual compares the current fitness value with the fitness value of the neighborhood-optimal spider individual, and based on the current individual position in the structural parameter space, performs a round of position update operation according to the guidance direction of adjacent individuals; S52. The position of each spider individual represents a set of structural parameters of the residual neural network model. The spider individual updates its position according to its own and the optimal fitness values of adjacent individuals: ; Among them, is the position of spider individual p in the t-th generation, is the best position obtained by spider individual q in history, is the globally best position obtained in the entire spider population, , and are acceleration constants, is the position of spider individual p in the (t + 1)-th generation, is the set of neighborhood individuals of spider individual p. p is the individual number in the spider population, and g is the number of the currently globally optimal spider individual. is the coupling influence weight of spider individual q on p in the current iteration, is the estimation of the long-term average drift trend of historical position changes, is the position change amount of spider individual p in the i-th iteration. t is the current iteration round number, is the number of historical iterations; S53. The spider individual updates its position through multiple iterations and adjusts the search direction according to the recognition accuracy after each iteration until the population reaches the maximum number of iterations, which is 100. S54. During each iteration, the spider individual updates its position based on the recognition accuracy of the dredging area and the initial mask map, and simultaneously monitors the improvement of the recognition effect of the dredging area. S55. After the iteration is completed, the optimized residual neural network model structure parameters are output.

[0017] Optionally, the specific steps of S6 are as follows: S61. Based on the optimized residual neural network model, perform layer-by-layer convolution operations on the triangular mesh ground model of the dredging area to obtain the classification results of each pixel point, and output a binary mask map. Each pixel position only takes two labels: the dredging area or the non-dredging area. S62. Extract all the pixel points marked as the dredging area in the binary mask map, and match them with the corresponding coordinate points in the ground point cloud data to construct a spatial point set of the dredging area. The spatial point set of the dredging area consists of the three-dimensional coordinates corresponding to all the points determined to be in the dredging area in the binary mask map. S63. Divide the three-dimensional space with a fixed side length to construct a regular voxel grid. Project and map the spatial point set of the dredging area into the voxel grid, and determine whether each voxel contains at least one valid spatial point, and count the number of all occupied voxel units. S64. Calculate the volume of each voxel according to the side length of the voxel unit, and sum up the volumes of all occupied voxel units to obtain the three-dimensional total volume of the dredging area: ; Where, is the three-dimensional total volume of the dredging area, K is the total number of all divided spatial voxels, is the spatial division accuracy threshold, is the index of a valid spatial point in the kth spatial voxel unit, k is the kth spatial voxel unit obtained by dividing the space, is the th volume weighting factor of the point in the kth spatial voxel, is the spatial volume of the kth spatial voxel unit, is the number of valid points contained in the kth spatial voxel, is the position of the spider individual p in the (t + 1)th generation; S65. Based on the edge pixel points of the dredging area in the binary mask map, extract the boundary contour of the dredging area and correspond it with the three-dimensional spatial point set coordinates to form the boundary map of the dredging area. S66. Organize the recognized binary mask image, the boundary map of the dredging area, and the three-dimensional total volume of the dredging area into a unified data structure, and output it as the accounting result of the dredging project volume.

[0018] Optionally, the S7 specifically includes: S71. Output the binary mask image in the accounting result of the dredging project volume as two-dimensional image data. Specifically, encode the binary mask image into a layer image format file, and perform spatial registration and color overlay with the image obtained by aerial survey of the highway construction area before dredging by the UAV to generate visual two-dimensional image data. S72. Package the boundary map of the dredging area, the three-dimensional total volume of the dredging area, and the visual two-dimensional image data together into a structured data result file. S73. Generate a unique task identification number for each group of data in the structured data result file, bind the structured data result file with the task identification number, and upload it to the remote database. S74. Index in the remote database according to the task identification number, load the structured data result file on the user interface, display the visual two-dimensional image data as the background map, overlay and display the boundary map of the dredging area in the form of a wireframe, and display the statistical result of the three-dimensional total volume of the dredging area on the numerical panel. The three together constitute the result visualization display structure. S75. Generate a result preview number for each group of data in the structured data result file according to the dredging identification time, and package the visual two-dimensional image data, the boundary map of the dredging area, and the three-dimensional total volume of the dredging area together with the corresponding task identification number and result preview number to generate a result preview file. S76. The result preview file is displayed in the accounting interface in a unified format, and supports online viewing, downloading and archiving, and task archiving operations through the result preview number.

[0019] The beneficial effects of the present invention are: The present invention proposes an automatic accounting method and tool for dredging project volume in highway construction, which systematically solves the prominent problems existing in the traditional dredging project volume accounting process, such as low efficiency, poor accuracy, large potential safety hazards, and data non-traceability, and significantly improves the scientific and intelligent level of the project volume calculation in the dredging construction link. By integrating key technical means such as UAV aerial survey, high-precision point cloud acquisition, triangular network ground modeling, residual neural network recognition, and spider colony algorithm optimization, the present invention constructs a full-process automated quantity calculation system from raw data acquisition, recognition and modeling, project volume accounting to result output, effectively reducing the dependence on manual operations in the traditional quantity calculation method, and improving the data processing efficiency and the objectivity of the calculation results.

