Bridge crack detection method and system combining unmanned aerial vehicle and computer vision

Abstract

The invention discloses a bridge crack detection method and system combining an unmanned aerial vehicle and computer vision, and the method comprises the steps: employing the unmanned aerial vehicle to carry a laser radar to obtain bridge point cloud data, and constructing a bridge point cloud model; based on the point cloud model and the spatial constraint condition, planning an unmanned aerial vehicle flight route close to the bridge surface; a camera is carried by the unmanned aerial vehicle, and bridge images are collected along the flight route of the unmanned aerial vehicle; and fusing transfer learning and dynamic snakelike convolution to construct and train a crack recognition deep learning model, and performing crack recognition on a bridge image acquired by the unmanned aerial vehicle. According to the method, the crack recognition deep learning model based on transfer learning and dynamic snake-shaped convolution is constructed, so that full-process automation, intelligence and high precision of bridge crack detection are realized, the detection efficiency and quality are effectively improved, the labor cost and the safety risk are reduced, and the method has wide application value and popularization prospect.

Description

Technical Field

[0001] The present invention belongs to the technical field of bridge detection, and specifically relates to a bridge crack detection method and system combining unmanned aerial vehicle (UAV) and computer vision. Background Art

[0002] Concrete bridges are critical infrastructure for modern transportation, and their safety is directly linked to the safety of people's lives and property, as well as the sustainable development of society and the economy. As bridges age and are impacted by the external environment, they inevitably develop various structural defects, of which cracks are the most common and representative. The presence of cracks in bridges can reduce the structure's load-bearing capacity and durability, seriously threatening its safe operation. Therefore, timely, accurate, and efficient detection of bridge cracks is crucial for ensuring bridge safety, extending its lifespan, and reducing maintenance costs.

[0003] Traditional bridge crack detection methods rely primarily on manual visual inspection, supplemented by simple tools such as magnifying glasses and crack rulers. This method has the following drawbacks: (1) low detection efficiency and time-consuming and labor-intensive; (2) detection results are affected by the subjective experience of the inspectors, making accuracy and reliability difficult to guarantee; and (3) high-rise bridges and bridges across rivers are difficult to inspect manually, posing safety risks.

[0004] In recent years, with the rapid development of unmanned aerial vehicle (UAV) and computer vision technologies, bridge crack detection methods based on UAVs and computer vision have gradually become a research hotspot. UAVs offer advantages such as high maneuverability, high efficiency, and low cost, enabling rapid and flexible acquisition of bridge surface image data. Computer vision technology, particularly deep learning, has achieved breakthroughs in image recognition, object detection, and semantic segmentation, providing strong technical support for the automatic identification of bridge cracks. However, existing UAV- and computer vision-based bridge crack detection methods still have some shortcomings: (1) Insufficient automation of data collection: In current UAV bridge crack detection, data collection still relies heavily on manual remote control, resulting in a low degree of automation, low efficiency, and difficulty in ensuring data quality.

[0005] (2) Limited intelligence level of route planning: Existing UAV route planning methods are mainly based on simple geometric rules or manually preset waypoints, which are difficult to adapt to complex bridge structures and environments.

[0006] (3) The accuracy and robustness of crack recognition need to be improved: The apparent crack images of bridges have the characteristics of complex background, diverse crack morphology, small target ratio, and are easily disturbed by factors such as lighting, shadows, and water stains. The recognition accuracy and robustness of the existing deep learning crack recognition model in complex scenes still need to be improved, and there are problems of false detection, missed detection, and inaccurate segmentation. Summary of the Invention

