A method and system for planning a UAV inspection route based on a three-dimensional model

Abstract

The present application relates to the technical field of unmanned aerial vehicle inspection, in particular to a kind of unmanned aerial vehicle inspection route planning method and system based on three-dimensional model, method includes the following steps, obtain the surface block of the voxel grid composition of pipe gallery model intersection and extract geometric features, according to the depth variance and feature split cross block into visible and hidden sub-region, utilize virtual cone model to emit reverse ray to extract visible channel and occlusion direction, based on hidden coverage demand to establish supplementary viewpoint and generate planning result in combination with basic navigation point.The present application, by the feature discrimination of pipe gallery block, effectively identifies the complex cross surface difficult to distinguish, combined with topological analysis, finely disassembles the occlusion state to reveal the deep hidden area covered, reversely calculates the space blocking intersection and accurately quantifies the actual coverage demand of hidden part, eliminates the gap detection blind area and avoids collision danger by integrating safety supplementary viewpoint, improves the completeness of pipe gallery inspection space coverage range and the safety of route arrangement.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) inspection technology, and in particular to a method and system for planning UAV inspection routes based on a three-dimensional model. Background Technology

[0002] The field of UAV inspection technology mainly involves using unmanned aerial vehicles (UAVs) equipped with visible light cameras, infrared thermal imagers, and lidar as sensors to perform patrol and inspection tasks on specific targets or areas. Its core aspects include UAV platform control, multi-source sensor data acquisition, flight path setting, and target status monitoring. Flight tasks are set via a ground station control console, and the UAV is guided by a satellite positioning system to fly sequentially along preset geographical coordinates to acquire images or 3D spatial point cloud data of the inspected targets. Traditional UAV inspection route planning methods, which address the generation of 3D spatial flight paths and the arrangement of inspection points during UAV patrol missions, are typically accomplished through manual on-site flight marking or waypoint selection based on 2D orthophoto maps.

[0003] Current UAV inspection planning methods often divide the surface of a 3D model into units using a fixed-size grid and determine grid occlusion based on ray tracing. When encountering densely intersecting pipe corridors, even small pipe gaps can cause a single grid to span both sides of the gap. The single binary visibility determination mechanism cannot distinguish the true occlusion status of sub-regions within the spanning unit. Misjudging visibility will cause areas hidden deep within the line of sight to be uncovered, thus obscuring occlusion information. Misjudging occlusion will induce the generation algorithm to increase dangerous waypoints. Fixed grid resolution is difficult to handle local occlusion problems, thus creating detection blind spots in the planning. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a method and system for planning unmanned aerial vehicle (UAV) inspection routes based on a three-dimensional model.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for planning unmanned aerial vehicle (UAV) inspection routes based on a three-dimensional model, comprising the following steps: S1: Obtain the triangular mesh model file of the pipe gallery in the area to be inspected, map the model space coordinates recorded in the triangular mesh model file to the local coordinate system of the project to generate a global pipe gallery entity model, calculate the extreme coordinates of the three-dimensional contour boundary corresponding to the global pipe gallery entity model to obtain the global axial bounding box, divide the three-dimensional space associated with the global axial bounding box into a three-dimensional voxel mesh according to the bounding box cutting size threshold, extract the intersecting meshes in the three-dimensional voxel mesh that intersect with the surface of the global pipe gallery entity model as initial surface blocks, extract the geometric features of the initial surface blocks and associate and encapsulate them to generate a block geometric feature dataset; S2: Call the block geometric feature dataset, classify the initial surface blocks based on the local depth variance, normal angle scattering range and ray intersection spatial fault features, and generate surface block classification results; S3: Extract the complex cross-connected blocks in the surface block classification results, combine the topological connection relationship of the patches to perform region growth clustering algorithm processing, and obtain the visible surface sub-regions in front and the hidden surface sub-regions behind based on the semantic splitting of the independent cluster surface hierarchy. Extract the independent centroid coordinates, true normal vectors and boundary contour point set attributes of the visible surface sub-regions in front and the hidden surface sub-regions behind respectively to generate a local sub-region attribute dataset. S4: Call the local sub-region attribute dataset, construct a virtual detection frustum model based on the hardware parameters of the UAV-mounted sensor, divide candidate observation directions on the detection space sphere with the target detection anchor point as the center and emit reverse space detection rays using the virtual detection frustum model, calculate the surface obstruction intersection between the global tube gallery entity model and the reverse space detection rays, extract the safe visual channel interval and obstacle occlusion direction set according to the surface obstruction intersection and bind the target hidden sub-region, and establish an independent coverage requirement data package; S5: Generate a safety avoidance supplementary viewpoint based on the independent coverage requirement data packet, and construct a global spatial waypoint by combining the safety avoidance supplementary viewpoint with the set of regular basic inspection waypoints generated by the ordinary surface blocks in the surface block classification results. With the goal of minimizing the sum of path movement edge weights, the inspection route planning result is generated by processing with the traveling salesman problem optimization algorithm and the polynomial curve smoothing function.

[0006] The present invention is improved in that the specific steps of S1 are as follows: S111: Obtain the triangular mesh model file of the utility tunnel, map the model space coordinates recorded in the triangular mesh model file of the utility tunnel to the local coordinate system of the project to generate a global utility tunnel entity model, calculate the extreme coordinates of the three-dimensional contour boundary corresponding to the global utility tunnel entity model, and obtain the global axial bounding box. S112: Call the bounding box cutting size threshold, divide the three-dimensional space associated with the global axial bounding box into a three-dimensional voxel mesh according to the bounding box cutting size threshold, extract the intersecting meshes in the three-dimensional voxel mesh that intersect with the surface of the global tube gallery entity model, and generate initial surface blocks. S113: Extract the original model vertex coordinates and triangular patch indices inside the initial surface block, calculate the spatial mean of the original model vertex coordinates to obtain the block geometric centroid coordinates, calculate the mean of the patch normal vectors based on the original model vertex coordinates and the triangular patch indices, extract the block average normal vector, and associate and encapsulate the initial surface block, the original model vertex coordinates, the triangular patch indexes, the block geometric centroid coordinates, and the block average normal vector to generate a block geometric feature dataset.

[0007] The present invention is improved in that the specific steps of S2 are as follows: S211: Extract the initial surface blocks, original model vertex coordinates, triangular patch indexes, block geometric centroid coordinates, and block average normal vectors from the block geometric feature dataset. Establish a local spatial tangent plane with the block geometric centroid coordinates as the origin. Calculate the vertical projection distance of the vertex based on the original model vertex coordinates and the local spatial tangent plane. Calculate the statistical variance corresponding to the vertical projection distance of the vertex and generate the local depth variance. S212: Extract the facets of the utility tunnel entity model based on the triangular facet index, calculate the spatial angle between the facet normal vector associated with the facet of the utility tunnel entity model and the block average normal vector, extract the scattering range corresponding to the spatial angle, and obtain the normal angle scattering range. S213: A spatial detection ray is emitted from the virtual detection hemisphere associated with the initial surface block in front of the initial surface block into the interior of the initial surface block. The Euclidean distance between the intersection points of the pipe gallery surfaces between adjacent spatial detection rays is calculated. A gap width threshold is obtained. The Euclidean distance between the intersection points of the pipe gallery surfaces that is greater than the gap width threshold is extracted. A spatial fault feature of the ray intersection point is established. S214: Obtain the jump threshold and the continuous scattering threshold. Based on the local depth variance greater than the jump threshold, the normal angle scattering range greater than the continuous scattering threshold, and the spatial fault characteristics of the ray intersection, separate complex bridging blocks and ordinary surface blocks from the initial surface blocks to generate surface block classification results.

[0008] The present invention is improved in that the process of generating the surface block classification results is specifically as follows: Obtain the preset equipment ranging error parameters and the pipeline design wall thickness parameters, calculate the sum of the values ​​of the pipeline design wall thickness parameters and the preset equipment ranging error parameters, and use the sum of the values ​​as the jump threshold; Extract the standard radius of curvature of the pipe and the surface sampling grid size, calculate the normal deflection limit value based on the ratio of the surface sampling grid size to the standard radius of curvature of the pipe, and determine the normal deflection limit value as the dispersion continuity threshold; If the local depth variance corresponding to the initial surface block is greater than the jump threshold, and the normal angle scattering range is greater than the scattering continuity threshold, and the ray intersection spatial fault feature exists, the initial surface block is classified as the complex bridging block. If not all of the initial surface blocks meet the above judgment conditions, the initial surface blocks will be classified as ordinary surface blocks. Integrate all the complex cross-connection blocks and the ordinary surface blocks to generate the surface block classification result.

