A road route selection survey method and system based on machine vision

Abstract

This invention discloses a road alignment survey method and system based on machine vision, relating to the field of geological exploration. The system consists of several functional modules, including: a data acquisition module: acquiring a sequence of multi-angle oblique images collected by a UAV acquisition device along the proposed road corridor; a point cloud generation module: generating a color three-dimensional point cloud model of the proposed road corridor based on the multi-angle oblique image sequence; a structural surface interpretation module: extracting the attitude data of rock mass structural surfaces from the color three-dimensional point cloud model and establishing a structural surface attitude vector field for the proposed road corridor; and a natural risk assessment module: extracting natural slope units from the color three-dimensional point cloud model and calculating the orientation of the natural slope surface; and determining the natural geological hazard avoidance zone within the proposed road corridor based on the spatial intersection relationship between the structural surface attitude vector field and the orientation of the natural slope surface.

Description

Technical Field

[0001] This invention relates to the field of geological exploration, specifically to a road alignment survey method and system based on machine vision. Background Technology

[0002] Road alignment is a core component of linear transportation engineering projects such as highways and railways, directly determining the project's investment scale, operational safety, and geological hazard risk level throughout its entire lifecycle. The complex terrain and geological conditions in the mountainous areas of western and southern my country, characterized by deep canyons, fractured rock masses, and well-developed faults and folds, make road alignment extremely challenging. Traditional road alignment surveys primarily rely on geological engineers conducting field investigations: using compasses, rangefinders, and other tools to measure exposed rock mass surfaces along the route point by point, manually recording occurrence data, and delineating unfavorable geological areas on topographic maps. However, key areas such as steep slopes and deeply incised canyons are often inaccessible to humans, resulting in significant gaps in survey coverage. This leads to the omission of hidden dangerous rock masses and creeping landslides during the survey phase, creating serious safety hazards for construction and operation.

[0003] In recent years, UAV oblique photography and machine vision technologies have been gradually introduced into the field of road surveying to acquire high-resolution images along the route and reconstruct three-dimensional models of the ground surface. However, existing technical solutions still have the following unresolved problems: image acquisition is mostly based on uniform flight paths with fixed height and overlap, failing to adapt to the complexity of rock mass structure development, resulting in insufficient capture of structural surface information of key parts such as steep walls and obstructed areas; rock mass structure information extracted from point clouds or images is only used to identify existing natural geological hazard risks and delineate avoidance zones accordingly, leaving engineering route selection in a passive avoidance state, lacking a dynamic collaborative analysis mechanism based on a unified geological data source between avoidance boundaries and engineering excavation schemes; stability analysis of bedding rock cut slopes is usually conducted independently only after the basic alignment is determined, making it impossible to predict the engineering excavation risks caused by different alignment schemes during the route selection stage, and adjustments are subject to many limitations when problems are discovered, making it difficult to optimize the spatial morphology of the route from the source to reduce the risk of disasters induced by the engineering excavation itself. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution: A machine vision-based road alignment and survey system includes: Data acquisition module: Acquires multi-angle oblique image sequences collected by a drone acquisition device along the selected road corridor; Point cloud generation module: Generates a color 3D point cloud model of the proposed road corridor based on a multi-angle tilted image sequence; Structural surface interpretation module: Extracts the attitude data of rock mass structural surfaces from the color 3D point cloud model and establishes the structural surface attitude vector field of the proposed road corridor; Natural risk assessment module: Extracts natural slope units from the color 3D point cloud model and calculates the orientation of the natural slope surface; determines the natural geological hazard avoidance zone within the proposed road corridor based on the spatial intersection relationship between the structural surface attitude vector field and the orientation of the natural slope surface. Excavation simulation module: Obtains the horizontal alignment data and longitudinal slope data of candidate route schemes from the external design platform, simulates the spatial morphology of the artificially excavated slope, and outputs the spatial location data of the artificially excavated slope. Dynamic risk assessment module: From the structural plane attitude vector field, query the structural plane attitude data corresponding to the spatial location data of the artificially excavated slope, calculate the bedding slip risk index between the structural plane attitude data corresponding to the spatial location data of the artificially excavated slope, and generate a dynamic bedding risk field. Overlay display and scheme generation module: Based on natural geological disaster avoidance zones and dynamic bedding risk fields, generate target route schemes.

[0005] Furthermore, the acquisition process for multi-angle tilt image sequences is as follows: Obtain topographic data of the proposed road corridor and identify its topographic relief features and surface texture complexity; An adaptive 3D flight path is generated based on the topographic relief features. Based on the complexity of the surface texture, a preset first overlap is set for the potential development area of ​​the structural surface, and a preset second overlap is set for the non-development area. The first overlap is higher than the second overlap. The system performs quality checks on the preview images transmitted in real time during flight. When blurry or obstructed images are detected, a supplementary flight path is generated and reshooting is triggered. When exposed rock face features are detected, hovering and multi-angle rotation shooting is triggered. All acquired images are linked with their corresponding shooting positions and attitude angle data to form a multi-angle tilted image sequence with spatial index.

[0006] Furthermore, the process of generating a colored 3D point cloud model of the proposed road corridor is as follows: Feature point extraction and multi-view matching are performed on multi-angle tilted image sequences to obtain the mapping relationship of corresponding feature points between images; Based on the mapping relationship of the same feature points, the pose of the UAV acquisition device is recovered by the structure of motion recovery algorithm and a sparse point cloud is generated. After optimization by global bundle adjustment, the camera pose set and sparse 3D points are obtained. A multi-view stereo matching algorithm is used to match and fuse disparity maps of multiple frames of images pixel by pixel. The matching pixels are back-projected into three-dimensional space through triangulation to generate a dense three-dimensional point cloud. For each 3D point in the dense 3D point cloud, trace back the color of its projected pixels on each visible frame of the image, and assign color values ​​using a weighted fusion strategy based on the shooting angle to generate a colored 3D point cloud model with realistic color texture.

[0007] Furthermore, the process of extracting the attitude data of rock mass structural planes from the color 3D point cloud model is as follows: The color 3D point cloud model is preprocessed to calculate the local surface normal vector of each 3D point and unify its orientation to construct a continuous normal vector field. Based on the mutation feature of normal vectors and the region growing algorithm, several independent structural surface patches are segmented from the normal vector field; For each structural surface patch, perform planar fitting, calculate the dip and dip angle based on the normal vector of the fitted plane, generate the structural surface attitude vector attached to the corresponding surface patch, and collect them to establish the structural surface attitude vector field.

