Intelligent visual mountainous highway construction road rapid design method and system

Abstract

This invention provides a smart and visualized method and system for rapid design of construction access roads for mountainous highways. First, it acquires topographic maps of mountainous roads, oblique photogrammetry data from drones, and information on special areas. Then, it optimizes a multi-class support vector machine (SVM) based on this data. Using the optimized SVM, it calibrates digital fences and hierarchical information in an initial geographic model to obtain an oblique photogrammetry geographic model. Next, it initializes multiple alternative routes for the construction access road and uses a genetic algorithm for iterative optimization to select the optimal route. This invention uses drone oblique photogrammetry technology to solve the problems of low efficiency, large errors, and insufficient model accuracy in traditional surveying. Simultaneously, it uses a dynamic fitness function to overcome the limitations of traditional design that only focuses on static costs. Combined with a genetic algorithm for multi-scheme iteration, it uses dynamic risk factors to predict potential disruption events throughout the construction cycle, achieving a balance between short-term construction costs and long-term construction risks.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional design technology, and in particular to a smart visualization method and system for rapid design of access roads for mountain highway construction. Background Technology

[0002] Mountain highway surveying and design is a highly specialized and complex subfield within the field of highway construction. Its core task is to develop highway construction plans that balance safety, economy, ecology, and constructability in mountainous environments with complex topography and varied geological conditions through scientific surveying and systematic design. Among these, construction access roads, as "preliminary infrastructure" for mountain highway construction, are crucial channels for the entry of construction equipment, the transportation of building materials, personnel access, and subsequent operation and maintenance. Their design efficiency and quality directly determine the overall project's start time, construction costs, and safety risks.

[0003] Traditional technical methods have long faced multi-dimensional bottlenecks: On the one hand, conventional remote sensing methods, which are relied upon for topographic surveys, are difficult to accurately capture the microscopic features and hidden geological conditions in complex mountainous terrains, resulting in incomplete topographic information and an inability to provide sufficient detailed support for design; on the other hand, manual on-site verification is limited by objective conditions such as rugged mountainous terrain and variable climate, making it difficult to achieve comprehensive coverage and prone to information deviations due to differences in human judgment. Therefore, the design of construction access roads using existing technologies often results in unreasonable design issues. Summary of the Invention

[0004] This invention provides a smart and visualized method and system for rapid design of construction access roads for mountainous highways, aiming to solve the problem of unreasonable design of construction access roads caused by incomplete presentation of terrain information in the prior art.

[0005] A first aspect of this invention provides a method for rapid design of intelligent, visualized access roads for mountain highway construction, comprising: Acquire topographic maps of mountain roads, oblique photography data from drones, and information on special areas; Based on the topographic map of mountain roads, oblique photogrammetry data from UAVs, and information on special areas, the multi-class support vector machine is optimized. Based on the optimized multi-class support vector machine, digital fences and hierarchical information are calibrated in the initial geographic model to obtain the oblique photogrammetry geographic model. Based on the oblique photogrammetry geographic model, multiple alternative construction access road routes are initialized, and a genetic algorithm is used for iterative optimization to select the optimal construction access road route. In the iterative process of the genetic algorithm, a dynamic fitness function is constructed, which includes a static cost term and a dynamic risk term. The static cost term is used to evaluate the performance of each alternative scheme under undisturbed conditions based on static indicators. The dynamic risk term is used to predict the expected risk costs that may be caused by potential blocking events in the future construction phase.

[0006] In one possible implementation, the dynamic fitness function is constructed as follows: Based on the oblique photogrammetry geographic model, a fixed cost assessment is conducted on the alternative solutions. The fixed costs include at least one of the following: transportation costs for excavation and filling operations, usage costs, and scheduling loss costs. Cost estimates should be made for potential risks in the future construction phase. Potential risks include at least: costs of delays at blocking points, costs of slope instability, and costs of structural conflict rectification. The static cost items and dynamic risk items are weighted and merged to form a dynamic fitness function; The weights of the static cost item and the dynamic risk item are dynamically adjusted based on the terrain complexity index fed back by the oblique photogrammetry geographic model.

[0007] In one possible implementation, the weight adjustment strategy is as follows: Based on the oblique photogrammetry geographic model, extract and quantify at least one terrain complexity index; Among them, the terrain complexity indicators include the proportion of high embankment sections to the total route length, the spatial density of potential blockage points, and the area proportion of the route crossing geologically unstable areas. Based on the mapping relationship between terrain complexity index and weight allocation, the weights of static cost items and dynamic risk items are determined.

[0008] In one possible implementation, based on an oblique photogrammetry geographic model, several alternative routes for the construction access road are initialized, including: Based on the digital fence and hierarchical information calibrated in the oblique photogrammetry geographic model, the preset basic constraints that the construction access road route must meet are defined. The basic constraints include at least: avoiding prohibited crossing areas, meeting the minimum turning radius limit, and meeting the maximum longitudinal slope limit. Under the premise of meeting the preset basic constraints, multiple initial construction access road routes are automatically generated in the oblique photogrammetry geographic model as multiple alternative solutions. Each alternative is encoded as a chromosome in a genetic algorithm. The chromosome contains a set of genes, which are used to characterize at least: the coordinates of the route control points, the boundaries of the cut and fill areas, and the parameters of the slope support structure.

[0009] In one possible implementation, a genetic algorithm is used for iterative optimization to select the optimal construction access road scheme, including: Before each iteration of the genetic algorithm, the Long Short-Term Memory Network (LSTM) time-series prediction model is called to determine the dynamic risk prediction results for the current construction stage. The LSTM model integrates real-time updated data from the oblique photogrammetry geographic model with historical construction time-series data to predict the risk characteristics of at least one potential blocking event. The risk characteristics include the expected duration of each blocking point, the probability of slope instability in the high fill area, and the scope of impact. For each chromosome in the current population, the fitness value is calculated based on the dynamic fitness function representing the alternative scheme. Based on the calculated fitness values, selection, crossover, and mutation operations are performed to generate the next generation population; among them, the mutation operation adopts a directed mutation mechanism. When the preset termination condition is met, the iteration terminates and the optimal construction access road solution is output.

[0010] In one possible implementation, a multi-class support vector machine is optimized based on topographic maps of mountain roads, oblique photogrammetry data from UAVs, and information on special areas. Then, based on the optimized multi-class support vector machine, digital fences and hierarchical information are labeled in the initial geographic model to obtain an oblique photogrammetry geographic model, including: The mountain road survey topographic map, UAV oblique photography data, and special area information are spatiotemporally registered and fused to generate an enhanced optimization set; the enhanced optimization set includes a basic feature layer, an engineering feature layer, and a boundary calibration layer; An initial geographical model was established based on the topographic map of the mountain road survey. The multi-class support vector machine is optimized based on the enhanced optimization set, and digital fences and hierarchical information are calibrated in the initial geographic model based on the optimized multi-class support vector machine to obtain the oblique photogrammetry geographic model.

[0011] In one possible implementation, topographic maps of mountain roads, oblique photogrammetric data from UAVs, and information on special areas are spatiotemporally registered and fused to generate an enhanced optimized set, including: Aerial triangulation and dense matching are performed on the oblique photography data of UAVs to generate 3D point clouds and digital orthophoto maps. Vectorize the topographic map of the mountain road survey to extract contour lines, elevation points and ground features; Using the coordinate system of the topographic map of the mountain road survey as the reference coordinate system, at least three ground control points are selected, and the three-dimensional point cloud and digital orthophoto map are geometrically corrected and coordinate transformed according to the least squares method to align with the spatial coordinate system of the reference coordinate system. The spatiotemporally registered 3D point cloud, digital orthophoto map, and vectorized topographic map data are overlaid to obtain a fused data volume. Based on the fused data volume, a basic feature layer, an engineering feature layer, and a boundary calibration layer are constructed. A unique identifier is established for each geographic unit in the basic feature layer, engineering feature layer, and boundary calibration layer, and the feature data of the three layers are spatially correlated to obtain an enhanced optimization set.

[0012] In one possible implementation, the spatiotemporally registered 3D point cloud, digital orthophoto map, and vectorized topographic map data are overlaid to obtain a fused data volume, including: Using vectorized topographic map data as the bottom layer, digital orthophoto overlay as the middle layer, and 3D point cloud overlay as the top layer, a three-dimensional overlay structure with topographic, texture, and geometric information is obtained. The system handles elevation conflicts, feature outline conflicts, and element attribute conflicts in the 3D overlay structure to obtain a fused data volume. The accuracy, completeness, and consistency of the generated fused data volume are verified.

[0013] In one possible implementation, topographic maps of mountain roads, oblique photogrammetry data from drones, and information on special areas are acquired, including: Select topographic map data sources that meet the project's timeliness requirements, and verify the accuracy of the topographic map data to ensure that its horizontal and vertical accuracy meets the error threshold for subsequent model construction; and convert the topographic map data into a standardized format compatible with subsequent data processing workflows, and extract the core topographic and feature elements. Based on the terrain features of the survey area and the preset model resolution requirements, the flight path and parameters of the UAV are planned; data acquisition is performed during the window period that meets the preset environmental conditions, and optical images and position and attitude data are acquired simultaneously; and the raw data is preprocessed, including image correction, point cloud generation and coordinate system 1, to generate the initial 3D point cloud and digital orthophoto map. Collect background information related to geological hazards, engineering constraints, ecology, and hydrology in the survey area; conduct on-site surveys and supplements in key areas to verify or refine the background information; and standardize and encode all collected and surveyed special area information according to a preset data structure, which includes at least information type, spatial range, attribute description, and geographic coordinates.

[0014] A second aspect of the present invention provides a collaborative design system, including a drone and an electronic device; the method of the first aspect above is used to run on the electronic device.