[0009] The present invention is improved in that the specific steps of S3 are as follows: S311: Extract complex cross-connection blocks from the surface block classification results, obtain the topological connection relationship of the facets corresponding to the global utility tunnel entity model, extract the utility tunnel entity model faces contained in the complex cross-connection blocks, map the unassigned utility tunnel entity model faces to clustering seed faces associated with the region growing clustering algorithm, calculate the facet spatial Euclidean distance and the facet normal angle between the clustering seed facets and the neighboring utility tunnel faces, obtain the distance continuity threshold and the smooth transition threshold, filter the neighboring utility tunnel faces whose facet spatial Euclidean distance is less than the distance continuity threshold and whose facet normal angle is less than the smooth transition threshold according to the region growing clustering algorithm, merge the filtered neighboring utility tunnel faces, and generate independent surface clusters; S312: Obtain the local coordinate system of the project, calculate the surface spatial depth value of the independent surface cluster in the local coordinate system of the project, establish the surface hierarchical semantics of the independent cluster based on the surface spatial depth value and the topological connection relationship of the patch, and split the independent surface cluster according to the surface hierarchical semantics of the independent cluster to obtain the visible surface sub-region in front and the hidden surface sub-region behind. S313: For the visible surface sub-region in front and the hidden surface sub-region behind, calculate the independent centroid coordinates, the true normal vector, and the boundary contour point set of the sub-region based on the visible surface sub-region in front and the hidden surface sub-region behind, respectively. Map the independent centroid coordinates, the true normal vector, and the boundary contour point set of the sub-region to the corresponding visible surface sub-region in front and the hidden surface sub-region behind, respectively, to generate a local sub-region attribute dataset.

[0010] The present invention is improved in that the process of generating independent surface clusters is specifically as follows: Calculate the average side length of the grid of the pipe gallery solid model patch, and configure the product of the average side length of the grid and the preset spatial tolerance coefficient as the distance continuity threshold. Obtain the standard curvature parameters of the pipe gallery pipeline, calculate the theoretical maximum normal deflection angle based on the standard curvature parameters of the pipe gallery pipeline and the average side length of the grid, and set the theoretical maximum normal deflection angle as the smooth transition threshold. If the extracted Euclidean distance of the patch space is less than the distance continuity threshold, and the included angle of the patch normal vector is less than the smooth transition threshold, the neighborhood tube gallery patch is added to the current cluster set. The most recently added neighborhood tube gallery patch in the current cluster set is selected as the iteratively derived seed patch for outward comparison until there are no adjacent patches that meet the numerical constraints. All the neighboring tube gallery patches recorded within the current cluster set are spliced ​​together to generate the independent surface cluster.

[0011] The present invention is improved in that the specific steps of S4 are as follows: S411: Obtain the horizontal and vertical field-of-view parameters corresponding to the sensor hardware on the UAV. Construct a virtual detection frustum model based on the horizontal and vertical field-of-view parameters. Call the local sub-region attribute dataset and extract the hidden surface sub-region, independent centroid coordinates of the sub-region, boundary contour point set of the sub-region, and true normal vector of the sub-region contained in the local sub-region attribute dataset. Use the independent centroid coordinates of the sub-region and the boundary contour point set of the sub-region as the target detection anchor point and obtain the detection distance. On the detection space sphere with the target detection anchor point as the center and the detection distance as the radius, use the virtual detection frustum model to select the candidate observation viewpoint direction that meets the field-of-view constraints. Emit a space detection ray along the candidate observation viewpoint direction toward the target detection anchor point to obtain the reverse space detection ray. S412: Calculate the surface obstruction intersection between the global tube gallery entity model and the reverse space detection ray using the hierarchical bounding box tree structure algorithm, and extract the safe visible channel interval and the set of obstacle occlusion directions based on the surface obstruction intersection. S413: For the rear hidden surface sub-region, select the target hidden sub-region corresponding to the missing effective ray visual path from the rear hidden surface sub-region according to the surface blocking intersection, and bind the independent centroid coordinates of the sub-region, the true normal vector of the sub-region, the obstacle occlusion direction set and the safe visual channel interval to the target hidden sub-region to establish an independent coverage requirement data packet.

[0012] The present invention is improved in that the specific steps of S5 are as follows: S511: Read the independent centroid coordinates of the sub-region, the set of obstacle occlusion directions, and the safe visibility channel interval associated with the independent coverage requirement data packet record. Using the independent centroid coordinates of the sub-region as a reference point, obtain the detection distance. Calculate the three-dimensional coordinates of the candidate supplementary viewpoint in reverse along the safe visibility channel interval within the detection distance range. Obtain the UAV body size parameters and flight control error margin parameters. Calculate the safety expansion radius based on the UAV body size parameters and flight control error margin parameters. Establish a UAV safety expansion sphere with the three-dimensional coordinates of the candidate supplementary viewpoint as its center based on the safety expansion radius. Call the global utility tunnel entity model to calculate the interference collision space distance between the UAV safety expansion sphere and the global utility tunnel entity model. When the interference collision space distance is less than the set physical safety margin, use the set of obstacle occlusion directions to guide the three-dimensional coordinates of the candidate supplementary viewpoint to be finely adjusted towards the spatial safety area to obtain a safe avoidance supplementary viewpoint. S512: Extract ordinary surface blocks from the surface block classification results, generate a set of regular basic inspection waypoints based on the ordinary surface blocks, merge the set of regular basic inspection waypoints with the safety avoidance supplementary viewpoints to construct global spatial waypoints, map the global spatial waypoints to waypoint graph structure nodes associated with the traveling salesman problem optimization algorithm, calculate the node spatial Euclidean distance and node altitude change values ​​between the waypoint graph structure nodes, obtain the maximum flight speed and UAV climb rate, calculate the level flight time cost based on the maximum flight speed and the node spatial Euclidean distance, calculate the UAV altitude adjustment time cost based on the UAV climb rate and node altitude change values, and establish path movement edge weights based on the level flight time cost and the UAV altitude adjustment time cost; S513: With minimizing the sum of the weights of the path movement edges as the optimization objective, the Traveling Salesman Problem optimization algorithm is used in conjunction with the weights of the path movement edges to traverse and solve all the waypoint graph structure nodes to obtain the optimal sorted path sequence of the UAV. The polynomial curve smoothing function is then used to fit the flight trajectory associated with the optimal sorted path sequence of the UAV to generate the inspection route planning result.

[0013] The present invention is improved in that the process of using the obstacle occlusion direction set to guide the three-dimensional coordinates of the candidate supplementary viewpoint to be finely adjusted towards the spatial safety area to obtain a safe avoidance supplementary viewpoint is specifically as follows: Extract the three-dimensional normal vector of the main body occlusion recorded within the set of obstacle occlusion directions; Calculate the absolute reverse avoidance vector that is opposite to the three-dimensional normal vector space of the occlusion of the main body; Obtain the basic translation step size, and adjust the three-dimensional coordinates of the candidate supplementary viewpoint by extending it outward along the direction of the absolute reverse avoidance vector according to the basic translation step size; Obtain the equipment emergency braking distance parameter and the gust offset compensation parameter, calculate the sum of the values ​​of the equipment emergency braking distance parameter and the gust offset compensation parameter, and configure the sum of the values ​​as the physical safety margin; Recalculate the interference collision space distance corresponding to the three-dimensional coordinates of the candidate supplementary viewpoint after the extension and adjustment. If the interference collision space distance is less than the physical safety margin, perform the extension and adjustment operation repeatedly. Until the distance between the interference and collision spaces corresponding to the extended and adjusted three-dimensional coordinates is greater than the physical safety margin, the qualified three-dimensional coordinate nodes are extracted to generate the safety avoidance supplementary viewpoint.