[0008] Furthermore, the process of establishing the structural surface orientation vector field of the proposed road corridor is as follows: For each structural facet, a vector element is constructed. The vector element includes the facet identifier, spatial location range, dip value, tilt angle value, and area parameter. All vector elements are mapped to a stereographic projection map, and high-density clustering areas are identified through adaptive density clustering. Dominant joint groups are automatically divided and their affiliations are labeled. A spatial grid index structure is constructed, and a hierarchical storage mechanism is established that includes the attitude index of dominant joint groups and the attitude index of single facets, forming a structural facet attitude vector field that can be queried in space.

[0009] Furthermore, the process of determining the natural geological hazard avoidance zone within the proposed road corridor is as follows: The colored 3D point cloud model is divided into several natural slope units, and the slope orientation of each natural slope unit is calculated. For each natural slope unit, the associated structural surface attitude data can be queried through the hierarchical spatial index of the structural surface attitude vector field; Based on preset kinematic criteria, slope units with potential damage risks are identified according to slope orientation and associated structural surface attitude data. Slope units identified as having risks are spatially aggregated and classified into different levels to generate natural geological disaster avoidance zones.

[0010] Furthermore, the process of simulating the spatial morphology of an artificially excavated slope is as follows: Obtain the horizontal alignment data and vertical slope data of the candidate route schemes, and establish the three-dimensional route design axis; A sequence of cross-sections is generated along the three-dimensional route design axis and intersected with the color three-dimensional point cloud model to obtain the original terrain profile at each cross-section. At each cross section, based on the design elevation and preset slope parameters, the excavation slope line on the cut side is generated, forming a parametric cross section model. All cross-sectional models are longitudinally connected along the route to construct a three-dimensional surface model of the artificially excavated slope that describes the spatial orientation of each section of the excavated slope.

[0011] Furthermore, the process of generating a dynamic layered risk field is as follows: The spatial morphology of the artificially excavated slope is spatially partitioned to generate several evaluation units and spatial query points within each unit. The spatial index of the structural surface attitude vector field is used to query the structural surface attitude data associated with each evaluation unit. Based on the angle relationship and dip comparison relationship between the orientation of the artificially excavated slope surface of each assessment unit and the queried structural surface attitude data, the bedding sliding risk index is calculated. By combining the bedding slip risk indices of each assessment unit, a dynamic bedding risk field distributed along the route is generated, and the bedding risk field is updated in real time according to the adjustment of candidate route schemes.

[0012] Furthermore, the process of generating the target route plan is as follows: The natural geological hazard avoidance zone and the dynamic bedding risk field are overlaid on the same display interface. The avoidance zone is rendered using a preset first visual coding scheme, and the bedding risk field is rendered using a preset second visual coding scheme. In response to adjustments to the horizontal alignment data or longitudinal slope data of candidate route schemes, the display of the bedding risk field is updated in real time, and the spatial distance between the route alignment and the boundary of the avoidance zone is monitored. When the adjusted candidate route meets the preset safety conditions, it is determined as the target route and output. The safety conditions include that the route does not intrude into the avoidance zone and the entire layer-by-layer risk index is lower than the preset threshold.

[0013] A road alignment survey method based on machine vision includes the following steps: Step 1: Acquire a sequence of multi-angle oblique images collected along the selected road corridor using a drone acquisition device; Step 2: Generate a color 3D point cloud model of the proposed road corridor based on the multi-angle tilted image sequence; Step 3: Extract the attitude data of rock mass structural surfaces from the color 3D point cloud model and establish the attitude vector field of the structural surfaces of the proposed road corridor; Step 4: Extract natural slope elements from the color 3D point cloud model and calculate the orientation of the natural slope surface; determine the natural geological hazard avoidance zone within the proposed road corridor based on the spatial intersection relationship between the structural surface attitude vector field and the orientation of the natural slope surface. Step 5: Obtain the horizontal alignment data and longitudinal slope data of the candidate route scheme from the external design platform, simulate the spatial morphology of the artificially excavated slope, and output the spatial location data of the artificially excavated slope. Step 6: From the structural plane attitude vector field, query the structural plane attitude data corresponding to the spatial location data of the artificially excavated slope, and calculate the bedding slip risk index between the structural plane attitude data to generate a dynamic bedding risk field. Step 7: Generate the target route plan based on the natural geological disaster avoidance zone and the dynamic bedding risk field.

[0014] The present invention provides a road alignment survey method and system based on machine vision, which has the following beneficial effects: (1) This invention constructs a structural surface attitude vector field as a unified geological data base for the entire corridor, and incorporates the disaster risk identification of natural slopes and the engineering risk pre-assessment of artificially excavated slopes into the same data source for collaborative analysis. Compared with the existing technology where geological disaster identification and road cut slope stability analysis belong to different stages, use different data sources, and are separated from each other, this scheme enables route selection engineers to observe the linkage changes between the external natural geological disaster avoidance zone and the internal bedding risk field in real time when adjusting the horizontal alignment or longitudinal slope of the route. This allows them to seek a dynamic balance between avoiding external disaster sources and optimizing internal excavation, and upgrades the traditional passive route selection mode of avoiding a single disaster to an active route finding mode of multi-objective collaborative optimization.

[0015] (2) This invention establishes a closed-loop response mechanism between route adjustment and bedding risk assessment by combining the structural plane attitude vector field with the dynamic simulation of artificially excavated slopes. When the alignment or slope parameters of the candidate route change, the system automatically re-simulates the spatial morphology of the artificially excavated slope and re-queries the associated data from the structural plane attitude vector field to update the bedding risk index in real time. In the existing method, the slope stability analysis is carried out independently only after the route selection is completed, which causes a lag. This scheme advances the prediction of engineering excavation risk to the route selection stage, so that each fine adjustment of the alignment can promptly reflect its impact on slope stability, reduce the risk of bedding slip from the source, and avoid design rework and investment waste caused by geological problems after the route selection is finalized.

[0016] (3) This invention displays natural geological hazard avoidance zones and dynamic bedding risk fields on the same interface with differentiated visual codes, so that the two types of risk information with different properties are precisely matched in space and clearly distinguished in visual hierarchy; the avoidance zone is marked with static color blocks to indicate the insurmountable hazard boundary, and the bedding risk field is marked with dynamic heat maps along the route to indicate the potential risk distribution of the engineering excavation; in the prior art, various geological maps and route design maps are separated and require manual comparison and analysis, which is a cumbersome process. This solution enables route selection engineers to clearly perceive the comprehensive risk situation of the route plan in the global scope, and to intuitively observe the real-time response of risk heat through interactive operations such as dragging the line position. This greatly reduces the difficulty of manually searching for the optimal route plan under the constraints of multiple factors and improves the efficiency and scientific nature of route selection decision. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the system of the present invention; Figure 2 This is a schematic diagram of the overall method of the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0019] Example 1: Please see Figure 1 Embodiment 1 of this application provides a road alignment survey system based on machine vision, the system comprising: Data acquisition module: Acquires multi-angle oblique image sequences collected by a drone acquisition device along the selected road corridor; An integrated airborne payload system for performing multi-angle tilted image acquisition tasks is achieved by using a drone acquisition device mounted on a drone flight platform. This device integrates at least an imaging sensor for converting ground object reflected light signals into digital images, a gimbal stabilization system for isolating flight attitude changes to maintain image stability, a positioning and attitude determination unit for recording the geographical location and attitude information at the moment each image is captured, and a data transmission link for local storage of image data and real-time preview transmission. Through the collaborative work of the above components, a multi-angle tilted image sequence with accurate spatiotemporal position and attitude metadata is obtained, providing raw image data with multi-view coverage and spatial solvability for the subsequent point cloud generation module.