[0020] In terms of efficiency, the present invention uses a drone system equipped with a lidar or an image sensor to quickly cover and conduct aerial surveys of the dredging area, achieving the rapid acquisition of large-scale and high-density point cloud data. Compared with the traditional GPS-RTK based point collection method, the aerial survey operation not only significantly reduces the labor input, but also greatly shortens the field measurement time. At the same time, in the in-house processing stage, the present invention combines deep learning recognition and three-dimensional modeling analysis to achieve efficient and batch calculation of the dredging project quantity, improving the overall project scheduling efficiency.

[0021] In terms of precision control, the present invention conducts multi-level recognition of the spatial features in the triangulated irregular network (TIN) ground model based on a residual neural network, and adaptively optimizes the network structure parameters through the spider swarm algorithm, enabling the model to maintain strong recognition generalization ability when facing different terrain structures and different dredging scenarios. The optimized neural network structure makes the boundary extraction of the dredging area more accurate, and the three-dimensional volume calculation results completed in combination with the point cloud coordinate system are also more consistent and stable, effectively avoiding the problem of underestimated or overestimated volume caused by empirical judgment deviation or data collection error in traditional manual calculations.

[0022] In terms of safety guarantee, the present invention fully considers that the dredging operation area is mostly a high-risk environment such as muddy and waterlogged areas, and uses remote aerial surveys to replace the traditional manual on-site measurement method, greatly reducing the necessity of personnel entering dangerous areas and improving the safety protection level during the operation. At the same time, the original data collected and the calculation results are uploaded to the cloud computing platform throughout the process, with automatic data recording and transmission, having good traceability and integrity, and avoiding the risks of data fraud, omission or tampering caused by human intervention. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification, and are used to explain the present invention together with the embodiments of the present invention, and do not constitute a limitation to the present invention. In the drawings: Figure 1 is a flowchart of an automatic calculation method for the dredging project quantity in highway construction proposed by the present invention; Figure 2 is a schematic diagram of an automatic calculation method for the dredging project quantity in highway construction proposed by the present invention; Figure 3 is a data flow diagram of an automatic calculation method for the dredging project quantity in highway construction proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0024] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0025] Reference Figures 1-3 , an automatic accounting method for the amount of silt cleaning work in highway construction, comprising the following steps: S1. Conduct aerial surveys of the highway construction area before and after silt cleaning by using an unmanned aerial vehicle (UAV), collect original data, and collect range files defining the boundaries of the silt cleaning area; S2. Preprocess the original data and perform range file verification to generate ground point cloud data of the silt cleaning area; S3. Construct a triangular network ground model based on the range file and the ground point cloud data, input the triangular network ground model into a residual neural network model, and output an initial mask map of the silt cleaning area. Each residual module of the residual neural network model extracts the features of the silt cleaning area through a convolutional layer and retains the spatial boundary information through a skip connection technique; S4. Adopt a spider swarm algorithm to guide the search direction of optimal parameters according to the recognition accuracy of the silt cleaning area and the initial mask map, where each spider individual in the spider swarm algorithm corresponds to a set of residual network structure parameters; S5. The spider individuals update their positions according to the optimal parameters of themselves and neighboring spider individuals until the population converges, and output the optimized residual neural network model structure parameters; S6. Based on the optimized residual neural network model, perform pixel-level segmentation on the silt cleaning area, calculate the three-dimensional total volume of the silt cleaning area in combination with the point cloud spatial coordinate information, and output the accounting result of the amount of silt cleaning work; S7. Generate a result preview file based on the accounting result.

[0026] The present invention optimizes the residual neural network structure parameters by using a spider swarm algorithm, guides the structure search direction by simulating the cooperative vibration behavior between spider individuals, and realizes the adaptive structure optimization of the network in the silt cleaning area recognition task. The system constructs a triangular network ground model based on the point cloud data obtained by UAV aerial survey. The residual neural network extracts spatial features layer by layer and retains boundary information through skip connections, and outputs an initial mask map. Combining pixel-level recognition and three-dimensional space reconstruction technologies, the volume calculation of the silt cleaning area is accurately completed. An automatic result preview file is generated, comprehensively improving the intelligent, automated, and precise level of engineering quantity calculation.

[0027] In this embodiment, the original data specifically includes RGB images, lidar point cloud data, ground check point data, spatial coordinates, inertial measurement units, flight path planning, and timestamp markings.