[0007] The main purpose of the present invention is to overcome the shortcomings and deficiencies of the existing technology and provide a bridge crack detection method and system that combines drones and computer vision. The purpose is to solve the problems existing in the existing technology, such as low degree of automation, limited intelligent level of route planning, and the need to improve the accuracy and robustness of crack identification. The method realizes the automation and intelligence of the bridge crack detection process, improves the detection efficiency and quality, and reduces labor costs and safety risks.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions: One aspect of the present invention provides a bridge crack detection method combining drones and computer vision, comprising the following steps: Use UAV equipped with LiDAR to obtain bridge point cloud data and build a bridge point cloud model; Based on the point cloud model and spatial constraints, a UAV flight route close to the bridge surface is planned. Specifically, the method includes: using the bridge point cloud model, a skeleton-guided spatial decomposition method is used to decompose the bridge structure into multiple subspaces; within each subspace, a bidirectional ray method is used to generate an initial viewpoint set based on the bridge skeleton and spatial constraints; a gravity-like model is used to optimize the viewpoints, dynamically adjust the field of view, and remove redundant viewpoints to generate an optimal viewpoint set that effectively covers the bridge surface; for each subspace, local viewpoint paths are planned in parallel, and a global optimization algorithm is used to integrate the local paths to generate a globally optimal UAV flight route; Using a camera mounted on a drone to collect bridge images along the drone's flight path; By integrating transfer learning and dynamic snake convolution, a deep learning model for crack recognition is constructed and trained to identify cracks in bridge images collected by drones.

[0009] As a preferred technical solution, the construction of the bridge point cloud model specifically includes: Design a target bridge LiDAR scanning route, use a drone equipped with LiDAR to execute the route mission, and scan and obtain bridge point cloud data; Perform PPK calculation on the collected bridge point cloud data to reconstruct the bridge point cloud model; The bridge point cloud model is subjected to denoising, registration and streamlining processing to obtain processed bridge point cloud data.

[0010] As a preferred technical solution, the skeleton-guided spatial decomposition method based on the bridge point cloud model is used to decompose the bridge structure into multiple subspaces, specifically including: 1) Extract the bridge skeleton: Calculate the normal vector of each point in the bridge point cloud data to obtain the oriented point cloud PO ; The random sampling method was used to P O Downsampled to P D , and is calculated based on the iterative algorithm of plane cutting P D Each point p The ROSA point is denoted as r p =( x p ,v p ),in x p and v p Respectively represent the position and direction of the ROSA point; From the initial direction v p 0 Starting from, use iterative optimization to gradually adjust v p : ; In the formula, the update direction of each iteration is v p , and using the neighborhood N p i Normal covariance matrix of the interior points To guide optimization; After determining the optimal direction, calculate the position of the ROSA point according to the following formula to optimize the target x p : ; In the formula, each neighborhood point p k With normals n ( p k ), used for calculation of ROSA points; Finally, all ROSA points are smoothed by one-dimensional moving least squares method to generate skeleton lines that express the main structural features of the bridge point cloud and store them in an undirected graph. G In the vertex set V s Represents all ROSA points, edges E s Describe the connection relationship between ROSA points; 2) Decomposition of the bridge skeleton: The extracted bridge skeleton is decomposed into several geometric branches, each of which is represented by an undirected graph GThe decomposed branches are composed of a set of edges in B Represents, and meets the following constraints: Define joints: transform the undirected graph G Vertices with degree greater than 2 are defined as joints J , and each branch starts from a joint and ends at another joint or leaf vertex; Branch direction consistency: The edges in the same branch must have similar directions, and the similar directions are determined by a preset threshold. θ Control; if the direction change in a branch is greater than θ , it will be recursively decomposed into smaller branches with more consistent directions until all branches meet the direction consistency requirements; 3) Allocate the bridge point cloud space and generate point cloud subspaces: Plane definition: Discretize the edge of each branch to obtain a series of directed points, whose direction is the direction of the edge where the directed point is located. Each directed point determines a plane Γ; Point cloud assignment: For each plane Γ, from the oriented point cloud P O Search for a point near the plane; if there is a point p ∈ P O If the point is near multiple planes at the same time, p Assigned to the plane closest to its directed point; Subspace generation: assign points that are searched on the same branching plane A Contained in the same subspace, thus completing all P O The distribution of midpoints; wherein, the edge of each branch is discretized to obtain a series of directed points, whose direction is the direction of the edge where the directed point is located.

