An unmanned aerial vehicle autonomous cruising method and system for complex construction environment

Abstract

The application discloses a kind of unmanned aerial vehicle autonomous cruise method and system for construction complex environment, it is related to unmanned aerial vehicle autonomous cruise technical field, its method includes data acquisition, offline mapping, regular training, line skeleton extraction, anchor point screening, global planning, repositioning, dynamic obstacle avoidance and re-planning, its system includes data acquisition module, three-dimensional Gaussian mapping module, topological deconstruction module, planning module, positioning fusion module, obstacle avoidance re-planning module and update module;The application is inhibited by anisotropic regular constraint three-dimensional Gaussian field, inhibits the elongation of Gauss body, is deconstructed into lightweight line skeleton and visual anchor point, realizes high-fidelity map compression, and combines point-line fusion repositioning and factor graph multi-source fusion, output high-precision pose under no GPS, in addition, cooperate local ESDF dynamic obstacle avoidance and track regression, form robust closed loop, significantly improve the safety and reliability of unmanned aerial vehicle autonomous cruise in construction complex environment.

Description

Technical Field

[0001] This invention relates to the field of autonomous drone navigation technology, and in particular to an autonomous drone navigation method and system for complex construction environments. Background Technology

[0002] Construction environments (such as construction sites, steel structure workshops, bridge construction areas, and underground utility tunnels) present typical complexities and challenges. First, there are numerous slender structures (scaffolding, steel beams, cables, and reinforcing bars), which traditional point clouds or mesh maps struggle to represent efficiently. Second, surfaces are often low-texture or repetitive (cement walls, undecorated steel frames), making pure visual SLAM prone to drift or even loss of lock due to feature loss or ambiguous matching. Third, severe environmental occlusion and dynamic obstacles (personnel, vehicles, tower cranes, transported materials) necessitate real-time obstacle avoidance capabilities for UAVs. Finally, indoor or steel structure areas often lack GPS signals, and fusion navigation relying on IMU and ranging faces multipath effects and integral drift.

[0003] Existing technologies each have shortcomings in addressing the aforementioned challenges. Pure visual SLAM (such as ORB-SLAM and VINS-Mono) exhibits poor robustness in low-texture / repetitive texture environments; LiDAR SLAM suffers from severe point cloud noise under steel mirror reflections, and its equipment weight and cost are relatively high; wireless ranging (UWB / RSSI) is prone to measurement jumps in complex electromagnetic environments. In recent years, 3D Gaussian fields (3DGS) have achieved high-fidelity scene reconstruction by optimizing anisotropic covariance and density control, but its output is a set of millions of Gaussian volumes, lacking a topological structure directly applicable to collision lookup and path planning, and therefore cannot be directly applied to UAV navigation. Therefore, this invention proposes an autonomous UAV navigation method and system for complex construction environments to solve the problems existing in the prior art. Summary of the Invention

[0004] To address the aforementioned problems, the present invention aims to propose an autonomous cruise method and system for unmanned aerial vehicles (UAVs) in complex construction environments. This method involves offline construction of a three-dimensional Gaussian field prior map constrained by anisotropic regularization, and inversely deconstructing the Gaussian field into a linear skeleton topology map and sparse visual anchor points. By combining point-line fusion relocation, factor graph multi-source fusion, and local ESDF dynamic obstacle avoidance, a highly robust autonomous cruise and map compression can be achieved in complex construction environments.

