Unmanned aerial vehicle navigation method and system based on semantic BIM and multi-sensor fusion

Abstract

The present application belongs to the technical field of unmanned aerial vehicle navigation, in order to solve the problem that the existing complex tunnel environment unmanned aerial vehicle path planning fails to mine BIM semantic information, etc., an unmanned aerial vehicle navigation method and system based on semantic BIM and multi-sensor fusion are proposed, the tunnel original road BIM model is subjected to lightweight processing, the geometric entity extraction and attribute set are adopted to analyze the component data of the tunnel BIM model; the three-dimensional grid map is used as a global map template, the current collected point cloud data is used to adjust the unmanned aerial vehicle attitude, the unmanned aerial vehicle accurate pose is obtained, and then a dynamic environment map is constructed; based on the static semantic cost, the dynamic obstacle space-time prediction cost and the channel passing cost, a path search algorithm is used for path planning. The present application combines priori semantics and real-time obstacle avoidance requirements for path optimization, realizing high-precision and high-safety flight inspection of the unmanned aerial vehicle in the complex tunnel environment.

Description

Technical Field

[0001] This invention belongs to the field of UAV navigation technology, and particularly relates to a UAV navigation method and system based on semantic BIM and multi-sensor fusion. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] As an indispensable infrastructure, the importance of tunnel safety inspection and maintenance is self-evident. With the continuous development of technology, drones, with their superior flexibility and maneuverability, are gradually becoming an important tool in the field of tunnel inspection, providing strong support for the safe operation of tunnels. However, the internal environment of tunnels is extremely complex, exhibiting many characteristics that are unfavorable to the autonomous navigation and path planning of drones.

[0004] Currently, several methods exist for UAV path planning, but all have certain limitations. One traditional and common method is based on pre-programmed waypoint flight; another relies on real-time obstacle avoidance algorithms using onboard sensors (such as vision and lidar). Furthermore, while there have been preliminary attempts to use BIM models for UAV path planning, these attempts are mostly superficial, treating BIM merely as a visual background or simple static obstacles, failing to delve into the semantic information contained within the BIM model, such as the spatial logical relationships between lanes and maintenance access routes. This semantic information is crucial for reasonable UAV path planning. Simultaneously, existing methods have failed to effectively integrate BIM models with real-time sensing data. Because the generated paths are often static, they cannot be adjusted in a timely manner according to dynamic environmental changes during inspections, significantly limiting their practicality. Summary of the Invention

[0005] To overcome the shortcomings of the existing technologies, this invention provides a UAV navigation method and system based on semantic BIM and multi-sensor fusion. It integrates prior semantics and real-time obstacle avoidance requirements for path optimization, enabling UAVs to perform autonomous flight inspections with high precision, high safety, and high adaptability in complex tunnel environments.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a UAV navigation method based on semantic BIM and multi-sensor fusion, comprising: The tunnel BIM model is lightweighted, and the component data of the tunnel BIM model is extracted using geometric entities and attribute set parsing to identify semantic entities with navigation significance. The tunnel space is processed into a three-dimensional raster and the attributes of each cell in the three-dimensional raster map are filled to obtain a three-dimensional raster map rich in semantic attributes. The discrete raster attributes are then converted into a continuous semantic cost field. Using a 3D raster map rich in semantic attributes as a global map template, the attitude of the UAV is adjusted based on the currently collected point cloud data to obtain the precise pose of the UAV in BIM coordinates. The precise pose of the drone is integrated into the global map template to construct a dynamic environment map; Based on the semantic cost field as the static semantic cost basis for path planning, and according to the movement pattern of dynamic obstacles along the tunnel axis in the dynamic environment map, a spatiotemporal prediction cost for dynamic obstacles is constructed. Based on the tunnel design axis and safety clearance zone in the BIM model, a channel cost function is constructed to guide the UAV to stay in the center area of ​​the tunnel. Based on the static semantic cost, the spatiotemporal prediction cost of dynamic obstacles, and the channel passage cost, a path search algorithm is used to perform path planning and obtain the path navigation result of the UAV.

[0007] Secondly, the present invention provides an unmanned aerial vehicle (UAV) navigation system based on semantic BIM and multi-sensor fusion, comprising: The semantic extraction module is configured to: perform lightweight processing on the tunnel BIM model, use geometric entity extraction and attribute set parsing to analyze the component data of the tunnel BIM model, and identify semantic entities with navigation significance; The rasterization module is configured to: perform three-dimensional rasterization on the tunnel space, fill the attributes of each cell in the three-dimensional raster map to obtain a three-dimensional raster map rich in semantic attributes, and convert the discrete raster attributes into a continuous semantic cost field. The pose module is configured to: use a 3D raster map rich in semantic attributes as a global map template, adjust the UAV's attitude based on the currently collected point cloud data, and obtain the UAV's accurate pose in BIM coordinates; The dynamic environment map building module is configured to: integrate the precise pose of the UAV into the global map template to build a dynamic environment map; The navigation module is configured to: use a semantic cost field as the static semantic cost basis for path planning; construct a spatiotemporal prediction cost for dynamic obstacles based on the movement pattern of dynamic obstacles along the tunnel axis in the dynamic environment map; construct a channel cost function to guide the UAV to stay in the center area of ​​the tunnel based on the tunnel design axis and safety clearance zone in the BIM model; and use a path search algorithm to perform path planning based on the static semantic cost, the spatiotemporal prediction cost of dynamic obstacles, and the channel passage cost to obtain the path navigation result of the UAV.

[0008] Thirdly, the present invention provides an electronic device including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the method described in the first aspect.

[0009] Fourthly, the present invention provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in the first aspect.

[0010] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the method described in the first aspect.

[0011] The above one or more technical solutions have the following beneficial effects: In this invention, the original BIM model of the tunnel is lightweighted. Geometric entity extraction and attribute set analysis are used to analyze the component data of the tunnel BIM model, identifying semantic entities with navigational significance. A 3D raster map rich in semantic attributes is used as a global map template, and the real-time acquired precise pose of the UAV is integrated into the global map template to construct a dynamic environment map. This achieves a deep fusion of static prior knowledge and dynamic environmental information. By identifying navigational semantic entities in the BIM, a physically meaningful cost field is generated, enabling path planning to follow a reasonable passage logic within the tunnel. Based on this, a comprehensive cost function is formed by superimposing the passage cost and the dynamic obstacle spatiotemporal prediction cost, using the semantic cost field as a static foundation. The passage cost guides the UAV to fly along the central area of ​​the tunnel, effectively avoiding the risk of drifting to the sidewalls due to local obstacle avoidance, which is a risk associated with traditional methods. The dynamic obstacle spatiotemporal prediction cost predicts the movement trend of moving obstacles based on their movement along the tunnel axis, enabling early avoidance. This achieves high-precision, high-safety, and highly adaptive autonomous flight inspection of the UAV in complex tunnel environments.