[0028] In the present invention, the original data includes RGB images, lidar point cloud data, ground check point data, spatial coordinates, inertial measurement unit information, flight path planning, and timestamp markings, to construct a data foundation for multi-source fusion. By introducing a spatio-temporal collaborative calibration mechanism, precise alignment of images and point clouds in the same coordinate system is achieved, enhancing the spatial consistency of subsequent modeling and recognition. Utilizing the timestamp and path information, the acquisition process can be effectively traced, supporting data backtracking and verification, improving the accuracy and controllability of overall data processing, and providing a reliable guarantee for the high-precision automatic calculation of dredging quantities.

[0029] In this embodiment, the dredging area features specifically include height changes, regional undulations, and dredging traces.

[0030] In the present invention, the dredging area features include height changes, regional undulations, and dredging traces, to construct a multi-dimensional spatial feature expression system. By capturing terrain height differences and local slope information, combined with the surface disturbance characteristics after dredging operations, precise extraction of the dredging area boundary is achieved. This method avoids relying solely on a single height value for judgment, improving the robustness and integrity of area recognition. After multi-feature fusion is input into the residual neural network, the model's perception ability of complex geomorphic changes is enhanced, effectively improving the accuracy of dredging recognition, and providing an accurate basis for volume calculation and construction assessment.

[0031] In this embodiment, S2 specifically includes: S21: Convert the format of the original data into a common.las ground point cloud data format; S22: Conduct quality inspection on the ground point cloud data, use the collected ground check point data for precise inspection and calculation of the point cloud, and remove noise points and abnormal data through filtering algorithms; S23: Based on the scope file, the construction unit checks whether the scope is qualified according to the set measurement criteria including regional dimensions, construction depth, and volume calculation standards.

[0032] The present invention realizes data quality guarantee before calculating dredging quantities by constructing a point cloud data standardization and automatic verification process. The original data is uniformly converted into a common.las format to ensure multi-platform compatibility, and precise verification is carried out in combination with ground check points, effectively eliminating abnormalities and non-ground points. The system uses filtering and classification algorithms for noise removal and thinning processing to extract stable ground point clouds, improving the integrity and accuracy of ground modeling. At the same time, the scope file submitted by the construction unit is automatically compared with the measurement criteria in the quantity calculation platform, covering the regional scope, depth setting, and volume calculation standards, to achieve precise verification of the calculation boundary. This method improves the data consistency, accuracy, and intelligence level in the point cloud preprocessing stage, provides high-quality basic data support for subsequent recognition and volume calculation, and ensures that the quantity calculation is more standardized and reliable.

[0033] In this embodiment, S3 specifically includes: S31. Read the specified area boundary data from the range file to determine the ground area range to be processed; S32. Use the triangulation algorithm to triangulate the processed ground point cloud data, connect every three adjacent ground points into a triangle, and generate a triangular mesh ground model covering the ground; S33. Trim the triangular mesh ground model according to the boundary provided by the range file, remove the part outside the boundary, and only retain the ground data within the dredging area; S34. For the trimmed triangular mesh ground model, eliminate triangles with an area greater than 5 square meters and acute triangles with a minimum interior angle less than 20 degrees by adjusting the size and angle of the triangles; S35. Input the adjusted triangular mesh ground model into the residual neural network model to output the initial mask map of the dredging area.

[0034] The present invention realizes the spatial structured expression of ground point cloud data by constructing a triangular mesh ground model, providing geometric support for the identification of the dredging area. The system reads the dredging boundary information from the range file, accurately locates the area to be processed, and uses the triangulation algorithm to convert the ground point cloud into a grid structure, improving the data processing efficiency and the integrity of spatial expression. By automatically trimming the model boundary, only the valid data within the dredging area range is retained, avoiding interference from irrelevant terrain and enhancing the recognition focus. At the same time, the system optimizes the morphology of the triangular units, eliminates acute angles and long structures, and improves the stability in the subsequent recognition stage. The residual neural network takes the triangular mesh model as the input, deeply extracts spatial features, and realizes the accurate identification of the dredging area in complex landforms. This method improves the spatial modeling accuracy, data processing standardization and recognition effect, laying a high-quality foundation for the calculation of the dredging volume.

[0035] In this embodiment, S35 specifically includes: S351. The triangular mesh ground model is passed into the residual neural network model as input data, and the convolutional layer uses nodes to extract the dredging area features from the input triangular mesh ground model; ; Among them, is the dredging area feature of node i in the l-th layer of the residual neural network model, is the set of neighbor nodes of node i, i is a node in the triangular mesh ground model, is the layer's convolutional kernel weight matrix for weighting nodes i and j, is the degree of node i, is the degree of node j, is the The dredging area feature of node j in the layer is the bias term of the layer; S352. In the convolution operation, adopt the skip connection technology to retain the spatial boundary information of the dredging area. The skip connection allows the input data to be directly transmitted to the output data stream; S353. Generate multiple feature maps through convolution operations. The feature maps contain different spatial features of the dredging area, including the morphology of the dredging area, the fineness of the boundary, and the depth change; S354. Flatten the two-dimensional or three-dimensional feature maps generated by the convolution operation, that is, convert the pixel values of each feature map into a one-dimensional vector in sequence, and then use the one-dimensional vector as the input to be transmitted to the fully connected layer; S355. The fully connected layer performs matrix multiplication operations on each received input node and the set weight matrix and generates an initial mask map of the dredging area.