[0011] As a preferred technical solution, in each subspace, based on the bridge skeleton and spatial constraints, a bidirectional ray method is used to generate an initial viewpoint set, specifically including: 1) Use volume maps to represent the interior space. Each voxel is initialized to an "idle" state. If the voxel contains point cloud data P O , the status is updated to "occupied"; 2) For each plane γ i ∈Γ, from the skeleton point O i Towards its corresponding target point A iEmit a ray. Before encountering an "occupied" voxel, all the voxels that the ray passes through are marked as "interior space", indicating that it is before an obstacle or scene surface, and the part after passing through the "occupied" voxel belongs to "free space"; 3) Using a bidirectional ray casting method, traverse from both ends of the ray simultaneously, and optimize viewpoint generation through spatial constraints to set the constraint area S , filter out safe viewpoint positions: v i ∈ S i ,{ S i ∈ S , v i ∈ V}; Where, S i is the sub-region of the spatial constraint, v i For the viewpoints to be screened, eliminate the viewpoints located in the dangerous area; The viewpoint posture is iteratively updated through a gravity-like model. The viewpoint posture is defined as follows: vp =[ p , θ , ϕ , id ]; Where, p represents the three-dimensional position of the viewpoint, θ is the elevation angle of the viewpoint, ϕ is the yaw angle of the viewpoint, parameter id represents the bridge subspace to which the viewpoint belongs; Viewpoint sampling ray r vs The definition is as follows: sr =&[ x sr , y sr , z sr ]; dr =&[ nx dr , ny dr , nz dr ]; Where, sr Sampling rays for the viewpoint r vs The three-dimensional starting point coordinates, dris a three-dimensional direction vector; Distance along the ray D For sampling, the viewpoint coordinates and angle calculation formulas are as follows: p = sr + D * dr ; θ = arcsin ((- nz dr ) ⁄ || dr ||2); ϕ = arctan (- ny dr ⁄ - nx dr ); The initial viewpoint set obtained is recorded as vp ini .

[0012] As a preferred technical solution, the gravity-like model is used to optimize viewpoints, dynamically adjust the field of view, and remove redundant viewpoints to generate an optimal viewpoint set that can effectively cover the bridge surface. Specifically, the method includes: Calculate the coverage of each viewpoint through bidirectional ray casting and optimize the viewpoint allocation strategy: 1) If a voxel is observed by multiple viewpoints, it is preferentially assigned to the viewpoint that covers the most voxels to achieve optimal allocation; 2) For viewpoints that do not cover any voxels, remove them from the initial viewpoint set vp ini Deleted to reduce redundancy; 3) Build kd Tree for multidimensional data query: Initial viewpoint set vp ini Build kd Tree T ini , and set the "active" or "dormant" state for each viewpoint, and the initial state is "active"; Adjust the viewpoint positions based on the gravity-like model, merging viewpoints with a coverage range smaller than the preset value into viewpoints with a coverage range larger than the preset value. The specific process is as follows: Viewpoints are processed in order of decreasing number of covered voxels, and each viewpoint vp q exist kd Tree T ini Query its neighborhood viewpoint set VP q , query radius r qBy maximum visible distance d v and the field of view FOV, the field of view FOV includes the horizontal viewing angle f h and vertical viewing angle f w : r q = d v *tan( min ( f h , f w )⁄2); Viewpoint VP q The posture of the VP a Affected, and the position is updated according to the following formula: ; Where, c q and c a Respectively represent viewpoints vp q and the active viewpoint set VP a The number of voxels covered by each viewpoint in vp q The updated position of p q ;After the update, the pitch and yaw angles of the viewpoint are also adjusted accordingly; The subspace to which the viewpoint belongs is given by kd The nearest initial viewpoint in the tree determines the neighboring viewpoints after the update. VP a The state is set to "sleep" and no longer participates in subsequent updates; the process is repeated until all voxels are covered by the viewpoint; Finally, all viewpoints that are still in the "active" state are assigned to the corresponding subspace according to their numbers, and the final viewpoint set is obtained: V ={ VP 1,…, VP N}.