[0014] A UAV inspection route planning system based on a 3D model, wherein the 3D model-based UAV inspection route planning system is used to implement the aforementioned 3D model-based UAV inspection route planning method, and the system includes: The spatial modeling module acquires the triangular mesh model file of the utility tunnel in the area to be inspected, maps the model space coordinates recorded in the triangular mesh model file to the local coordinate system of the project to generate a global utility tunnel entity model, calculates the extreme coordinates of the three-dimensional contour boundary corresponding to the global utility tunnel entity model to obtain a global axial bounding box, divides the three-dimensional space associated with the global axial bounding box into a three-dimensional voxel mesh according to the bounding box cutting size threshold, extracts the intersecting meshes in the three-dimensional voxel mesh that intersect with the surface of the global utility tunnel entity model as initial surface blocks, extracts the geometric features of the initial surface blocks and performs associated encapsulation to generate a block geometric feature dataset; The feature interpretation module calls the block geometric feature dataset and classifies the initial surface blocks based on the local depth variance, the range of normal angle dispersion, and the spatial tomographic features of ray intersections, generating surface block classification results. The topology parsing module extracts complex cross-connected blocks within the surface block classification results, performs region growing clustering algorithm processing based on the topological connection relationship of the patches, and obtains the visible surface sub-regions in front and the hidden surface sub-regions behind based on the semantic segmentation of the independent cluster surface hierarchy. The independent centroid coordinates, true normal vectors and boundary contour point set attributes of the visible surface sub-regions in front and the hidden surface sub-regions behind are extracted respectively to generate a local sub-region attribute dataset. The occlusion calculation module calls the local sub-region attribute dataset, constructs a virtual detection frustum model based on the hardware parameters of the sensors on the UAV, uses the virtual detection frustum model to divide candidate observation directions on the detection space sphere with the target detection anchor point as the center and emits reverse space detection rays, calculates the surface obstruction intersection between the global tube gallery entity model and the reverse space detection rays, extracts the safe visual channel interval and obstacle occlusion direction set based on the surface obstruction intersection and binds the target hidden sub-region, and establishes an independent coverage requirement data package; The trajectory arrangement module generates a safety avoidance supplementary viewpoint based on the independent coverage requirement data packet, and constructs a global spatial waypoint by combining the safety avoidance supplementary viewpoint with the set of regular basic inspection waypoints generated by the ordinary surface blocks in the surface block classification results. With the goal of minimizing the sum of path movement edge weights, the inspection route planning results are generated through the traveling salesman problem optimization algorithm and the polynomial curve smoothing function.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, initial surface blocks are formed by extracting the intersecting three-dimensional voxel meshes from the surface of the utility tunnel model. Multidimensional interpretation is performed based on local depth variance and spatial fault characteristics of ray intersections to identify complex cross-connection blocks. Region clustering analysis is performed in combination with the topological connection relationship of the patches to decompose the cross-connection unit into visible and hidden sub-regions to reveal the internal occlusion state. A virtual detection frustum model is used to emit reverse detection rays in the detection space to calculate the surface blocking intersections and quantify the occlusion requirements. Based on the hidden coverage requirements, a safety avoidance supplementary viewpoint is generated and incorporated into the regular basic waypoint set. After path movement edge weight optimization calculation, gap blind spots are eliminated and dangerous routes are avoided, thereby improving the integrity and safety of global spatial inspection coverage. Attached Figure Description

[0016] Figure 1 This is the main flowchart for the pipeline inspection route planning of this invention; Figure 2 This is a flowchart of the surface segmentation and classification process of the present invention; Figure 3 A flowchart for generating hidden surface coverage requirements in this invention; Figure 4 This invention provides a safety supplement to the viewpoint and route generation flowchart. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Please see Figures 1 to 4This embodiment provides a method and system for planning UAV inspection routes based on a 3D model. In practical applications, such as the actual operation of route planning for the inspection of the outer surface of urban underground utility tunnels, the triangular mesh model of the tunnel is output by the engineering modeling platform, the local engineering coordinate system is maintained by the tunnel inspection engineering coordinate reference, and the UAV is equipped with a visible light camera or rangefinder to observe the tunnel surface, pipe joints, support and hanger obstructions, and the hidden surfaces. The route planning process revolves around the same tunnel entity model, the same local engineering coordinate system, and the same waypoint data link. It includes a method for planning UAV inspection routes based on a 3D model, comprising the following steps: S1: Obtain the triangular mesh model file of the utility tunnel in the area to be inspected. The triangular mesh model file is a data file that records the vertex coordinates, triangular facet indices, and model space coordinate reference of the utility tunnel surface. After mapping the model space coordinates to the engineering local coordinate system, a global utility tunnel solid model is formed. The global utility tunnel solid model is a three-dimensional mesh object that bears the surface geometry of the utility tunnel outer wall, pipe bridging components, supports, and shielding components under a unified coordinate reference. The extreme coordinates of the three-dimensional contour boundary are generated by the axial boundary positions of each vertex in the global utility tunnel solid model in the engineering local coordinate system. The global axial bounding box is a spatial boundary data object that encloses all surface vertices of the model. The bounding box cutting size threshold is an engineering configuration field. The field is derived from the model mesh resolution, sensor detection distance requirements, and waypoint planning granularity configuration, stored in the inspection task parameter table, and called during the voxel segmentation stage. The 3D voxel mesh is a set of spatial units formed by regularly dividing the global axial bounding box associated space. Intersecting meshes are voxel units that geometrically cross or occupy the surface of the global utility tunnel solid model. The initial surface block is a set of local surface patches carried by the intersecting mesh. The block geometric feature dataset is a structured data package for subsequent classification and occlusion analysis, containing the initial surface block identifier, vertex coordinate set, triangular patch index set, block geometric centroid coordinates, block average normal vector, data verification status, and geometric source record.

[0019] S111: After obtaining the triangular mesh model file of the utility tunnel, parse the coordinate reference field, vertex record field, and facet index field in the file header. When mapping the model space coordinates to the engineering local coordinate system, read the coordinate translation record, orientation alignment record, and axial direction record from the engineering coordinate reference table, and write the model vertex positions into the global vertex cache under the engineering local coordinate system. If the model file is missing a coordinate reference field, the parsing status is marked with a coordinate verification flag, and the project coordinate reference record stored in the engineering modeling platform is called to complete the mapping source; if the vertex field and facet index field cannot reference each other, the corresponding facet enters the geometric anomaly cache and is not included in the subsequent voxel intersection extraction.

[0020] After the global utility tunnel solid model is generated, the valid vertices in the global vertex cache are traversed. A valid vertex is defined as a vertex with complete spatial coordinates, indexable by at least one triangle, not occupied by duplicate facet conflicts, and not falling into the model resolution anomaly cache. The axial boundary positions of the valid vertices in the local engineering coordinate system are extracted to form the extreme coordinates of the 3D contour boundary, and this boundary is written to the global axial bounding box field. The global axial bounding box is stored together with the model version identifier, coordinate reference identifier, and inspection task identifier for S112 to read.

[0021] S112: The bounding box cutting size threshold field is invoked to split the 3D space associated with the global axial bounding box into a continuous 3D voxel mesh. Each 3D voxel mesh records its spatial boundary, center position, adjacent mesh indices, and occupancy status. Intersection detection uses triangular facets as objects, verifying the geometric relationship between the facet boundaries, the plane containing the facet, and the voxel space boundary. Voxels where facets cross, facet vertices fall into, or facet boundaries fit are marked as intersecting meshes. Intersecting meshes record the corresponding triangular facet index, vertex reference path, and voxel occupancy status.

[0022] When intersecting meshes are merged into initial surface blocks, they are merged according to the spatial continuity of adjacent voxels and the set of referenced faces. When the voxel segmentation size threshold is updated, the initial surface block version is triggered to update, and the task traceability record between the old and new versions is retained. If a voxel is only associated with a parsing anomaly face, the voxel is written into the mesh state to be repaired and is not output as an initial surface block. If the model surface and voxel boundary are attached, resulting in duplicate assignment, the face index unique assignment rule is used to bind the same face to the initial surface block with a more complete spatial coverage, and the shared boundary relationship is recorded in the adjacent block reference table.

[0023] S113: After extracting the original model vertex coordinates and triangular patch indices within the initial surface blocks, the block's geometric centroid coordinates are generated based on the spatial distribution of the vertex coordinates. The block's geometric centroid coordinates represent the concentrated spatial positions of effective vertices within the initial surface blocks, used in S2 to establish the local spatial tangent plane, and for subsequent generation of routine basic inspection waypoints. The average patch normal vector is determined by the triangular patch vertex rotation order and the outer direction of the engineering local coordinate system. The block's average normal vector represents the direction of the initial surface block's surface orientation. If the patch rotation order is inconsistent, the patch normal vector direction is first unified based on the entity's external direction before extracting the block's average normal vector to prevent confusion between internal and external directions from entering the classification judgment.

[0024] The initial surface blocks, original model vertex coordinates, triangular patch indices, block geometric centroid coordinates, and block average normal vectors are associated and encapsulated into a block geometric feature dataset. This dataset includes a block identifier field, a geometry field, an orientation field, a source field, and an anomaly field. The block identifier field is used for subsequent classification result write-back; the geometry field stores vertex and patch references; the orientation field stores the average normal vector and normal consistency status; the source field records the voxel mesh and model version; and the anomaly field records missing vertices, duplicate patches, normal conflicts, and boundary misalignment. The encapsulated dataset then enters the feature interpretation process in S2.