[0036] In the present invention, by using the triangular network ground model as the input and combining the convolution structure of the residual neural network model to extract the dredging area features, the depth recognition of the dredging boundary in complex landforms is realized. The convolution layer aggregates the spatial adjacent node features at multiple levels in the network, and uses the skip connection technology to enhance the ability to maintain boundary information and prevent features from being lost during the downsampling process. The multi-dimensional feature maps output by the network comprehensively express the morphological structure and edge changes of the dredging area. Through the fully connected layer, the multi-dimensional features are flattened and matrix operations are performed to extract significant features and generate an initial mask map. When constructing the dredging area recognition model, this method not only considers the spatial structure, but also combines the expression ability of different level features, enhances the recognition robustness of the model to irregular terrain, local perturbations and regional boundaries, significantly improves the recognition accuracy and spatial consistency, and provides a stable input for subsequent volume calculation.

[0037] In this embodiment, the specific steps of S4 include: S41. Each spider individual corresponds to a set of residual network structure parameters, including the convolution kernel size, the number of residual modules, the number of feature channels, and the learning rate; S42. The fitness value of each spider individual is determined by comparing the ratio of the intersection and union between the generated initial mask map and the manually marked dredging area. If the ratio is greater than 0.5, it indicates that the recognition result of the residual neural network model is close to the actual dredging area and the fitness value is high: ; where IoU is the fitness value, Z is the initial mask map, and G is the manually marked dredging area; S43. The spider swarm simulates the vibration signal propagation behavior of social spiders according to the fitness value of each spider individual, enabling the spider individuals with high fitness values to guide the spiders to move towards the area with high fitness values; S44. The spider individual guides the search direction of the optimal parameters by guiding the spiders to move towards the area with high fitness values: ; Among them, is the vibration signal propagation state within the spider swarm, is the search guidance direction vector of spider individual p under the propagation state , is the set of all nodes in the residual neural network model graph structure. i is a node in the triangular mesh ground model, and N(i) is the set of adjacent nodes of node i. is the dredging area feature of node i in the -th layer of the residual neural network model, is the dredging area feature of node j in the -th layer of the residual neural network model. j is a neighbor node adjacent to node i. is the guidance weight of node j to node i under the propagation state .

[0038] The present invention introduces a spider colony algorithm to intelligently optimize the structural parameters of the residual neural network model. Each spider individual corresponds to a set of structural parameter combinations, covering core elements such as the convolutional kernel size, the number of residual modules, the number of feature channels, and the learning rate. By calculating the intersection over union between the initial mask graph and the manually labeled dredging area, a fitness evaluation mechanism is established. Individuals with high fitness will generate stronger vibration signals, simulating the information propagation behavior of social spiders, and guiding the remaining individuals to move towards the optimal solution area. Spider individuals perform search and update based on the feature expression of the residual neural network model in the graph structure, dynamically adjusting the structural combination direction, and achieving global convergence and adaptive optimization of the structural parameters. This method avoids the uncertainty and low efficiency of manual parameter adjustment, improves the robustness and generalization ability of the dredging area recognition model in different scenarios, and provides a more accurate recognition result for subsequent volume calculation.

[0039] In this embodiment, the specific steps of S5 are as follows: S51. In each round of iteration, the spider individual compares the current fitness value with the fitness value of the optimal spider individual in the neighborhood, and based on the current individual's position in the structural parameter space, performs a round of position update operation according to the guidance direction of neighboring individuals; S52. The position of each spider individual represents a set of structural parameters of the residual neural network model. The spider individual updates its position according to the optimal fitness values of itself and neighboring individuals: ; wherein, is the position of spider individual p in the t-th generation, is the best position obtained by spider individual q in history, is the globally best position obtained in the entire spider population, , and are acceleration constants, is the position of spider individual p in the (t + 1)-th generation, is the set of neighboring individuals of spider individual p. p is the individual number in the spider population, and g is the individual number of the currently globally optimal spider, is the coupling influence weight of spider individual q on p in the current iteration, is the estimation of the long-term average drift trend of the historical position change, is the position change amount of spider individual p in the i-th iteration. t is the current iteration round number, is the number of historical iterations; S53. The spider individuals update their positions through multiple iterations and adjust the search direction according to the recognition accuracy after each iteration until the population reaches the maximum number of iterations of 100; S54. In each iteration process, the spider individuals update their positions according to the dredging area recognition accuracy and the initial mask map, and at the same time monitor the improvement of the dredging area recognition effect; S55. After the iteration is completed, the optimized residual neural network model structure parameters are output.