[0013] As a preferred technical solution, for each subspace, local viewpoint paths are planned in parallel, and the global optimization algorithm is used to integrate the local paths to generate the globally optimal UAV flight route, specifically including: 1) Global path sequence planning: Calculate the centroid of the viewpoint in each subspace and simplify it to a single point; Build the Euclidean distance matrix M G , record the distance between the center of mass of each subspace and the current position of the drone; Using the LKH solver, M G For input, get the global access sequence [ g 1,…, g n ], that is, the order of visiting subspaces; 2) Local boundary viewpoint selection: Sort all centroids according to the global sequence to form a global sequence: S eqC =[ k 0,…, k n ]; Where, k 0 represents the current position of the drone, and the rest are sorted centroids; Select the i The starting point of the subspace and End Viewpoint : ; ; For the last subspace, no end viewpoint is set; 3) Parallel planning of local paths: If there is a subspace R viewpoints, the cost matrix M L The definition is as follows: ; in: ; Ang ( a 1, a 2)= min (| a 1- a 2|,2 π -| a 1- a 2|); in, L (.) indicates the viewpoints found by the A* algorithm. v i To Viewpoint v j Path length;v max is the maximum speed limit, ω max are the maximum limits of pitch angle and yaw angle respectively; when processing the last subspace, the cost matrix M L All values in The item is set to , represents the cost between the viewpoints; for each viewpoint vp i ∈ V Use multi-threaded programming to independently and parallelize the search for local collision-free paths P C ; 4) Local path optimization: Adopt local optimization strategy to bridge the global coverage path of 3D scene P C , sub-branch set B and intersection radius r jc As input, identify all junction points and construct kd Tree TC To speed up the search; For each intersection z , perform radius search to obtain adjacent viewpoints v jc ; conduct K Rounds of iterations, randomly removing viewpoints in each round v 1. Select a nearby viewpoint v 2 Make subsequent or predecessor adjustments and randomly select a third viewpoint from the remaining viewpoints v 3; Finally, the 2-opt algorithm is used to optimize the three paths and generate the optimized path P R ; 5) Collision-free path optimization: The path P R Divided into M Segment trajectory, and generate a three-dimensional safe flight corridor for ensuring flight safety, wherein each convex flight corridor corresponds to a set of continuous trajectory segments; tp i ( t )∈ CP ( i ),∀ t ∈[0, T i ],∀1≤ i ≤ M ; in, tp i It is i Segment trajectory, T i is the duration of the trajectory, CP ( i ) indicates that the tp i Convex flight corridor; ; ; ; ; Where, v max 、 α max and j max represent the dynamic limits of velocity, acceleration and jerk respectively; Finally, the generated UAV flight path is converted into a UAV flight route.

[0014] As a preferred technical solution, the crack identification deep learning model includes an encoding layer, a feature fusion layer, and a decoding layer, specifically: The encoding layer has a dual encoder structure, including a global general feature encoder and a geometric feature encoder; the global general feature encoder is used to extract the global semantic features of the crack image; the geometric feature encoder uses dynamic snake convolution to focus on the crack area and capture local geometric structure features; The feature fusion layer includes a splicing layer, a group normalization layer, and a ReLU activation function connected in sequence. The splicing layer is used to perform a splicing operation on the features extracted by the global general feature encoder and the geometric feature encoder; the group normalization layer and the ReLU activation function are used to perform nonlinear transformation and feature normalization on the spliced feature map; The decoding layer includes an upsampling module for decoding the feature map output by the feature fusion layer, gradually restoring the spatial resolution of the feature map, and finally generating a pixel-level crack semantic segmentation result.

[0015] As a preferred technical solution, the global universal feature encoder adopts the encoder of the EfficientSAM model pre-trained on the image dataset. The encoder of the EfficientSAM model freezes the underlying parameters, and during the training process, the weights remain fixed, and only the last layer of the network is unfrozen.