[0025] S2: The initial surface blocks are classified based on the local depth variance, normal angle dispersion range, and ray intersection spatial tomography features, generating classification results. Local depth variance is the vertical deviation dispersion of vertices within the initial surface block relative to the local spatial tangent plane, used to identify surface depth jumps caused by cross-connections, pipeline overlaps, and occlusion layers. Normal angle dispersion range is the directional dispersion of the patch normal vector within the block relative to the block's average normal vector, used to identify surface orientation continuity. Ray intersection spatial tomography features are the discontinuous records of intersections formed after emitting probe rays from the virtual probe hemisphere into the block, used to identify spatial cracks, occlusion intervals, or discontinuous surfaces within the same voxel block. The surface block classification results include complex cross-connection blocks and ordinary surface blocks. Complex cross-connection blocks proceed to S3, while ordinary surface blocks proceed to S5 for the generation of regular basic inspection waypoints.

[0026] S211: Extract the initial surface blocks, original model vertex coordinates, triangular patch indices, block geometric centroid coordinates, and block average normal vector from the block geometric feature dataset. Establish a local spatial tangent plane with the block geometric centroid coordinates as the origin. The local spatial tangent plane is a local geometric reference plane that passes through the block geometric centroid coordinates and uses the block average normal vector as its vertical reference. The vertex vertical projection distance is generated by the spatial deviation relationship of the original model vertex projected along the direction of the block average normal vector to the local spatial tangent plane. The field records the depth distribution of each vertex relative to the local spatial tangent plane.

[0027] The local depth variance is generated from the discrete state of the vertex vertical projection distance set and stored in the block classification cache. This field is not used as a separate classification output, but is entered into S214 together with the normal angle scattering range and the spatial fault features of the ray intersection. If the initial surface block vertices are insufficient to form a stable local spatial tangent plane, the shared boundary normal record of the adjacent blocks is called to complete the direction; blocks that still cannot form a local spatial tangent plane after completion are written into the pending review state and enter the basic waypoint generation link as ordinary surface blocks, while the reason for the geometric incompleteness of the block is recorded in the inspection task log.

[0028] S212: Extract facets from the utility tunnel solid model based on triangular facet indexing, and read the vertex winding order, adjacent facet identifiers, and outer direction records for each facet. The spatial angle between the facet normal vector and the block average normal vector is generated by the orientation difference between the two types of orientation fields. The normal angle distribution range is the discrete boundary field covered by the orientation difference of all valid facets within the block, used to characterize whether the block surface maintains continuous curvature, bend transition, or multi-layer bridging.

[0029] When the normal angle dispersion range is written to the classification cache, the normal consistency status of the facets is recorded simultaneously. If a facet normal vector points into the entity due to an incorrect winding order, it is first corrected based on the topological direction of adjacent faces and the external direction of the entity. Facets that fail to correct are not included in the dispersion range generation, and the normal conflict status is recorded in the block anomaly field. This anomaly status is used as an auxiliary verification field during S214 classification to prevent misjudgments of complex cross-connected blocks caused by model data errors.

[0030] S213: A space probe ray is emitted into the interior of the initial surface block from a virtual probe hemisphere associated with the initial surface block. The virtual probe hemisphere is a set of candidate probe directions expanded outward from the average normal vector of the block, centered on the geometric centroid coordinates of the block, used to simulate the incident path of the inspection camera observing the local surface from the front of the block. The intersection points of the space probe ray and the surface of the pipe gallery solid model are recorded in the ray intersection record according to the ray direction, block affiliation, and adjacent ray order.

[0031] The Euclidean distance between the intersection points of adjacent spatial detection rays on the pipe gallery surfaces is converted into an intersection interval state. The gap width threshold is an engineering configuration field derived from the requirements for identifying the detailed structure of the pipe gallery surface, the model mesh granularity, and the sensor cone coverage configuration, and is stored in the inspection task parameter table. When the intersection interval state falls into a category exceeding the threshold boundary, a spatial fault feature of the ray intersection is established. This feature records the fault location, the direction of the participating rays, the corresponding patch index, and the block identifier, and is read by S214. If the ray does not produce a valid intersection with the model surface, the ray record enters a no-intersection state; when the no-intersection state occurs mainly at the block edges, it is recorded as a boundary spatial domain and does not directly constitute a spatial fault feature.

[0032] S214: Obtain the jump threshold and the dispersion continuity threshold. The jump threshold is a determination field for the depth discrete state, derived from the engineering configuration of the preset equipment ranging error parameters and the pipeline design wall thickness parameters. The preset equipment ranging error parameters are determined by the sensor factory configuration, calibration records, and inspection task calibration records. The pipeline design wall thickness parameters are provided by the pipe gallery design data or the engineering model attribute table. The two types of fields are combined to form the depth jump determination boundary and stored in the classification rule table. The dispersion continuity threshold is a normal continuity determination field, derived from the standard radius of curvature of the pipeline and the surface sampling grid size. The standard radius of curvature of the pipeline is read from the pipe gallery component design attribute table, and the surface sampling grid size is generated by the model mesh analysis results. After entering the rule table, the two form the normal deflection boundary, used to determine whether the surface orientation within the same block remains continuous.

[0033] The classification process employs a fixed, progressive judgment order. First, the local depth variance state is read; then, the normal angle scattering range state is read; and finally, the ray intersection spatial fault characteristics are verified. If the local depth variance state falls outside the jump threshold boundary, the normal angle scattering range state falls outside the scattering continuity threshold boundary, and ray intersection spatial fault characteristics exist, the initial surface block is classified as a complex bridging block. If any judgment item does not fall into the corresponding anomaly category, the initial surface block is classified as a normal surface block. If any of the three fields are missing or fail verification, the classification result is recorded as a normal surface block with an additional "to be reviewed" marker. This block is retained in subsequent basic waypoint generation to avoid waypoint link interruptions. All complex bridging blocks and normal surface blocks are integrated to form the surface block classification result, which is written back to the classification result storage area according to the block identifier for use by S3 and S5 respectively.

[0034] S3: Extract complex cross-connected blocks from the surface block classification results and process them using a region growing clustering algorithm based on the topological connectivity of facets. The topological connectivity of facets is the adjacency data structure formed between triangular facets in the global utility tunnel entity model through shared edges, shared vertices, and continuous surface directions, used to define the expansion boundary of the region growing clustering. In this scheme, the region growing clustering algorithm refers to using unassigned facets as clustering seeds, expanding to neighboring facets according to the spatial continuity and normal smoothness of facets until no more adjacent facets conforming to the rules exist, outputting a deterministic facet merging process for independent surface clusters. Independent surface clusters are sets of geometrically continuous local surfaces with consistent orientation transitions within complex cross-connected blocks. The surface hierarchy semantics of independent clusters are the front-to-back hierarchy identifiers of surfaces formed based on the depth relationship of the engineering local coordinate system and the topological connectivity of facets, used to separate the visible surface sub-regions in front from the hidden surface sub-regions behind. The local sub-region attribute dataset is a structured data package carrying the independent centroid coordinates of the sub-region, the true normal vector of the sub-region, and the boundary contour point set of the sub-region, which is then used for occlusion measurement in S4.

[0035] S311: Extract the pipe gallery entity model patches contained in complex cross-connection blocks and read the topological connection relationships of the patches corresponding to the global pipe gallery entity model. Unassigned pipe gallery entity model patches are mapped to cluster seed patches associated with the region growing clustering algorithm. The patch spatial Euclidean distance between the cluster seed patch and the neighboring pipe gallery patches is used to describe the spatial proximity of the two patches in the local engineering coordinate system, and the angle between the patch normal vectors is used to describe the transition state of the surface orientation of the two patches. The distance continuity threshold is a patch spatial connection determination field, derived from the average edge length of the grid of the pipe gallery entity model patch and the preset spatial fault tolerance coefficient. The average edge length of the grid is generated by the effective patch boundary record in the model analysis stage, and the preset spatial fault tolerance coefficient is maintained by the inspection task parameter table. Together, they form a lookupable spatial continuity boundary. The smooth transition threshold is a normal connection determination field, derived from the standard curvature parameters of the pipe gallery pipeline and the average edge length of the grid, stored in the clustering rule table, and used to identify the difference between the continuous transition of the pipe gallery surface and the cross-connection fault.

[0036] When the spatial proximity of a neighboring utility tunnel surface falls within the continuous distance boundary and its normal transition state falls within the smooth transition boundary, that neighboring utility tunnel surface is added to the current cluster set. The most recently added neighboring utility tunnel surface in the current cluster set continues to be used as an iteratively derived seed surface for comparison until none of the surrounding adjacent surfaces can meet the above two boundary conditions. After the expansion is completed, all neighboring utility tunnel surfaces recorded in the current cluster set are stitched together into an independent surface cluster. If there is a break in the topological connection relationship of the surfaces, the surfaces on both sides of the break will not be forcibly merged, and the break information is written to the clustering anomaly record; if the same surface is requested to be merged by multiple seeds, the cluster affiliation of the first to complete the topological closure is retained first, and the conflict affiliation is recorded in the surface tracing field.