[0040] In the present invention, the structure parameters of the residual neural network are iteratively optimized through the spider swarm algorithm to construct a multi-dimensional parameter search mechanism driven by fitness. In each round of iteration, the spider individuals perform position update operations along the optimal direction in the structure parameter space according to the fitness difference between themselves and the individuals in the neighborhood, forming the coordination of local guidance and global convergence. The individual position represents a set of network structure parameter combinations, including key factors such as convolution kernels and number of channels, and dynamically adjusts the search path in combination with the historical drift trend. The system continuously monitors the intersection over union ratio of the mask map and the manually labeled area by setting the maximum number of iterations and the recognition accuracy threshold to ensure that the network parameters are optimized towards high accuracy. The finally output structure parameters can be directly used for dredging area recognition, significantly improving the structural adaptability and generalization ability of the model, and providing robust and accurate recognition support for the automatic accounting system.

[0041] In this embodiment, the S6 specifically includes: S61. Based on the optimized residual neural network model, perform layer-by-layer convolution operations on the triangular mesh ground model of the dredging area to obtain the classification result of each pixel point, and output a binary mask map. Each pixel position only takes two labels: dredging area or non-dredging area; S62, extracting all pixel points marked as desilting areas in the binary mask image, and matching them with corresponding coordinate points in the ground point cloud data, to construct a desilting area spatial point set, wherein the desilting area spatial point set is composed of the three-dimensional coordinates corresponding to all points determined as desilting areas in the binary mask image; S63, dividing the three-dimensional space with a fixed side length, constructing a regular voxel grid, projecting and mapping the dredging area spatial point set to the voxel grid, determining whether each voxel contains at least one valid spatial point, and counting the number of all occupied voxel units; S64. Calculate the volume of each voxel according to the side length of the voxel unit, and sum up the volumes of all occupied voxel units to obtain the three-dimensional total volume of the desilting area: ; in, is the total three-dimensional volume of the desilting area, K is the total number of all divided spatial voxels, is the spatial partition accuracy threshold, is the index of a valid spatial point in the spatial voxel unit k, k is the kth spatial voxel unit obtained by dividing the space, is the kth spatial voxel The volume weighting factor of each point, is the spatial volume of the kth spatial voxel unit, is the number of valid points contained in the k-th spatial voxel, is the position of spider individual p in the t+1 generation; S65, based on the edge pixel points of the desilting area in the binary mask image, extract the boundary contour of the desilting area, and correspond it to the coordinates of the three-dimensional space point set to form a boundary map of the desilting area; S66, organizing the identified binary mask map, the boundary map of the dredging area and the three-dimensional total volume of the dredging area into a unified data structure, and outputting it as the calculation result of the dredging engineering quantity.

[0042] The present invention performs layer-by-layer convolution on the triangulated ground model through the optimized residual neural network, outputs a binary mask map, and accurately identifies the scope of the dredging area. The system matches the dredging labels in the mask map with the point cloud coordinates one by one, constructs a three-dimensional dredging point set, and realizes the integration of image recognition and spatial modeling. By dividing the three-dimensional space by fixed side lengths, constructing a regular voxel grid, mapping the point set to the voxel unit, counting the number of occupied voxels and calculating the total volume in combination with the side length, the structural accuracy and computational efficiency of volume estimation are improved. The three-dimensional coordinates corresponding to the edge pixels are further extracted, and a boundary map is generated to realize the spatial visualization expression of the dredging area. Finally, the mask map, boundary map and volume data are integrated into a unified accounting result file to ensure the integrity of the results and the readability of the platform, and significantly improve the calculation accuracy and engineering adaptability of the dredging engineering quantity.

[0043] In this embodiment, the S7 specifically includes: S71. Output the binary mask image in the accounting result of the dredging work volume as two-dimensional image data. Specifically, encode the binary mask image into a layer image format file, and perform spatial registration and color overlay with the image obtained by aerial survey of the highway construction area before dredging by the unmanned aerial vehicle to generate visual two-dimensional image data; S72. Package the boundary line map of the dredging area, the three-dimensional total volume of the dredging area and the visual two-dimensional image data into a structured data result file; S73. Generate a unique task identification number for each group of data in the structured data result file, bind the structured data result file with the task identification number, and upload it to the remote database; S74. Index according to the task identification number in the remote database, load the structured data result file on the user terminal interface, display the visual two-dimensional image data as the background image, overlay and display the boundary map of the dredging area in the form of a wireframe, and display the statistical result of the three-dimensional total volume of the dredging area in the numerical panel. The three together constitute the result visualization display structure; S75. Generate a result preview number for each group of data in the structured data result file according to the dredging identification time, and package the visual two-dimensional image data, the boundary map of the dredging area and the three-dimensional total volume of the dredging area together with the corresponding task identification number and result preview number to generate a result preview file; S76. The result preview file is displayed in the accounting interface in a unified format, and supports online viewing, downloading and archiving operations through the result preview number.