[0012] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0013] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0014] Figure 1 This is an overall block diagram of the UAV navigation method based on semantic BIM and multi-sensor fusion in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the edge folding operation in Embodiment 1 of the present invention. Detailed Implementation

[0015] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0016] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0017] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0018] Example 1 The interior of tunnels is a typical GPS signal rejection zone. Traditional GPS-based navigation methods are completely ineffective within tunnels, and drones cannot obtain their own location information using GPS signals. Secondly, the repetitive structural features of tunnels, with their long, straight, and symmetrical layout, result in low spatial feature differentiation, posing significant challenges to the drone's environmental perception and positioning. Furthermore, the variable lighting conditions within tunnels, from bright entrances to dim interiors, severely impact the performance of the drone's onboard sensors, further increasing the difficulty of autonomous navigation.

[0019] Currently, several methods exist for UAVs in tunnel path planning, but all have certain limitations. The method based on pre-programmed waypoints is simple to operate; by setting the UAV's flight path points in advance, it flies along a pre-defined route. However, it lacks adaptability to dynamic environmental changes. In actual tunnel inspections, various unforeseen obstacles may appear, such as temporarily parked engineering vehicles and construction areas. If these situations are encountered, the pre-programmed path cannot be adjusted in real time, and the UAV will be unable to effectively avoid these obstacles, leading to mission failure or even serious collisions that damage the UAV and tunnel facilities, posing a significant threat to the safe operation of the tunnel. Another approach relies on real-time obstacle avoidance algorithms using onboard sensors (such as vision and lidar). This algorithm uses sensors onboard the UAV to perceive the surrounding environment in real time and dynamically adjusts the flight path accordingly to avoid obstacles. Tunnels are characterized by their long, straight, symmetrical, and monotonous textures. As the UAV's flight distance increases, sensor measurement errors accumulate, causing positioning deviations in the UAV. Moreover, due to the similarity of tunnel structures, sensors have difficulty accurately distinguishing between obstacles and normal tunnel wall structures ahead. This may cause drones to sway unpredictably during flight or fly too close to the tunnel walls, seriously affecting flight safety and efficiency.

[0020] Building Information Modeling (BIM), as a digital representation containing rich geometric and semantic information, provides high-precision prior 3D models for tunnels. While there have been preliminary attempts to use BIM models for UAV path planning, these efforts have largely remained superficial, treating BIM merely as a visual background or simple static obstacles. They have failed to delve deeper into the semantic information contained within the BIM model, such as the spatial logical relationships between lanes and maintenance access routes—information crucial for effective UAV path planning. Furthermore, existing methods have not achieved effective integration of BIM models with real-time sensing data. Because the generated paths are often static, they cannot be adjusted in a timely manner according to dynamic environmental changes during inspections, significantly limiting their practicality.

[0021] As an indispensable infrastructure, tunnel safety inspection is crucial. Drones, with their flexibility, are ideal tools, but the characteristics of tunnels—GPS rejection, repetitive structures, and variable lighting—pose significant challenges to autonomous navigation. Existing methods have obvious limitations: pre-programmed waypoint methods cannot adapt to dynamic obstacles; pure sensor-based real-time obstacle avoidance methods accumulate errors and are prone to path jitter in long, straight tunnels; existing BIM-assisted methods only treat BIM as a static background, failing to extract semantic information and cannot be deeply integrated with real-time perception to achieve dynamic replanning.

[0022] Therefore, this embodiment discloses a UAV navigation method based on semantic BIM and multi-sensor fusion, including: The tunnel BIM model is lightweighted, and the component data of the tunnel BIM model is extracted using geometric entities and attribute set parsing to identify semantic entities with navigation significance. The tunnel space is processed into a three-dimensional raster and the attributes of each cell in the three-dimensional raster map are filled to obtain a three-dimensional raster map rich in semantic attributes. The discrete raster attributes are then converted into a continuous semantic cost field. Using a 3D raster map rich in semantic attributes as a global map template, the attitude of the UAV is adjusted based on the currently collected point cloud data to obtain the precise pose of the UAV in BIM coordinates. The precise pose of the drone is integrated into the global map template to construct a dynamic environment map; Based on the semantic cost field as the static semantic cost basis for path planning, and according to the movement pattern of dynamic obstacles along the tunnel axis in the dynamic environment map, a spatiotemporal prediction cost for dynamic obstacles is constructed. Based on the tunnel design axis and safety clearance zone in the BIM model, a channel cost function is constructed to guide the UAV to stay in the center area of ​​the tunnel. Based on the static semantic cost, the spatiotemporal prediction cost of dynamic obstacles, and the channel passage cost, a path search algorithm is used to perform path planning and obtain the path navigation result of the UAV.

[0023] This embodiment constructs a complete technical closed loop consisting of three main parts. The first part involves offline processing of the original BIM model to extract its geometric and semantic information, building a lightweight, intelligent navigation knowledge base. The second part, during online runtime, utilizes this knowledge base as an "intelligent template," fusing it with real-time sensor data to achieve precise positioning and perceive dynamic changes, constructing a digital twin of the dynamic environment. The third part, based on this digital twin, performs spatiotemporal joint path optimization by integrating prior semantic rules and a cost function that incorporates real-time obstacle avoidance requirements, generating a safe, efficient, and executable flight trajectory.

[0024] The following is combined Figures 1-2 This embodiment provides a detailed description of the UAV navigation method based on semantic BIM and multi-sensor fusion. Step 1: Offline processing and semantic information extraction of the prior BIM model.

[0025] To achieve intelligent navigation of drones within tunnels, the information-rich but massive original BIM model must be transformed into a lightweight, semantically rich navigation knowledge base that can be invoked by algorithms in real time. This section will achieve this goal in three steps: First, the BIM model will be geometrically lightweighted to address its excessive computational burden; second, its inherent semantic information will be deeply analyzed and structured, enabling the machine to "understand" the functions and risks of the tunnel environment; finally, the geometric and semantic information will be integrated to generate a semantic 3D map that supports online positioning and path planning.