[0020] The system acquires rough terrain data of the proposed road corridor. Based on existing digital elevation models or satellite imagery, and according to this rough terrain, it expands a predetermined width to both sides of the initial horizontal alignment of the proposed road to form a flight corridor with a certain buffer zone. The system automatically identifies the terrain undulation features within the corridor, including ridgelines, valley orientations, and the distribution of steep slopes. Based on these terrain features, it generates one or more three-dimensional flight paths adapted to the terrain, enabling the UAV to maintain a predetermined relative height and distance from the slope during flight. For steep canyon areas, a layered looping strategy is adopted for the flight paths: flight paths at different elevations are set at the valley bottom, the middle of the slope, and near the platform at the top of the slope, ensuring that the image sequence can cover the same slope area from multiple angles such as overhead, eye level, and upward, avoiding visual obstruction of the rock structure due to a single perspective.

[0021] After the flight path planning is completed, image overlap parameters are set for each flight strip. The overlap parameters include directional overlap along the flight direction and lateral overlap between adjacent flight strips. The specific values ​​are not fixed, but are set differently according to the degree of slope fragmentation. The system first estimates the complexity of the surface texture within the corridor. If a certain area shows steep slope changes and dense gullies in the terrain data, the area is marked as a potential structural surface development area. For these potential structural surface development areas, the system automatically increases the directional overlap and lateral overlap to a preset first threshold range, so that the same surface feature point can be captured by dozens of images from different angles. For areas with gentle and complete terrain, the overlap is reduced to a preset second threshold range to balance data acquisition efficiency and 3D reconstruction accuracy. The overlap value in the first threshold range is higher than that in the second threshold range.

[0022] During the drone's flight path, the onboard image acquisition device transmits low-resolution preview image streams in real time. The system performs real-time blur and occlusion detection on the preview image stream. If motion blur or large-scale shadow occlusion is detected in the current image, the system immediately records the image's shooting pose information and automatically generates a supplementary flight path for the area after completing the current flight path, triggering the drone to return and reshoot until a clear, unobstructed, and effective image is obtained. If a large area of ​​exposed rock face is found in the preview image, the system determines that there may be a steep and dangerous rock mass. It then automatically hovers over the area and triggers a multi-angle rotation shooting sequence. That is, while maintaining the hovering position, the drone rotates the gimbal camera from the normal direction facing the slope to the left and right in sequence according to the preset angle step size to obtain a detailed multi-angle tilt image sequence of the dangerous rock mass area, thereby forming image data enhancement for local high-risk areas.

[0023] After the UAV flight concludes, all collected raw oblique images, recaptured images, and rotating image sequences are compiled to form a raw image dataset. For each frame in the raw image dataset, the geographic coordinates at the time of capture, the gimbal attitude angle, and camera intrinsic parameters are extracted, and this metadata is written into the image's attribute information. The images are arranged and numbered according to the capture time sequence and flight path, creating an index table containing image file names, capture locations, attitude angles, and the corresponding flight path number. This index table is the final output multi-angle oblique image sequence. Its data structure allows the subsequent point cloud generation module to quickly retrieve image pairs that need feature matching based on the spatial adjacency and overlap between images, completing the transformation from an unordered image set to an ordered sequence.

[0024] Point cloud generation module: Generates a color 3D point cloud model of the proposed road corridor based on a multi-angle tilted image sequence; The system receives a multi-angle tilted image sequence and its index table output by the data acquisition module; it iterates through each frame of the tilted image sequence, uses a scale-invariant feature transform algorithm to extract local feature points in each frame, and generates a feature description vector for each feature point; the feature description vector is a quantitative representation of the gradient direction distribution in the neighborhood of the feature point, and has scale and rotation invariance; after completing feature extraction for all images, it determines candidate image pairs with overlapping relationships based on the image spatial adjacency and overlap information recorded in the index table; for each candidate image pair, it calculates the Euclidean distance between the feature description vectors of the two frames, and uses the distance ratio method to select feature point matching pairs that meet the preset similarity threshold; if the number of successfully matched feature point pairs for an image pair exceeds the preset minimum matching threshold, the image pair is confirmed as a valid associated image pair, and the mapping relationship of the same-name feature points between the two is recorded.

[0025] Based on the mapping relationship of corresponding feature points among all valid associated image pairs, a global feature trajectory is constructed. Each feature trajectory corresponds to a set of projection points of an object point in 3D space onto different images. An initial image pair is selected, and the fundamental or essential matrix between the image pairs is estimated using epipolar geometry constraints and random sampling consensus algorithms. The relative motion parameters of the image pairs, namely rotation matrix and translation vector, are then decomposed. The 3D coordinates of the commonly visible object points of the image pairs are recovered using triangulation to form an initial sparse point cloud. An incremental reconstruction strategy is adopted to gradually register the remaining images into the reconstructed coordinate system: for each frame to be registered, the perspective n-point mapping relationship between it and the reconstructed point cloud is solved to obtain the camera pose of the image. The newly registered object points visible in the newly registered images are then added to the sparse point cloud through triangulation. After all images are registered, global bundle adjustment optimization is performed. The objective function is to minimize the sum of reprojection errors of all object points on all images. The camera pose parameters of all images and the 3D coordinates of all object points are jointly optimized to obtain an accurate sparse point cloud and camera pose set.

[0026] Using bundle adjustment-optimized images and their camera poses as input, pixel-by-pixel stereo matching is performed. A multi-view stereo matching algorithm is employed to select several neighboring frames with similar spatial positions and viewpoints for each reference frame in the image sequence, forming a matching set. At each pixel position in the reference image, the best matching pixel is searched for in each neighboring frame of the matching set along the corresponding epipolar direction, generating a disparity map for each pixel. The disparity maps of multiple frames are fused, and each valid matching pixel is back-projected into 3D space using triangulation to generate dense 3D point coordinates. During this process, outliers caused by mismatches are eliminated through adjacent pixel disparity smoothness constraints and left-right disparity consistency checks.