[0015] Compared to traditional technologies, this invention provides a smart and visualized method for rapid design of construction access roads for mountainous highways. First, it acquires topographic maps of the mountainous area, oblique photogrammetry data from drones, and information on special regions. Then, it optimizes a multi-class support vector machine (SVM) based on these data. Using the optimized SVM, it calibrates digital fences and hierarchical information in an initial geographic model to obtain an oblique photogrammetry geographic model. Finally, based on this model, it initializes multiple alternative routes for the construction access road and uses a genetic algorithm for iterative optimization to select the optimal route. During the iteration of the genetic algorithm, a dynamic fitness function is constructed, comprising a static cost term and a dynamic risk term. The static cost term evaluates the performance of each alternative route under undisturbed conditions based on static indicators. The dynamic risk term predicts the expected risk costs that may arise from potential obstruction events during future construction phases. This invention optimizes a multi-class support vector machine to calibrate digital fences and hierarchical information in the initial geographic model to obtain a high-precision oblique photogrammetry geographic model. This solves the problems of low efficiency, large errors, and incomplete terrain information presentation in traditional surveys, which cannot provide sufficient design support and lead to unreasonable construction access road designs. At the same time, by constructing a dynamic fitness function and combining iterative optimization of alternative construction access road routes with a genetic algorithm, a balance is achieved between the short-term construction cost of the access road and the potential risks throughout the construction cycle. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the implementation of the intelligent and visual rapid design method for construction access roads in mountainous areas provided in this embodiment of the invention. Detailed Implementation

[0017] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart illustrating the implementation of the intelligent, visualized, rapid design method for construction access roads in mountainous areas, provided in this embodiment of the invention. Figure 1 As shown, the method includes: Acquire topographic maps of mountain roads, oblique photography data from drones, and information on special areas; Based on the topographic map of mountain roads, oblique photogrammetry data from UAVs, and information on special areas, the multi-class support vector machine is optimized. Based on the optimized multi-class support vector machine, digital fences and hierarchical information are calibrated in the initial geographic model to obtain the oblique photogrammetry geographic model. Based on the oblique photogrammetry geographic model, multiple alternative construction access road routes are initialized, and a genetic algorithm is used for iterative optimization to select the optimal construction access road route. In the iterative process of the genetic algorithm, a dynamic fitness function is constructed, which includes a static cost term and a dynamic risk term. The static cost term is used to evaluate the performance of each alternative scheme under undisturbed conditions based on static indicators. The dynamic risk term is used to predict the expected risk costs that may be caused by potential blocking events in the future construction phase.

[0019] In this embodiment of the invention, an optimal solution with dual control of cost and risk is selected through iterative optimization using a genetic algorithm. Based on the constraints of the oblique photogrammetry geographic model and design specifications, 50-100 alternative solutions containing route and engineering parameters are generated in a "random + constrained" manner. Invalid solutions are eliminated, and compliant solutions are retained. A dynamic fitness function containing static cost items and dynamic risk items is constructed. The static cost item calculates the costs of excavation, filling, transportation, materials, and equipment scheduling based on model data. The dynamic risk item predicts the expected risk costs of future construction delays, slope instability, and structural conflicts. Through multiple rounds of iteration using the genetic algorithm, solutions with better fitness values ​​are retained. After cross-fusion of gene fragments and targeted mutation for high-risk areas, the optimal construction access road solution that satisfies both "controllable short-term costs and controllable long-term risks" is finally output, breaking through the limitation of traditional design that only focuses on static costs.

[0020] In some embodiments, a multi-class support vector machine (SVM) is optimized based on topographic maps of mountain roads, oblique photogrammetry data from UAVs, and information on special areas. Based on the optimized SVM, digital fences and hierarchical information are then calibrated in the initial geographic model to obtain an oblique photogrammetry geographic model. This includes: spatiotemporally registering and fusing the topographic maps of mountain roads, oblique photogrammetry data from UAVs, and information on special areas to generate an enhanced optimization set; wherein the enhanced optimization set includes a basic feature layer, an engineering feature layer, and a boundary calibration layer; establishing an initial geographic model based on the topographic maps of mountain roads; optimizing the multi-class SVM based on the enhanced optimization set; and calibrating digital fences and hierarchical information in the initial geographic model based on the optimized SVM to obtain the oblique photogrammetry geographic model. The enhanced optimization set is a structured dataset used to adjust model parameters and reduce prediction / classification errors, containing basic feature data and initial model error feedback data. In addition, the enhanced optimization set adds an initial model classification error feedback layer to the original basic feature layer, engineering feature layer, and boundary calibration layer. This layer records the initial SVM's misclassification of samples, accurately locating the model's weaknesses. Each geographic unit is given a unique identifier, linking the three layers of feature data with error feedback information to form a "feature-error-optimization parameter" mapping relationship, making model optimization more targeted and avoiding blind adjustments. During the optimization process, the multi-class SVM first adjusts the kernel function parameters based on basic and engineering features, and then optimizes the classification threshold for geologically complex areas by combining the error feedback layer data. After verification with 20% of reserved samples, it can be used for digital fence calibration only after meeting the standards. The LSTM, on the other hand, updates the training data using a sliding window method with a 7-day optimization cycle, integrating real-time data from the oblique photogrammetry geographic model (such as slope displacement and meteorological data) with historical construction time series data to dynamically adjust network weights, improving the timeliness and accuracy of risk prediction.

[0021] The historical construction timeline data sources are: publicly available construction logs, supervision records, geological monitoring reports, meteorological observation data, blocking event ledgers, and slope instability case databases of mountainous highway construction projects in the same region / geological type over the past 3 years; the selection criteria are: selecting project data with a similarity of ≥80% to the survey area of ​​this project in terms of terrain slope, geological lithology, and climate conditions, and slicing them into fixed time windows (7 days / week). Each slice includes meteorological indicators for the time period, construction progress indicators, slope displacement data, and blocking event records; the LSTM model uses this standardized historical data as the training set, integrates oblique photogrammetry geographic model data updated in real time, and predicts the risk characteristics of at least one potential blocking event. The risk characteristics include the expected duration of each blocking point, the probability of slope instability in high embankment areas, and the scope of impact.

[0022] In some embodiments, the topographic map of mountain road survey, the oblique photogrammetry data of UAV and the information of special areas are spatiotemporally registered and fused to generate an enhanced optimization set, including: performing aerial triangulation and dense matching on the oblique photogrammetry data of UAV to generate a three-dimensional point cloud and a digital orthophoto map. Vectorize the topographic map of the mountain road survey to extract contour lines, elevation points and ground features; Using the coordinate system of the topographic map of the mountain road survey as the reference coordinate system, at least three ground control points are selected, and the three-dimensional point cloud and digital orthophoto map are geometrically corrected and coordinate transformed according to the least squares method to align with the spatial coordinate system of the reference coordinate system. The spatiotemporally registered 3D point cloud, digital orthophoto map, and vectorized topographic map data are overlaid to obtain a fused data volume. Based on the fused data volume, a basic feature layer, an engineering feature layer, and a boundary calibration layer are constructed. A unique identifier is established for each geographic unit in the basic feature layer, engineering feature layer, and boundary calibration layer, and the feature data of the three layers are spatially correlated to obtain an enhanced optimization set.

[0023] In this embodiment of the invention, the oblique photogrammetry data from the UAV is first processed specifically. Specifically, aerial triangulation encryption technology is used to construct an aerial triangulation network containing internal and external orientation elements of the images, utilizing camera parameters, position information, and image overlap relationships recorded during UAV flight. This accurately calculates the spatial position and attitude of each image. Based on this, dense matching is performed, matching each pixel of the encrypted images point-by-point to generate 3D point cloud data that reflects the three-dimensional shape of the terrain. Simultaneously, orthorectification is used to eliminate distortions caused by terrain undulations and image tilt, outputting a digital orthophoto map with clear texture and accurate coordinates. This provides a data source with both geometric accuracy and visual detail for subsequent fusion.

[0024] Next, the topographic maps of the mountain road survey were converted into vector data. Specifically, using professional geographic information software, the topographic maps were converted into vector data format. Through a combination of human-computer interaction and automatic recognition, the core topographic and feature information in the topographic maps was accurately extracted. The topographic information includes contour lines indicating elevation changes and elevation points marking specific elevation values, while the feature information covers the outlines and attributes of existing artificial or natural features such as roads, waterways, buildings, and bridges, forming structured vector topographic map data to ensure that the topographic map information can be efficiently accessed and correlated.

[0025] Subsequently, spatial alignment and calibration of multi-source data were performed. Using the coordinate system adopted for topographic maps of mountain road surveys as a unified reference coordinate system, at least three evenly distributed, topographically significant, and coordinate-known ground control points (such as mountaintops, corners of fixed buildings, and road intersections) were selected within the survey area. Based on the least squares principle, a mathematical transformation model was established between the 3D point cloud, digital orthophoto map, and the reference coordinate system. By calculating the optimal transformation parameters, geometric correction and coordinate transformation were performed on the spatial coordinates of the 3D point cloud and the pixel positions of the digital orthophoto map, enabling precise spatial alignment between the two types of data and the topographic map under the reference coordinate system. This eliminated coordinate deviations between different data sources and ensured spatial consistency for subsequent overlay and fusion.

[0026] After spatial alignment, the registered 3D point cloud, digital orthophoto map, and vectorized topographic map data are layered and integrated: the vectorized topographic map data is used as the bottom layer, retaining its core topographic framework information such as contour lines and elevation points; the digital orthophoto map is superimposed on the bottom layer to ensure that the image texture and the outline of the topographic map features are accurately matched, supplementing the intuitive feature details (such as vegetation cover and surface material differences) that are lacking in the topographic map; finally, the 3D point cloud is superimposed on the middle layer image, and the density and three-dimensionality of the point cloud are used to refine the elevation transition details between the contour lines of the topographic map, forming a three-in-one fused data body of "topographic framework-image texture-three-dimensional geometry", which fully integrates the advantages of the three types of data.

[0027] Based on the comprehensive information of the fused data volume, a three-layer core structure of the enhanced optimization set is further constructed. Basic information such as terrain elevation, image RGB values, and point cloud geometric coordinates are extracted from the fused data volume to form a basic feature layer reflecting the basic characteristics of the region. Combined with engineering design requirements, engineering parameters such as slope, aspect, and soil type are calculated, and information on geological hazard risks and underground pipeline constraints in special areas is integrated to construct an engineering feature layer supporting engineering decisions. According to preset rules (such as slope thresholds and risk level standards), key boundaries such as cut-and-fill boundaries, risk control areas, and engineering constraint areas are identified and marked in the fused data volume, forming a boundary marking layer with attribute annotations, thus realizing the transformation of data from "comprehensive fusion" to "classified structuring."

[0028] Finally, a unified association system is established for the three feature layers: For each geographic unit (such as a region or natural feature unit divided by a fixed grid) covered by the basic feature layer, engineering feature layer, and boundary calibration layer, a unique identifier is generated. The identifier contains key information such as layer type, spatial coordinate range, and data generation time to ensure that each geographic unit can be accurately located and distinguished. Using the unique identifier and geographic coordinates as dual indexes, the parameter information (such as the elevation value, slope parameter, and boundary attribute of a certain grid) of the same geographic unit in the three feature layers is associated and bound to establish a cross-layer data mapping relationship, ultimately forming an enhanced optimization set with complete structure, information association, and direct support for subsequent model optimization.

[0029] In some embodiments, the spatiotemporally registered 3D point cloud, digital orthophoto map, and vectorized topographic map data are overlaid to obtain a fused data volume, including: Using vectorized topographic map data as the bottom layer, digital orthophoto overlay as the middle layer, and 3D point cloud overlay as the top layer, a three-dimensional overlay structure with topographic, texture, and geometric information is obtained. The system handles elevation conflicts, feature outline conflicts, and element attribute conflicts in the 3D overlay structure to obtain a fused data volume. The accuracy, completeness, and consistency of the generated fused data volume are verified.