[0037] S312: After obtaining the local coordinate system of the project, calculate the surface spatial depth value of the independent surface clusters within the local coordinate system. The surface spatial depth value is the front and back position field formed by the facets within the cluster along the inspection and observation direction of the pipe gallery. The data comes from the centroid coordinates of the facets, the local observation direction, and the topological connection relationship. The surface hierarchy semantics of the independent clusters are generated by the depth order, adjacency relationship, and occlusion relationship of each facet within the same cluster, including the front layer identifier, back layer identifier, inter-layer connection identifier, and occlusion boundary identifier.

[0038] When splitting independent surface clusters, a front-to-back hierarchy is first formed based on the surface spatial depth values. Then, the topological connectivity of the patches is used to verify whether the hierarchy boundaries are supported by continuous patches. The set of patches located on the front side of the viewing direction and connected to the external visible boundary forms the front visible surface sub-region, while the set of patches located on the back side of the front visible surface, connected to the occlusion boundary, and lacking direct line of sight in the normal viewing direction forms the rear hidden surface sub-region. If there are intersections in the hierarchical relationships within the clusters, the priority splitting position is determined according to the connectivity of the occlusion boundary, and patches that cannot be stably assigned are written into the hierarchical pending verification state. This state is passed to S4 along with the sub-region attribute dataset for subsequent reverse spatial probe ray verification.

[0039] S313: For both the visible surface sub-region in front and the hidden surface sub-region behind, calculate the independent centroid coordinates, the true normal vector, and the boundary contour point set of the sub-region. The independent centroid coordinates of the sub-region are the spatially concentrated representative points of the effective patch centroids within the corresponding sub-region, used to generate target detection anchor points in S4. The true normal vector of the sub-region is the representative direction of the sub-region patch orientation after unifying the outer direction of the entity, used to limit the candidate observation direction and the sensor's view frustum orientation. The boundary contour point set of the sub-region is an ordered set of boundary points extracted along the outer edge of the sub-region, the occlusion boundary, and the topological break boundary, used to supplement edge details that cannot be covered by the centroid anchor points.

[0040] The independent centroid coordinates of each sub-region, the true normal vector of the sub-region, and the set of contour points of the sub-region boundary are mapped to the corresponding visible surface sub-regions in front and the hidden surface sub-regions behind, forming a local sub-region attribute dataset. This dataset includes sub-region identifiers, parent block identifiers, hierarchical semantic fields, centroid fields, normal fields, contour fields, and anomaly fields. If there are breakpoints in the contour point set of the sub-region boundary, contour completion is first performed based on the shared edges of adjacent patches; breakpoints that fail to complete are written to the contour anomaly field and used as trigger records for adding object detection anchor points in S4. The local sub-region attribute dataset is stored in the occlusion calculation cache and is read item by item by S4 according to the hidden surface sub-regions behind.

[0041] S4: Utilize the local sub-region attribute dataset to construct a virtual detection frustum model based on the hardware parameters of the UAV's onboard sensors. The virtual detection frustum model is a spatial coverage object defined by horizontal field-of-view parameters, vertical field-of-view parameters, detection distance, and sensor installation orientation. It is used to determine whether the candidate viewpoint's observation direction can cover the target detection anchor point and its boundary contour. The target detection anchor point is the boundary representative point in the set of independent centroid coordinates and boundary contour points of the hidden surface sub-region behind it. The detection space sphere is the set of candidate observation positions with the target detection anchor point as the center and the detection distance field as the radius. The reverse spatial detection ray is a spatial line segment pointing towards the target detection anchor point along the candidate observation viewpoint direction, used to check whether the global utility tunnel entity model obstructs this line of sight. The safe visible passage interval is the candidate direction range where the reverse spatial detection ray is not obstructed by the utility tunnel entity model and satisfies the frustum coverage rule. The obstacle occlusion direction set is a set of direction data recording the direction of the model facet causing occlusion, the intersection position, and the normal of the occluding surface. The independent coverage requirement data packet generates input for the binding target hidden sub-region, visual channel, and supplementary waypoints in the occlusion direction.

[0042] S411: Obtain the horizontal and vertical field-of-view parameters corresponding to the sensor hardware on the UAV. These parameters are sensor hardware configuration fields, derived from the device's factory parameter table, mission calibration records, and installation attitude records. These fields are written to the sensor configuration cache when the inspection mission is established. After constructing a virtual detection frustum model based on these fields, the local sub-region attribute dataset is called to extract the hidden surface sub-region, the sub-region's independent centroid coordinates, the sub-region's boundary contour point set, and the sub-region's true normal vector. The sub-region's independent centroid coordinates and boundary contour point set are used as target detection anchor points. The detection distance field is jointly provided by the sensor's effective observation configuration and the inspection safety distance configuration, and stored in the waypoint generation parameter table.

[0043] On a detection space sphere centered at the target detection anchor point and with the detection distance as the radius, a virtual detection frustum model is used to screen candidate observation viewpoint directions that meet the field-of-view constraints. The screening order for candidate observation viewpoint directions is as follows: first, verify whether the frustum orientation can cover the target detection anchor point; second, verify whether the boundary contour of the target detection anchor point falls within the sensor coverage area; and finally, verify whether the observation direction forms a positive observation relationship with the true normal vector of the sub-region. Candidate observation records are generated based on the screened directions, and spatial detection rays are emitted towards the target detection anchor point along the candidate observation viewpoint directions to obtain reverse spatial detection rays. If the boundary contour point set has a contour anomaly field, the anomalous boundary points are added to the target detection anchor point queue, ensuring that subsequent occlusion calculations cover both centroid and boundary positions.

[0044] S412: Calculate the surface blocking intersection points between the global utility tunnel solid model and the reverse spatial detection ray using a hierarchical bounding box tree structure algorithm. In this scheme, the hierarchical bounding box tree structure algorithm refers to organizing the faces of the global utility tunnel solid model into multiple layers of candidate collision indices according to spatial bounding relationships, and then sequentially performing bounding space elimination, candidate face selection, and real face intersection verification on the reverse spatial detection ray. The surface blocking intersection point is the first model surface intersection point encountered by the reverse spatial detection ray as it travels from the candidate observation direction towards the target detection anchor point. The fields include the spatial location of the intersection point, the index of the occluding face, the normal of the occluding surface, the ray direction identifier, and the target detection anchor point identifier.

[0045] The safe visible channel interval and obstacle occlusion direction set are extracted based on surface obstruction intersections. When no obstruction intersection occurs before the target detection anchor point and the candidate observation direction satisfies the view frustum coverage rule, the ray direction is written into the safe visible channel interval. When an obstruction intersection occurs before the target detection anchor point, the occlusion patch normal, intersection position, and ray direction are written into the obstacle occlusion direction set. If the hierarchical bounding box tree structure index fails, a spatial index is re-established based on the current global tunnel entity model version; if intersection verification still cannot be completed after reconstruction, the relevant target detection anchor point is written into the occlusion calculation pending verification status and is used as a candidate for missing valid ray visible paths in S413.

[0046] S413: For the hidden surface sub-regions behind the target, select the target hidden sub-regions corresponding to the missing effective ray visual path from the hidden surface sub-regions behind the target based on the intersection of surface obstructions. The target hidden sub-regions are those where a stable visual path cannot be formed by conventional candidate observation directions, and which require additional viewpoint coverage. The selection process first reads the safe visual channel interval corresponding to each target detection anchor point, then reads the obstacle occlusion direction set, and finally verifies the visibility status of the centroid anchor point and boundary anchor point within the same hidden surface sub-region behind the target. When the centroid anchor point or boundary anchor point lacks an effective visual path, the hidden surface sub-region behind the target is recorded as a target hidden sub-region.

[0047] The independent centroid coordinates of the sub-region, the true normal vector of the sub-region, the set of obstacle occlusion directions, and the safe view channel interval are bound to the target hidden sub-region to establish an independent coverage requirement data packet. This data packet contains the target hidden sub-region identifier, parent block identifier, target detection anchor point field, view channel field, occlusion direction field, normal field, anomaly field, and subsequent waypoint generation status field. If the safe view channel interval is empty, the data packet still retains the set of obstacle occlusion directions and the target detection anchor point field for S5 to perform spatial safe region fine-tuning; if the set of obstacle occlusion directions is empty but the view frustum coverage is not satisfied, the data packet records the view frustum constraint insufficiency status and regenerates candidate supplementary viewpoint directions in S5.