[0027] After obtaining the dense 3D point coordinates, each 3D point needs to be assigned realistic color information to support the subsequent visual interpretation of rock mass structure surfaces. The specific operation is as follows: for each 3D point in the dense point cloud, trace back its projected pixel position on each visible image frame and extract the color value of the projected pixel; if a 3D point is visible in multiple images simultaneously, a preset color fusion strategy is used to determine its final color value; this strategy can be a weighted average of the color values ​​of multiple frames, with the weight determined based on the angle between the 3D point and the camera's optical axis and the distance to the camera when each image is captured, prioritizing the color of orthogonally captured images; the fused color value is assigned to the corresponding 3D point; the original colorless geometric point cloud is then transformed into a colored 3D point cloud model with realistic color texture information. This model not only preserves the three-dimensional shape of the terrain and features, but also carries the surface visual attributes such as exposed rock surfaces and vegetation cover, providing basic data with both geometric and visual information for the subsequent structure surface interpretation module.

[0028] Structural surface interpretation module: Extracts the attitude data of rock mass structural surfaces from the color 3D point cloud model and establishes the structural surface attitude vector field of the proposed road corridor; The process of extracting the attitude data of rock mass structural planes is as follows: The system receives a color 3D point cloud model output from the point cloud generation module; performs statistical outlier removal on the point cloud model, i.e., calculates the average distance from each 3D point to its neighboring points within a preset neighborhood; if the average distance exceeds a preset multiple of the global average distance standard deviation, it is removed as a noise point; constructs a kd-tree spatial index structure for the filtered point cloud to accelerate subsequent neighborhood queries; calculates the local surface normal vector for each 3D point in the point cloud: with the point as the center, searches for all neighboring points within a preset radius, performs principal component analysis on these neighboring points, calculates the covariance matrix, and solves for the eigenvalues ​​and eigenvectors of the covariance matrix; uses the eigenvector corresponding to the smallest eigenvalue as the initial normal vector direction for the point; uses a preset minimum spanning tree method to globally reorient all initial normal vectors to ensure that the normal vectors of adjacent points have the same orientation, thus constructing a continuous normal vector field covering the entire corridor.

[0029] Based on the normal vector field, discrete discontinuities in the rock mass are identified. The angular deviation of the normal vector between each 3D point and its neighboring points is calculated. If the angular deviation between the normal vector of a point and its neighboring points is less than a preset smoothing threshold, the two points are determined to belong to the same smooth surface region. If the angular deviation exceeds a preset abrupt change threshold, a discontinuity boundary is marked between the two points. The point with the smallest angular deviation is selected as the seed point, and the region growth process is initiated. Starting from the seed point, neighboring points that meet the conditions of normal vector smoothing and color / texture consistency are gradually included in the current growth region. When the growth front touches the boundary of a discontinuity surface or the edge of a point cloud cavity, the expansion in that direction is stopped. After the growth is completed, if the number of points in the current region is greater than the preset minimum number of patch points, it is recorded as an independent structural patch and assigned a unique identifier. If the number of points is insufficient, the fragment region is discarded. The point with the smoothest normal vector among those not belonging to any patch is iteratively selected as the new seed point, and the growth process is repeated until all 3D points are processed, resulting in several independently segmented structural patches.

[0030] For each independent structural surface patch obtained from the segmentation, the best plane is fitted from the set of three-dimensional points contained within it. With the objective of minimizing the sum of the squared vertical distances from each point within the patch to the fitted plane, the least-squares solution of the plane is obtained using eigenvalue decomposition, yielding the plane equation parameters representing the spatial orientation of the structural surface, i.e., the normal vector coefficients. Based on the plane normal vector, the attitude data of the structural surface is calculated, including dip and dip angle. The dip is defined as the azimuth angle of the projection of the plane normal vector onto the horizontal plane, and the dip angle is defined as the angle between the plane normal vector and the vertical line. The calculated dip and dip angle are assigned to the corresponding structural surface patch, generating a structural surface attitude vector attached to each patch. The attitude vectors of all structural surface patches are collected to establish a structural surface attitude vector field for the proposed road corridor. Each vector element in this field contains a unique identifier for the structural surface patch, its three-dimensional spatial location range, dip value, and dip angle value, providing a unified geological data foundation for the subsequent natural risk assessment module and dynamic risk evaluation module.

[0031] The process of establishing the structural surface attitude vector field of the proposed road corridor is as follows: Following the previous processing step, several independently segmented structural surface patches and their fitted plane equations have been obtained. For each structural surface patch, based on its plane equation parameters, the attitude data of the patch is calculated, including dip and dip angle values. The dip direction is defined as the clockwise angle between the projection of the plane normal vector onto the horizontal plane and the true north direction, and the dip angle is defined as the angle between the plane normal vector and the vertical direction. A structural surface attitude vector element is constructed for each structural surface patch. The data structure of this vector element includes: a unique identifier for the patch, the spatial range of the three-dimensional point set contained in the patch, the dip value, the dip angle value, and the area parameter of the patch. The area parameter is estimated by the ratio of the number of three-dimensional points within the patch to the average density of the point cloud. The discrete structural surface patches are transformed into vector data units with clear geomechanical meaning.

[0032] The dip and dip angle of all structural surface attitude vector elements are mapped onto a stereographic projection map, with each vector corresponding to a pole in the projection map. A preset adaptive density clustering algorithm is used to perform cluster analysis on the poles in the stereographic projection map, automatically identifying high-density clusters of pole distribution. Each high-density cluster corresponds to a group of dominant joints, and its cluster center represents the dominant dip and dip angle of the structural surface in that group. During the clustering process, preset cluster radius thresholds and minimum pole number thresholds are used to mark discrete poles that do not belong to any high-density cluster as random structural surfaces. A dominant joint group affiliation label is added to each structural surface attitude vector element. If the vector belongs to a dominant joint group, its group number is recorded; if it belongs to a random structural surface, it is marked as unaffiliated.

[0033] All structural surface attitude vector elements that have completed clustering and attribution labeling are spatially organized to establish a structural surface attitude vector field covering the proposed road corridor. The specific construction method of this vector field is as follows: based on the three-dimensional spatial range of the proposed road corridor, a spatial grid index structure is constructed. Each grid cell stores an index list pointing to several structural surface attitude vector elements. The index list records the exposed or potentially influential structural surface vector identifiers within the spatial range of that grid cell. To support efficient querying by the subsequent natural risk assessment module and dynamic risk evaluation module, a hierarchical storage mechanism is established for the spatial grid index: the first level is the dominant joint group attitude index, storing the dominant tendency, dominant dip angle, and spatial influence range of each dominant joint group; the second level is the single-surface attitude index, storing the detailed attitude parameters and spatial location of each structural surface facet. Thus, the structural surface attitude vector field forms a three-layer data system containing a discrete vector set, grouping and attribution labels, and a spatial hierarchical index, allowing downstream modules to quickly obtain structural surface attitude information at any location based on any three-dimensional spatial query coordinates.