[0044] The present invention realizes the standardized output and visual management of the dredging work volume accounting result by constructing a unified result generation and display process. The system registers and overlays the identified binary mask image with the unmanned aerial vehicle image to generate an image layer, enhancing the intuitive presentation of the spatial recognition result. The dredging boundary map and the volume statistical result are packaged into a structured data file together and uploaded to the platform database by binding a unique task number, realizing the task-level archival management of the result. The platform automatically indexes the task number, loads the layer, the boundary line and the numerical information, and constructs a graphic and data integrated result display interface. The result image generates a unique preview number according to the resolution and the recognition time, and is uniformly incorporated into the platform preview and download mechanism, supporting multi-task comparison, archiving and re-call, comprehensively improving the traceability, interactivity and engineering delivery efficiency of the dredging accounting result.

[0045] Example 1: To verify the feasibility of the present invention in implementation, the present invention is applied to the accounting task of a certain section of dredging project. Due to long-term waterlogging, sediment accumulation and slope erosion in this section, multiple ditches, ponds and undulating terrain areas have been formed, seriously affecting the construction traffic and the flatness of the roadbed. It is urgent to carry out dredging construction and synchronously conduct the accounting and quality assessment of the dredging workload. The traditional method of workload accounting requires staff to be equipped with RTK equipment to go deep into the site to collect cross-section points. The points are sparse, the risk is high and the processing cycle is long, which is difficult to meet the dual requirements of the construction unit for data accuracy and accounting efficiency. Therefore, this project is selected as the field test demonstration section of the automatic accounting method of the present invention.

[0046] In this scenario, the operators used a multi-rotor UAV equipped with a lidar and a high-resolution RGB image sensor to conduct two flight surveys of this section before and after dredging. The flight height was 80 meters each time, the flight line overlap was 80%, and the single operation coverage area was about 1.6 square kilometers. Each flight collected about 48GB of original data, including point cloud data, RGB images, flight path planning data, inertial measurement unit records and acquisition timestamps. All data was automatically uploaded to the traffic construction accounting platform deployed in Province A after the flight ended, and the data preprocessing and dredging area identification process began to be executed.

[0047] First, the original data is converted into the standard.las format, and the point cloud is filtered, denoised and ground point extraction is performed. The verification result of the check point shows that the ground error of the point cloud is controlled within ±3 cm, meeting the requirements of engineering survey. Subsequently, the system extracts the point cloud within the range according to the uploaded construction range file and constructs a triangular mesh ground model, and trims the model boundary and optimizes the triangle shape. By loading the residual neural network structure parameters optimized by the present invention, layer-by-layer convolution operation is performed on the ground model to automatically identify the dredging area and generate an initial mask map. By comparing this image with the manually labeled samples, the intersection over union reaches 89.72%, which is significantly higher than 76.85% of the traditional threshold method.

[0048] To further verify the effectiveness of the network structure, the spider swarm algorithm is introduced to iteratively optimize the neural network parameters. During 100 rounds of evolutionary iteration, the recognition accuracy is steadily improved. The final combination of structure parameters includes a convolution kernel size of 5×5, a residual module depth of 7, a feature channel number of 96, and a learning rate of 0.0008. The recognition accuracy of this structure on the validation set reaches 91.08%, which is about 5.4% higher than that of the unoptimized network.

[0049] Finally, by aligning the mask map with the point cloud data, the spatial point set of the dredging area is extracted and projected into the constructed three-dimensional voxel grid to complete the dredging volume accounting. The calculated dredging volume is 4,825.34 cubic meters. After the boundary map is fused with the aerial survey image, it is displayed on the quantity calculation platform, and the result preview file is automatically generated and uploaded to the project supervision side and the design unit synchronously.

[0050] Compared with the traditional method, the data collection stage of the present invention only takes 3 hours, the data processing and recognition calculation takes 5 hours, and the overall accounting cycle is compressed to within 8 hours. In contrast, the traditional measurement method requires about 5 people to spend 3 days to complete, with less than 200 data collection points, low volume calculation accuracy, and obvious errors. In terms of ensuring operation safety, using drones to complete all data collection eliminates the need for personnel to enter the silt area, avoiding high-risk operation scenarios. Moreover, the whole-process data records are traceable and reviewable, greatly improving data transparency and credibility.

[0051] Table 1 Comparison of the optimization effects of the automatic accounting method for the dredging work volume in highway construction ; Table 1 shows the full-process automation, refinement, and structuring of the dredging work volume accounting from data collection to result generation, solving the problems of low efficiency, poor accuracy, and high safety risks in the traditional quantity calculation method, verifying the high adaptability and engineering feasibility of the present invention in complex terrain environments, and having good promotion and application value.

[0052] The above is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, shall be covered by the protection scope of the present invention.