[0048] S5: Based on the independent coverage requirement data package, generate supplementary viewpoints for safety avoidance. Combine these supplementary viewpoints with the set of regular basic inspection waypoints generated from ordinary surface blocks in the surface block classification results to construct global spatial waypoints. Supplementary viewpoints for safety avoidance are candidate observation positions generated for target hidden sub-regions, supplemented by checks on the visual channel, occlusion direction, and UAV safety inflation. The set of regular basic inspection waypoints is the set of basic observation positions for ordinary surface blocks, with data sources including the block geometric centroid coordinates, average normal vector, and sensor detection distance configuration of the ordinary surface blocks. Global spatial waypoints are the set of waypoint nodes merged from the regular basic inspection waypoints and supplementary viewpoints for safety avoidance. Path movement edge weights are the movement cost fields connecting two waypoints in the waypoint graph structure, derived from configuration fields such as node spatial Euclidean distance, node altitude change value, maximum flight speed, and UAV climb rate. The Traveling Salesman Problem optimization algorithm in this scheme refers to the path solving process of sorting all waypoint graph structure nodes using closed or open traversal, with the sorting objective being to minimize the sum of path movement edge weights. The polynomial curve smoothing function is a trajectory smoothing processing object that performs continuous trajectory fitting on discrete waypoint sequences and outputs the inspection route planning results.

[0049] S511: Read the independent centroid coordinates of the sub-region, the set of obstacle occlusion directions, and the safe visibility channel interval associated with the independent coverage requirement data packet record. Using the independent centroid coordinates of the sub-region as the reference point, call the detection distance field and reverse-calculate the 3D coordinates of the candidate supplementary viewpoints within the detection distance range along the safe visibility channel interval. The 3D coordinates of the candidate supplementary viewpoints are written to the candidate waypoint cache and bound to the target hidden sub-region identifier. If the safe visibility channel interval is empty, read the outer avoidance direction of the occlusion direction in the obstacle occlusion direction set to generate the initial 3D coordinates of the candidate supplementary viewpoints, and then proceed to collision verification.

[0050] The system obtains the UAV's airframe dimensions and flight control error margin parameters. The UAV airframe dimensions are derived from the UAV equipment configuration table, while the flight control error margin parameters are derived from the inspection mission control strategy table and flight control calibration records. The safety expansion radius, a collision detection field that combines the airframe's occupied space and the flight control offset boundary, is stored in the waypoint safety rule table. Based on the safety expansion radius, a UAV safety expansion sphere is established with the candidate supplementary viewpoint's 3D coordinates as its center. The system then uses the global utility tunnel solid model to calculate the interference collision space distance between the UAV safety expansion sphere and the utility tunnel solid model. This interference collision space distance is then used for physical safety margin rule verification. The physical safety margin is derived from the equipment emergency braking distance parameters and gust offset compensation parameters, both provided by the flight control safety configuration and mission environment configuration.

[0051] When the interference collision spatial distance falls within the physical safety margin boundary, the 3D coordinates of the candidate supplementary viewpoint are fine-tuned towards the spatial safety area using the obstacle occlusion direction set. The fine-tuning process extracts the main occlusion 3D normal vector recorded within the obstacle occlusion direction set. This main occlusion 3D normal vector represents the direction in the occlusion patch normal that aligns with the candidate line-of-sight blocking relationship. An absolute reverse avoidance vector is generated spatially in the reverse direction along the main occlusion 3D normal vector, and the basic translation step size field is used to extend and adjust the candidate supplementary viewpoint's 3D coordinates outwards. The basic translation step size is a waypoint adjustment configuration field derived from the waypoint generation parameter table. After each adjustment, the interference collision spatial distance verification between the UAV safety expansion sphere and the global utility tunnel entity model is re-executed until the adjusted 3D coordinates enter a safe state outside the physical safety margin boundary. The qualified 3D coordinate nodes are written into the safe avoidance supplementary viewpoint field, retaining the adjustment source, occlusion direction, and collision verification status.

[0052] S512: Extract ordinary surface blocks from the surface block classification results, and generate a set of routine basic inspection waypoints based on these blocks. The routine basic inspection waypoints use the geometric centroid coordinates of the ordinary surface blocks as the observation target, determine candidate waypoint positions using the outer direction of the block's average normal vector and the detection distance field, and perform safety verification using the UAV's safety expansion sphere. The routine basic inspection waypoints record the block identifier, waypoint 3D coordinates, observation direction, sensor cone coverage status, and safety verification status. If an ordinary surface block has a verification mark, the corresponding basic waypoint retains redundant records of the observation direction, enabling the inspection execution terminal to transmit verification marks back during on-site data collection.

[0053] A global spatial waypoint is constructed by merging the set of regular basic inspection waypoints with supplementary viewpoints for safety avoidance. The global spatial waypoints are mapped to nodes in the waypoint graph structure associated with the Traveling Salesman Problem (TSP) optimization algorithm. Each waypoint graph structure node records the node's 3D coordinates, node height, observed target, node type, and safety status. The node spatial Euclidean distance is the spatial displacement relationship between two waypoint graph structure nodes, and the node height change value is the height difference between the two nodes in the vertical direction of the local engineering coordinate system. The maximum flight speed is derived from the UAV flight control configuration, and the climb rate is derived from the flight control performance configuration. The level flight time cost is generated jointly by the maximum flight speed field and the node spatial Euclidean distance field, while the altitude adjustment time cost is generated jointly by the UAV climb rate field and the node height change value field. The level flight time cost and the altitude adjustment time cost are merged into the path movement edge weight and written to the edge field of the waypoint graph structure. If a waypoint's safety status fails, the node is not included in the final sorting queue; the reason for the anomaly is written back to the waypoint safety log, and S511 is triggered to fine-tune the corresponding supplementary viewpoint.

[0054] S513: With the optimization objective of minimizing the sum of path movement edge weights, a Traveling Salesman Problem (TSP) optimization algorithm combined with path movement edge weights is used to traverse all waypoint graph structure nodes to obtain the optimal sorted path sequence for the UAV. The solution process reads the waypoint graph structure nodes, path movement edge weights, takeoff point constraints, return point constraints, and node type fields. Regular basic inspection waypoints corresponding to ordinary surface blocks and safety avoidance supplementary viewpoints corresponding to target hidden sub-regions are all considered as waypoint nodes that must be traversed and entered into the sorting process. During sorting, the closer the edge weight state is to the low-cost category, the higher the priority of the corresponding connection relationship in the candidate path. If a candidate path causes continuous waypoint type conflicts or local spatial backtracking, the sorting process calls the waypoint neighborhood constraint record for replacement until an optimal sorted path sequence for the UAV covering all valid waypoint nodes is formed.

[0055] The process employs a polynomial curve smoothing function to fit the flight trajectory associated with the optimal sorted path sequence of the UAV. The fitting process reads the 3D coordinates, observation direction, safety expansion status, and node type fields of the waypoints. It then performs continuous processing on the discrete polyline trajectories between adjacent waypoints, generating smooth trajectory segments while maintaining the waypoint observation direction and safety boundary status. If a smooth trajectory segment conflicts with the global utility tunnel entity model due to safety expansion, the conflicting trajectory segment is written back to the waypoint adjustment queue, and S511 and S512 are called again to update the corresponding waypoints and edge weights. After verification, the inspection route planning results are output. These results include the optimal sorted path sequence of the UAV, the smoothed flight trajectory, the observation target corresponding to each waypoint, the sensor cone coverage status, the safety verification status, and anomaly tracing records. These results are stored in the inspection task result database for use in UAV flight control task issuance and inspection data collection and feedback.

[0056] Corresponding to the aforementioned UAV inspection route planning method based on 3D models, the UAV inspection route planning system based on 3D models carries out data acquisition, feature interpretation, topology analysis, occlusion calculation, and trajectory arrangement in the same embodiment. The spatial modeling module connects to the engineering modeling platform and the inspection task parameter table, receiving the triangular mesh model file of the utility tunnel, coordinate reference records, and bounding box cutting size threshold fields, and outputting the global utility tunnel entity model, global axial bounding boxes, initial surface blocks, and block geometric feature datasets. The feature interpretation module reads the block geometric feature dataset, classification rule table, and ray intersection records, outputs the surface block classification results, and transmits complex cross-connection blocks and ordinary surface blocks to the topology analysis module and trajectory arrangement module, respectively.