[0034] Natural risk assessment module: Based on the spatial intersection relationship between the structural surface attitude vector field and the natural slope orientation, determine the natural geological hazard avoidance zone within the proposed road corridor; The system receives a colored 3D point cloud model output from the point cloud generation module and a structural surface attitude vector field output from the structural surface interpretation module. It distinguishes between natural slopes and artificial objects in the colored 3D point cloud model. Specifically, based on the initial plane alignment of the proposed road, a strip-shaped area extending to both sides of a preset width is marked as the internal area of ​​the road corridor and is temporarily left unprocessed. For the terrain point cloud outside this strip-shaped area, a region segmentation algorithm based on slope and elevation continuity is used to divide the terrain surface into several independent natural slope units. For each natural slope unit, its boundary contour and internal 3D point set are extracted. The slope orientation of each natural slope unit is calculated, including two parameters: slope dip and slope inclination angle. The slope dip is the projection azimuth angle of the normal vector of the overall slope of the slope unit onto the horizontal plane, and the slope inclination angle is the angle between the slope normal vector and the vertical line, obtained by performing plane fitting on the internal 3D point set of the unit and solving the fitted plane equation.

[0035] For each identified natural slope unit, its spatial location range is used as the query condition to perform spatial retrieval of the structural surface attitude vector field. The spatial retrieval is performed using a hierarchical spatial grid index established in the structural surface attitude vector field: first, the spatial grid units covered by the natural slope unit are queried; then, the index list of structural surface attitude vector elements stored in these grid units is extracted; finally, the complete structural surface attitude vector data is obtained based on the index list, including the dip value, dip angle value, and dominant joint group affiliation label of each structural surface; if a natural slope unit is associated with multiple structural surface attitude vectors, all associated structural surface attitude vectors are collected into a set of structural surfaces of the slope unit, and the spatial relative positional relationship between each structural surface is preserved.

[0036] For each natural slope unit, its slope orientation parameters and associated set of structural surfaces are input into a pre-defined slope kinematics judgment model; this judgment model incorporates criteria for three failure modes: The first criterion is planar sliding: it is determined whether there exists a structural surface in the set of structural surfaces whose angle difference with the slope inclination is less than a preset allowable sliding deviation threshold, and whose inclination angle is less than the slope inclination angle, and whose inclination angle is greater than its own preset internal friction angle; if all conditions are met, the slope element is marked as having a risk of planar sliding.

[0037] The second criterion is the wedge failure criterion: it is determined whether there are two structural surfaces with different attitudes in the set of structural surfaces, and the difference between the angle between the dip of their intersection line and the dip of the slope surface is less than the preset wedge slip tolerance threshold, and the dip angle of the intersection line is less than the dip angle of the slope surface; if both conditions are met, the slope unit is marked as having a risk of wedge failure.

[0038] The third criterion is the toppling failure criterion: determine whether there is a steeply inclined structural surface in the set of structural surfaces, whose inclination is approximately opposite to the slope inclination, and whose inclination angle plus the slope inclination angle exceeds a preset toppling critical angle threshold; if the condition is met, then the slope unit is marked as having a toppling failure risk.

[0039] Based on the judgment results of step three, all slope units marked as having at least one risk of damage are identified as potential hazard sources. These potential hazard sources are then spatially aggregated: if multiple adjacent potential hazard sources are spatially continuous, they are merged into a continuous hazard impact area. For each hazard impact area, a preset risk level is assigned based on the damage mode type and distribution density of the potential hazard sources within it. For example, areas with two or more damage modes and a structural surface density exceeding a preset density threshold are marked as high-risk avoidance areas; areas with only a single damage mode and a sparse distribution are marked as medium-risk avoidance areas. The boundary lines of the delineated hazard impact areas at each level are output to form natural geological hazard avoidance areas within the proposed road corridor. These avoidance areas are stored in the form of spatial polygon layers, with each polygon recording its boundary coordinates, risk level, and main hazard type, serving as one of the input constraints for subsequent overlay display and scheme generation modules.

[0040] Excavation simulation module: acquires the horizontal alignment data and longitudinal slope data of candidate route schemes, and simulates the spatial morphology of artificially excavated slopes; The system obtains complete spatial parameters of the candidate route schemes to be evaluated from the road alignment design platform or interactive design interface. It establishes a data communication interface with the road alignment design platform or interactive route design software, receiving the spatial geometric parameters of the candidate route schemes initially drafted by the alignment engineer in the current design session. The horizontal alignment data describes the geometric orientation of the candidate route on the horizontal plane, specifically including the three-dimensional coordinate sequence of the horizontal control points, the type of connecting curves between each control point, and their radius of curvature. The longitudinal slope data describes the elevation changes of the candidate route along the mileage direction, specifically including the mileage markers of the slope change points, the design elevation, and the longitudinal slope values ​​between adjacent slope change points. This system does not autonomously generate candidate route schemes internally; instead, it uses the route geometric data output from the external design platform as input to drive subsequent manual excavation slope simulation, bedding risk assessment, and dual-layer overlay display. The scheme optimization feedback forms a data loop between external route design and internal geological risk assessment, enabling route selection engineers to iteratively adjust the alignment or slope based on the risk visualization results fed back by this system within the original design platform. The spatial parameters consist of two parts: the first part is the horizontal alignment data, which includes the geometric shape description of the route scheme on the horizontal plane, specifically the three-dimensional spatial coordinate sequence of several horizontal control points, the connection curve type parameters between each control point, and the corresponding radius of curvature; the second part is the longitudinal slope data, which includes the mileage station numbers of slope change points distributed along the route mileage direction, the design elevation at each slope change point, and the longitudinal slope value between adjacent slope change points. The horizontal alignment data and the longitudinal slope data are spatially synchronized and correlated, that is, under the same route mileage station system, a correspondence between horizontal coordinates and design elevations is established, forming a continuous route design axis with three-dimensional spatial location attributes.

[0041] Based on the 3D route design axis generated in step one, starting from the route starting point, the design elevation and spatial tangent vector of each sampling point are sequentially obtained along the axis direction according to the preset cross-section sampling step size. At each sampling point, the horizontal projection direction is calculated based on the spatial tangent vector of that point, thereby determining the normal direction of the cross-section at that sampling point, which is the horizontal direction perpendicular to the route's forward direction. With that sampling point as the center, a preset half-width is extended to the left and right sides along the normal direction to form a horizontal cross-section. The color 3D point cloud model is spatially intersected with this horizontal cross-section to obtain the original surface elevation profile line at the location of the cross-section, which is a series of original terrain elevation points distributed along the direction of the cross-section. This operation is performed on all sampling points to generate a series of original terrain cross-sections distributed along the entire length of the route.