[0005] To achieve the objectives of this invention, the invention is implemented through the following technical solution: an autonomous cruise method for unmanned aerial vehicles (UAVs) in complex construction environments, comprising the following steps: S1, Data Acquisition The drone is controlled to perform coverage flights over the target construction area, collecting multi-view image sequences and synchronous IMU data for subsequent offline mapping and anchor point selection. This step is fundamental to data acquisition; multi-view images provide observations for subsequent 3D reconstruction, and IMU provides prior information for motion constraints. The collected data will be used throughout the entire offline mapping and anchor point selection process. S2, Offline Mapping Based on multi-view image sequences and IMU data, a three-dimensional Gaussian field prior map of the target scene is constructed using a differentiable rendering three-dimensional Gaussian mapping framework. This three-dimensional Gaussian field is composed of multiple Gaussian volumes, each of which has a covariance matrix, opacity, and color attributes. The three-dimensional Gaussian mapping framework contains an optimizable objective function. This step transforms the acquired data into a high-fidelity three-dimensional Gaussian field representation. Each Gaussian volume describes its shape through covariance and its appearance through opacity and color, providing raw geometric and appearance information for subsequent structure extraction. S3, Regular Expression Training An anisotropic regularization penalty term is added to the objective function to suppress the infinite elongation of a single Gaussian body along the principal axis, and to make the slender structure fit with a distribution of several scale-controlled Gaussian bodies along the line, so as to obtain a three-dimensional Gaussian field after regularization constraint. This step prevents the Gaussian body from becoming too slender through regularization constraint, and forces the slender object (such as scaffolding steel pipe) to be fitted with multiple short and thick Gaussian bodies in series, thereby providing a stable geometric basis for subsequent skeleton extraction. S4, Wire Skeleton Extraction For each Gaussian body in the regularized Gaussian field, the covariance matrix eigenvalue is decomposed. The linear factor is calculated based on the eigenvalues ​​and linear Gaussian bodies are identified. The selected linear Gaussian bodies are clustered and fitted with the central axis to generate a linear skeleton topology map. This step transforms the dense Gaussian field into a lightweight linear skeleton map. Linear structures are identified through eigenvalue decomposition, and after clustering, the central axis is fitted and topological connections are constructed to obtain a navigation map that can be directly used for global planning. S5, Anchor Point Filtering Based on the acquired multi-view image sequence and the regularized three-dimensional Gaussian field, the Gaussian volume is back-projected onto the multi-view image sequence to calculate the photometric variance. Gaussian volumes with photometric variance below a set threshold are selected as visual anchors to form a visual anchor set. The color and spherical harmonic function properties of the visual anchors are preserved. This step selects a small number of Gaussian volumes with stable appearance and distinct features from the Gaussian field as visual anchors. These anchors are preferentially distributed in areas with rich texture, providing a high-discrimination prior feature library for online relocalization. S6, Global Planning A traversable graph is constructed based on a wire skeleton topology graph. A graph search algorithm is used to generate global paths and optimize the global paths to generate executable flight paths for UAVs. This step uses a lightweight topology graph for global path planning. After obtaining discrete path points through graph search, trajectory smoothing and dynamic constraint optimization are performed to output a smooth flight path that the UAV can fly directly. S7, Relocation During the drone's cruise, onboard camera images, IMU data, and wireless ranging data are acquired in real time. Based on the visual anchor point set and the skeleton topology graph, feature points and edge lines are extracted from the real-time images. Point-line fusion matching is performed to obtain the pose correction. A factor graph framework is used to fuse and optimize the visual relocalization results, IMU pre-integration data, and wireless ranging data to output a high-precision pose. This step uses both visual anchor points and skeletons for relocalization. Point features provide texture matching, and line features provide structural constraints. IMU and wireless ranging (RSSI) are fused in the backend factor graph, enabling the output of a highly robust pose even without GPS. S8, Dynamic Obstacle Avoidance and Replanning During the drone's cruise, the surrounding environment is perceived in real time by onboard sensors, and a local Euclidean symbolic distance field or occupied grid map is constructed. The executable trajectory is used as a global reference, and the high-precision pose is used as the current state. When a dynamic obstacle is detected, local replanning is triggered to generate a detour trajectory. After detouring around the obstacle, the drone returns to the executable trajectory based on the visual anchor point set or the skeleton topology map. This step realizes dynamic obstacle avoidance and trajectory return. The local map perceives obstacles in real time, and local replanning is used to detour when an obstacle is encountered. After detouring around the obstacle, the global trajectory is quickly restored using the anchor points or skeleton in the prior map, forming a closed loop.

[0006] A further improvement is made in S3, where the anisotropic regularization penalty term is expressed as: in, l i1 and l i3 The first i The maximum and minimum eigenvalues ​​of the eigenvalue decomposition of the Gaussian covariance matrix, where ε is the numerical stability constant. r max The maximum anisotropy ratio is preset; when the anisotropy ratio of a single Gaussian body exceeds... r max At that time, the penalty term produces a non-zero loss, which is suppressed by gradient descent. l i1 and rise l i3The paper presents the specific mathematical form of the regularization term and its gradient properties. By limiting the ratio of the largest eigenvalue to the smallest eigenvalue (condition number), it suppresses the infinite elongation behavior of the Gaussian body from an optimization perspective.

[0007] A further improvement lies in the following: In step S4, the specific steps for clustering and fitting the central axis of the selected linear Gaussian bodies are as follows: Clustering is performed based on the spatial proximity and principal axis orientation of linear Gaussian bodies to obtain several linear clusters; For each Gaussian body in the linear cluster, the weights are calculated based on its opacity, linearity, and cylindricity. Weighted least squares fitting is then used to obtain the direction vector and position point of the initial central axis. The initial center axis is iteratively reweighted least squares optimized by using Huber weights to suppress the influence of outlier Gaussian bodies, resulting in the optimized center axis. The two endpoints of the axis are determined based on the projection range of the Gaussian body onto the optimized central axis, and a line segment representation is generated. Detect the endpoint proximity and intersection relationships between different line segments, construct topological connections, and form a line skeleton topology graph; The fitting process from linear Gaussian clusters to the central axis has been refined, including weighted least squares to obtain the initial axis, IRLS robust optimization to remove outliers, endpoint determination and topological connection, to ensure the accuracy and robustness of skeleton extraction.