[0030] When handling elevation conflicts, a permissible deviation threshold based on the nominal accuracy of the data source is set. For conflict points exceeding this threshold, the terrain slope at the point is considered: if the slope is gentle, 3D point cloud data is retained to capture subtle terrain changes; if the slope is steep, the point cloud data is smoothed by referring to the contour lines on the topographic map. When handling feature outline conflicts, the texture clarity and edge sharpness of the digital orthophoto map are evaluated. Image outlines with edge sharpness exceeding the set threshold are adopted, and the corresponding features on the topographic map are updated. When handling feature attribute conflicts, the authority and timeliness of information in specific areas are prioritized. When high-priority information such as ecological protection red lines conflict with basic feature attributes, the high-priority information overrides the basic attributes, and conflict markers are added.

[0031] In this embodiment of the invention, a layered overlay construction is first performed to build a three-dimensional overlay structure containing terrain, texture, and geometric information. Vectorized topographic map data serves as the bottom layer, which has already undergone pre-processing to extract contour lines, elevation points, and feature elements (such as existing roads, water systems, and building outlines). This layer clearly presents the macroscopic topographic framework and key feature distribution of the surveyed area, providing a stable spatial reference and topographic skeleton for the entire overlay structure. A digital orthophoto map is overlaid on top of this bottom layer as the middle layer. This image map has undergone geometric correction and color optimization, and can intuitively present surface texture details—such as the density of vegetation cover, the shape of exposed rock areas, and differences in surface materials (such as the visual distinction between dirt roads and paved surfaces). Furthermore, it accurately corresponds to the feature outlines of the bottom topographic map (e.g., shadows). The road edges in the image overlap with the road lines marked on the topographic map by ≥95%, supplementing the macro-topographic framework with concrete visual information of the land surface. Finally, a three-dimensional point cloud is superimposed on the middle layer as the top layer. The three-dimensional point cloud is composed of a large number of spatial points containing X, Y, and Z coordinates, which can accurately restore the three-dimensional geometric shape of the terrain, such as the steepness of the slope, the depth and width of the gully, and the undulating outline of the hill. It can fill the elevation transition gaps between contour lines on the topographic map and the lack of three-dimensional spatial information in the image map, ultimately forming a three-layer three-dimensional superimposed structure of "bottom-layer terrain skeleton - middle-layer texture details - top-layer three-dimensional geometry", realizing the comprehensive integration of the three core information types of terrain, texture, and geometry.

[0032] Next, specific processing was carried out to address three types of data conflicts that may occur in the 3D overlay structure, eliminating information contradictions. For elevation conflicts, i.e., discrepancies between the Z-value (elevation) of the 3D point cloud and the elevation points and contour lines of the vectorized topographic map, a comprehensive judgment was made considering both data accuracy and timeliness: if the discrepancy is small, the average of the two was taken as the final elevation value, balancing data stability and precision; if the discrepancy is large, the elevation data of the UAV 3D point cloud was used first (because UAV data has higher resolution and the shooting time is closer to the current survey node, which can better reflect the latest terrain status), and the location and value of the deviation were marked for subsequent traceability; if the 3D point cloud data contains abnormal elevations caused by local noise points, the noise points were removed by statistical filtering algorithms before elevation comparison. For conflicts in feature outlines, such as inconsistencies between the feature shapes presented in the digital orthophoto map (e.g., newly added temporary buildings, widened gullies) and the feature outlines marked on the vectorized topographic map, verification was required through field sampling. Select 3-5 feature points in the conflict area (such as building corners, gully edges). If the measurement results match the outlines of features in the imagery, the topographic map data is considered outdated and the corresponding feature outlines on the topographic map need to be updated; if the measurement results do not match the topographic map data, the topographic map data is considered lagging and the corresponding feature outlines on the topographic map need to be updated. Figure 1If the misjudgment is determined to be due to perspective deviation or partial occlusion during image capture, local correction of the corresponding area in the image map is required. For conflicts in element attributes, such as overlap between the "geological hazard zone" marked in special area information and the "forest land" attribute marked on the vectorized topographic map, confirmation must be made in conjunction with the geological survey report and on-site investigation. If the area is a "landslide hazard zone under forest cover," it should be assigned a dual attribute of "forest land + landslide hazard zone," preserving the land feature type information while clearly defining engineering risk constraints, avoiding information loss due to a single attribute.

[0033] Finally, the processed fused data volume undergoes three core verifications: accuracy, completeness, and consistency, to ensure data quality meets standards. For accuracy verification, 10-15 evenly distributed ground control points (such as mountain tops, valley bottoms, and road intersections) are selected, and their actual coordinates and elevations are measured and compared with the coordinates and elevations of corresponding points in the fused data volume to ensure the data meets the accuracy requirements for mountain highway surveying and design. For completeness verification, it is checked whether the fused data volume fully covers the entire survey section, whether it includes all key geographic features (such as bridges, tunnel entrances and exits, large gullies, and boundaries of ecologically sensitive areas), and whether there are any missing or blank areas. For consistency verification, using the spatial query function of geographic information software, 20 geographic units (such as 10m×10m grids) are randomly selected to verify whether their information in the three-layer overlay structure is logically consistent. For example, if a geographic unit is marked as "steep slope" (with dense contour lines) on the bottom-level topographic map, the slope calculated in the top-level 3D point cloud should be ≥25°, and the surface should appear as a steep shape in the middle-level image map. If there is a logical contradiction (such as the topographic map marking a steep slope but the 3D point cloud computing slope is only 10°), it is necessary to go back to the conflict handling stage to correct it again until all verification items meet the standards and finally form a qualified fused data body.

[0034] In some embodiments, based on the fused data volume, a basic feature layer, an engineering feature layer, and a boundary calibration layer are constructed, including: Extract multidimensional basic parameters related to terrain, texture, and geometry from the fused data volume; associate the multidimensional basic parameters according to geographic coordinates to form a basic feature layer containing at least terrain basic parameters, texture feature parameters, and geometric fine parameters; Based on the basic feature layer and the fused data volume, calculate or extract engineering parameters related to engineering design and construction; the engineering parameters include at least engineering topographic parameters, engineering constraint parameters and ecological engineering parameters; associate the engineering parameters with the corresponding geographic units to form the engineering feature layer; According to the preset boundary identification rules, at least the cut and fill boundary, risk control boundary, and engineering constraint boundary are identified and marked in the fused data volume and / or engineering feature layer; the identified boundaries are assigned type, level and associated attribute information to form a boundary marking layer.

[0035] In this embodiment of the invention, a basic feature layer is first constructed, focusing on the extraction and association of three core basic information types—terrain, texture, and geometry—from the fused data volume. From the bottom-layer vectorized topographic map of the fused data volume, contour line elevation values ​​reflecting terrain undulations, coordinates of elevation points (X, Y, Z) indicating specific heights, and basic terrain parameters such as existing road widths and water system directions are extracted. From the middle-layer digital orthophoto map, surface texture feature parameters are extracted using image processing techniques. For example, the texture contrast and correlation between vegetation-covered areas and exposed rock areas are calculated using a gray-level co-occurrence matrix, and surface materials are distinguished by RGB three-color channel values ​​(e.g., red clay areas with high red channel values, and forest areas with high green channel values). From the top-layer three-dimensional point cloud, fine geometric parameters reflecting spatial morphology are extracted. These include dense Z-values ​​of ground points (sampled at 0.1m intervals) and point cloud contour coordinates of feature edges (e.g., continuous point cloud sequences at the bottom of gullies, and three-dimensional coordinates of slope inflection points). After extraction, using geographic coordinates as a unified link, the basic terrain parameters, texture feature parameters, and geometric fine parameters of the same geographic unit (such as a 10m×10m grid) are bound one by one to ensure that the multi-dimensional basic information of each unit is complete and traceable, and finally form a basic feature layer that covers the survey area and is structured.

[0036] Next, based on the basic feature layer and the fused data volume, engineering parameters are further transformed and extracted to construct the engineering feature layer. For engineering terrain parameters, the slope of each geographic unit is calculated (accuracy retained to 0.1°) by combining the contour elevation values ​​and 3D point cloud Z values ​​of the basic feature layer and using the slope calculation formula (slope = arctan(ΔZ / ΔX)). The slope aspect is determined by the azimuth algorithm, and the soil type is determined by associating it with the soil type annotated in the vectorized topographic map in the fused data volume (e.g., sandy soil 1.6 g / cm³, clay soil 1.8 g / cm³). For engineering constraint parameters, the lithology (e.g., granite, shale), groundwater level depth, and the burial depth and diameter data of existing pipelines are extracted from the special area information contained in the fused data volume and added to the corresponding geographic units. For ecological engineering parameters, the vegetation coverage is calculated by threshold segmentation based on the image RGB values ​​of the basic feature layer, and the distribution areas of rare plants, wildlife habitats, and other ecologically sensitive areas recorded in the fused data volume are marked. The three types of engineering parameters are integrated in the form of "geographic unit-parameter group". Each unit is labeled with engineering adaptation attributes (such as "slope 25° + granite + low vegetation cover - suitable for excavation"), forming an engineering feature layer that directly serves design decisions.

[0037] Finally, based on the preset boundary identification rules, key boundaries are identified and calibrated in the fused data volume and engineering feature layer, and a boundary calibration layer is constructed. For cut and fill boundaries, the slope parameters of the engineering feature layer are used as the core judgment basis, combined with the edge contours of the 3D point cloud in the fused data volume (such as the continuous coordinates of slope change points), to generate closed boundary polygons, and the boundary type is labeled. For risk control boundaries, based on the geological hazard areas (such as landslides and debris flows) recorded in the special area information in the fused data volume, areas with abnormal vegetation coverage (such as vegetation fault zones) and areas with sudden slope changes in the engineering feature layer are superimposed, and the three-dimensional contours of the boundaries (such as the slope undulations of landslide bodies) are verified through the 3D point cloud, generating risk boundaries and assigning warning levels (red - prohibition of point selection, yellow - caution in line selection, blue - normal area). For engineering constraint boundaries, the contour coordinates of existing buildings (such as houses and bridges) and cultural relics protection units in the fused data volume are extracted, combined with the prohibited construction area in the special area information, to generate constraint boundaries and label the constraint type. Supplement all boundaries with spatial attributes (such as boundary area, perimeter, and center point coordinates) and associated parameters (such as the basic value of the earthwork volume estimation corresponding to the cut and fill boundary) to ensure that each boundary can be associated with the corresponding geographical unit of the basic feature layer and the engineering feature layer, and finally form a boundary marking layer with clear boundaries, attributes and levels.