[0042] For the original topographic cross-section at each sampling point, based on the design elevation, longitudinal slope data, and preset roadbed width, slope grade, and slope parameters for each grade, the morphology of the artificially excavated slope is simulated. Specifically, using the road surface center point corresponding to the design elevation as a reference, the top surface outline of the roadbed platform is constructed to the left and right sides respectively. When the cut-fill junction point at the edge of the roadbed platform is higher than the original topography, that side is in a cut state. Starting from the cut-fill junction point, the slope is constructed according to the preset slope parameters for each grade. The excavation slope line is generated diagonally upwards at a predetermined rate until it intersects with the original terrain surface. This intersection point is the top boundary point of the excavated slope. When the cut-fill junction is lower than the original terrain, this side is in a fill state. The fill slope line is generated diagonally downwards according to the preset fill slope rate until it intersects with the original terrain surface. Thus, a parameterized artificial excavated slope cross-sectional model is generated at each sampling point, where the slope line sequence on the cut side constitutes the outline of the artificial excavated slope on the cross-section.

[0043] The cross-sectional models of the artificially excavated slopes at all sampling points are longitudinally connected to form a continuous spatial surface of the artificially excavated slopes. Specifically, along the route, the corresponding slope shoulders and toe points of the same level on the excavated slopes between adjacent sampling points are connected sequentially to construct a three-dimensional triangular network model of the graded slopes. The three-dimensional triangular network model accurately describes the spatial location, slope dip, and slope inclination of each section of the road cut slope. The spatial morphology of the artificially excavated slopes that strictly corresponds to the current candidate route scheme is obtained. This spatial morphology is stored in the form of a three-dimensional curved surface geometric model, in which each triangular facet carries the spatial orientation information of its location. This provides the target slope data on the engineering excavation side for the subsequent dynamic risk assessment module to query the structural surface attitude vector field and perform bedding-parallel sliding risk calculation.

[0044] Dynamic risk assessment module: From the structural plane attitude vector field, query the structural plane attitude data corresponding to the spatial location of the artificially excavated slope, calculate the bedding slip risk index between the artificially excavated slope and the queried structural plane attitude data, and generate a dynamic bedding risk field. The system receives the spatial morphology of the artificially excavated slope, i.e., a three-dimensional triangular mesh model, output by the excavation simulation module. This model is then divided into several continuous evaluation units along the route mileage direction according to a preset longitudinal segment length. Each evaluation unit corresponds to an independent artificially excavated slope. For each evaluation unit, the vertex coordinates of all triangular facets within it are extracted, and the overall slope orientation of the evaluation unit is calculated, including the average slope dip and average slope angle. A preset number of spatial query points are selected within the evaluation unit, and these query points are evenly distributed across the slope area of ​​the evaluation unit. Their three-dimensional coordinates serve as the location index for retrieving structural surface data from the structural surface attitude vector field.

[0045] For the coordinates of the spatial query points generated for each evaluation unit, the hierarchical spatial grid index of the structural surface attitude vector field is used for retrieval. Specifically, based on the three-dimensional spatial location of the query point, its corresponding spatial grid unit is located; the second-level single-surface attitude index stored in the grid unit is read to obtain all structural surface attitude vector elements existing within a preset influence radius near the spatial location; each retrieved structural surface attitude vector element contains the unique identifier of the structural surface, dip value, dip angle value, and dominant joint group affiliation label; the structural surface attitude vectors retrieved from all query points within the evaluation unit are merged and deduplicated to form the structural surface attitude dataset of the evaluation unit; this dataset will serve as the basic basis for the geological conditions of the artificially excavated slope section.

[0046] For each assessment unit, its structural plane attitude dataset is compared with the orientation of the artificially excavated slope obtained from the excavation simulation, and the bedding-parallel sliding risk index is calculated. The specific calculation logic is as follows: traverse each structural plane attitude vector or the representative attitude of its dominant joint group in the structural plane attitude dataset of the assessment unit, and calculate the angle difference between the dip of the structural plane and the dip of the artificially excavated slope; if the angle difference is less than the preset bedding-parallel angle threshold, it indicates that the strike of the structural plane is nearly parallel to the strike of the artificially excavated slope, and then further determine the relationship between the dip angle of the structural plane and the dip angle of the artificially excavated slope; if the dip angle of the structural plane is less than the dip angle of the artificially excavated slope, and the dip angle of the structural plane is greater than the preset minimum internal friction... If the angle threshold is met, it is marked as high bedding-parallel sliding risk, and a corresponding bedding-parallel sliding risk index value is generated. This index value is proportional to the reciprocal of the difference between the dip angle of the structural surface and the dip angle of the slope, and inversely proportional to the difference in the angle between the dip direction of the structural surface and the dip direction of the slope. If the angle difference is less than the bedding-parallel angle threshold, but the dip angle of the structural surface is greater than or equal to the dip angle of the slope, it is marked as medium bedding-parallel sliding risk, and a lower bedding-parallel sliding risk index value is generated. If there are multiple risky structural surfaces in the assessment unit, the bedding-parallel sliding risk index value corresponding to the most unfavorable structural surface is taken as the comprehensive bedding-parallel sliding risk index of the assessment unit. If there are no structural surfaces in the assessment unit that meet the bedding-parallel angle threshold condition, its bedding-parallel sliding risk index is set to zero.

[0047] The comprehensive bedding-slip risk index of all assessment units is correlated with their spatial location to form a bedding-risk data sequence distributed along the route direction. A spatial interpolation algorithm is used to transform the discrete assessment unit risk index into a continuous spatial scalar field, i.e., a dynamic bedding-risk field. This bedding-risk field is plotted with the route mileage as the horizontal axis and the risk index as the vertical axis, and is rendered as a heat map through color mapping, so that high-risk sections are presented with a first preset hue, such as red, and low-risk sections are presented with a second preset hue, such as green. The dynamism is reflected in the fact that when the horizontal alignment data or longitudinal slope data of the candidate route scheme are adjusted, the excavation simulation module will regenerate the spatial morphology of the artificial excavation slope, repeat the process, and update the distribution of the bedding-risk field in real time, thereby providing the overlay display module with a risk assessment view that is strictly synchronized with the current route scheme.