Claims

1. An automatic calculation method for the amount of silt cleaning work in highway construction, characterized in that, It includes the following steps: S1. Conduct aerial surveys before and after dredging in the highway construction area by using a drone, collect original data and collect range files defining the boundaries of the dredging area; S2. Preprocess the original data and perform range file verification to generate ground point cloud data of the dredging area; S3. Construct a triangular network ground model based on the range file and the ground point cloud data, input the triangular network ground model into a residual neural network model, and output an initial mask map of the dredging area. Each residual module of the residual neural network model extracts dredging area features through a convolutional layer and retains spatial boundary information through a skip connection technique; S4. Adopt a spider swarm algorithm to guide the search direction of optimal parameters according to the dredging area recognition accuracy and the initial mask map. Each spider individual in the spider swarm algorithm corresponds to a set of residual network structure parameters; S5. The spider individuals update their positions according to the optimal parameters of themselves and neighboring spider individuals until the population converges, and output the optimized residual neural network model structure parameters; S6. Perform pixel-level segmentation on the dredging area based on the optimized residual neural network model, calculate the three-dimensional total volume of the dredging area in combination with the point cloud spatial coordinate information, and output the accounting result of the dredging project volume; S7. Generate a result preview file based on the accounting result.

2. The automatic calculation method for the dredging work volume in highway construction according to claim 1, wherein, The original data specifically includes RGB images, lidar point cloud data, ground check point data, spatial coordinates, inertial measurement units, flight path planning, and timestamp markings.

3. The automatic calculation method for the dredging work volume in highway construction according to claim 1, characterized in that, The dredging area features specifically include height changes, regional undulations, and dredging traces.

4. The automatic calculation method for the dredging work volume in highway construction according to claim 1, characterized in that, The S2 specifically includes: S21. Convert the format of the original data into a common.las ground point cloud data format; S22. Conduct quality inspection on the ground point cloud data, use the collected ground check point data for precise point cloud accounting, and remove noise points and abnormal data through a filtering algorithm; S23. The construction unit checks whether the range is qualified based on the range file according to the set measurement criteria including regional dimensions, construction depth, and volume calculation standards.

5. The automatic accounting method for the dredging work volume in highway construction according to claim 1, wherein, The S3 specifically includes: S31. Read the specified regional boundary data from the range file to determine the ground area range to be processed; S32. Use a triangulation algorithm to triangulate the processed ground point cloud data, connect every three adjacent ground points into a triangle, and generate a triangular network ground model covering the ground; S33. Crop the triangular network ground model according to the boundaries provided by the range file, remove the parts outside the boundaries, and only retain the ground data within the dredging area; S34. For the cropped triangular network ground model, eliminate triangles with an area greater than 5 square meters and acute triangles with a minimum interior angle less than 20 degrees by adjusting the size and angle of the triangles; S35. Input the adjusted triangular network ground model into the residual neural network model and output an initial mask map of the dredging area.

6. The automatic calculation method for the dredging work volume in highway construction according to claim 5, wherein, The S35 specifically includes: S351. The triangular network ground model is passed into the residual neural network model as input data, and the convolutional layer uses nodes to extract dredging area features from the input triangular network ground model; ； Among them, is the dredging area feature of node i in the l-th layer residual neural network model, is the set of neighbor nodes of node i, and i is a node in the triangular mesh ground model, is the convolution kernel weight matrix for weighting nodes i and j in the layer, is the degree of node i, is the dredging area feature of node j in the layer, is the bias term of the S352. In the convolution operation, the skip connection technology is adopted to retain the spatial boundary information of the dredging area. The skip connection allows the input data to be directly transmitted to the output data stream; S353. Multiple feature maps are generated through the convolution operation. The feature maps contain different spatial features of the dredging area, including the shape of the dredging area, the fineness of the boundary, and the depth change; S354. Flatten the two-dimensional or three-dimensional feature maps generated by the convolution operation, that is, convert the pixel values of each feature map into a one-dimensional vector in sequence, and then use the one-dimensional vector as the input to be transmitted to the fully connected layer; S355. The fully connected layer performs matrix multiplication on each received input node and the set weight matrix and generates an initial mask map of the silt removal area.

7. The automatic calculation method for the dredging work volume in highway construction according to claim 1, characterized in that, The specific content of S4 includes: S41. Each spider individual corresponds to a set of residual network structure parameters, including the convolution kernel size, the number of residual modules, the number of feature channels, and the learning rate; S42. The fitness value of each spider individual is determined by comparing the ratio of the intersection and union between the generated initial mask map and the manually marked dredging area. If the ratio is greater than 0.5, it indicates that the recognition result of the residual neural network model is close to the actual dredging area, and the fitness value is high; S43. The spider population simulates the vibration signal propagation behavior of social spiders according to the fitness value of each spider individual, so that the spider individuals with high fitness values guide the spiders to move towards the areas with high fitness values; S44. The spider individuals guide the spiders to move towards the areas with high fitness values, guiding the search direction of the optimal parameters: ； Among them, is the propagation state of vibration signals within the spider population, is in the propagation state The search guidance direction vector of spider individual p, is the set of all nodes in the residual neural network model graph structure. i is a node in the triangular mesh ground model, and N(i) is the set of adjacent nodes of node i. is at the The silt removal area feature of node i in the layer of the residual neural network model, is at the The silt removal area feature of node j in the layer of the residual neural network model. j is a neighbor node adjacent to node i. is the propagation state The guidance weight of node j to node i under the condition.