[0038] In some embodiments, a unique identifier is established for each geographic unit in the basic feature layer, engineering feature layer, and boundary calibration layer, and the feature data of the three layers are spatially correlated to obtain an enhanced optimization set, including: The survey area is divided into multiple spatially discrete geographic units; a unique identifier is generated for each geographic unit to uniquely identify its spatial location and the feature layer to which it belongs; Using unique identifiers and / or geographic coordinates as indexes, feature parameters corresponding to the same geographic unit in the basic feature layer, engineering feature layer, and boundary calibration layer are associated to establish a mapping relationship between cross-layer feature parameters; Verify the uniqueness of the unique identifier and the logical consistency of cross-layer feature parameters; integrate the verified related data into a structured dataset, form corresponding data description information, and generate an enhanced optimization set.

[0039] In this embodiment of the invention, the geographical unit division and unique identifier generation of the survey area are carried out first. Based on the terrain complexity and engineering design accuracy requirements of the survey area, a fixed grid method (such as a 10m×10m or 5m×5m grid) is used to divide the entire survey section into multiple spatially discrete, non-overlapping, and fully covered geographical units. The boundary of each unit is clearly defined by geographical coordinates to ensure that the spatial range of the unit is unique and traceable. Subsequently, a unique identifier is generated for each geographic unit. The identifier adopts a combined structure of "layer type code + grid coordinate code + time sequence code". The layer type code uses 2 letters to distinguish the feature layer ("JC" for the basic feature layer, "GC" for the engineering feature layer, and "QJ" for the boundary calibration layer); the grid coordinate code uses 12 digits to record the coordinate range of the unit (e.g., X300000-X300010 is recorded as "300000", Y200000-Y200010 is recorded as "200000", and the combination is "300000200000"); the time sequence code uses 4 digits to mark the week number generated by the data.

[0040] Next, using unique identifiers and geographic coordinates as dual indexes, cross-layer association of the three-layer feature data is achieved. First, the feature parameters of the same geographic unit in the basic feature layer, engineering feature layer, and boundary calibration layer are bound to the corresponding unique identifiers. For example, the "elevation value 580m, texture contrast 25, point cloud density 100 points / m²" of a unit in the basic feature layer, the "slope 22°, soil density 1.7g / cm³, groundwater depth 3m" of the same unit in the engineering feature layer, and the "excavation boundary attribute, risk level yellow" of the same unit in the boundary calibration layer are all associated with the same set of unique identifiers. Using geographic coordinates as an auxiliary index, the spatial overlay function of geographic information software is used to verify whether the coordinate range of the same unit in the three layers of data is completely matched, ensuring that there is no misalignment caused by coordinate offset. At the same time, a mapping relationship table of cross-layer feature parameters is established to clearly mark the corresponding logic of basic feature parameters, engineering feature parameters, and boundary attributes (such as "basic feature layer elevation difference > 5m → engineering feature layer is judged as steep slope → boundary marking layer is classified as excavation boundary"), so that the three layers of data form a complete association chain of "parameter-attribute-boundary".

[0041] Subsequently, a dual verification of the uniqueness of the identifier and the logical consistency of cross-layer feature parameters is performed. For the uniqueness of the identifier, the database deduplication function is used to traverse the identifiers of all geographic units to check for duplicate codes (such as the same grid coordinate codes within the same layer, or incorrect coding formats between different layers). If duplicates or errors are found, the code generation process is immediately traced back to correct them, ensuring that each identifier is globally unique across the three feature layers. To ensure logical consistency of parameters across layers, a combination of "sampling verification + rule verification" is adopted: 20% of the geographical units are randomly selected, and the parameters of the three layers are checked for contradictions based on preset engineering logic rules (e.g., when the slope of the basic feature layer is greater than 30°, the engineering feature layer should be marked "unsuitable for filling", and the boundary marking layer should not be classified as a filling boundary). For example, if the slope of the basic feature layer of a certain unit is calculated to be 35°, but the engineering feature layer is marked "suitable for filling" and the boundary marking layer is classified as a "fill boundary", it is determined to be a logical conflict. It is necessary to trace back to the feature layer construction stage to check whether the slope calculation is incorrect and whether the boundary judgment rule is misused, until the parameter logic of all sampled units is consistent. Non-sampled units are batch verified through automated rule scripts to ensure that there are no hidden conflicts.

[0042] Finally, the validated associated data is integrated to generate an enhanced optimization set with data description information. The associated geographic unit data from the three feature layers is imported into a structured database according to the field format of "unique identifier - basic feature parameters - engineering feature parameters - boundary attributes" to form a standardized data table, facilitating quick retrieval and querying during subsequent model optimization. Detailed data description information is also added to the dataset, including: data source (e.g., UAV oblique photogrammetry data, field survey report), geographic unit division standard (e.g., 10m×10m grid), definition and unit of each parameter (e.g., "slope: unit is degrees, calculation method is arctan(ΔZ / ΔX)"), validation results (e.g., identifier duplication rate 0%, parameter logical consistency rate 99.8%), data generation time and validity period, ensuring the interpretability and reusability of the dataset. The final enhanced optimization set contains complete parameters from the three feature layers, achieves cross-layer association through unique identifiers, and is quality-verified to ensure reliability, making it directly usable for multi-class support vector machine optimization.

[0043] In some embodiments, based on the optimized multi-class support vector machine, digital fences and hierarchical information are calibrated in the initial geographic model to obtain an oblique photogrammetry geographic model, including: The optimized multi-class support vector machine model is integrated with the initial geographic model, and the mapping relationship of the enhanced optimization set is imported as the classification benchmark. The initial geographic model is spatially discretized, the first feature parameter of each discrete unit is extracted, and the first feature parameter is input into a multi-class support vector machine model. The model outputs the classification results of the boundary type and risk level of each discrete unit. Based on the classification results, discrete units with the same or similar boundary types are spatially aggregated to generate closed digital fences, and each digital fence is assigned corresponding engineering attributes and risk level information. The digital fence and its associated engineering attributes and risk level information are fused with the initial geographic model and oblique photogrammetry data to generate an oblique photogrammetry geographic model.

[0044] In some embodiments, digital fences and their associated engineering attributes and risk level information are fused with an initial geographic model and oblique photogrammetry data to generate an oblique photogrammetry geographic model, including: The oblique photogrammetry data is geometrically optimized and visually consistent, and the engineering attributes and risk level information of the digital fence are structured and encoded to generate standardized fusion input data. Using the coordinate system of the initial geographic model as a reference, the preprocessed oblique photogrammetric data is spatially registered with high precision to ensure spatial alignment with the initial geographic model. The structured digital fence is used as a vector feature with attribute information and embedded into the initial geographic model according to its geographic coordinates. The attribute information of the digital fence is then associated with the corresponding terrain unit in the model. The texture and geometric information in the spatially registered oblique photogrammetric data are mapped onto the surface of the initial geographic model to generate a three-dimensional geographic model that integrates real-world terrain, digital fences, and attribute information. Integrate at least one interactive functional module for highway design into the 3D geographic model.

[0045] In this embodiment of the invention, the first step is to integrate and adapt the optimized model with the initial geographic model: the multi-class support vector machine model (which has been validated through the enhanced optimization set and has the ability to identify boundary types and determine risk levels) is embedded into the computational framework of the initial geographic model through a programming interface, ensuring that the model can call the basic data such as spatial coordinates and terrain elevation of the initial geographic model in real time; at the same time, the mapping relationship of "basic feature parameters - engineering feature parameters - boundary attributes" in the enhanced optimization set (such as "elevation difference > 8m + soil density < 1.6g / cm³ → boundary type is fill area, low risk level") is imported as a classification benchmark to provide a unified judgment standard for subsequent unit classification and avoid the classification logic from being disconnected from the optimization stage.

[0046] Next, the initial geographic model is spatially discretized and its features are classified. Following a grid division standard consistent with the enhanced optimization set (e.g., 5m×5m or 10m×10m), the initial geographic model is divided into several spatially independent, coordinate-defined discrete units. For each discrete unit, the first feature parameter corresponding to the model optimization is extracted. This includes basic parameters such as terrain elevation and contour curvature obtained from the initial geographic model, combined with related parameters such as surface texture grayscale values ​​and soil type supplemented from previous survey data, forming a standardized feature vector. The feature vector is input into a multi-class support vector machine model. By comparing the feature parameters with a preset classification benchmark, the model automatically outputs the classification result for each discrete unit, clearly labeling its boundary type (e.g., cut area, fill area, geological risk area, pipeline constraint area) and risk level (e.g., red high risk, yellow medium risk, blue low risk), achieving refined attribute labeling of the initial geographic model.

[0047] The multi-class support vector machine model is constructed using a one-to-one model, and the training process is as follows: Using the geographic units corresponding to the basic feature layer, engineering feature layer, and boundary calibration layer as sample carriers, the topographic elevation, slope, aspect, surface texture contrast, 3D point cloud geometric features, geological stability parameters, vegetation coverage, and groundwater depth of each geographic unit are extracted as model input features. Then, the boundary attributes and risk levels of the boundary calibration layer, pre-annotated manually and verified on-site, are used as training labels. Boundary type labels include five categories: excavation area, fill area, geological risk area, engineering constraint area, and ecologically sensitive area; risk level labels include three levels: red (high risk), yellow (medium risk), and blue (low risk). Finally, the appropriate techniques are selected. The radial basis function (RBF) is used for multi-class classification. The kernel width γ and the penalty factor C are combined using a grid search method to traverse the combination [0.001, 0.01, 0.1, 1, 10, 100]. The parameter combination that achieves the highest classification accuracy is selected through 5-fold cross-validation. All samples are then randomly divided into training and validation sets in an 8:2 ratio. The training set is used for model learning, and the validation set is used to evaluate generalization ability. Finally, when the model achieves a boundary type classification accuracy of ≥95% and a risk level determination accuracy of ≥92% on the validation set, the model is considered to have converged, resulting in an optimized multi-class support vector machine model that can be used for digital fence labeling.

[0048] The model compares the first feature parameter of the input with the feature-label mapping relationship learned in the training phase, and automatically outputs the boundary type classification result and risk level classification result corresponding to each discrete unit according to the maximum confidence criterion.

[0049] In addition, areas with typical topographic, geological, and land feature characteristics within the survey scope were selected as training sample collection areas, including: typical excavation areas, fill areas, semi-fill and semi-excavation areas, steep slope areas, gentle slope areas, geological hazard hazard areas, geologically stable areas, ecologically sensitive areas, ordinary construction areas, and underground pipeline / structure constraint areas, etc.