[0026] Step 11: Lightweight processing of BIM model.

[0027] While the original BIM model contains detailed structural and semantic information, its massive data volume would impose a huge computational burden if used directly for real-time path planning. Therefore, it must be lightweighted. Traditional simplification methods (such as randomly deleting vertices or faces), although simple, severely damage the topology of the BIM model, leading to geometric distortions such as holes and cracks in the mesh, which cannot meet the basic requirements of navigation algorithms for model integrity and accuracy.

[0028] To address the aforementioned issues, this embodiment employs an edge-folding algorithm as the core lightweighting method. Its core idea is to iteratively shrink an edge into a single vertex and remove adjacent faces. While preserving the tunnel's macroscopic geometric contours and key structural features to the greatest extent possible, this significantly reduces the number of triangular faces and vertices in the model, generating a lightweight BIM 3D model and laying the foundation for subsequent real-time matching and calculation.

[0029] Mathematical model of edge folding algorithm: (1) BIM model conversion.

[0030] The surface of any complex BIM component (wall, slab, space) is discretized in the computer into a mesh composed of countless triangles. The geometry in the BIM model is diverse, and triangular mesh is the standard and universal format for processing complex curved surfaces in computer graphics. After being converted into a mesh, all subsequent algorithms can process it uniformly.

[0031] Set up a triangular mesh:

[0032] in: V = {v 1 ，v 2 ...v n } It is the set of all the vertices of the triangular facets in the original BIM model, with each vertex corresponding to a specific sampling point of the tunnel lining or component surface in the global coordinate system; E It is a set of edges connecting these vertices, where each edge represents the geometric connection between two adjacent vertices in the BIM model surface mesh; F It is a collection of triangular facets, each consisting of three sides, and is the smallest unit that makes up the visualized surface of a BIM model.

[0033] (2) Fold the edges.

[0034] To reduce the number of triangles in a BIM model, a controllable basic operation that preserves the mesh's fundamental structure needs to be defined. Edge collapsing, by merging adjacent vertices to delete edges and faces, is an atomic operation that achieves progressive simplification. Instead of brute-force deletion, it "merges," thus preserving the mesh's topology (connectivity) as much as possible without drastic changes.

[0035] Set edge folding operation:

[0036] It doesn't simply delete an edge, but rather adds the two endpoints v of that edge. i and v j Merge into a new vertex v ij .

[0037] The process is as follows: ① Select an edge: Select one edge in the grid (assuming the two endpoints of the edge are v). i and v j ); ② Create a new vertex: Generate a new vertex v at a certain location. ij to replace the original v i and v j ; ③ Delete geometric element: Delete this edge (v i , v j ), and all triangular faces containing this edge; ④ Reconnect the mesh: Reconnect the mesh that was originally connected to v i or v j All other edges are reconnected to the new vertex v. ij superior.

[0038] (3) Secondary error measurement.

[0039] The key to edge folding operations lies in deciding "which edge to fold first" and "where to place the new vertex." To this end, this embodiment introduces a Quadric Error Metric (QEM) as the decision-making basis. The core idea of ​​QEM is very intuitive: it treats each vertex on the model surface as being defined by its adjacent triangular planes. QEM quantifies the "reasonableness" or "importance" of a vertex in its current position by calculating the sum of the squared distances from a vertex to all its surrounding relevant planes. The smaller this value, the better the vertex fits the surrounding geometry.

[0040] For each vertex , where x, y, z are the three-dimensional coordinates of the vertex in the BIM global coordinate system.

[0041] Define its error matrix:

[0042] in: , represents the plane equation containing all triangles of vertex v in the original BIM model. ax+by+cz+d=0 ,and a 2 +b 2 +c 2 =1Q(v) is a 4×4 symmetric matrix that quantifies the sum of squared distances from vertex v to all its surrounding triangular faces. The smaller the value, the lower the contribution of the vertex to maintaining the original geometry of the BIM model, and the more likely it is to be collapsed.

[0043] side （v i ，v j ） The cost of folding is: (4) in: and The vertices are calculated according to formula (3). and The error matrix.

[0044] The cost calculation for edge folding aims to find an optimal merging scheme that minimizes geometric distortion.

[0045] (4) Solving for the optimal contraction position.

[0046] The optimal contraction position v* is obtained by solving a system of linear equations.

[0047] in: q ij It is a matrix + Corresponding position (number) i Line number j The matrix (column) contains the values ​​of two adjacent vertices on the surface of a BIM component.

[0048] After processing, based on the experience of lightweighting typical tunnel BIM models, the number of triangular facets can usually be reduced by 70%-90%, and the data volume can be reduced by 1-2 orders of magnitude, laying a solid foundation for subsequent real-time matching and calculation. Through lightweight processing, we obtained a BIM model that retains key geometric features and has high computational efficiency. However, a model with only a geometric shell is far from sufficient for intelligent navigation; it is also necessary to explore its inherent meaning.

[0049] Step 12: Semantic information extraction and classification system.

[0050] The core task of this stage is semantic information extraction and classification system construction. The BIM model is not just a three-dimensional shell, but also contains rich component attribute information, such as "driveway", "maintenance access", "ventilation system" etc. By parsing the component data of BIM (such as the attribute set in IFC format), these semantic entities with navigation significance can be identified.

[0051] (1) Semantic parsing framework.

[0052] The semantic parsing framework aims to systematically extract and structure navigation-related information from BIM models, transforming raw building data into computable navigation knowledge. The core of the framework is the deep parsing of the Industry Foundation Class (IFC) standard format model, employing a dual-path parallel processing mechanism of geometric entity extraction and attribute set parsing.

[0053] In terms of geometric entity extraction, to achieve accurate identification and classification of key structural components of the tunnel, this embodiment adopts a semantic-geometric joint parsing method based on the Industrial Foundation Class (IFC) standard, which specifically includes the following sub-steps: (a) IFC entity type filtering: By parsing the IFC file, all entities in the model are traversed, and the initial filtering is performed based on the IFC entity name. Entities of the type as tunnel lining (IFCWall), roadway slab and maintenance road (IFCSlab), beams and columns as supporting structures (IFC Beam / Column), and three-dimensional space (IFC-Space) defining functional areas are extracted to construct a candidate component set.

[0054] (b) Geometric feature quantification and role determination: For candidate components, their three-dimensional geometric expression is extracted, and their specific functional role in the tunnel is determined by calculating geometric feature quantities.