[0048] Overlay display and scheme generation module: Overlay display of natural geological hazard avoidance areas and dynamic bedding risk fields, guiding the adjustment of horizontal alignment data or longitudinal slope data based on the distribution of bedding risk fields in areas outside the avoidance areas, and generating target route schemes; The system receives the natural geological hazard avoidance zone layer output by the natural risk assessment module and the dynamic bedding risk field layer output by the dynamic risk assessment module. The two layers are uniformly registered in a spatial reference system to ensure that they are accurately aligned in the same display coordinate system. For the natural geological hazard avoidance zone layer, a first visual coding scheme is assigned based on the risk level of each avoidance zone polygon: high-risk avoidance zones are rendered with a first preset color, such as red, and their filled areas are rendered with a preset first transparency; medium-risk avoidance zones are rendered with a second preset color, such as orange, and their boundary lines are outlined with solid lines. For the dynamic bedding risk field layer, a second visual coding scheme is assigned based on the bedding sliding risk index distributed along the route: the risk index values ​​are mapped to a color bar that transitions from a third preset color, such as green, to a fourth preset color, such as red. Sections with a risk index of zero or below a preset low-risk threshold are displayed in the third preset color, and sections with a risk index above a preset high-risk threshold are displayed in the fourth preset color. Intermediate values ​​are mapped to transition colors linearly or non-linearly to form a continuously changing banded heat map along the route.

[0049] The two visually encoded layers are overlaid and rendered on the same display interface. The bottom layer is a colored 3D point cloud model generated by the point cloud generation module or a large-scale topographic base map derived from it. The middle layer is a natural geological hazard avoidance zone layer, and the top layer is a dynamic bedding risk field layer. The dynamic bedding risk field layer is displayed in the center along the alignment of the candidate route. The display interface simultaneously shows the horizontal alignment and longitudinal slope line of the current candidate route. The overlay display allows users to observe the bedding risk of the excavated slope of the route itself at the same time. Through the vertical distribution of the heat map colors and the relative spatial relationship between the route and the surrounding natural hazard avoidance zone, and through the horizontal distribution of the avoidance zone color blocks, a comprehensive perception of both internal and external risks can be achieved.

[0050] The system monitors user adjustments to candidate route plans on the display interface. These adjustments include dragging route alignment control points to change alignment data and dragging slope change points to change longitudinal profile slope data. When an adjustment is detected, the following closed-loop response process is executed: the adjusted alignment data or longitudinal profile slope data is captured and sent to the excavation simulation module to regenerate the spatial morphology of the artificially excavated slope; the dynamic risk assessment module re-queries the structural plane attitude vector field based on the new slope morphology and calculates the bedding slip risk index to generate an updated dynamic bedding risk field; the updated dynamic bedding risk field is sent back to this module, and the heat map rendering on the display interface is refreshed according to the visual coding scheme in step one; the spatial distance between the adjusted route alignment and the boundary of the natural geological hazard avoidance zone is calculated in real time. If the alignment intrudes into or approaches the preset warning distance threshold of the avoidance zone boundary, a visual or auditory warning is given on the display interface to alert the user that there is a safety hazard in the adjusted direction.

[0051] After the user repeatedly adjusts the horizontal alignment data and longitudinal slope data and observes the linkage changes of the two layers, when the candidate route scheme meets the preset generation conditions, it is determined as the target route scheme and output. The preset generation conditions are: the entire route does not intrude into any natural geological disaster avoidance zone, and there are no sections in the dynamic bedding risk field along the route with a bedding slip risk index higher than the preset high-risk threshold. If the above conditions cannot be fully met due to terrain limitations, a preset comprehensive objective function is used for quantitative comparison. The comprehensive objective function uses the length of the avoidance zone intrusion and the integral of the bedding risk index along the route as the cost term, and the candidate scheme with the smallest weighted sum of the cost terms is taken as the target route scheme. The final output target route scheme includes horizontal alignment data, longitudinal slope data, and natural disaster avoidance zone bypass verification report and bedding risk assessment report associated with the scheme.

[0052] Example 2: Please see Figure 2Based on Example 1, Example 2 of this application also provides a road alignment survey method based on machine vision, including the following specific steps: Step 1: Acquire a sequence of multi-angle oblique images collected along the selected road corridor using a drone acquisition device; Step 2: Generate a color 3D point cloud model of the proposed road corridor based on the multi-angle tilted image sequence; Step 3: Extract the attitude data of rock mass structural surfaces from the color 3D point cloud model and establish the attitude vector field of the structural surfaces of the proposed road corridor; Step 4: Extract natural slope elements from the color 3D point cloud model and calculate the orientation of the natural slope surface; determine the natural geological hazard avoidance zone within the proposed road corridor based on the spatial intersection relationship between the structural surface attitude vector field and the orientation of the natural slope surface. Step 5: Obtain the horizontal alignment data and longitudinal slope data of the candidate route scheme from the external design platform, simulate the spatial morphology of the artificially excavated slope, and output the spatial location data of the artificially excavated slope. Step 6: From the structural plane attitude vector field, query the structural plane attitude data corresponding to the spatial location data of the artificially excavated slope, and calculate the bedding slip risk index between the structural plane attitude data to generate a dynamic bedding risk field. Step 7: Generate the target route plan based on the natural geological disaster avoidance zone and the dynamic bedding risk field.

[0053] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0054] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0055] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A road alignment and survey system based on machine vision, characterized in that, The system includes: Data acquisition module: Acquires multi-angle oblique image sequences collected by a drone acquisition device along the selected road corridor; Point cloud generation module: Generates a color 3D point cloud model of the proposed road corridor based on a multi-angle tilted image sequence; Structural surface interpretation module: Extracts the attitude data of rock mass structural surfaces from the color 3D point cloud model and establishes the structural surface attitude vector field of the proposed road corridor; Natural risk assessment module: Extracts natural slope units from the color 3D point cloud model and calculates the orientation of the natural slope surface; determines the natural geological hazard avoidance zone within the proposed road corridor based on the spatial intersection relationship between the structural surface attitude vector field and the orientation of the natural slope surface. Excavation simulation module: Obtains the horizontal alignment data and longitudinal slope data of candidate route schemes from the external design platform, simulates the spatial morphology of the artificially excavated slope, and outputs the spatial location data of the artificially excavated slope. Dynamic risk assessment module: From the structural plane attitude vector field, query the structural plane attitude data corresponding to the spatial location data of the artificially excavated slope, calculate the bedding slip risk index between the structural plane attitude data, and generate a dynamic bedding risk field. Overlay display and scheme generation module: Based on natural geological disaster avoidance zones and dynamic bedding risk fields, generate target route schemes.