[0050] Spatial aggregation is performed based on the classification results to generate a closed digital fence. Specifically, using the spatial topology analysis function of the Geographic Information System (GIS), the classification results of all discrete units are filtered and merged. Discrete units with identical boundary types (e.g., both being "high-risk landslide areas") or highly similar (e.g., "general fill areas" and "priority fill areas" both belonging to the fill category) and spatially adjacent are outlined and smoothed according to their coordinate boundaries to eliminate jagged edges caused by grid cutting, forming a continuous, closed polygonal area, i.e., a digital fence. For example, 20 adjacent "cut area" discrete units are aggregated to generate a complete cut area digital fence covering the region. Its boundary coordinates are formed by connecting the edge points of the outermost unit, ensuring that the fence range matches the actual terrain features.

[0051] Simultaneously, each digital fence is assigned corresponding engineering attributes and risk level information: combining the classification results with the engineering parameters of the enhanced optimization set, specific attributes are added to the digital fence. For example, the fence in the excavation area is labeled "average excavation depth 3.5m, rock ratio 60%", the fence in the fill area is labeled "recommended fill type, maximum fill height 2m", and the fence in the risk area is labeled "risk triggers (such as landslides) and protection measures recommendations (such as adding anti-slide piles)". The risk level information is directly related to the classification results. For example, the red fence corresponds to "point selection is prohibited, detour is required", the yellow fence corresponds to "careful line selection, enhanced protection is required", and the blue fence corresponds to "normal design, conventional protection". All attribute information is bound to the spatial coordinates of the digital fence, forming a complete data chain of "fence boundary - engineering attributes - risk constraints".

[0052] For oblique photogrammetry data, a 3D point cloud denoising algorithm is used to remove discrete noise points (such as redundant points caused by birds and cloud interference) while preserving the micro-geometric details of ground features (such as rock joints and gully textures). At the same time, the digital orthophoto map is segmented for color equalization to eliminate image color differences caused by uneven lighting in mountainous areas and ensure visual consistency. For digital fence data, its engineering attributes and risk level information are structured in a key-value pair format and a vector file format compatible with the initial geographic model is generated to facilitate subsequent embedding operations.

[0053] Using the coordinate system adopted by the initial geographic model as a reference, 6-8 evenly distributed ground control points (such as mountain tops or corners of fixed buildings) are selected in the survey area to obtain the precise coordinates of the control points. Based on these control points, the coordinate transformation parameters of the oblique photography data (3D point cloud, digital orthophoto map) are calculated using the least squares method, and geometric correction is performed to ensure that the spatial position deviation between the oblique photography data and the initial geographic model is ≤0.1m, laying the coordinate foundation for subsequent fusion.

[0054] The structured digital fences are used as vector elements and precisely embedded into the initial geographic model based on their geographic coordinates. Spatial association technology binds the fence attribute information to the corresponding terrain units within the model—designers can click on a fence area to view its engineering attributes and risk level. Subsequently, registered oblique photogrammetric data is overlaid and fused with the initial geographic model. Digital orthophoto maps are used as texture maps and mapped onto the terrain surface of the initial geographic model according to projection matching principles, ensuring that the image texture is consistent with the terrain undulations (e.g., no stretching or distortion in steep slopes). At the same time, the dense geometric information of the 3D point cloud is integrated into the model to supplement the three-dimensional details of the terrain surface (e.g., protruding rocks, depressions, and gullies), improving the 3D visualization accuracy of the model.

[0055] Finally, interactive functional modules for highway design are integrated into the fusion model. At least one practical function is embedded, such as a "dynamic fence attribute query" module (supporting clicking on fences to view cut and fill volumes, risk levels, and protection recommendations), a "design conflict pre-detection" module (automatically popping up warnings when the drawn route crosses a high-risk fence), and a "rapid engineering quantity estimation" module (calculating earthwork volumes based on fence area and average cut and fill depth). This ensures the model not only has the ability to present realistic scenery but also directly supports design decisions. Through these steps, an oblique photogrammetry geographic model is finally generated, integrating real-world terrain, digital fences, and attribute information, and possessing interactive design capabilities.

[0056] In some embodiments, acquiring topographic maps of mountain roads, oblique photogrammetry data from drones, and information on special areas includes: Select topographic map data sources that meet the project's timeliness requirements, and verify the accuracy of the topographic map data to ensure that its horizontal and vertical accuracy meets the error threshold for subsequent model construction; and convert the topographic map data into a standardized format compatible with subsequent data processing workflows, and extract the core topographic and feature elements. Based on the terrain features of the survey area and the preset model resolution requirements, the flight path and parameters of the UAV are planned; data acquisition is performed during the window period that meets the preset environmental conditions, and optical images and position and attitude data are acquired simultaneously; and the raw data is preprocessed, including image correction, point cloud generation and coordinate system 1, to generate the initial 3D point cloud and digital orthophoto map. Collect background information related to geological hazards, engineering constraints, ecology, and hydrology in the survey area; conduct on-site surveys and supplements in key areas to verify or refine the background information; and standardize and encode all collected and surveyed special area information according to a preset data structure, which includes at least information type, spatial range, attribute description, and geographic coordinates.

[0057] The standardized coding rules are as follows: The "Information Type" field uses a two-letter capitalization, such as "DZ" for geological hazards, "GX" for engineering constraints, "ST" for ecological sensitivity, and "SW" for hydrology. The "Spatial Scope" field is described according to the region's shape: regular regions use the center point coordinates plus length and width, while irregular regions use a sequence of consecutive coordinates. The "Attribute Description" field uses a key-value pair format of "Attribute Name: Attribute Value," with multiple key-value pairs separated by semicolons. The "Geographic Coordinates" field uses the same coordinate system format as the baseline topographic map. All field information is separated by a vertical bar "|", forming a unique coded string for each specific region.

[0058] In this embodiment of the invention, the first step is to screen and verify the accuracy of the topographic map data source. Based on the project design cycle and the dynamic characteristics of terrain changes, at least five evenly distributed ground control points (such as mountaintops, corners of fixed buildings, and road intersections) are selected. Their coordinates and elevations are measured in the field and compared with the corresponding point data on the topographic map. The planar position error is verified to meet the standard. If the error exceeds the threshold, the data provider must be contacted for correction or a replacement data source to ensure that the accuracy of the topographic map meets the requirements for subsequent initial geographic model construction.

[0059] Next, the topographic map data format was converted and core elements were extracted. The filtered topographic maps were converted into a standardized format compatible with GIS systems, and redundant information (such as unmarked markers and temporary measurement marks) was removed using professional geographic information software. The core topographic and feature elements were extracted: topographic elements included contour lines (marking elevation values), elevation points (recording precise Z coordinates), and topographic slope aspect markings; feature elements included the center lines and widths of existing roads, the outlines of water systems (rivers, gullies), the locations of bridges and tunnels, and the boundaries of buildings and vegetation cover areas. All extracted elements retained their original coordinate information, forming a structured topographic map base dataset.

[0060] The first step is to plan the UAV flight path. Based on the terrain features of the surveyed area (such as slope and elevation difference) and the preset model resolution requirements (usually requiring an image ground resolution of ≥5cm / pixel), determine the UAV flight parameters: the flight altitude is calculated using the formula "resolution = camera focal length × flight altitude / sensor size" (e.g., for a 20-megapixel camera with a focal length of 16mm, 5cm resolution corresponds to a flight altitude of approximately 80m). The forward overlap is set to 70%-80%, and the lateral overlap is set to 60%-70% to ensure sufficient overlap between images for subsequent matching. For complex terrain sections such as steep slopes and deep canyons, additional intersecting flight paths perpendicular to the main flight path are added to avoid image blind spots. At the same time, no-fly zones (such as under high-voltage lines and military control zones) are marked, and a complete flight path planning document is generated.

[0061] The second step is to select a suitable window for data acquisition. Monitor the weather conditions in the survey area and launch the flight when there is no strong wind (wind speed ≤ 5m / s), no rainfall, and visibility ≥ 5km, avoiding strong midday sunlight (which easily produces strong shadows) and dense fog in the early morning (which affects image clarity). During the flight, the UAV's five-lens tilt camera (front, rear, left, right, and downward view) simultaneously acquires optical images, and the IMU (Inertial Measurement Unit) and GNSS module record the camera position (latitude, longitude, and elevation) and attitude parameters (roll angle, pitch angle, and heading angle) in real time to ensure that each image is associated with accurate spatial position information. After the flight, check the integrity of the images. If there are any missed areas (such as blank areas caused by flight path deviation), a follow-up flight must be carried out immediately.

[0062] The third step involves preprocessing the raw data. Professional photogrammetry software is used to correct distortion in the acquired optical images, eliminating image deformation caused by camera lens errors and flight attitude fluctuations. Based on image overlap and GNSS / IMU data, an aerial triangulation network is constructed using aerial triangulation encryption technology to calculate the precise interior and exterior orientation elements of each image. A dense matching algorithm is then used to perform pixel-by-pixel matching on the encrypted images, generating 3D point cloud data containing a massive number of spatial points (X, Y, Z coordinates). The 3D point cloud is then denoised (removing interference points such as birds and clouds), and orthorectified to generate a digital orthophoto map. Finally, the coordinate system of the 3D point cloud and the digital orthophoto map is aligned with the terrain. Figure 1 A reference coordinate system (such as CGCS2000) is established to form an initial oblique photogrammetry dataset that can be directly used for subsequent registration.

[0063] In the process of acquiring and standardizing information on special areas, the first step is to collect background information through multiple channels. This includes obtaining geological hazard survey reports from local natural resources departments to clarify the approximate scope and risk level of potential hazard areas such as landslides, debris flows, and karst formations; retrieving data on the direction, depth, and diameter of underground pipelines (water supply, gas, and electricity) from housing and construction and transportation departments, as well as design and maintenance records of existing roads, bridges, and tunnels; applying to the ecological and environmental departments for vector maps of the distribution of ecologically sensitive areas such as forest land, wetlands, and habitats of rare plants and animals; and obtaining hydrological data from the hydrological departments for the past 10 years, including river flood levels and gully catchment areas, to ensure coverage of four key categories of information: geological hazards, engineering constraints, ecology, and hydrology.

[0064] Following this, on-site surveys and supplementary verifications were conducted. For areas where the background information was ambiguous or controversial (such as unclear boundaries of geological hazard zones or potential discrepancies between pipeline location markings and actual locations), the survey team conducted on-site verification: ground-penetrating radar was used to detect the distribution of underground karst caves and the actual location of pipelines; rock weathering and soil compaction were analyzed through borehole sampling; total stations were used to measure the boundary coordinates of ecologically sensitive areas; and temporary water level observation points were set up at the bottom of gullies to record flood season water levels. Information discrepancies discovered during the verification (such as the actual pipeline burial depth being 0.5m shallower than the archival record) were corrected, and missing information (such as small areas of unmarked weak interlayers) was supplemented to ensure the accuracy of information in special areas.