[0008] A further improvement is made in S5, where the specific steps for forming the visual anchor point set are as follows: Using nodes in the line skeleton topology map generated in S4 as spatial prior locations, the nodes include line segment intersections and line segment endpoints. Within the local search neighborhood centered on each node, Gaussian bodies in the 3D Gaussian field prior map are back-projected onto the multi-view image sequence acquired in S1. The photometric variance of each Gaussian body under different views is calculated. Gaussian bodies with photometric variance below a preset threshold are extracted as candidate anchor points, and the color attributes and spherical harmonic function parameters of the candidate anchor points are retained to form a visual anchor point set. Gaussian bodies not in the node neighborhood are not retained as visual anchor points. This clarifies the spatial guidance strategy for anchor point selection—prioritizing the selection of Gaussian bodies with small photometric variance in the neighborhood of nodes (intersections / endpoints) in the line skeleton map, so that anchor points are naturally distributed in visually rich areas, improving the relocation success rate.

[0009] An autonomous navigation system for unmanned aerial vehicles (UAVs) designed for complex construction environments includes a data acquisition module, a 3D Gaussian mapping module, a topology deconstruction module, a planning module, a positioning fusion module, an obstacle avoidance and replanning module, and an update module, wherein: The data acquisition module is used to control the UAV to acquire multi-view image sequences and synchronized IMU data, and to acquire airborne sensor data in real time. The 3D Gaussian mapping module is connected to the data acquisition module. It is used to construct a 3D Gaussian field prior map based on the acquired multi-view image sequences and IMU data. An anisotropic regularization penalty term is added to the mapping optimization objective, and the 3D Gaussian field after regularization constraint is output. The topology deconstruction module, connected to the 3D Gaussian mapping module, is used to inversely deconstruct the 3D Gaussian field and output a line skeleton topology map and a set of visual anchor points. The topology deconstruction module includes a covariance decomposition unit, a line discrimination unit, a filtering unit, a clustering unit, and a line segment weighted fitting unit. The planning module, connected to the topology deconstruction module, is used for global path planning and trajectory optimization based on the skeleton topology graph, generating executable tracks. The positioning fusion module is connected to the data acquisition module and the topology deconstruction module respectively. It is used to perform point-line fusion repositioning based on the visual anchor point set and the wire skeleton topology map. It also uses factor graph optimization to fuse the visual repositioning results, IMU pre-integration data and wireless ranging data to output high-precision pose. The positioning fusion module uses a robust kernel function for the wireless ranging factor to suppress multipath jump outpoints. The obstacle avoidance and replanning module connects the planning module and the positioning fusion module. It is used to construct a local Euclidean symbolic distance field or occupy grid map based on airborne sensors. It uses the executable trajectory as a global reference and high-precision pose as the current state. When a dynamic obstacle is detected, it triggers local replanning and generates a detour trajectory. After detour, it guides the UAV back to the global trajectory generated by the planning module. The update module connects the obstacle avoidance replanning module and the 3D Gaussian mapping module. It is used to write the stable new obstacle information from multiple consistent observations back to the 3D Gaussian field prior map to achieve incremental map updates.

[0010] A further improvement is that the factor graph optimization structure in the positioning fusion module includes the following factor nodes: The visual repositioning factor, whose residuals are constructed based on the reprojection error of the visual anchor point and the geometric constraints of the line segment; IMU pre-integration factor, used to constrain relative motion between adjacent keyframes; The wireless ranging factor, whose residual is constructed based on the difference between the measured distance and the predicted distance between the UAV and the known location anchor point; The robust kernel function is either the Huber kernel or the Cauchy kernel, which is used to weight the residuals of the wireless ranging factor to reduce the impact of measurement jumps caused by multipath effects on the optimization results. The specific node types of the factor graph were refined, and a robust kernel was explicitly adopted for the wireless ranging factor to suppress ranging jumps caused by multipath effects in the construction environment and improve the stability of fusion positioning.