[0055] (c) Verification of spatial topology relationship: Verify the rationality of classification by calculating the spatial topology relationship between components, such as detecting the “beam-column” support relationship and whether IfcSpace is enclosed by the wall.

[0056] Through the above sub-steps, the system can automatically identify key structural components such as lining, driveway slabs, maintenance access, beams, columns, and functional spaces from the original BIM model.

[0057] In terms of attribute set parsing, rich semantic information is extracted from the metadata of each component. For example, by analyzing the construction attribute set (Pset-Construction), the construction stage and quality status of the component can be obtained; the maintenance attribute set (Pset-Maintenance) reveals the equipment maintenance cycle and history; while attribute sets such as traffic management (Pset-TrafficManagement), ventilation (Pset-Ventilation), and lighting (Pset-Lighting) directly provide key environmental parameters that affect flight safety and path planning.

[0058] The specific implementation of the semantic parsing framework described above is shown in the table below. It systematically establishes a mapping relationship from the IFC model to navigation knowledge. As shown in Table 1, the left column "Geometric Entity Extraction" ensures that the machine can "see" the correct objects, while the right column "Attribute Set Parsing" allows the machine to "understand" the attributes of these objects. The combination of the two constitutes complete prior knowledge.

[0059] Table 1: IFC Model Analysis

[0060] (2) Semantic classification system.

[0061] Based on the semantic parsing framework, this embodiment establishes a refined semantic classification system. This system aims to quantitatively assess the friendliness and risk level of different areas within a tunnel to UAV flights, directly serving the cost calculation for path planning. The system divides the environmental areas into six categories and defines quantified behavioral guidance parameters for each category.

[0062] Flightable passageways (such as above lanes and maintenance access areas) are assigned the lowest navigation cost (1-10) and high priority, making them the preferred areas for path planning. Low-risk areas (such as lighting and signage areas) are allowed to approach but require caution. Medium-risk areas (equipment areas and near ventilation openings) have significantly higher costs (50-200) and require maintaining a large safety distance (1.0-2.0m). High-risk areas (electrical distribution cabinets and fire-fighting facilities) are considered areas that are generally prohibited from entering, and their extremely high costs (200-1000) will form a strong constraint at the algorithm level. Structural surfaces (tunnel walls and road surfaces) are explicitly defined as physical obstacles, and their infinitely high costs ensure absolute avoidance of obstacles. Reserved spaces (emergency parking lanes, etc.) serve as alternative passageways with medium priority.

[0063] As shown in Table 2, this classification system, through a comprehensive definition of three dimensions—navigation cost, safe distance, and priority—transforms complex engineering specifications and safety principles into quantifiable rules that are understandable and executable by machines. It serves as the decision-making cornerstone for achieving safe, compliant, and intelligent autonomous flight.

[0064] It should be noted that the semantic classification system shown in Table 2 is not a manually set empirical value, but rather an automatic conversion of the geometric entities extracted in step (1) into corresponding semantic categories through preset mapping rules. The specific mapping rules are as follows: (a) Direct mapping: For components with clearly defined functions, their semantic categories are directly inherited from the entity type. For example, an entity identified as a "lane" is automatically mapped to "flyable passage" within a certain height range above it.

[0065] (b) Attribute-driven mapping: The semantic category of certain components is determined by the content of their attribute set. For example, an entity identified as "equipment" is mapped to a "medium-risk area" if its attribute set contains the "ventilation vent" identifier; and it is mapped to a "high-risk area" if it contains the "fire protection facility" identifier.

[0066] (c) Spatial reasoning mapping: For non-physical spaces, geometric topological reasoning is used to obtain the mapping. For example, by detecting local recessed areas on the tunnel sidewall and combining their spatial dimensions, these areas are identified as "emergency parking lanes" and mapped to the "reserved space" category.

[0067] Table 2: Semantic Classification System

[0068] This classification system serves as a "traffic rule manual" for path planning algorithms. The quantitative parameters defined within the system are directly transformed into the decision-making basis for the algorithm; "navigation cost" is directly substituted into the evaluation function of path search, guiding the drone to prioritize low-cost areas; "safe distance" generates a repulsive field in the cost field to ensure that the drone maintains a safe distance from risk sources; and "priority level" is used to make the optimal decision among multiple feasible paths.

[0069] Step 13: Semantic grid map generation.

[0070] (1) Three-dimensional rasterization processing.

[0071] 3D rasterization is the process of discretizing a continuous 3D tunnel space into regularly arranged cubic units, which is the foundation for constructing digital environment representations.

[0072] Establish a global coordinate system W for the tunnel, and set the grid resolution to r = 0.1m (this value can be configured according to computing resources and accuracy requirements). Define a 3D grid map:

[0073] in, L, W, H These represent the number of raster cells in the three dimensions of a 3D raster map: vertical, horizontal, and vertical. The calculation method is as follows: L=[L tunnel / r]、W=[W tunnel / r]、H=[H tumnel / r] In the formula L tunnel 、W tunnel 、H tumnelThe actual physical dimensions of the tunnel are represented by its length, width, and height, in meters (m); r is the raster resolution, in meters (m), and in this embodiment, r = 0.1 m; R is the set of real numbers, representing the values ​​stored in the raster cells.

[0074] Each basic unit in a map, namely a grid g ijk It is no longer a simple occupancy identifier, but a data structure containing multi-dimensional attributes, thus forming an information-rich discretized environment model.

[0075] Specifically defined as: g ijk =(occupancy, semantic-class, cost--base, normal-vector, entity-id) Where: occupancy: whether the grid is occupied by an object; semantic-class: its semantic category (e.g., flyable area, risk area, etc.); cost-base: basic access cost; normal-vector: the normal vector of the surface it is located on, used for navigation orientation determination; entity-id: the ID of the associated original BIM component.

[0076] (2) Occupied state judgment algorithm.

[0077] To populate the attributes of each cell in the aforementioned 3D raster map, an occupancy status determination algorithm is executed. This algorithm determines the occupancy status of each raster center point. P ijk =(xi,yj,zk) Perform item-by-item calculations.

[0078] To fill in the attributes of each raster, the judgment algorithm is executed in the following order: Step ①: Occupancy Detection — Determines whether a grid is occupied by a BIM entity, providing the most basic obstacle information for navigation.

[0079] Step 2: Semantic Inheritance — If occupied, the semantic category of the component with the largest proportion is assigned to the raster by calculating the intersection volume of the raster and different components.