[0065] Finally, all information is standardized and coded. Following a pre-defined data structure (containing four main fields: "Information Type - Spatial Range - Attribute Description - Geographic Coordinates"), the collected and surveyed information on special areas is structured: the Information Type field is labeled with categories such as "Geological Hazards," "Underground Pipelines," "Ecologically Sensitive Areas," and "Hydrology"; the Spatial Range field uses coordinate polygons or a center point plus radius to describe the area boundary; the Attribute Description field records specific parameters (e.g., geological hazard areas are labeled "Landslide Risk Level: High, Sliding Surface Depth: 5m," and pipelines are labeled "Type: Gas, Pipeline Diameter: DN300, Burial Depth: 1.8m"); the Geographic Coordinates field uses a coordinate system consistent with topographic maps and oblique photography data; all coded information is imported into the database to form a standardized dataset of special area information, facilitating subsequent integration and retrieval with other data.

[0066] In some embodiments, the dynamic fitness function is constructed as follows: based on an oblique photogrammetry geographic model, a fixed cost assessment is performed on the alternative solutions. The fixed costs include at least one of the following: transportation costs for excavation and filling operations, usage costs, and scheduling loss costs. The quantitative assessment of the above static costs is completed based on a preset engineering unit price database and engineering quantity parameters extracted from the model. Specifically, the transportation costs for excavation and filling operations are calculated by associating the boundaries of the excavation and filling areas in the model with the earthwork allocation scheme, and by accumulating the unit price tables for different transport distance intervals. The usage costs are calculated based on the type of road surface structure (such as mud-bound gravel, graded gravel), thickness, and material unit price, combined with a loss coefficient adjusted according to the terrain slope. The scheduling loss costs are calculated by inputting parameters such as the route slope and curve length into a preset equipment energy consumption model, outputting the additional fuel or electricity consumption, and then multiplying it by the energy unit price.

[0067] Cost estimates are made for potential risks in the future construction phase. Potential risks include at least: costs of delays at blocking points, costs of slope instability, and costs of structural conflict rectification. Static cost items and dynamic risk items are weighted and integrated to form a dynamic fitness function. The weights of static cost items and dynamic risk items are dynamically adjusted based on the terrain complexity index fed back by the oblique photogrammetry geographic model.

[0068] In this embodiment of the invention, a fixed cost assessment is first conducted based on an oblique photogrammetry geographic model. Relying on the centimeter-level three-dimensional terrain parameters provided by the model (such as elevation difference, route length, and cut / fill area boundaries), precise quantitative calculation of various static costs is achieved, avoiding errors in traditional empirical estimations. Specifically, the transportation cost for cut / fill operations is calculated using the model to extract the cut and fill volumes and the spatial distance between cut / fill areas for alternative routes, calculated according to the formula: "(cut volume + fill volume) × average transport distance × unit earthwork transport cost". The usage cost is calculated based on the access road dimensions (length, width, thickness) and road surface structure type marked in the model, combined with the material unit price and the construction loss rate suitable for the terrain (e.g., the loss rate is higher on steep slopes than on gentle slopes), to determine the total cost of material procurement and laying. The scheduling loss cost is calculated by using parameters such as route slope and turning angle identified by the model, correlated with equipment energy consumption coefficients (e.g., fuel consumption is 20% higher on steep slopes of 5° or more than on gentle slopes), combined with the equipment rental unit price and the expected number of trips, to determine the additional loss cost of equipment scheduling.

[0069] Secondly, cost estimates are made for potential risks in the future construction phase. Based on risk labeling information (digital fences, risk classification) in the oblique photogrammetry geographic model, historical construction timeline data is integrated to predict risk costs. For delay costs at blockage points, the model is used to determine the location and extent of bridge and tunnel blockage points and temporary obstacle areas, and historical blockage event duration data is overlaid to estimate the number of days of construction delay. This is then calculated as "delay days × (daily equipment rental fee + labor idle fee + project overdue penalty)". For slope instability costs, the model focuses on high embankment risk areas (yellow / red areas), extracting parameters such as slope gradient, embankment height, and soil bearing capacity. Combined with regional meteorological data (such as rainfall during the rainy season), the instability probability is calculated. If the risk exceeds 30%, reinforcement costs, rework costs, and delay losses are calculated. For structural conflict rectification costs, the model's spatial analysis function is used to detect the distance between the temporary road route and underground pipelines and existing structures. If the distance is less than the safety threshold (e.g., distance to pipeline < 1.5m), rectification expenditures such as pipeline relocation costs and structure protection costs are estimated.

[0070] Finally, the static cost item and dynamic risk item are weighted and fused, and the weights are dynamically adjusted based on the terrain complexity index fed back by the oblique photogrammetry geographic model to form a complete dynamic fitness function. The weighted fusion follows the logic of "comprehensive evaluation value = static cost item × α + dynamic risk item × β" (α + β = 1), and the smaller the comprehensive evaluation value, the better the solution. The terrain complexity index is extracted from the model, including the proportion of high embankment sections, the density of potential blocking points, and the proportion of areas traversing geologically unstable areas. The complexity level is then divided according to the comprehensive score of the index: for low complexity (<40 points), α = 0.7 and β = 0.3 (prioritize cost control); for medium complexity (40-70 points), α = 0.5 and β = 0.5 (balance between cost and risk); for high complexity (>70 points), α = 0.3 and β = 0.7 (prioritize risk avoidance). This ensures that the function adapts to different terrain scenarios and achieves the solution evaluation goal of "controllable short-term costs and controllable long-term risks".

[0071] In some embodiments, the weight adjustment strategy is as follows: based on the oblique photogrammetry geographic model, at least one terrain complexity index is extracted and quantified; wherein, the terrain complexity index includes the proportion of high embankment sections to the total route length, the spatial density of potential blockage points, and the area proportion of the route crossing geologically unstable areas; and the weight of the static cost item and the weight of the dynamic risk item are determined according to the mapping relationship between the terrain complexity index and the weight allocation.

[0072] In this embodiment of the invention, firstly, the terrain complexity index is extracted and quantified based on an oblique photogrammetry geographic model. Leveraging the model's 3D visualization characteristics, precise spatial positioning function, and risk labeling information, efficient acquisition and objective quantification of index data are achieved, avoiding the subjective bias of traditional manual assessment. For the "proportion of high embankment sections to the total route length," the model identifies sections in the candidate alternative routes whose embankment height meets the high embankment standard, extracts the 3D length of each high embankment section, sums them, and then divides them by the total 3D length of the alternative route to obtain the percentage value of this index (e.g., in a certain scheme, the total length of the high embankment section is 200 meters, and the total route length is 1000 meters, the proportion is 20%). For the "spatial density of potential blocking points," the model calls the coordinates of pre-marked potential blocking points such as bridge and tunnel blocking points, underground pipeline intersections, and large structure conflict points, counts the number of blocking points within a 100-meter radius of the candidate route, and then divides by... The total route length is used to obtain a density value in units of "numbers / km" (e.g., if a route is 5 km long and has 8 surrounding blocking points, the density is 1.6 numbers / km). For the "area percentage of the route crossing geologically unstable areas", a geological risk thematic layer (e.g., landslides, karst caves, weak interlayer areas) is overlaid on the model to calculate the actual horizontal projected area of ​​the candidate route crossing such areas, and then divided by the total horizontal projected area of ​​the route (route length × design width) to obtain the percentage quantitative result of this indicator (e.g., if a route has a projected area of ​​5000 square meters and crosses an unstable area of ​​800 square meters, the percentage is 16%).

[0073] Secondly, the core connotation of the terrain complexity index and its impact on weight adjustment are clarified. Each index reflects the "cost-risk" balance requirement of mountainous terrain for access road design from different dimensions: the higher the proportion of high embankment sections, the greater the earthwork volume and the stronger the demand for slope support during construction, and the significantly increased risk of slope instability in the later stage. It is necessary to increase the weight of dynamic risk items to prioritize the avoidance of safety hazards; the higher the spatial density of potential blocking points, the greater the probability of construction delays and equipment idleness caused by blocking in the future. The importance of dynamic risk cost in the scheme evaluation needs to be increased accordingly; the higher the proportion of the route passing through geologically unstable areas, the higher the risk of structural conflicts (such as collision with underground caves) and geological disasters (such as landslides) faced during the construction and use of the access road. It is necessary to strengthen the consideration of dynamic risks through weight adjustment to avoid neglecting long-term safety due to excessive focus on static costs.

[0074] Finally, a mapping relationship between terrain complexity indicators and weight allocation is established. The specific weight values ​​(α+β=1) of the static cost item (α) and the dynamic risk item (β) are determined through a three-level logic of "indicator scoring - comprehensive grading - weight matching". First, each quantitative indicator is graded and scored: <10% high embankment section is low (10 points), 10%-30% is medium (30 points), and >30% is high (50 points); potential blocking point density is <1 / km is low (10 points), 1-2 / km is medium (30 points), and >2 / km is high (50 points); the area traversing geologically unstable regions is <10% low (10 points), 10%-20% is medium (30 points), and >20% is high (50 points). The scores of the three indicators are then summed, and the terrain complexity is categorized according to the total score: a total score < 50 indicates low complexity (flat terrain, low risk), where static cost has a greater impact on the economic viability of the solution, with matching weights α=0.7 and β=0.3; a total score of 50-100 indicates medium complexity (undulating terrain, moderate risk), requiring a balance between cost and risk, with matching weights α=0.5 and β=0.5; a total score > 100 indicates high complexity (complex terrain, high risk), requiring priority control of risk and cost, with matching weights α=0.3 and β=0.7. This mapping relationship replaces subjective experience judgment with objective indicator data, ensuring that the weight adjustment is accurately adapted to the actual terrain conditions, making the evaluation logic of the dynamic fitness function more aligned with actual engineering needs.

[0075] In some embodiments, based on an oblique photogrammetry geographic model, multiple alternative construction access road routes are initialized, including: defining preset basic constraints that the construction access road routes must meet based on the digital fences and hierarchical information calibrated in the oblique photogrammetry geographic model; the basic constraints at least include: avoiding prohibited crossing areas, meeting minimum turning radius limits, and meeting maximum longitudinal slope limits; under the premise of meeting the preset basic constraints, multiple initial construction access road routes are automatically generated in the oblique photogrammetry geographic model as multiple alternative routes; each alternative route is encoded as a chromosome in a genetic algorithm, and the chromosome contains a set of genes, which at least represent: route control point coordinates, cut and fill area boundaries, and slope support structure parameters. The encoding process adopts a segmented real number encoding method. For the route control point coordinate gene, the three-dimensional spatial coordinate values ​​of each control point are directly concatenated into a real number string as gene position values ​​in the order of the route start and end points. For the cut and fill area boundary gene, the vertex coordinate sequence of each cut or fill area polygon is converted into a real number string. For the slope support structure parameter gene, the support type of each section requiring support is represented by a preset numerical code (for example, code one represents wire mesh shotcrete and anchor, and code two represents gravity retaining wall), and continuous parameters such as anchor spacing and retaining wall thickness are directly appended as real values ​​after the type code. All gene segments are spliced ​​together in a predetermined order to form a complete chromosome sequence.