[0016] As a preferred technical solution, the dynamic snake convolution is specifically as follows: Gradually accumulate the offset Δ of the convolution kernel, x The axis directions are transformed as follows: ; Similarly, in y The transformation of the axis direction is: .

[0017] Another aspect of the present invention provides a bridge crack detection system combining drones and computer vision, which is applied to the above-mentioned bridge crack detection method combining drones and computer vision, and includes a data acquisition module, a model building module, a route planning module, and a crack identification module; The data acquisition module uses a drone equipped with a laser radar to obtain bridge point cloud data and collects bridge images according to the drone's flight path; The model building module is used to build a bridge point cloud model based on the bridge point cloud data; The route planning module is used to plan a UAV flight route close to the bridge surface based on the bridge point cloud model and spatial constraints. Specifically, the module includes: based on the bridge point cloud model, a skeleton-guided spatial decomposition method is used to decompose the bridge structure into multiple subspaces; within each subspace, an initial viewpoint set is generated using a bidirectional ray method based on the bridge skeleton and spatial constraints; a gravity-like model is used to optimize the viewpoints, dynamically adjust the field of view and remove redundant viewpoints to generate an optimal viewpoint set that can effectively cover the bridge surface; for each subspace, local viewpoint paths are planned in parallel, and a global optimization algorithm is used to integrate the local paths to generate a globally optimal UAV flight route; The crack recognition module is used to integrate transfer learning and dynamic snake convolution to build and train a crack recognition deep learning model to identify cracks in bridge images collected by drones.

[0018] Compared with the prior art, the present invention has the following advantages and beneficial effects: (1) Based on UAV and deep learning technology, the present invention integrates transfer learning and dynamic snake convolution to construct and train a deep learning model for crack recognition, providing a technical solution for bridge crack detection. This method can realize the automation, intelligence and high precision of the entire process of bridge crack detection, effectively improve the detection efficiency and quality, reduce labor costs and safety risks, and has broad application value and promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of a bridge crack detection method combining drones and computer vision according to an embodiment of the present invention; Figure 2This is a structural diagram of a bridge crack detection system that combines drones and computer vision according to an embodiment of the present invention. DETAILED DESCRIPTION

[0020] In order to enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0021] Example: like Figure 1 As shown, this embodiment provides a bridge crack detection method combining drones and computer vision, including the following steps: S1: Use a drone equipped with a lidar to obtain high-precision 3D bridge point cloud data and construct a bridge point cloud model; S2: Based on the point cloud model and spatial constraints, the drone flight path is planned to stay close to the bridge surface. S3: using a drone equipped with a high-definition camera to automatically capture bridge images along the drone's flight path; S4: We integrate transfer learning with dynamic snake convolution to build and train a deep learning model for crack recognition, and perform crack recognition on bridge images captured by drones.

[0022] Furthermore, the specific implementation process of step S1 is as follows: S11: Design a target bridge LiDAR scanning route in the UAV remote controller, execute the route mission with the UAV equipped with LiDAR, and scan to obtain bridge point cloud data; S12: Export the bridge point cloud data from the LiDAR memory card to the workstation, and perform PPK calculation on it through the modeling software to reconstruct the bridge point cloud model; S13: De-noising, registration, and streamlining are performed on the bridge point cloud model using Cloud Compare point cloud processing software to obtain pre-processed bridge point cloud data, which serves as a basis for subsequent route planning.

[0023] Furthermore, the specific implementation process of step S2 is as follows: S21: First, the bridge skeleton is extracted. This patent uses the generalized rotational symmetry axis (ROSA) point to represent the bridge skeleton structure. By preprocessing the bridge point cloud data and calculating the normal vector of each point, a directional point cloud is obtained. P O In order to improve the computational efficiency, a random sampling method is used to P ODownsampled to P D , and is calculated based on the iterative algorithm of plane cutting P D Each point p The generalized rotational symmetry axis (ROSA) point of r p =( x p ,v p ),in x p and v p Respectively represent the position and direction of the ROSA point; For each point in the point cloud p , its ROSA point r p Acts on itself and its direction v p The plane formed by the impact of the adjacent area N p ROSA point in N p The interior shows a high degree of symmetry, so the direction v p and location x p The determination of v p and N p Minimize the angular variance between internal normals; second, position x p As close as possible N p The extension line of the inner normal. The optimization of direction and position can be reduced to a quadratic programming problem, that is, from the initial direction v p 0 Starting from, use iterative optimization to gradually adjust v p : ; In the formula, the update direction of each iteration is v p , and using the neighborhood N p i Normal covariance matrix of the interior points To guide optimization.