[0080] Step 3: Base Cost Calculation — Based on the inherited semantic category, query the preset cost table, and make fine adjustments considering local curvature to generate the base access cost of the raster.

[0081] ① Determining spatial relationships.

[0082] Spatial relationship determination is used to determine whether a grid is occupied by an entity. The logic is: if there exists any component entity E in the BIM model such that the distance from the center point of the grid to the entity is less than half a grid resolution (i.e., ... If the value is 0, the grid is considered occupied (occupancy = 1); otherwise, it is considered free (occupancy = 0). This provides basic physical occupancy information for subsequent calculations.

[0083]

[0084] in, p ijk As the center point of the grid, The shortest distance from a point to an entity. Entities For the set of all BIM entities.

[0085] ② Semantic category inheritance.

[0086] The component with the largest intersection volume with the grid is determined by ray casting, and its semantic category is assigned to the grid. The calculation is as follows: (8) in, Indicates volume, It is a grid cube. To belong to category A collection of entities.

[0087] Specific steps: Iterate through all possible categories c (e.g., walls, pipes, etc.); for the current category c, retrieve all component entities belonging to that category (i.e., Entities). c ); Calculate the grid g ijk The intersection volume with each of these entities is calculated, and then summed. If the raster and an entity do not overlap at all, the intersection volume is 0. The total intersection volume under different categories c is compared, and the category with the largest total intersection volume is taken as the class. ijk .

[0088] ③ Calculation of basic costs.

[0089] After obtaining the semantic category of each grid, it needs to be converted into a quantifiable passage cost so that the path planning algorithm can evaluate the passage quality of different areas.

[0090] Formula (9) is the basic cost calculation. Its function is to obtain the baseline cost by querying the preset cost table according to the semantic category of the raster, and to use the local curvature factor to reward (reduce the cost) the flat area, thereby generating the basic passage cost of the raster.

[0091] This cost not only serves as the attribute storage for discrete grid maps, but also as the core weight for the subsequent construction of the continuous semantic cost field formula (10), and is directly used in online path search and dynamic obstacle avoidance to realize the quantitative mapping of prior knowledge to navigation decisions.

[0092] Based on the fundamental cost of query semantic categories, a local curvature factor is introduced to fine-tune the cost, encouraging drones to fly in flat areas. The calculation is as follows:

[0093] in, f class This is a mapping from semantics to baseline cost. α As curvature weight, curvature ijk For local curvature, it is usually normalized to the [0, 1] interval to align with curvature weights. α Used in conjunction with: In flat areas (such as straight sections of tunnels), curvature ≈ 0, factor The cost remains the same; in curved areas (such as tunnel bends and pipe surfaces), curvature > 0, factor > 1, and the cost increases, thus guiding the drone to choose flat areas for flight as much as possible.

[0094] (3) Semantic cost field generation.

[0095] Through the above steps, a discrete semantic raster map, M-semantic, is obtained. However, in order to perform smooth optimization during path search, a cost function defined in a continuous three-dimensional space is needed. Therefore, the discrete raster attributes are transformed into a continuous semantic cost field, C-semantic(p), using the following formula, which can calculate the passage cost of any point p in the space.

[0096] Define the semantic cost function C semantic :R 3 →R + :

[0097] in, For any point in space p The semantic cost; p Coordinates of a point in three-dimensional space (x,y, z) ; N The number of basis functions is usually related to the total number of grid cells or keypoints; w i For the first i The weights of each basis function are determined by the semantic category corresponding to that basis function, as shown in Table 4; i For the first i One basis function, which can be a Gaussian function, an inverse distance function, or a signed distance function; This is the set of parameters for the basis functions; different basis functions contain different parameters.

[0098] basis functions Different functions can be selected as needed, as shown in Table 3.

[0099] Table 3: Different functions

[0100] The default and universally accepted choice is the Gaussian function, which provides good smoothness and natural risk decay. When safety avoidance is emphasized, the inverse distance function is preferred, as it has an immediate repelling effect on high-risk areas and obstacles. When geometric accuracy is paramount, the signed distance function should be considered, as it provides the strongest support for BIM-driven precision navigation.

[0101] Weight The semantic categories are determined as shown in Table 4.

[0102] Table 4: Semantic Categories

[0103] At this point, the offline processing flow of prior BIM knowledge has been completed, yielding two core outputs: a 3D raster map rich in semantic attributes (M-semantic) and a continuous semantic cost field (C-semantic).

[0104] In the next step, step 2, M-emantic will serve as a global map template to correct the drone's real-time positioning within the tunnel. In the third part, C-semantic will serve as the core cost function of the path planning algorithm, directly generating safe and efficient flight paths. Prior knowledge is thus activated, becoming the "intelligent brain" for the drone to fly autonomously in complex tunnel environments.

[0105] Step 2: Online positioning and environmental perception based on multi-sensor fusion.

[0106] After completing the offline processing and semantic information extraction of the prior BIM model, the system enters the online operation phase. The core task of this part is how to activate this static knowledge for online guidance of UAV perception and positioning.

[0107] The result of step 1—the semantic raster map M-semantic—is regarded as a global intelligent template that never drifts. By deeply fusing it with real-time sensor data, two major online challenges—accumulated odometer error and dynamic environment perception—are solved.

[0108] To achieve this goal, this embodiment constructs an online perception and positioning system driven by semantic features and using prior BIM as a global template.

[0109] Step 21: Tight coupling of multiple sensors and real-time semantic feature extraction.

[0110] The drone's LiDAR, IMU, and visual camera begin working synchronously. After synchronization, denoising, and distortion compensation, the raw data is first used for front-end pose calculation through a tightly coupled laser-inertial odometry (LIO) scheme. The principle is that the IMU provides high-frequency motion prediction, while the LiDAR corrects IMU drift by performing precise geometric matching with the local map using a point-to-surface iterative closest point (ICP) algorithm. The goal of this ICP algorithm is to find an optimal transformation matrix that minimizes the matching error between the current point cloud and the map; its cost function is:

[0111] in, T It is the rigid body transformation matrix to be solved; p i The first in the current frame of the LiDAR point cloud i A point, in the local map, and The nearest point; n i point q i The normal vector of the local plane. N It represents the total number of point pairs participating in the matching.

[0112] The constraints of IMU and LiDAR are optimized in a unified factor graph model, ultimately outputting a high-frequency, low-drift local pose estimate T. k lio However, in long, straight tunnels, no odometry can avoid cumulative drift. Therefore, a semantic feature-driven global localization correction mechanism is introduced.