[0076] In this embodiment of the invention, firstly, based on the digital fences and hierarchical information calibrated in the oblique photogrammetry geographic model, preset basic constraints are precisely defined to delineate "insurmountable red lines" and "must-meet technical thresholds" for route generation. Among these, the "avoid prohibited crossing zones" constraint is directly related to the red digital fences in the model (such as ecological core areas and major geological disaster hazard points). Through the model's spatial topology analysis function, a minimum safe distance (e.g., ≥50 meters) is set between the route and the boundary of the prohibited zone to ensure that the generated route does not touch ecological, safety, or engineering red lines. The "comply with minimum turning radius limits" constraint addresses the passage requirements of construction vehicles in mountainous areas (such as heavy dump trucks and road rollers), and differentiates the limits based on the terrain of the areas the route traverses (such as canyon sections and foothill sections) marked in the model. —In flat and open areas, the maximum longitudinal slope should be no less than 15 meters. In narrow canyon areas, due to space constraints, the maximum longitudinal slope can be appropriately relaxed to 12 meters, but the additional explanation of "requiring special vehicle dispatch" must be marked simultaneously. The constraint of "meeting the maximum longitudinal slope limit" is based on the elevation data extracted from the model and sets a threshold for different road sections. Generally, it should not exceed 18%. If the route needs to cross steep slope areas (marked as yellow risk areas in the model), the maximum longitudinal slope can be relaxed to 20%, but the engineering measure of "requiring the addition of anti-slip mat layer" must be linked simultaneously to ensure that the constraint conditions meet the construction requirements and are adapted to the actual terrain.

[0077] Secondly, under the premise of meeting the above-mentioned basic constraints, based on the three-dimensional spatial data of the oblique photogrammetry geographic model, multiple initial construction access road routes are automatically generated through path planning algorithms, forming a diverse alternative solution library. The generation logic adopts a "start-end anchoring + intermediate path random sampling" mode: First, the start point (such as the entrance and exit of the construction camp) and the end point (such as the main construction work surface) of the access road are locked in the model as fixed anchor points for route generation; then, the improved A* algorithm is called, using the elevation data, digital fence, and risk classification in the model as cost functions (low cost for crossing low-risk areas, high cost for approaching high-risk areas), to plan the basic path between the start and end points; on this basis, by randomly perturbing the coordinates of intermediate control points (the perturbation range does not exceed the constraints), 50-100 differentiated routes are generated. For example, for a certain route that needs to bypass a high embankment area, some schemes choose "smooth route along the foot of the mountain", some schemes choose "short-distance crossing by cutting the slope but increasing the support", and some schemes choose "route around the valley but extending the distance", ensuring that the alternative schemes form significant differences in direction, length, and engineering measures, providing sufficient solution space for subsequent genetic algorithm iteration optimization. During the generation process, the model performs constraint verification on each route in real time, automatically eliminating invalid routes that violate prohibited areas, have insufficient turning radius, or exceed longitudinal slope standards. The final selected alternatives all meet the basic constraints.

[0078] Finally, each compliant alternative is encoded as a chromosome that can be recognized by the genetic algorithm, realizing the mapping between "route features and algorithm parameters" and laying a data foundation for iterative optimization. The chromosome adopts a "multi-gene segment combination" structure, with each gene segment corresponding to a core feature of the route: First, the "route control point coordinate gene" consists of the three-dimensional coordinates (X, Y, Z) of 20-30 key control points evenly distributed along the route, accurately representing the spatial orientation of the route (e.g., the coordinates of the 5th control point are (32500.5, 51200.3, 850.2), which represents the planar position and elevation of that point); Second, the "cut and fill area boundary gene" defines the spatial range of the cut area (the area where earth and stone need to be removed) and the fill area (the area where earth and stone need to be filled) through the coordinates of the polygon vertices, and the cut and fill volumes can be directly calculated by combining the model elevation data; Third, the "slope support structure parameter gene" contains quantitative indicators such as slope gradient (e.g., 1:1.5), anchor spacing (e.g., 2 meters), and retaining wall thickness (e.g., 0.8 meters). For high-risk slope sections marked in the model (red grading area), this gene segment will automatically add feature parameters such as "densified anchors" and "lengthened retaining walls". Through this encoding method, all engineering features of each candidate solution are transformed into a computable and heritable numerical sequence, enabling the genetic algorithm to iteratively optimize the solution through operations such as gene recombination and mutation, and finally select the optimal solution.

[0079] In some embodiments, a genetic algorithm is used for iterative optimization to select the optimal construction access road scheme, including: before each iteration of the genetic algorithm, a long short-term memory network time-series prediction model is invoked to determine the dynamic risk prediction results for the current construction stage; an LSTM model integrates real-time updated data from an oblique photogrammetry geographic model with historical construction time-series data to predict the risk characteristics of at least one potential blocking event, including the expected duration of each blocking point, the probability of slope instability in high-fill areas, and the scope of impact; for each chromosome in the current population, the fitness value of the candidate scheme is calculated based on a dynamic fitness function; based on the calculated fitness value, selection, crossover, and mutation operations are performed to generate the next generation population; wherein, the mutation operation adopts a directed mutation mechanism; when a preset termination condition is met, the iteration is terminated and the optimal construction access road scheme is output. The input data serialization processing of the LSTM model is as follows: historical construction time-series data is sliced ​​according to a fixed time window (e.g., weekly), and each slice contains meteorological indicators, construction progress indicators, and corresponding blocking event records for that period. Real-time updated data is merged with data from the current time window. The model outputs a multi-dimensional risk prediction vector for each potential blocking point. Each component of this vector corresponds to risk characteristics such as expected duration, instability probability, and impact range. This LSTM model is trained using a historical dataset labeled with actual blocking events and their characteristics. During training, hyperparameters such as the number of network layers, the number of hidden nodes, and the learning rate are adjusted to minimize the prediction error.

[0080] The implementation rules of the targeted mutation mechanism are as follows: During the mutation phase, all risk prediction results output by the current round of the LSTM model are first sorted according to their severity. Then, based on the set mutation probability, the chromosomal gene segments associated with the top-ranked risk prediction results are selected as mutation targets in descending order of severity. For mutations of the "route control point coordinate gene," control point coordinates are randomly inserted or fine-tuned on the route segment corresponding to the risk area to generate bypass or adjustment paths. For mutations of the "slope support structure parameter gene," the support level is adjusted to a higher level or the support parameter values ​​are increased. For mutations of the "cut and fill area boundary gene," the boundary range of the area is narrowed to avoid risks. This process ensures that mutation operations are prioritized and specifically improve the highest-risk aspects of the plan.

[0081] In this embodiment of the invention, the genetic algorithm iterative optimization takes "risk prediction in advance - dynamic evaluation and adaptation - targeted mutation to control risk" as its core logic. It deeply integrates the dynamic risk prediction of the Long Short-Term Memory Network (LSTM) into the iterative process to ensure that each round of optimization is consistent with the actual risks in the construction stage, and finally selects the optimal solution with "low cost and low risk". The specific process is as follows: Before the fitness evaluation of each iteration of the genetic algorithm begins, the LSTM time series prediction model is called first to generate dynamic risk prediction results that are adapted to the current construction stage, thus solving the limitation of the traditional genetic algorithm that is "based only on static risk assessment".

[0082] The LSTM model uses a dual-source fusion mode of "real-time + historical" for data input: On the one hand, it accesses real-time updated data from the oblique photogrammetry geographic model, including the latest topographic changes (e.g., circumferential displacement of a high embankment area reaches 5mm) supplemented by IoT sensors (e.g., tunnel construction progress is 2 days behind schedule) and meteorological data (e.g., probability of rainy season in the next 15 days); on the other hand, it imports historical construction time series data, covering records of blocking events in similar mountainous projects in the past 3 years (e.g., average duration of bridge and tunnel blockages, probability of blockage relief under different geological conditions) and high embankment slope instability cases (e.g., correlation data between water content and instability probability).

[0083] Through dual-source data fusion optimization, the LSTM model focuses on predicting the "risk characteristics of potential blocking events," specifically outputting three core results: First, the expected duration of the blocking point, such as predicting that the blocking of a bridge construction will be extended from the original 10 days to 18 days due to complex geology, or that the blocking of a tunnel entrance can be lifted 3 days ahead of schedule due to progress exceeding the target; second, the probability of slope instability in high embankment areas, calculated by combining real-time displacement data and historical instability thresholds (e.g., 65%); and third, the scope of risk impact, such as predicting that if slope instability occurs, it will affect the passage of a 150-meter section of the temporary road, requiring temporary detours of approximately 200 square meters. These prediction results directly serve as the basis for calculating the "dynamic risk term" in the subsequent dynamic fitness function, ensuring a high degree of match between the risk assessment and the current construction reality.

[0084] For each chromosome (i.e. each alternative construction access road) in the current genetic algorithm population, the fitness value is calculated by combining the dynamic risk prediction results output by LSTM with the dynamic fitness function, thereby achieving a collaborative quantitative evaluation of "static cost + dynamic risk".

[0085] The calculation logic consists of two steps: First, the static cost items are calculated by extracting the corresponding engineering parameters of the chromosomes based on the oblique photogrammetry geographic model (such as the coordinates of the route control points → calculating the excavation and filling distance, slope support parameters → calculating material costs), and calculating fixed costs such as excavation and filling transportation, equipment scheduling, and material procurement; Second, the dynamic risk items are calculated by converting the risk characteristics predicted by LSTM into costs, such as "expected duration of the blockage point is 18 days" → calculating the delay cost based on "delay days × average daily equipment idle cost × labor cost", "probability of instability of high fill is 65%" → calculating the potential instability cost based on "instability probability × reinforcement engineering cost × rework loss", and "risk impact range is 200 square meters" → calculating the emergency cost based on "temporary diversion engineering cost".

[0086] Subsequently, based on the previously determined terrain complexity weights (e.g., dynamic risk item weight of 0.7 and static cost item weight of 0.3 under high-complexity terrain), the final value is calculated according to "fitness value = static cost item × α + dynamic risk item × β"—the smaller the fitness value, the better the overall "cost-risk" performance of the solution, providing a screening basis for subsequent genetic operations.

[0087] Based on the fitness values ​​of each chromosome, the next generation population is generated through three genetic operations: selection, crossover, and mutation. The focus is on strengthening the adaptability of the scheme to dynamic risks through "directed mutation" to avoid the blindness of traditional random mutation.