[0011] A further improvement is made in that the obstacle avoidance replanning module includes a local map construction unit and a replanning triggering unit. The local map construction unit constructs a local Euclidean symbolic distance field centered on the current position of the UAV based on a depth map generated in real time by an onboard depth sensor or binocular camera. This distance field stores the signed distances to the nearest obstacle in a spatial grid. When the replanning triggering unit detects that the signed distance within a predetermined range in front of the UAV in the local Euclidean symbolic distance field is less than a safety threshold, it triggers local replanning. The local replanning uses A... The algorithm generates a detour path on the obstacle expansion layer of the local Euclidean symbolic distance field, and then generates a smooth detour trajectory through Minimum-Snap optimization. After the detour is completed, the obstacle avoidance replanning module matches the current pose of the UAV with the skeleton topology map output by the topology deconstruction module, and guides the UAV back to the global track of the planning module. This refines the specific implementation process of local obstacle avoidance: the obstacle distance field is expressed with ESDF, A generates the detour path, Minimum-Snap smooths the trajectory, and after the obstacle is bypassed, the global track is re-anchored using the skeleton topology map.

[0012] Further improvements are made in the following aspects: The update module includes a consistency judgment unit and a map fusion unit. The consistency judgment unit is used to continuously track the spatial position stability of the obstacle in multiple consecutive cruise cycles or multi-frame sensor observations after the obstacle avoidance and replanning module detects a new obstacle. When the number of times the obstacle at the same location is consistently detected exceeds a threshold within a preset time window, it is determined to be a stable new obstacle. The map fusion unit is used to convert the spatial occupancy information of the stable new obstacle into a new Gaussian volume or to correct the opacity of the original Gaussian volume, and send the updated map data to the 3D Gaussian mapping module to update the 3D Gaussian field prior map and the wire skeleton topology map output by the topology deconstruction module. This realizes online incremental update of the map: by filtering temporary dynamic objects through multi-frame consistency judgment, only stable obstacles are fused into the prior map to ensure the long-term availability and security of the map.

[0013] The beneficial effects of this invention are as follows: Addressing the challenges of complex construction environments, such as the absence of GPS signals, low-texture repetitive textures, frequent dynamic obstacles, and dense slender structures, this invention constructs an offline 3D Gaussian field prior map constrained by anisotropic regularization. This effectively suppresses the infinite elongation of the Gaussian volume, providing a stable geometric foundation for slender structures. Furthermore, it deconstructs the map into a linear skeleton topology map and sparse visual anchor points, achieving lightweight map compression while maintaining high-fidelity scene representation. This significantly reduces online computation and storage burden. Simultaneously, in the online phase, global planning is performed based on the linear skeleton, integrating real-time relocalization of visual anchor points and line features, as well as multi-source fusion of factor maps (visual / IMU / wireless ranging). Even in GPS-free environments, it can still output high-precision robust pose. Combined with local ESDF dynamic obstacle avoidance and rapid regression to the global trajectory using the prior map after obstacle avoidance, a complete closed loop of planning-localization-obstacle avoidance-update is formed, achieving highly robust and safe autonomous navigation of UAVs in complex construction environments. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the autonomous cruise method for unmanned aerial vehicles (UAVs) in complex construction environments according to the present invention. Figure 2 This is a schematic diagram of the framework structure of the UAV autonomous cruise system for complex construction environments according to the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] It should be noted that the technical means not described in detail in the following embodiments are all conventional means in the art, are not the key points of the invention, and will not be elaborated upon.

[0017] Example 1 See Figure 1 This embodiment provides an autonomous drone navigation method for complex construction environments, applied to a typical construction scenario (high-rise building steel structure construction area, including dense scaffolding, multiple tower cranes, and no GPS signal). The specific implementation steps are as follows: S1, Data Acquisition A drone equipped with an RGB-D camera (such as the Intel RealSense D455) and a six-axis IMU was used to perform a tic-tac-toe coverage flight over the target area. The flight speed was 0.5–2 m / s, the frame rate was 30 fps, and the image resolution was 640×480. The acquired multi-view image sequences and synchronous IMU data were stored at a ground station for offline mapping.

[0018] S2, Offline Mapping Offline mapping is performed using a differentiable rendering-based 3D Gaussian mapping framework (such as SplaTAM-like methods). Parameter settings: initial learning rate of 0.01, 30k iterations, with density control every 2k iterations. An initial Gaussian volume is generated by backprojection from the first frame's depth map. RGB-D predictions are calculated using differentiable rendering to optimize color and depth consistency loss, and backpropagation optimizes the covariance matrix Σ of each Gaussian volume. i Opacity α i The spherical harmonic coefficients. The covariance matrix of each Gaussian body is a 3×3 positive definite matrix, which determines the shape and orientation of the Gaussian body. The final result is a dense three-dimensional Gaussian field prior map containing approximately 500,000 to 2 million Gaussian bodies.