8. The automatic calculation method for the dredging work volume in highway construction according to claim 1, characterized in that, The specific content of S5 includes: S51. In each iteration, the spider individual compares the current fitness value with the fitness value of the neighboring optimal spider individual, and in the structural parameter space, based on the current individual position, performs a round of position update operation according to the guiding direction of the neighboring individuals; S52. The position of each spider individual represents a set of structural parameters of the residual neural network model. The spider individual updates its position according to the optimal fitness values of itself and the neighboring individuals; ； Among them, is the position of spider individual p in the t-th generation, is the best position obtained by spider individual q in history, is the global best position obtained among all spider populations, , and are acceleration constants, is the position of spider individual p in the (t + 1)-th generation, is the set of neighboring individuals of spider individual p. p is the individual number in the spider population, and g is the individual number of the currently globally optimal spider, is the coupling influence weight of spider individual q on p in the current iteration, is the estimation of the long-term average drift trend of historical position changes, is the position change amount of spider individual p in the i-th iteration. t is the current iteration round number, is the number of historical iterations; S53. The spider individuals update their positions through multiple iterations, and adjust the search direction according to the recognition accuracy after each iteration until the population reaches the maximum number of iterations of 100; S54. In each iteration process, the spider individual updates its position according to the dredging area recognition accuracy and the initial mask map, and at the same time monitors the improvement of the dredging area recognition effect; S55. After the iteration is completed, output the optimized residual neural network model structure parameters.

9. The automatic calculation method for the dredging work volume in highway construction according to claim 1, characterized in that The specific content of S6 includes: S61. Based on the optimized residual neural network model, perform a layer-by-layer convolution operation on the triangular mesh ground model of the dredging area to obtain the classification result of each pixel point, and output a binary mask map. Each pixel position only takes two labels: dredging area or non-dredging area; S62. Extract all the pixel points marked as the dredging area in the binary mask map, and match them with the corresponding coordinate points in the ground point cloud data to construct a spatial point set of the dredging area. The spatial point set of the dredging area consists of the three-dimensional coordinates corresponding to all the points determined to be the dredging area in the binary mask map; S63. Divide the three-dimensional space with a fixed side length to construct a regular voxel grid, project and map the spatial point set of the dredging area into the voxel grid, judge whether each voxel contains at least one valid spatial point, and count the number of all occupied voxel units; S64. Calculate the volume of each voxel according to the side length of the voxel unit, and sum up the volumes of all occupied voxel units to obtain the three-dimensional total volume of the dredging area: ； Among them, is the three-dimensional total volume of the dredging area, K is the total number of all divided spatial voxels, is the spatial division accuracy threshold, is the index of a valid spatial point in the spatial voxel unit k, and k is the k-th spatial voxel unit obtained by division in the space, is the volume weighting factor of the -th point in the k-th spatial voxel, is the spatial volume of the k-th spatial voxel unit, is the number of valid points contained in the k-th spatial voxel, is the position of the spider individual p in the (t + 1)-th generation; S65. Based on the edge pixel points of the dredging area in the binary mask map, extract the boundary contour of the dredging area and correspond it to the three-dimensional space point set coordinates to form the boundary map of the dredging area; S66. Organize the recognized binary mask map, the boundary map of the dredging area and the three-dimensional total volume of the dredging area into a unified data structure and output it as the accounting result of the dredging work volume.

10. The automatic accounting method for the dredging work volume in highway construction according to claim 1, wherein, The specific content of S7 includes: S71. Output the binary mask map in the accounting result of the dredging work volume as two-dimensional image data. Specifically, encode the binary mask map into a layer image format file, and perform spatial registration and color overlay with the image obtained by the UAV's aerial survey of the highway construction area before dredging to generate visual two-dimensional image data; S72. Package the boundary map of the dredging area, the three-dimensional total volume of the dredging area and the visual two-dimensional image data into a structured data result file; S73. Generate a unique task identification number for each group of data in the structured data result file, bind the structured data result file with the task identification number, and upload it to the remote database; S74. Index according to the task identification number in the remote database, load the structured data result file on the user interface, display the visual two-dimensional image data as the background map, overlay and display the boundary map of the dredging area in the form of a wireframe, and display the statistical result of the three-dimensional total volume of the dredging area in the numerical panel. The three together constitute the result visualization display structure; S75. Generate a result preview number for each group of data in the structured data result file according to the dredging recognition time, and package the visual two-dimensional image data, the boundary map of the dredging area and the three-dimensional total volume of the dredging area together with the corresponding task identification number and result preview number to generate a result preview file; S76. The result preview file is displayed in the accounting interface in a unified format, and supports online viewing, downloading and archiving, and task archiving operations through the result preview number.

Images