[0088] Selection operation: Using the "tournament selection method", 5-8 chromosomes are randomly selected from the current population to form a group, and the 2 chromosomes with the lowest fitness values ​​in the group are selected to enter the next generation; repeat this process until the number of chromosomes in the next generation population (e.g., 50) reaches the target, to ensure that high-quality solutions (low fitness values) are retained first, and to maintain the overall optimization trend of the population.

[0089] Crossover operation: A "two-point crossover" strategy is adopted. Two crossover sites are randomly selected from the chosen parent chromosomes (such as chromosome A and chromosome B), and gene segments between the two sites are exchanged. For example, the coordinates of the 5th to 10th control points in the "route control point coordinate gene segment" are exchanged, or the anchor spacing parameter in the "slope support parameter gene segment" is exchanged, generating a offspring chromosome that combines the advantages of the parent chromosome (such as the combination of A's short transport distance and B's low instability risk). After crossover, it is necessary to verify whether the offspring scheme meets the basic constraints (such as turning radius and longitudinal slope limits), and eliminate non-compliant schemes.

[0090] Targeted mutation operations: Based on the dynamic risk prediction results of LSTM, targeted mutation rules are designed to adjust gene segments in high-risk areas, rather than random mutation. For example, if LSTM predicts that the blockage of a bridge or tunnel will be prolonged by 18 days, the segment of the "route control point coordinate gene" in the chromosome close to the blockage point is mutated to add control points for "temporary connecting roads" (e.g., adding 3 control points outside the blockage point to form an bypass or connecting route). If the probability of instability in a high embankment area is predicted to be 65%, the "slope support parameter gene" is mutated to reduce the anchor spacing from 2 meters to 1.5 meters and increase the retaining wall thickness from 0.8 meters to 1.2 meters to strengthen the support capacity. If a pipeline conflict risk is predicted in a certain area, the "cut-fill area boundary gene" is mutated to reduce the excavation range in that area to avoid pipelines. Targeted mutation enables the offspring scheme to actively avoid high risks in the current stage, significantly improving the overall quality of the population.

[0091] When the preset iteration termination condition is met, the genetic algorithm iteration stops, and the optimal construction access road scheme is selected and output from the final population.

[0092] The preset termination conditions include two categories: First, the "stability condition," which states that if the fluctuation range of the optimal fitness value of the population for three consecutive generations is less than 5% (e.g., the optimal value is 120 in the 10th generation, 118 in the 11th generation, and 119 in the 12th generation, with a fluctuation of <5%), it indicates that the scheme has become stable, and further iteration is unlikely to improve the optimization effect; Second, the "efficiency condition," which states that if the number of iterations reaches the preset upper limit (e.g., 50 generations), the iteration will be terminated even if the stability condition is not met, in order to avoid excessive computation time consumption.

[0093] After the iteration terminates, the chromosome with the smallest fitness value is extracted from the final population and decoded into a specific construction access road scheme: a three-dimensional route is generated through the "route control point coordinate gene", the "cut and fill area boundary gene" determines the cut and fill range and workload, and the "slope support parameter gene" clarifies the details of the support measures; at the same time, the quantitative indicators of the scheme (such as total cost, construction period, and instability risk rate) are output and visualized in combination with the oblique photogrammetry geographic model for confirmation by the design and construction parties. The final output of the optimal scheme not only meets the basic constraints, but also achieves the goal of "controllable static cost and lowest dynamic risk", which can directly guide the on-site construction.

[0094] In some embodiments, the collaborative design system is characterized by including a drone and an electronic device; the method of the above embodiments is used to run on the electronic device.

[0095] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0096] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A smart visual mountain road construction access road rapid design method, characterized in that, include: Acquire topographic maps of mountain roads, oblique photography data from drones, and information on special areas; The mountain road survey topographic map, UAV oblique photography data, and special area information are spatiotemporally registered and fused to generate an enhanced optimization set; wherein, the enhanced optimization set includes a basic feature layer, an engineering feature layer, and a boundary calibration layer; Based on the topographic map of the mountain roads, an initial geographical model was established; The multi-class support vector machine is optimized based on the enhanced optimization set, and digital fences and hierarchical information are calibrated in the initial geographic model based on the optimized multi-class support vector machine to obtain the oblique photogrammetry geographic model. Based on the oblique photogrammetry geographic model, multiple alternative routes for the construction access road are initialized, and a genetic algorithm is used for iterative optimization to select the optimal construction access road route. In the iterative process of the genetic algorithm, a dynamic fitness function is constructed, which includes a static cost term and a dynamic risk term. The static cost term is used to evaluate the performance of each alternative scheme under undisturbed conditions based on static indicators. The dynamic risk term is used to predict the expected risk costs that may be caused by potential blocking events in the future construction phase.

2. The intelligent visual mountainous highway construction access road rapid design method according to claim 1, characterized in that, The dynamic fitness function is constructed as follows: Based on the oblique photogrammetry geographic model, a fixed cost assessment is performed on the alternative solutions. The fixed cost includes at least one of the following: transportation costs for excavation and filling operations, usage costs, and scheduling loss costs. Cost estimates are made for potential risks in the future construction phase, including at least: costs of delays at blocking points, costs of slope instability, and costs of structural conflict rectification. The static cost item and the dynamic risk item are weighted and fused to form the dynamic fitness function; The weights of the static cost item and the dynamic risk item are dynamically adjusted based on the terrain complexity index fed back by the oblique photogrammetry geographic model.

3. The intelligent visual mountainous highway construction access road rapid design method according to claim 2, characterized in that, The weighting adjustment strategy is as follows: Based on the oblique photogrammetry geographic model, at least one terrain complexity index is extracted and quantified; The terrain complexity indicators include the proportion of high embankment sections to the total route length, the spatial density of potential blockage points, and the area proportion of the route crossing geologically unstable regions. Based on the mapping relationship between terrain complexity index and weight allocation, the weights of the static cost item and the dynamic risk item are determined.

4. The intelligent visual mountainous highway construction access road rapid design method according to claim 3, characterized in that, Based on the oblique photogrammetry geographic model, several alternative routes for the construction access road are initialized, including: Based on the digital fences and hierarchical information calibrated in the oblique photogrammetry geographic model, the preset basic constraints that the construction access road route must meet are defined. The basic constraints include at least: avoiding prohibited crossing areas, meeting the minimum turning radius limit, and meeting the maximum longitudinal slope limit. Under the premise of meeting the preset basic constraints, multiple initial construction access road routes are automatically generated in the oblique photogrammetry geographic model as multiple alternative solutions. Each alternative is encoded as a chromosome in a genetic algorithm, the chromosome containing a set of genes that at least characterize: the coordinates of the route control points, the boundaries of the cut and fill areas, and the parameters of the slope support structure.

5. The intelligent visual mountainous highway construction access road rapid design method according to claim 4, characterized in that, The optimal construction access road scheme is selected through iterative optimization using a genetic algorithm, including: Before each iteration of the genetic algorithm, a long short-term memory network time-series prediction model is invoked to determine the dynamic risk prediction results for the current construction stage. The LSTM model integrates real-time updated data from the oblique photogrammetry geographic model with historical construction time-series data to predict the risk characteristics of at least one potential blocking event. The risk characteristics include the expected duration of each blocking point, the probability of slope instability in the high embankment area, and the scope of impact. For each chromosome in the current population, the fitness value of the alternative scheme is calculated based on the aforementioned dynamic fitness function. Based on the calculated fitness values, selection, crossover, and mutation operations are performed to generate the next generation population; wherein the mutation operation employs a directed mutation mechanism. When the preset termination condition is met, the iteration terminates and the optimal construction access road solution is output. 6.The intelligent visual mountain road construction access way rapid design method according to claim 1, characterized in that, The mountain road survey topographic maps, UAV oblique photogrammetry data, and information on special areas are spatiotemporally registered and fused to generate an enhanced optimized set, including: Aerial triangulation and dense matching are performed on the oblique photography data of the UAV to generate three-dimensional point clouds and digital orthophoto maps. The topographic map of the mountain road survey is vectorized to extract contour lines, elevation points and ground features; Using the coordinate system of the mountain road survey topographic map as the reference coordinate system, at least three ground control points are selected, and the three-dimensional point cloud and the digital orthophoto map are geometrically corrected and coordinate transformed according to the least squares method to align with the spatial coordinate system of the reference coordinate system. The spatiotemporally registered 3D point cloud, digital orthophoto map, and vectorized topographic map data are overlaid to obtain a fused data volume. Based on the fused data volume, a basic feature layer, an engineering feature layer, and a boundary calibration layer are constructed. A unique identifier is established for each geographic unit in the basic feature layer, engineering feature layer, and boundary calibration layer, and the feature data of the three layers are associated in spatial location to obtain an enhanced optimization set.

7. The intelligent visual mountainous highway construction access road rapid design method according to claim 6, characterized in that, The spatiotemporally registered 3D point cloud, digital orthophoto map, and vectorized topographic map data are overlaid to obtain a fused data volume, including: Using the vectorized topographic map data as the bottom layer, the digital orthophoto map as the middle layer, and the three-dimensional point cloud as the top layer, a three-dimensional overlay structure with topographic, texture, and geometric information is obtained. The elevation conflicts, feature outline conflicts, and element attribute conflicts in the three-dimensional overlay structure are processed to obtain the fused data volume; The accuracy, completeness, and consistency of the generated fused data volume are verified. 8.The intelligent visual mountain highway construction access road rapid design method according to claim 1, characterized in that, Acquire topographic maps of mountain roads, oblique photogrammetry data from drones, and information on special areas, including: Select topographic map data sources that meet the project's timeliness requirements, and verify the accuracy of the topographic map data to ensure that its horizontal and vertical accuracy meets the error threshold for subsequent model construction; and convert the topographic map data into a standardized format compatible with subsequent data processing procedures, and extract the core topographic and feature elements. Based on the terrain features of the survey area and the preset model resolution requirements, the flight path and parameters of the UAV are planned; data acquisition is performed during the window period that meets the preset environmental conditions, and optical images and position and attitude data are acquired simultaneously; and the collected raw data is preprocessed, including image correction, point cloud generation and coordinate system one, to generate an initial three-dimensional point cloud and digital orthophoto map. Collect background information related to geological hazards, engineering constraints, ecology, and hydrology in the survey area; conduct on-site surveys and supplements in key areas to verify or refine the background information; and standardize and encode all collected and surveyed special area information according to a preset data structure, which includes at least information type, spatial range, attribute description, and geographic coordinates.

9. A co-design system characterized by, Includes drones and electronic devices; the method described in any one of claims 1-8 is used to operate on the electronic device.

Images