[0057] The topology parsing module reads complex cross-connection blocks, patch topology connections, clustering rule tables, and the depth field of the local engineering coordinate system, outputting a local sub-region attribute dataset. The occlusion calculation module reads the local sub-region attribute dataset, sensor configuration cache, waypoint generation parameter table, and global utility tunnel entity model, outputting an independent coverage requirement data package. The trajectory orchestration module reads the independent coverage requirement data package, ordinary surface blocks, safety rule table, and flight control configuration, outputting the inspection route planning results. Each module maintains data consistency through inspection task identifiers, model version identifiers, block identifiers, sub-region identifiers, and waypoint node identifiers. Input parsing failures, missing coordinate references, patch topology breaks, insufficient view frustum constraints, failed collision checks, and trajectory smoothing conflicts are all written to the corresponding exception fields. These exception fields are carried into subsequent modules along with the data package and form traceable records in the inspection task result database.

[0058] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for planning unmanned aerial vehicle (UAV) inspection routes based on a 3D model, characterized in that, Includes the following steps: S1: Obtain the triangular mesh model file of the pipe gallery in the area to be inspected, map the model space coordinates recorded in the triangular mesh model file to the local coordinate system of the project to generate a global pipe gallery entity model, calculate the extreme coordinates of the three-dimensional contour boundary corresponding to the global pipe gallery entity model to obtain the global axial bounding box, divide the three-dimensional space associated with the global axial bounding box into a three-dimensional voxel mesh according to the bounding box cutting size threshold, extract the intersecting meshes in the three-dimensional voxel mesh that intersect with the surface of the global pipe gallery entity model as initial surface blocks, extract the geometric features of the initial surface blocks and associate and encapsulate them to generate a block geometric feature dataset; S2: Call the block geometric feature dataset, classify the initial surface blocks based on the local depth variance, normal angle scattering range and ray intersection spatial fault features, and generate surface block classification results; S3: Extract the complex cross-connected blocks in the surface block classification results, combine the topological connection relationship of the patches to perform region growth clustering algorithm processing, and obtain the visible surface sub-regions in front and the hidden surface sub-regions behind based on the semantic splitting of the independent cluster surface hierarchy. Extract the independent centroid coordinates, true normal vectors and boundary contour point set attributes of the visible surface sub-regions in front and the hidden surface sub-regions behind respectively to generate a local sub-region attribute dataset. S4: Call the local sub-region attribute dataset, construct a virtual detection frustum model based on the hardware parameters of the UAV-mounted sensor, divide candidate observation directions on the detection space sphere with the target detection anchor point as the center and emit reverse space detection rays using the virtual detection frustum model, calculate the surface obstruction intersection between the global tube gallery entity model and the reverse space detection rays, extract the safe visual channel interval and obstacle occlusion direction set according to the surface obstruction intersection and bind the target hidden sub-region, and establish an independent coverage requirement data package; S5: Generate a safety avoidance supplementary viewpoint based on the independent coverage requirement data packet, and construct a global spatial waypoint by combining the safety avoidance supplementary viewpoint with the set of regular basic inspection waypoints generated by the ordinary surface blocks in the surface block classification results. With the goal of minimizing the sum of path movement edge weights, the inspection route planning result is generated by processing with the traveling salesman problem optimization algorithm and the polynomial curve smoothing function.

2. The UAV inspection route planning method based on a three-dimensional model according to claim 1, characterized in that, The specific steps of S1 are as follows: S111: Obtain the triangular mesh model file of the utility tunnel, map the model space coordinates recorded in the triangular mesh model file of the utility tunnel to the local coordinate system of the project to generate a global utility tunnel entity model, calculate the extreme coordinates of the three-dimensional contour boundary corresponding to the global utility tunnel entity model, and obtain the global axial bounding box. S112: Call the bounding box cutting size threshold, divide the three-dimensional space associated with the global axial bounding box into a three-dimensional voxel mesh according to the bounding box cutting size threshold, extract the intersecting meshes in the three-dimensional voxel mesh that intersect with the surface of the global tube gallery entity model, and generate initial surface blocks. S113: Extract the original model vertex coordinates and triangular patch indices inside the initial surface block, calculate the spatial mean of the original model vertex coordinates to obtain the block geometric centroid coordinates, calculate the mean of the patch normal vectors based on the original model vertex coordinates and the triangular patch indices, extract the block average normal vector, and associate and encapsulate the initial surface block, the original model vertex coordinates, the triangular patch indexes, the block geometric centroid coordinates, and the block average normal vector to generate a block geometric feature dataset.

3. The UAV inspection route planning method based on a three-dimensional model according to claim 2, characterized in that, The specific steps of S2 are as follows: S211: Extract the initial surface blocks, original model vertex coordinates, triangular patch indexes, block geometric centroid coordinates, and block average normal vectors from the block geometric feature dataset. Establish a local spatial tangent plane with the block geometric centroid coordinates as the origin. Calculate the vertical projection distance of the vertex based on the original model vertex coordinates and the local spatial tangent plane. Calculate the statistical variance corresponding to the vertical projection distance of the vertex and generate the local depth variance. S212: Extract the facets of the utility tunnel entity model based on the triangular facet index, calculate the spatial angle between the facet normal vector associated with the facet of the utility tunnel entity model and the block average normal vector, extract the scattering range corresponding to the spatial angle, and obtain the normal angle scattering range. S213: A spatial detection ray is emitted from the virtual detection hemisphere associated with the initial surface block in front of the initial surface block into the interior of the initial surface block. The Euclidean distance between the intersection points of the pipe gallery surfaces between adjacent spatial detection rays is calculated. A gap width threshold is obtained. The Euclidean distance between the intersection points of the pipe gallery surfaces that is greater than the gap width threshold is extracted. A spatial fault feature of the ray intersection point is established. S214: Obtain the jump threshold and the continuous scattering threshold. Based on the local depth variance greater than the jump threshold, the normal angle scattering range greater than the continuous scattering threshold, and the spatial fault characteristics of the ray intersection, separate complex bridging blocks and ordinary surface blocks from the initial surface blocks to generate surface block classification results.

4. The UAV inspection route planning method based on a three-dimensional model according to claim 3, characterized in that, The process of generating surface block classification results is as follows: Obtain the preset equipment ranging error parameters and the pipeline design wall thickness parameters, calculate the sum of the values ​​of the pipeline design wall thickness parameters and the preset equipment ranging error parameters, and use the sum of the values ​​as the jump threshold; Extract the standard radius of curvature of the pipe and the surface sampling grid size, calculate the normal deflection limit value based on the ratio of the surface sampling grid size to the standard radius of curvature of the pipe, and determine the normal deflection limit value as the dispersion continuity threshold; If the local depth variance corresponding to the initial surface block is greater than the jump threshold, and the normal angle scattering range is greater than the scattering continuity threshold, and the ray intersection spatial fault feature exists, the initial surface block is classified as the complex bridging block. If not all of the initial surface blocks meet the above judgment conditions, the initial surface blocks will be classified as ordinary surface blocks. Integrate all the complex cross-connection blocks and the ordinary surface blocks to generate the surface block classification result.

5. The UAV inspection route planning method based on a three-dimensional model according to claim 3, characterized in that, The specific steps of S3 are as follows: S311: Extract complex cross-connection blocks from the surface block classification results, obtain the topological connection relationship of the facets corresponding to the global utility tunnel entity model, extract the utility tunnel entity model faces contained in the complex cross-connection blocks, map the unassigned utility tunnel entity model faces to clustering seed faces associated with the region growing clustering algorithm, calculate the facet spatial Euclidean distance and the facet normal angle between the clustering seed facets and the neighboring utility tunnel faces, obtain the distance continuity threshold and the smooth transition threshold, filter the neighboring utility tunnel faces whose facet spatial Euclidean distance is less than the distance continuity threshold and whose facet normal angle is less than the smooth transition threshold according to the region growing clustering algorithm, merge the filtered neighboring utility tunnel faces, and generate independent surface clusters; S312: Obtain the local coordinate system of the project, calculate the surface spatial depth value of the independent surface cluster in the local coordinate system of the project, establish the surface hierarchical semantics of the independent cluster based on the surface spatial depth value and the topological connection relationship of the patch, and split the independent surface cluster according to the surface hierarchical semantics of the independent cluster to obtain the visible surface sub-region in front and the hidden surface sub-region behind. S313: For the visible surface sub-region in front and the hidden surface sub-region behind, calculate the independent centroid coordinates, the true normal vector, and the boundary contour point set of the sub-region based on the visible surface sub-region in front and the hidden surface sub-region behind, respectively. Map the independent centroid coordinates, the true normal vector, and the boundary contour point set of the sub-region to the corresponding visible surface sub-region in front and the hidden surface sub-region behind, respectively, to generate a local sub-region attribute dataset.