[0024] After determining the optimal direction, calculate the position of the ROSA point according to the following formula to optimize the target x p : ; In the formula, each neighborhood point p k With normals n ( p k ), which is used to calculate the ROSA points. Finally, all ROSA points are smoothed by the one-dimensional moving least squares method to generate a skeleton line that can accurately express the main structural features of the bridge point cloud and store it in the undirected graph. G In the vertex set V s Represents all ROSA points, edges E s Describe the connection relationship between ROSA points; Subsequently, the bridge skeleton is decomposed. This patent decomposes the extracted bridge skeleton into several simple geometric branches, each of which is represented by an undirected graph. G The decomposed branches are composed of a set of edges in B This patent uses a depth-first search algorithm to decompose the skeleton, which is based on the following two main principles: 1) Define joints: transform the undirected graph G Vertices with degree greater than 2 are defined as joints J To satisfy the simple geometric properties of branches, each branch starts from a joint and ends at another joint or a leaf vertex (i.e., a vertex with degree 1).

[0025] 2) Branch direction consistency: The edges in the same branch need to have similar directions. This similarity is determined by a preset threshold θ Control. If the direction change in a branch is greater than θ , it will be recursively decomposed into smaller branches with more consistent directions until all branches meet the direction consistency requirements; Finally, the bridge point cloud space is allocated, and the point cloud subspace generation process is as follows: 1) Plane Definition: First, the edges of each branch are discretized to obtain a series of directed points, whose directions are the directions of the edges on which the directed points lie. Each directed point defines a plane Γ. This can be understood as defining a series of planes on each branch, which will be used for subsequent point cloud assignment.

[0026] 2) Point cloud assignment: Secondly, for each plane Γ, from the oriented point cloud P O Search for a point near the plane. If there is a point p ∈ P O If the point is near multiple planes at the same time,p Assigned to the plane closest to its directed point.

[0027] 3) Subspace generation: Finally, the distribution points searched on the same branch plane A Contained in the same subspace, thus completing all P O Midpoint allocation.

[0028] S22: This patent uses volume maps combined with ray casting technology to perform viewpoint sampling to initialize the viewpoint. The specific process is as follows: 1) Use volume maps to represent the interior space. Each voxel (3D grid cell) is initialized to an "idle" state. If the voxel contains point cloud data P O , the status is updated to "Occupied".

[0029] 2) For each plane γ i ∈Γ, from the skeleton point O i Towards its corresponding target point A i Shoot a ray. All voxels that the ray passes through before encountering an "occupied" voxel are marked as "interior space," indicating that they are before obstacles or scene surfaces. The portion after passing through the "occupied" voxel is considered "free space."

[0030] 3) To improve computational efficiency, this patent uses a bidirectional ray casting method to traverse from both ends of the ray simultaneously, reducing redundant computations and detecting obstacles earlier. However, this method may cause some viewpoints to fall into the obstacle area. Therefore, this patent optimizes viewpoint generation through spatial constraints and sets the constraint area. S , filter out safe viewpoint positions: v i ∈ S i ,{ S i ∈ S , v i ∈ V}; Where, S i is the sub-region of the spatial constraint, v i For the viewpoints to be screened, viewpoints in dangerous areas are eliminated to improve safety. The viewpoints finally generated not only cover rich bridge surface structure information, but also reduce redundant viewpoints and ensure their safety and rationality in the space around the bridge. In addition, since the sampling ray is composed of the bridge skeleton points, the bridge skeleton points are the same as the bridge skeleton points. Oi Pointing to the surface of the bridge scene, this method can effectively improve the accuracy of covering the bridge surface and ensure that the viewpoint belongs to the correct subspace.