[0113] Step 22: Semantic feature-driven global localization correction.

[0114] The core of this mechanism lies in fully utilizing the semantic information extracted in the first part. The system extracts features with clear semantic meaning from the LiDAR point cloud of the current frame in real time, such as "lane line edges" and "maintenance lane boundaries," forming a real-time semantic feature point set Fcurrent. Simultaneously, based on the initial pose provided by LIO, the system accurately indexes the prior semantic feature point set FBIM of the same category within the corresponding region of the semantic raster map M-semantic generated in step 1. Subsequently, the Iterative Closest Point (ICP) algorithm matches the two sets to solve for the pose correction matrix Tcorrect. Through the following transformation, the UAV obtains its accurate pose in the BIM global coordinate system:

[0115] in, The precise pose of the UAV in the BIM global coordinate system at time k; The pose correction matrix is ​​obtained by matching the semantic features of the current frame semantic point cloud with the prior BIM semantic map. Let be the local pose estimated by the laser-inertial odometry at time k.

[0116] This method greatly improves the discriminativeness and robustness of matching through semantic filtering, effectively avoiding mismatches caused by feature similarity in tunnels with repetitive geometric structures. Secondly, by matching high-level semantic features, it reduces the dependence on the integrity of environmental geometric details and improves the localization robustness in complex situations such as partial occlusion.

[0117] Step 23: Dynamic map construction guided by BIM prior knowledge.

[0118] Based on the obtained precise pose, the system continuously builds and updates a dynamic environment map M. global The map is initialized as a priori semantic map of the first part. Using T... k global Real-time LiDAR point clouds are integrated into this map. Grids marked as occupied in the M-semantic map are identified as static obstacles; while grids occupied in real-time perception but vacant in the corresponding positions in the prior map are identified as dynamic obstacles. More importantly, the prior semantic classification system can guide perception strategies, such as enhancing monitoring of high-risk areas, thereby constructing a vivid digital twin of the environment that integrates static prior knowledge with dynamic real-time perception.

[0119] Step 3: Path optimization by fusing semantic constraints of tunnel structure with dynamic spatiotemporal prediction After obtaining the accurate global pose and the environmental map integrating dynamic and static information through step 2, this section addresses the core questions of "where to go" and "how to get there." The structural topological constraints (tunnel design axis, safety clearance zone) inherent in the tunnel BIM model are deeply integrated with the motion patterns of dynamic obstacles (movement along the axial direction) into the cost function of path planning. This ensures that while avoiding dynamic obstacles, the UAV always adheres to the "reasonable passage space" stipulated by the tunnel engineering logic and possesses the ability to predictively avoid moving obstacles, rather than merely performing reactive local avoidance.

[0120] Step 31: Construction of a multi-layered fusion semantic cost field.

[0121] Based on the accurate environmental perception results obtained in step 2 above, the purpose of this invention is to make intelligent path decisions. The fundamental basis of path planning is the cost function. The core of this solution is to directly use the continuous semantic cost field C-semantic(p) generated offline in step 1 as the static semantic cost basis for path planning. On this basis, considering the uniqueness of the tunnel environment, the channel passage cost and the spatiotemporal prediction cost of dynamic obstacles are innovatively introduced.

[0122] (1) Static semantic cost function ; The decision-making basis for drone path search is deeply integrated from the very beginning with all the prior semantic rules contained in the BIM model. Therefore, the static semantic cost function The most direct definition is:

[0123] (2) Cost of passage ; Traditional dynamic costing only considers "distance from obstacles," which can cause drones to gradually drift towards the tunnel sidewalls due to local obstacle avoidance in long, straight tunnels, leading to wall-hugging flight or even collision risks. Therefore, this embodiment introduces a passageway cost. Utilizing the design axes and safety clearance zones in the BIM model, the tunnel center area is set as the recommended passageway, ensuring the drone stays in the center of the passageway as much as possible. The cost function is as follows:

[0124] in: For nodes Corresponding spatial point Lateral offset distance from the tunnel design axis; The safe passage width of this section is extracted from the BIM model (half the width of the passable area on both sides of the axis). This is the weighting coefficient for the passage cost.

[0125] This cost term penalizes off-center behavior, and the further off-center the behavior, the greater the penalty (growing quadratically). This is equivalent to creating a "virtual guidance channel" in the center of the tunnel. While avoiding obstacles, the drone is guided by a force that returns it to the center, thus preventing position drift during long-distance flights.

[0126] (3) Cost of spatiotemporal prediction of dynamic obstacles ; To address the strong regularity of movement of dynamic obstacles such as engineering vehicles and inspection personnel along the tunnel axis within the tunnel, this solution does not employ a traditional uniform repulsive force field but instead introduces a spatiotemporal prediction mechanism. Its cost function is as follows: (15) in: For dynamic obstacle cost weights; For nodes Euclidean distance to the nearest dynamic obstacle; The velocity vector of the dynamic obstacle is estimated through pose changes between consecutive frames. This is the unit vector along the tunnel axis. This represents the velocity projection of the obstacle along the tunnel axis; a positive value indicates movement toward the drone. This is an indicator function that takes the value 1 when the condition is met and 0 otherwise. To estimate the collision time, among which This represents the relative distance along the axis. The attenuation coefficient; This is the prediction gain factor, typically ranging from 0.5 to 1.0. It is a very small positive number.

[0127] When a dynamic obstacle moves towards the drone and the estimated collision time is short, the second term increases rapidly, causing the total cost to rise exponentially. This forces the drone to plan its response in advance (such as passing at a wide cross-section or slowing down and waiting). If the obstacle is stationary or far away from the drone, the prediction term approaches zero, and the cost degenerates into a traditional repulsive force field that depends only on distance. This design allows the drone not only to avoid the current obstacle position but also to predict its movement trend, making safer and more efficient obstacle avoidance decisions.

[0128] Step 32: Fusion of path search and cost function.

[0129] Using an improved A* algorithm, a global map M constructed in step 2, which integrates prior static information and real-time dynamic information, is generated. global A three-dimensional path search is performed.

[0130] The core evaluation function of the A* algorithm is... ,in From the starting point to the node The actual cost, From node The heuristic cost to the target point (usually Euclidean distance).