[0019] S3, Regular Expression Training An anisotropic regularization penalty term is added to the objective function of the graph construction optimization. For each Gaussian body, the covariance matrix Σ... i Perform eigenvalue decomposition to obtain eigenvalues ​​λ. i1 ≥λ i2 ≥λ i3 >0. Define anisotropy ratio. , where ε is the numerical stability constant (taken as 10). -6 Preset maximum anisotropy ratio ρ max =15 (can be adjusted within the range of 5-30 depending on the scenario). The anisotropic regularization penalty term is expressed as: in, l i1 and l i3 The first i The maximum and minimum eigenvalues ​​of the eigenvalue decomposition of the Gaussian covariance matrix, when ρ i >ρ max At that time, the regularization term produces a non-zero loss, and gradient descent suppresses λ. i1 , lift λ i3 This prevents the Gaussian body from stretching infinitely, forcing the slender structure to be fitted by multiple short and thick Gaussian bodies connected in series along the line.

[0020] S4, Wire Skeleton Extraction Perform eigenvalue decomposition on each Gaussian body to obtain three eigenvalues ​​λ.i1 , λ i2 , λ i3 (satisfies λ) i1 ≥λ i2 ≥λ i3 >0). Calculate the linear factor L. i and cylindricity C i The formula is: The criteria for screening linear Gaussian bodies are: L i >0.6 and λ i2 / λ i3 <1.5, while filtering opacity α i Gaussian bodies with a density <0.1 and locally low density are selected. These linear Gaussian bodies are then clustered: Euclidean clustering (distance threshold 0.2m) combined with a principal axis angle less than 30°, resulting in several linear clusters. For each linear cluster, the weight of each Gaussian body is calculated using the following formula: In this embodiment, γ=2, δ=1, θ=1, and η=0.5. Weighted least squares fitting is used to fit the initial central axis. Calculate the weighted centroid Weighted covariance matrix superscript This represents the transpose of the matrix, with the eigenvector u corresponding to the largest eigenvalue of S as the axis direction. Two rounds of iterative reweighted least squares (IRLS) are performed: the orthogonal distance (i.e., vector p) from each Gaussian body position to the current axis is calculated. i -c is the component length perpendicular to u), and the Huber weighting function is used to reduce the weights of outliers Gaussian bodies with large distances, updating the weights w. i Then recalculate c, S, and u. Finally, determine the endpoints of the line segments: calculate the projection of each Gaussian body onto the axis. Take the points corresponding to the minimum and maximum values ​​of the projection as the two endpoints of the line segment. , For all line segments, the distance between their endpoints is checked. If the distance is less than 0.1m, they are connected to form a line skeleton topology graph G=(V,E), where V is the set of nodes (including the endpoints and intersections of the line segments), and E is the set of edges (i.e., each line segment). The data volume of this topology graph is only 1 / 1000 to 1 / 10000 of the original Gaussian field.

[0021] S5, Anchor Point Filtering Centered on the nodes (intersections and endpoints of line segments) of the line skeleton topology graph G, a local search radius of 0.5m is established. Utilizing the projection mapping relationship between Gaussian bodies and image pixels established during the offline mapping phase (through a differentiable camera model), the Gaussian bodies within the search area are back-projected onto the multi-view image sequence S1. The projected colors at each viewpoint are obtained, and the color variance (luminance variance) is calculated. Gaussian bodies with variances below a threshold (e.g., 0.05) are selected as visual anchor points, retaining their position, spherical harmonic coefficients, and opacity. Those not within the node's neighborhood are not used as anchor points. The number of anchor points is approximately 0.5% to 2% of the total number of Gaussian bodies.

[0022] S6, Global Planning A walkable graph is constructed based on the wire skeleton topology graph G: nodes are walkable points, edges are walkable paths, and edges close to obstacles are deleted based on the 0.5m diameter expansion of the UAV. A... The algorithm searches for a global path, then uses Minimum-Snap trajectory optimization to generate a smooth, executable trajectory, which is then output to the flight controller.

[0023] S7, Relocation During online cruise, the onboard computer acquires real-time camera images (30fps), IMU (200Hz), and UWB / RSSI ranging (10Hz). ORB point features and LSD line features are extracted from each frame. The current frame features are then matched 2D-3D with the visual anchor point set: anchor points are projected and matched with the nearest neighbors of the feature point descriptors; line skeleton segments are projected and geometrically matched with the extracted line segments (angle difference <10°, distance difference <5 pixels). The initial pose is obtained through PnP. An optimization problem is constructed on the factor graph backend (e.g., GTSAM), including: ①Visual repositioning factor (reprojection error + line segment geometric constraints); ②IMU pre-integration factor; ③ Wireless ranging factor: The residual is based on the difference between the measured distance and the Euclidean distance, and the Huber kernel (threshold 0.5m) is used to suppress multipath jumps.