6. The UAV inspection route planning method based on a three-dimensional model according to claim 5, characterized in that, The process of generating independent surface clusters is as follows: Calculate the average side length of the grid of the pipe gallery solid model patch, and configure the product of the average side length of the grid and the preset spatial tolerance coefficient as the distance continuity threshold. Obtain the standard curvature parameters of the pipe gallery pipeline, calculate the theoretical maximum normal deflection angle based on the standard curvature parameters of the pipe gallery pipeline and the average side length of the grid, and set the theoretical maximum normal deflection angle as the smooth transition threshold. If the extracted Euclidean distance of the patch space is less than the distance continuity threshold, and the included angle of the patch normal vector is less than the smooth transition threshold, the neighborhood tube gallery patch is added to the current cluster set. The most recently added neighborhood tube gallery patch in the current cluster set is selected as the iteratively derived seed patch for outward comparison until there are no adjacent patches that meet the numerical constraints. By piecing together all the neighboring tube gallery patches recorded within the current cluster set, the independent surface cluster is generated.

7. The UAV inspection route planning method based on a three-dimensional model according to claim 5, characterized in that, The specific steps of S4 are as follows: S411: Obtain the horizontal and vertical field-of-view parameters corresponding to the sensor hardware on the UAV. Construct a virtual detection frustum model based on the horizontal and vertical field-of-view parameters. Call the local sub-region attribute dataset and extract the hidden surface sub-region, independent centroid coordinates of the sub-region, boundary contour point set of the sub-region, and true normal vector of the sub-region contained in the local sub-region attribute dataset. Use the independent centroid coordinates of the sub-region and the boundary contour point set of the sub-region as the target detection anchor point and obtain the detection distance. On the detection space sphere with the target detection anchor point as the center and the detection distance as the radius, use the virtual detection frustum model to select the candidate observation viewpoint direction that meets the field-of-view constraints. Emit a space detection ray along the candidate observation viewpoint direction toward the target detection anchor point to obtain the reverse space detection ray. S412: Calculate the surface obstruction intersection between the global tube gallery entity model and the reverse space detection ray using the hierarchical bounding box tree structure algorithm, and extract the safe visible channel interval and the set of obstacle occlusion directions based on the surface obstruction intersection. S413: For the rear hidden surface sub-region, select the target hidden sub-region corresponding to the missing effective ray visual path from the rear hidden surface sub-region according to the surface blocking intersection, and bind the independent centroid coordinates of the sub-region, the true normal vector of the sub-region, the obstacle occlusion direction set and the safe visual channel interval to the target hidden sub-region to establish an independent coverage requirement data packet.

8. The UAV inspection route planning method based on a three-dimensional model according to claim 7, characterized in that, The specific steps of S5 are as follows: S511: Read the independent centroid coordinates of the sub-region, the set of obstacle occlusion directions, and the safe visibility channel interval associated with the independent coverage requirement data packet record. Using the independent centroid coordinates of the sub-region as a reference point, obtain the detection distance. Calculate the three-dimensional coordinates of the candidate supplementary viewpoint in reverse along the safe visibility channel interval within the detection distance range. Obtain the UAV body size parameters and flight control error margin parameters. Calculate the safety expansion radius based on the UAV body size parameters and flight control error margin parameters. Establish a UAV safety expansion sphere with the three-dimensional coordinates of the candidate supplementary viewpoint as its center based on the safety expansion radius. Call the global utility tunnel entity model to calculate the interference collision space distance between the UAV safety expansion sphere and the global utility tunnel entity model. When the interference collision space distance is less than the set physical safety margin, use the set of obstacle occlusion directions to guide the three-dimensional coordinates of the candidate supplementary viewpoint to be finely adjusted towards the spatial safety area to obtain a safe avoidance supplementary viewpoint. S512: Extract ordinary surface blocks from the surface block classification results, generate a set of regular basic inspection waypoints based on the ordinary surface blocks, merge the set of regular basic inspection waypoints with the safety avoidance supplementary viewpoints to construct global spatial waypoints, map the global spatial waypoints to waypoint graph structure nodes associated with the traveling salesman problem optimization algorithm, calculate the node spatial Euclidean distance and node altitude change values ​​between the waypoint graph structure nodes, obtain the maximum flight speed and UAV climb rate, calculate the level flight time cost based on the maximum flight speed and the node spatial Euclidean distance, calculate the UAV altitude adjustment time cost based on the UAV climb rate and node altitude change values, and establish path movement edge weights based on the level flight time cost and the UAV altitude adjustment time cost; S513: With minimizing the sum of the weights of the path movement edges as the optimization objective, the Traveling Salesman Problem optimization algorithm is used in conjunction with the weights of the path movement edges to traverse and solve all the waypoint graph structure nodes to obtain the optimal sorted path sequence of the UAV. The polynomial curve smoothing function is then used to fit the flight trajectory associated with the optimal sorted path sequence of the UAV to generate the inspection route planning result.

9. The UAV inspection route planning method based on a three-dimensional model according to claim 8, characterized in that, The process of using the obstacle occlusion direction set to guide the three-dimensional coordinates of the candidate supplementary viewpoint to be finely adjusted towards the spatial safety area to obtain a safe avoidance supplementary viewpoint is as follows: Extract the three-dimensional normal vector of the main body occlusion recorded within the set of obstacle occlusion directions; Calculate the absolute reverse avoidance vector that is opposite to the three-dimensional normal vector space of the occlusion of the main body; Obtain the basic translation step size, and adjust the three-dimensional coordinates of the candidate supplementary viewpoint by extending it outward along the direction of the absolute reverse avoidance vector according to the basic translation step size; Obtain the equipment emergency braking distance parameter and the gust offset compensation parameter, calculate the sum of the values ​​of the equipment emergency braking distance parameter and the gust offset compensation parameter, and configure the sum of the values ​​as the physical safety margin; Recalculate the interference collision space distance corresponding to the three-dimensional coordinates of the candidate supplementary viewpoint after the extension and adjustment. If the interference collision space distance is less than the physical safety margin, perform the extension and adjustment operation repeatedly. Until the distance of the interference collision space corresponding to the extended and adjusted three-dimensional coordinates is greater than the physical safety margin, the qualified three-dimensional coordinate nodes are extracted to generate the safety avoidance supplementary viewpoint.

10. A UAV inspection route planning system based on a three-dimensional model, characterized in that, The system is used to implement the UAV inspection route planning method based on a three-dimensional model as described in any one of claims 1-9, and the system includes: The spatial modeling module acquires the triangular mesh model file of the utility tunnel in the area to be inspected, maps the model space coordinates recorded in the triangular mesh model file to the local coordinate system of the project to generate a global utility tunnel entity model, calculates the extreme coordinates of the three-dimensional contour boundary corresponding to the global utility tunnel entity model to obtain a global axial bounding box, divides the three-dimensional space associated with the global axial bounding box into a three-dimensional voxel mesh according to the bounding box cutting size threshold, extracts the intersecting meshes in the three-dimensional voxel mesh that intersect with the surface of the global utility tunnel entity model as initial surface blocks, extracts the geometric features of the initial surface blocks and performs associated encapsulation to generate a block geometric feature dataset; The feature interpretation module calls the block geometric feature dataset and classifies the initial surface blocks based on the local depth variance, the range of normal angle dispersion, and the spatial tomographic features of ray intersections, generating surface block classification results. The topology parsing module extracts complex cross-connected blocks within the surface block classification results, performs region growing clustering algorithm processing based on the topological connection relationship of the patches, and obtains the visible surface sub-regions in front and the hidden surface sub-regions behind based on the semantic segmentation of the independent cluster surface hierarchy. The independent centroid coordinates, true normal vectors and boundary contour point set attributes of the visible surface sub-regions in front and the hidden surface sub-regions behind are extracted respectively to generate a local sub-region attribute dataset. The occlusion calculation module calls the local sub-region attribute dataset, constructs a virtual detection frustum model based on the hardware parameters of the sensors on the UAV, uses the virtual detection frustum model to divide candidate observation directions on the detection space sphere with the target detection anchor point as the center and emits reverse space detection rays, calculates the surface obstruction intersection between the global tube gallery entity model and the reverse space detection rays, extracts the safe visual channel interval and obstacle occlusion direction set based on the surface obstruction intersection and binds the target hidden sub-region, and establishes an independent coverage requirement data package; The trajectory arrangement module generates a safety avoidance supplementary viewpoint based on the independent coverage requirement data packet, and constructs a global spatial waypoint by combining the safety avoidance supplementary viewpoint with the set of regular basic inspection waypoints generated by the ordinary surface blocks in the surface block classification results. With the goal of minimizing the sum of path movement edge weights, the inspection route planning results are generated through the traveling salesman problem optimization algorithm and the polynomial curve smoothing function.

Images