[0131] The innovation of this embodiment lies in reducing the static semantic cost derived from prior BIM knowledge. The passage cost originating from the topological constraints of the tunnel structure And the cost of dynamic obstacles derived from real-time perception and including motion prediction. The triple cost is deeply integrated into the evaluation function, thus forming a brand-new intelligent path evaluation system for overall evaluation. for:

[0132] in, , , These are the weighting coefficients for each cost item, and their values ​​are determined through analytic hierarchy process or offline simulation calibration based on tunnel cross-sectional dimensions, UAV dynamic characteristics, and mission safety level, satisfying the following requirements. .

[0133] The aforementioned triple cost fusion mechanism fundamentally transforms the decision-making mechanism of path search. Traditional path planning algorithms mostly focus only on the geometric shortest path (i.e., minimizing the heuristic function h(n), while this method, by introducing multi-dimensional costs, achieves a comprehensive trade-off between path safety, structural rationality, and dynamic adaptability.

[0134] Under this mechanism, the quality of a path is no longer determined solely by its geometric length: even if a path is a straight line in space ( (The value is small), but if it crosses a "high-risk zone" explicitly defined in the BIM model (leading to static semantic costs), it will cause significant damage. (Extremely high), its overall rating The value will become very large, and thus be automatically discarded by the algorithm; if a path is too close to the tunnel sidewall (tunnel passage cost), it will be discarded. Even if the path is geometrically short, it will still be penalized for deviating from the safe passage zone; if there is an oncoming moving obstacle ahead and the collision time is short (spatiotemporal prediction cost). (If the number of cases rises sharply), the algorithm will choose an alternative path in advance, rather than making an emergency avoidance when the situation is close to the peak.

[0135] This mechanism fundamentally addresses the limitations of existing methods described in the technical background: ① It overcomes the rigidity of the pre-programmed waypoint method: the spatiotemporal prediction cost of dynamic obstacles This enables drones not only to avoid dynamic obstacles in real time, but also to predict their movement trends, achieving proactive and smooth passing decisions.

[0136] ② It eliminates the blindness of pure sensor-based real-time obstacle avoidance methods: static semantic cost It provides globally optimal guidance, avoiding path jitter or wall-hugging issues caused by perceptual ambiguity in long, straight, and repetitive tunnels; tunnel passage cost This further ensures that drones always remain within a reasonable passage area, fundamentally solving the problem of drift accumulation.

[0137] The algorithm ultimately generates a comprehensive optimized path that conforms to prior safety standards globally, follows tunnel passage constraints structurally, can predict and avoid dynamic obstacles locally, and is nearly the shortest overall.

[0138] Step 33: Path post-processing and semantic compliance verification.

[0139] The original path generated by the A* algorithm is usually a polyline formed by connecting the center points of the grid, which is not smooth enough and does not meet the dynamic constraints of the UAV. Therefore, post-processing of the path is required.

[0140] The original path was smoothed using B-spline curve fitting technology. B-spline curves are defined by control points and basis functions, and have the advantages of high local support and continuity, which can generate very smooth trajectories that can be tracked by UAVs.

[0141] The smoothed path still needs to undergo semantic compliance verification: check point by point whether the fitted path enters the "high-risk area" or "structural surface" area marked in the BIM model. If there is a violation point, trigger local replanning with that point as the local target point to ensure that the final output path fully meets the prior semantic safety constraints.

[0142] The fitted smooth path is ultimately output to the UAV flight control system as a series of dense three-dimensional path point sequences, thereby driving the UAV to complete autonomous flight missions safely, smoothly, and efficiently.

[0143] This embodiment deeply mines and utilizes the geometric and semantic information in the BIM model to construct a semantically enhanced digital twin environment for tunnels, and deeply integrates it with the UAV's real-time perception, localization, and planning system. By introducing semantic-driven global positioning correction and dynamic path optimization that fuses prior semantic rules, the UAV ultimately achieves high-precision, high-safety, and highly adaptive autonomous flight inspection in complex tunnel environments.

[0144] Example 2 The purpose of this embodiment is to provide a drone navigation system based on semantic BIM and multi-sensor fusion, including: The semantic extraction module is configured to: perform lightweight processing on the tunnel BIM model, use geometric entity extraction and attribute set parsing to analyze the component data of the tunnel BIM model, and identify semantic entities with navigation significance; The rasterization module is configured to: perform three-dimensional rasterization on the tunnel space, fill the attributes of each cell in the three-dimensional raster map to obtain a three-dimensional raster map rich in semantic attributes, and convert the discrete raster attributes into a continuous semantic cost field. The pose module is configured to use a 3D raster map rich in semantic attributes as a global map template, adjust the UAV's attitude based on the currently collected point cloud data, and obtain the UAV's accurate pose in BIM coordinates. The dynamic environment map building module is configured to: integrate the precise pose of the UAV into the global map template to build a dynamic environment map; The navigation module is configured to: use a semantic cost field as the static semantic cost basis for path planning; construct a spatiotemporal prediction cost for dynamic obstacles based on the movement pattern of dynamic obstacles along the tunnel axis in the dynamic environment map; construct a channel cost function to guide the UAV to stay in the center area of ​​the tunnel based on the tunnel design axis and safety clearance zone in the BIM model; and use a path search algorithm to perform path planning based on the static semantic cost, the spatiotemporal prediction cost of dynamic obstacles, and the channel passage cost to obtain the path navigation result of the UAV.

[0145] In further embodiments, the following is also provided: An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When executed by the processor, the computer instructions perform the method described in Embodiment 1. For brevity, further details are omitted here.

[0146] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0147] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0148] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in Embodiment 1.

[0149] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0150] A computer program product includes a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0151] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in program modules, which execute in a device on a target real or virtual processor to perform the processes / methods described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided among program modules as needed. The machine-executable instructions for the program modules can execute within a local or distributed device. In a distributed device, the program modules can reside in both local and remote storage media.

[0152] The computer program code used to implement the methods of the present invention may be written in one or more programming languages. This computer program code may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the computer or other programmable data processing device, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a computer, partially on a computer, as a stand-alone software package, partially on a computer and partially on a remote computer, or entirely on a remote computer or server.

[0153] In the context of this invention, computer program code or related data may be carried by any suitable carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, and the like. Examples of signals may include electrical, optical, radio, sound, or other forms of propagation signals, such as carrier waves, infrared signals, etc.