[0024] Optimize output for high-precision pose (30Hz, position error <0.1m, attitude error <2°).

[0025] S8, Dynamic Obstacle Avoidance and Replanning An airborne depth sensor generates a local depth map (10m×10m×5m) in real time, constructing a local Euclidean symbolic distance field (ESDF) centered on the UAV's current location. The voxel resolution is 0.1m, and the symbolic distances are stored. A safe distance of 0.5m is maintained; replanning is triggered when the ESDF value within 2m in front of the UAV falls below the safe threshold. An algorithm is run on the ESDF obstacle expansion layer (+0.3m). A detour path is generated, with the local target point being the nearest unblocked waypoint on the global track. A smooth detour trajectory is then generated through Minimum-Snap optimization. After detour, the pose is aligned with the global track using the current observation anchor point or line segment matching, and the original track is returned to. All actions are constrained by geofencing (boundary 1m outward) and speed limits (≤3m / s).

[0026] Example 2 See Figure 2 This embodiment provides an autonomous drone navigation system for complex construction environments, used in conjunction with the autonomous drone navigation method for complex construction environments provided in Embodiment 1. The system is deployed on a six-rotor industrial drone (such as the DJI M350), with an onboard computing unit of NVIDIA Jetson Orin NX (16GB RAM). Sensor suite includes: a binocular camera (1280×720, 30fps), a six-axis IMU (1000Hz), a UWB module (10Hz), and optional LiDAR. The ground station is equipped with a high-performance GPU (such as an RTX 4090). The system consists of seven modules: a data acquisition module, a 3D Gaussian mapping module, a topology deconstruction module, a planning module, a localization fusion module, an obstacle avoidance and replanning module, and an update module. Data acquisition module: In the offline stage, it acquires and stores multi-view image sequences and IMU data; in the online stage, it sends data to the positioning fusion and obstacle avoidance replanning module at 30Hz (image), 200Hz (IMU), and 10Hz (range).

[0027] 3D Gaussian mapping module (ground station): Reads historical data, constructs a 3D Gaussian field constrained by anisotropic regularization using methods S2 and S3 in Example 1, and outputs a Gaussian field file.

[0028] Topology deconstruction module (ground station or airborne): Reads the Gaussian field, sequentially performs covariance decomposition, linear discrimination, filtering, clustering, weighted least squares + IRLS fitting, generates a line skeleton topology map and a set of visual anchor points, and stores them serially (<10MB).

[0029] Planning module (airborne): Reads the wire skeleton topology map, constructs a walkable map, A +Minimum-Snap generates a global track and sends waypoints to the flight controller.

[0030] Positioning fusion module (airborne): Subscribes to camera, IMU, UWB data and wire skeleton / anchor point set, and fuses visual repositioning factor, IMU pre-integration factor and wireless ranging factor (Huber kernel, threshold 0.5m) based on factor graph (GTSAM) to output high-precision pose at 20Hz.

[0031] Obstacle Avoidance Replanning Module (Airborne): The local map construction unit constructs an ESDF (10m×10m×5m, 0.1m resolution) in real time based on a binocular depth map; the replanning trigger unit monitors the ESDF value of a 2m sector in front, triggering replanning when it is less than 0.5m (A +Minimum-Snap). After the detour is completed, the point cloud and line skeleton are matched via ICP to correct the pose and return to the global track.

[0032] Update module (airborne, communicating with ground station): Consistency judgment unit tracks the spatial stability of newly added obstacles, and determines them as permanent obstacles after they stabilize; map fusion unit converts them into new Gaussian volume, writes them into the three-dimensional Gaussian field, and triggers the re-extraction of local line skeletons and anchor points, which are then synchronized to the airborne module.

[0033] During system operation: Before initially entering the construction area, the UAV performs a coverage flight, the ground station maps and generates a skeleton map and anchor points, which are then uploaded to the airborne system. After the cruise mission begins, the planning module generates a global flight path, the positioning fusion module continuously outputs high-precision pose, and the obstacle avoidance and replanning module monitors in real time. When there are no dynamic obstacles, the UAV flies along the flight path; when an obstacle is encountered, the obstacle avoidance and replanning module takes over, generates a detour trajectory, and returns to the global flight path after bypassing the obstacle. The update module monitors for newly added obstacles in the background and triggers a local map update. The system forms a closed loop of "mapping-planning-positioning-obstacle avoidance-updating".