2. The road alignment and survey system based on machine vision according to claim 1, characterized in that, The acquisition process for multi-angle tilt image sequences is as follows: Obtain topographic data of the proposed road corridor and identify topographic relief features and surface texture complexity; An adaptive 3D flight path is generated based on the topographic relief features, and a preset first overlap is set for the potential development area of ​​the structural surface and a preset second overlap is set for the non-development area based on the complexity of the surface texture; wherein, the first overlap is higher than the second overlap. The quality of the preview images transmitted in real time during the flight of the UAV data acquisition device is checked. When the image is blurred or obstructed, a supplementary flight path is generated and a reshoot is triggered. When exposed rock face features are detected, hovering and multi-angle rotation shooting is triggered. All acquired images are linked with their corresponding shooting positions and attitude angle data to form a multi-angle tilted image sequence with spatial index.

3. The road alignment and survey system based on machine vision according to claim 1, characterized in that, The process of generating a colored 3D point cloud model of the proposed road corridor is as follows: Feature point extraction and multi-view matching are performed on multi-angle tilted image sequences to obtain the mapping relationship of corresponding feature points between images; Based on the mapping relationship of the same feature points, the pose of the UAV acquisition device is recovered by the structure of motion recovery algorithm and a sparse point cloud is generated. After optimization by global bundle adjustment, the camera pose set and sparse 3D points are obtained. A multi-view stereo matching algorithm is used to match and fuse disparity maps of multiple frames of images pixel by pixel. The matching pixels are back-projected into three-dimensional space through triangulation to generate a dense three-dimensional point cloud. For each 3D point in the dense 3D point cloud, trace back the color of its projected pixels on each visible frame of the image, and assign color values ​​using a weighted fusion strategy based on the shooting angle to generate a colored 3D point cloud model with realistic color texture.

4. The road alignment and survey system based on machine vision according to claim 1, characterized in that, The process of extracting the attitude data of rock mass structural surfaces from a color 3D point cloud model is as follows: The color 3D point cloud model is preprocessed to calculate the local surface normal vector of each 3D point and unify its orientation to construct a continuous normal vector field. Based on the mutation feature of normal vectors and the region growing algorithm, several independent structural surface patches are segmented from the normal vector field; For each structural surface patch, perform planar fitting, calculate the dip and dip angle based on the normal vector of the fitted plane, generate the structural surface attitude vector attached to the corresponding surface patch, and collect them to establish the structural surface attitude vector field.

5. A road alignment and survey system based on machine vision according to claim 4, characterized in that, The process of establishing the structural surface attitude vector field of the proposed road corridor is as follows: For each structural facet, construct a vector element. The vector element shall include at least the facet identifier, spatial location range, dip value, tilt angle value, and area parameter. The vector elements are mapped to the stereographic projection map, and high-density clustering areas are identified through adaptive density clustering. The dominant joint groups are automatically divided and labeled. A spatial grid index structure is constructed, and a hierarchical storage mechanism is established that includes the attitude index of dominant joint groups and the attitude index of single facets, forming a structural facet attitude vector field that can be queried in space.

6. A road alignment and survey system based on machine vision according to claim 1, characterized in that, The process of determining the natural geological hazard avoidance zone within the proposed road corridor is as follows: The colored 3D point cloud model is divided into several natural slope units, and the slope orientation of each natural slope unit is calculated. By using the hierarchical spatial index of the structural surface attitude vector field, query the structural surface attitude data associated with each natural slope unit; Based on preset kinematic criteria, slope units with potential damage risks are identified according to slope orientation and associated structural surface attitude data. Slope units identified as having risks are spatially aggregated and classified into different levels to generate natural geological disaster avoidance zones.

7. The road alignment and survey system based on machine vision according to claim 1, characterized in that, The process of simulating the spatial morphology of an artificially excavated slope is as follows: Obtain the horizontal alignment data and vertical slope data of the candidate route schemes, and establish the three-dimensional route design axis; A sequence of cross-sections is generated along the three-dimensional route design axis and intersected with the color three-dimensional point cloud model to obtain the original terrain profile at each cross-section. At each cross section, based on the design elevation and preset slope parameters, the excavation slope line on the cut side is generated, forming a parametric cross section model. All cross-sectional models are longitudinally connected along the route to construct a three-dimensional surface model of the artificially excavated slope that describes the spatial orientation of each section of the excavated slope.

8. A road alignment and survey system based on machine vision according to claim 1, characterized in that, The process of generating a dynamic bedding risk field is as follows: The spatial morphology of the artificially excavated slope is spatially partitioned to generate several evaluation units and spatial query points within each evaluation unit. The spatial index of the structural surface attitude vector field is used to query the structural surface attitude data associated with each evaluation unit. Based on the angle relationship and dip comparison relationship between the orientation of the artificially excavated slope surface of each assessment unit and the queried structural surface attitude data, the bedding sliding risk index is calculated. By combining the bedding slip risk indices of each assessment unit, a dynamic bedding risk field distributed along the route is generated, and the bedding risk field is updated in real time according to the adjustment of candidate route schemes.

9. A road alignment and survey system based on machine vision according to claim 1, characterized in that, The process of generating the target route plan is as follows: The natural geological hazard avoidance zone and the dynamic bedding risk field are overlaid on the same display interface. The avoidance zone is rendered using a preset first visual coding scheme, and the bedding risk field is rendered using a preset second visual coding scheme. In response to adjustments to the horizontal alignment data or longitudinal slope data of candidate route schemes, the display of the bedding risk field is updated in real time, and the spatial distance between the route alignment and the boundary of the avoidance zone is monitored. When the adjusted candidate route meets the preset safety conditions, it is determined as the target route and output. The safety conditions include that the route does not intrude into the avoidance zone and the entire layer-by-layer risk index is lower than the preset threshold.

10. A road alignment survey method based on machine vision, applied to a road alignment survey system based on machine vision according to any one of claims 1-9, characterized in that, The method includes: Step 1: Acquire a sequence of multi-angle oblique images collected along the selected road corridor using a drone acquisition device; Step 2: Generate a color 3D point cloud model of the proposed road corridor based on the multi-angle tilted image sequence; Step 3: Extract the attitude data of rock mass structural surfaces from the color 3D point cloud model and establish the attitude vector field of the structural surfaces of the proposed road corridor; Step 4: Extract natural slope elements from the color 3D point cloud model and calculate the orientation of the natural slope surface; determine the natural geological hazard avoidance zone within the proposed road corridor based on the spatial intersection relationship between the structural surface attitude vector field and the orientation of the natural slope surface. Step 5: Obtain the horizontal alignment data and longitudinal slope data of the candidate route scheme from the external design platform, simulate the spatial morphology of the artificially excavated slope, and output the spatial location data of the artificially excavated slope. Step 6: From the structural plane attitude vector field, query the structural plane attitude data corresponding to the spatial location data of the artificially excavated slope, and calculate the bedding slip risk index between the structural plane attitude data to generate a dynamic bedding risk field. Step 7: Generate the target route plan based on the natural geological disaster avoidance zone and the dynamic bedding risk field.

Images