[0154] Those skilled in the art will recognize that the units and algorithm steps described in conjunction with the embodiments herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0155] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A UAV navigation method based on semantic BIM and multi-sensor fusion, characterized in that, include: The tunnel BIM model is lightweighted, and the component data of the tunnel BIM model is extracted using geometric entities and attribute set parsing to identify semantic entities with navigation significance. The tunnel space is processed into a three-dimensional raster and the attributes of each cell in the three-dimensional raster map are filled to obtain a three-dimensional raster map rich in semantic attributes. The discrete raster attributes are then converted into a continuous semantic cost field. Using a 3D raster map rich in semantic attributes as a global map template, the attitude of the UAV is adjusted based on the currently collected point cloud data to obtain the precise pose of the UAV in BIM coordinates. The precise pose of the drone is integrated into the global map template to construct a dynamic environment map; Based on the semantic cost field as the static semantic cost basis for path planning, and according to the movement pattern of dynamic obstacles along the tunnel axis in the dynamic environment map, a spatiotemporal prediction cost for dynamic obstacles is constructed. Based on the tunnel design axis and safety clearance zone in the BIM model, a channel cost function is constructed to guide the UAV to stay in the center area of ​​the tunnel. Based on the static semantic cost, the spatiotemporal prediction cost of dynamic obstacles, and the channel passage cost, a path search algorithm is used to perform path planning and obtain the path navigation result of the UAV.

2. The UAV navigation method based on semantic BIM and multi-sensor fusion as described in claim 1, characterized in that, The tunnel BIM model is lightweighted by using geometric entity extraction and attribute set parsing to analyze the component data and identify semantic entities with navigation significance. Specifically: A side-folding algorithm is used to lightweight the tunnel BIM model; Geometric entities are extracted from the three-dimensional space of the tunnel, and the mapping relationship between the geometric entities of the tunnel space and the attribute parsing of the standard format model of industrial basic class is constructed. Semantic information based on geometric entities quantifies the friendliness and risk level of different areas within the tunnel to drone flights.

3. The UAV navigation method based on semantic BIM and multi-sensor fusion as described in claim 1, characterized in that, The tunnel space is processed into a 3D raster, and the attributes of each cell in the 3D raster map are filled in to obtain a 3D raster map rich in semantic attributes. The discrete raster attributes are then converted into a continuous semantic cost field, specifically: The tunnel space is processed into a three-dimensional raster. Calculate each grid center point item by item to determine whether the grid is occupied by BIM entities. Based on the occupied intersection volume, fill the attributes of each unit in the 3D grid map to obtain a 3D grid map rich in semantic attributes. Based on the semantic categories of different regions, discrete raster attributes are transformed into a continuous semantic cost field.

4. The UAV navigation method based on semantic BIM and multi-sensor fusion as described in claim 1, characterized in that, Using a semantically rich 3D raster map as a global map template, the attitude of the UAV is adjusted based on the currently collected point cloud data to obtain the precise pose of the UAV in BIM coordinates, specifically: Extract entity features with clear semantic meaning from the current point cloud data to form a real-time semantic feature point set; Retrieve the set of prior semantic feature points of the same category within the corresponding area of ​​the global map template; The iterative nearest point algorithm is used to match the real-time semantic feature point set with the prior semantic feature point set, solve for the pose correction matrix, and then obtain the accurate pose of the UAV.

5. The UAV navigation method based on semantic BIM and multi-sensor fusion as described in claim 3, characterized in that, For each grid center point, calculate the values ​​to determine if the grid is occupied by a BIM entity. Based on the intersection volume of the occupied areas, fill in the attributes of each cell in the 3D grid map to obtain a 3D grid map rich in semantic attributes. Specifically: If there exists any component entity in the BIM model such that the distance from the center point of the grid to the entity is less than half the grid resolution, then the grid is considered occupied; otherwise, it is considered idle. If a grid is occupied, calculate the intersection volume between the grid and different components, and assign the semantic category of the component with the largest proportion to the corresponding grid. After obtaining the semantic category of each raster, the baseline cost is obtained by querying the preset cost table, and the flat area is rewarded by the local curvature factor to generate the basic passage cost of the raster, thus obtaining a three-dimensional raster map rich in semantic attributes.

6. The UAV navigation method based on semantic BIM and multi-sensor fusion as described in claim 1, characterized in that, Channel passage cost Cost of spatiotemporal prediction of dynamic obstacles Specifically: in: For nodes Corresponding spatial point Lateral offset distance from the tunnel design axis; The safe passage width for this section is extracted from the BIM model; This is the weighting coefficient for the passage cost; For dynamic obstacle cost weights; For nodes Corresponding spatial point Euclidean distance to the nearest dynamic obstacle; The velocity vector of the dynamic obstacle; This is the unit vector along the tunnel axis. This represents the velocity projection of the obstacle along the tunnel axis; a positive value indicates movement toward the drone. This is an indicator function that takes the value 1 when the condition is met and 0 otherwise. To estimate the collision time; , This represents the relative distance along the axis. The attenuation coefficient; For the prediction gain coefficient, It is a very small positive number.

7. A drone navigation system based on semantic BIM and multi-sensor fusion, characterized in that, include: The semantic extraction module is configured to: perform lightweight processing on the tunnel BIM model, use geometric entity extraction and attribute set parsing to analyze the component data of the tunnel BIM model, and identify semantic entities with navigation significance; The rasterization module is configured to: perform three-dimensional rasterization on the tunnel space, fill the attributes of each cell in the three-dimensional raster map to obtain a three-dimensional raster map rich in semantic attributes, and convert the discrete raster attributes into a continuous semantic cost field. The pose module is configured to use a 3D raster map rich in semantic attributes as a global map template, adjust the UAV's attitude based on the currently collected point cloud data, and obtain the UAV's accurate pose in BIM coordinates. The dynamic environment map building module is configured to: integrate the precise pose of the UAV into the global map template to build a dynamic environment map; The navigation module is configured to: use a semantic cost field as the static semantic cost basis for path planning; construct a spatiotemporal prediction cost for dynamic obstacles based on the movement pattern of dynamic obstacles along the tunnel axis in the dynamic environment map; construct a channel cost function to guide the UAV to stay in the center area of ​​the tunnel based on the tunnel design axis and safety clearance zone in the BIM model; and use a path search algorithm to perform path planning based on the static semantic cost, the spatiotemporal prediction cost of dynamic obstacles, and the channel passage cost to obtain the path navigation result of the UAV.

8. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, perform the method described in any one of claims 1-6.

10. A computer program product, characterized in that, Includes a computer program, which, when executed by a processor, implements the method described in any one of claims 1-6.

Images