[0034] The system has been implemented at a high-rise steel structure building construction site (area 5000m²). 2 (Dense scaffolding, multiple tower cranes working simultaneously) Verification: 120 minutes of flight time, no collisions occurred, average repositioning error was 0.08m, obstacle avoidance success rate was 100%.

[0035] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for autonomous navigation of unmanned aerial vehicles (UAVs) in complex construction environments, characterized in that: Includes the following steps: S1. Control the drone to perform coverage flight over the target construction area, and collect multi-view image sequences and synchronized IMU data; S2. Based on multi-view image sequences and IMU data, a three-dimensional Gaussian field prior map of the target scene is constructed using a differentiable three-dimensional Gaussian mapping framework. The three-dimensional Gaussian field is composed of several Gaussian volumes, each of which has a covariance matrix, opacity, and color attributes. The three-dimensional Gaussian mapping framework contains an optimizable objective function. S3. Add an anisotropic regularization penalty term to the objective function to suppress the infinite elongation of a single Gaussian body along the principal axis, and make the slender structure fit the distribution along the line of several scale-controlled Gaussian bodies to obtain the three-dimensional Gaussian field after regularization constraint. S4. Perform covariance matrix eigenvalue decomposition on each Gaussian body in the three-dimensional Gaussian field after regularization constraint, calculate the linear factor based on the eigenvalue and identify linear Gaussian bodies, and perform clustering and fitting of the selected linear Gaussian bodies with the central axis to generate a linear skeleton topology map. S5. Based on the acquired multi-view image sequence and the three-dimensional Gaussian field after regularization constraint, the Gaussian volume is back-projected onto the multi-view image sequence to calculate the photometric variance. Gaussian volumes with photometric variance below a set threshold are selected as visual anchor points to form a set of visual anchor points. The color and spherical harmonic function properties of the visual anchor points are retained. S6. Construct a traversable graph based on the wire skeleton topology graph, use a graph search algorithm to generate a global path, and optimize the trajectory of the global path to generate an executable flight path for the UAV. S7. During the drone's cruise, the onboard camera images, IMU data, and wireless ranging data are acquired in real time. Based on the visual anchor point set and the skeleton topology map, feature points and edge lines are extracted from the real-time images. Point-line fusion matching is performed to obtain the pose correction amount. The factor graph framework is used to fuse and optimize the visual relocalization results, IMU pre-integration data, and wireless ranging data to output a high-precision pose. S8. During the drone's cruise, the drone perceives the surrounding environment in real time based on onboard sensors, constructs a local Euclidean symbolic distance field or occupied grid map, uses the executable trajectory as a global reference, and uses high-precision pose as the current state. When a dynamic obstacle is detected, local replanning is triggered to generate a detour trajectory, and after detouring around the obstacle, the drone reverts to the executable trajectory based on the visual anchor point set or the wire skeleton topology map.

2. The method for autonomous drone navigation in complex construction environments according to claim 1, characterized in that: In S3, the anisotropic regularization penalty term is expressed as: in, λ i1 and λ i3 The first i The maximum and minimum eigenvalues ​​of the eigenvalue decomposition of the Gaussian covariance matrix, where ε is the numerical stability constant. ρ max The maximum anisotropy ratio is preset; when the anisotropy ratio of a single Gaussian body exceeds... ρ max At that time, the penalty term produces a non-zero loss, which is suppressed by gradient descent. λ i1 and rise λ i3 .

3. The method for autonomous drone navigation in complex construction environments according to claim 1, characterized in that: In step S4, the specific steps for clustering and fitting the central axis of the selected linear Gaussian bodies are as follows: clustering is performed based on the spatial proximity and consistency of the principal axis direction of the linear Gaussian bodies to obtain several linear clusters. For each Gaussian body in the linear cluster, the weight is calculated according to its opacity, linear intensity, and cylindricity. The direction vector and position point of the initial central axis are obtained by weighted least squares fitting. The initial central axis is iteratively re-weighted least squares optimized, and the influence of outlier Gaussian bodies is suppressed by Huber weights to obtain the optimized central axis. The two endpoints of the axis are determined according to the projection range of the Gaussian body on the optimized central axis, and line segment representation is generated. The endpoint proximity and intersection relationships between different line segments are detected, and topological connections are constructed to form a line skeleton topology map.

4. The method for autonomous drone navigation in complex construction environments according to claim 1, characterized in that: In step S5, the specific steps for forming the visual anchor point set are as follows: taking the nodes in the line skeleton topology map as spatial prior positions, the nodes include line segment intersections and line segment endpoints. In the local search neighborhood centered on each node, the Gaussian volume in the three-dimensional Gaussian field prior map is back-projected to the multi-view image sequence. The photometric variance of each Gaussian volume under different views is calculated. Gaussian volumes with photometric variance lower than a preset threshold are extracted as candidate anchor points, and the color attributes and spherical harmonic function parameters of the candidate anchor points are retained to form the visual anchor point set. Gaussian volumes that are not in the neighborhood of a node are not retained as visual anchor points.

5. An autonomous drone cruise system applied to the autonomous drone cruise method for complex construction environments as described in any one of claims 1-4, characterized in that, include: The data acquisition module is used to control the UAV to acquire multi-view image sequences and synchronized IMU data, and to acquire airborne sensor data in real time. The 3D Gaussian mapping module is used to construct a 3D Gaussian field prior map based on the acquired multi-view image sequence and IMU data, and adds an anisotropic regularization penalty term to the mapping optimization objective to output the 3D Gaussian field after regularization constraint. The topology deconstruction module includes a covariance decomposition unit, a linear discrimination unit, a filtering unit, a clustering unit, and a line segment weighted fitting unit, which are used to inversely deconstruct the three-dimensional Gaussian field and output a line skeleton topology map and a set of visual anchor points. The planning module is used to perform global path planning and trajectory optimization based on the skeleton topology map, and generate executable tracks. The positioning fusion module is used to perform point-line fusion repositioning based on the visual anchor point set and the wire skeleton topology map, and uses factor graph optimization to fuse the visual repositioning results, IMU pre-integration data and wireless ranging data to output high-precision pose. The positioning fusion module uses a robust kernel function for the wireless ranging factor to suppress multipath jump outpoints. The obstacle avoidance and replanning module is used to construct a local Euclidean symbolic distance field or occupy grid map based on airborne sensors. It uses the executable trajectory as a global reference and high-precision pose as the current state. When a dynamic obstacle is detected, it triggers local replanning and generates a detour trajectory. After detour, it guides the UAV back to the global trajectory generated by the planning module. The update module is used to write the stable new obstacle information observed consistently back to the three-dimensional Gaussian field prior map, so as to realize the incremental update of the map.

6. The autonomous cruise system for unmanned aerial vehicles (UAVs) in complex construction environments according to claim 5, characterized in that: The factor graph optimization structure in the positioning fusion module includes the following factor nodes: The visual relocation factor, whose residuals are constructed based on the reprojection error of the visual anchor point and the geometric constraints of the line segment; IMU pre-integration factor, used to constrain relative motion between adjacent keyframes; The wireless ranging factor, whose residual is constructed based on the difference between the measured distance and the predicted distance between the UAV and the known location anchor point; The robust kernel function, either the Huber kernel or the Cauchy kernel, is used to weight the residuals of the wireless ranging factor, reducing the impact of measurement jumps caused by multipath effects on the optimization results.

7. The autonomous cruise system for unmanned aerial vehicles (UAVs) oriented towards complex construction environments according to claim 5, characterized in that: The obstacle avoidance replanning module includes a local map building unit and a replanning triggering unit. The local map building unit constructs a local Euclidean symbolic distance field centered on the UAV's current position, based on a depth map generated in real-time by an onboard depth sensor or binocular camera. This distance field stores the signed distances to the nearest obstacle in a spatial grid. The replanning triggering unit triggers local replanning when it detects that a signed distance within a predetermined range in front of the UAV in the local Euclidean symbolic distance field is less than a safety threshold. The local replanning uses A... The algorithm generates a detour path on the obstacle expansion layer of the local Euclidean symbolic distance field, and then generates a smooth detour trajectory through Minimum-Snap optimization. After the detour is completed, the obstacle avoidance replanning module matches the current pose of the UAV with the wire skeleton topology map output by the topology deconstruction module, and guides the UAV back to the global track of the planning module.

8. The autonomous cruise system for unmanned aerial vehicles (UAVs) in complex construction environments according to claim 5, characterized in that: The update module includes a consistency judgment unit and a map fusion unit. The consistency judgment unit is used to continuously track the spatial position stability of the obstacle in multiple consecutive cruise cycles or multi-frame sensor observations after the obstacle avoidance and replanning module detects a new obstacle. When the number of times the obstacle at the same position is consistently detected exceeds a threshold within a preset time window, it is determined to be a stable new obstacle. The map fusion unit is used to convert the spatial occupancy information of the stable new obstacle into a new Gaussian volume or to correct the opacity of the original Gaussian volume, and send the updated map data to the three-dimensional Gaussian mapping module to update the three-dimensional Gaussian field prior map and the wire skeleton topology map output by the topology deconstruction module.

Images