[0031] Then, the viewpoint posture is iteratively updated through the gravity-like model. The initial viewpoint is oriented towards the corresponding bridge skeleton point. O i However, since the coverage of multiple viewpoints may overlap, it is necessary to optimize the viewpoint pose to ensure that the viewpoint can maximize the coverage of the bridge surface and reduce redundant viewpoints. In this patent, the viewpoint pose is defined as follows: vp =[ p , θ , ϕ , id ]; Where, p represents the three-dimensional position of the viewpoint, θ is the elevation angle of the viewpoint, ϕ is the yaw angle of the viewpoint, parameter id Represents the bridge subspace to which the viewpoint belongs.

[0032] Viewpoint sampling ray r vs The definition is as follows: r vs sr =&[ x sr , y sr , z sr ]; dr =&[ nx dr , ny dr , nz dr ]; Where, sr Sampling rays for the viewpoint r vs The three-dimensional starting point coordinates, dr is a three-dimensional direction vector.

[0033] Distance along the ray D For sampling, the viewpoint coordinates and angle calculation formulas are as follows: p = sr + D * dr ; θ = arcsin ((- nzdr ) ⁄ || dr ||2); ϕ = arctan (- ny dr ⁄ - nx dr ); The initial viewpoint set obtained is recorded as vp ini .

[0034] Next, the coverage of each viewpoint is calculated by bidirectional ray casting, and the viewpoint allocation strategy is optimized: 1) If a voxel is observed by multiple viewpoints, it is preferentially assigned to the viewpoint that covers the most voxels to achieve optimal allocation.

[0035] 2) For viewpoints that do not cover any voxels, remove them from vp ini Deleted to reduce redundancy.

[0036] To improve computing efficiency, this patent uses kd K-dimensional tree for efficient multi-dimensional data query. Initial viewpoint set vp ini Build kd Tree T ini , and set the "active" or "dormant" state for each viewpoint (all are initially "active"). Then, based on the gravity-like model, the viewpoint positions are adjusted, merging viewpoints with a coverage range smaller than the preset value into viewpoints with a coverage range larger than the preset value. The specific process is as follows: 1) Process the viewpoints in descending order of the number of covered voxels. Each viewpoint vp q exist kd Tree T ini Query its neighborhood viewpoint set VP q , query radius r q By maximum visible distance d v and FOV (including horizontal viewing angle f h and vertical viewing angle f w ) to determine: r q = d v *tan( min ( f h ,f w )⁄2); 2) Viewpoint VP q The posture of the VP a Affected, and the position is updated according to the following formula: ; Where, c q and c a Respectively represent viewpoints vp q and the active viewpoint set VP a The number of voxels covered by each viewpoint in vp q The updated position of p q After the update, the pitch and yaw angles of the viewpoint were adjusted accordingly to optimize the coverage effect within the field of view.

[0037] 3) The subspace to which the viewpoint belongs is kd The nearest initial viewpoint in the tree is used for decision. After updating, the neighboring viewpoints VP a The state is set to "sleep" and no longer participates in subsequent updates. This process is repeated until all voxels are covered by the viewpoint.

[0038] Finally, all viewpoints that are still in the "active" state are assigned to the corresponding subspace according to their numbers, and the final viewpoint set is obtained: V ={ VP 1,…, VP N}; S23: Plan the drone's flight path along the bridge, ensuring that its field of view fully covers the bridge's three-dimensional scene while maintaining the shortest possible path and avoiding collisions (i.e., avoiding the bridge structure). This involves the following five steps: (1) Step 1: Global path sequence planning; First, the centroid of the viewpoint in each subspace is calculated and simplified to a single point. Then, the Euclidean distance matrix is established. M G , record the distance between the centroid of each subspace and the current position of the drone. Finally, use the Lin-Kernighan-Helsgaun (LKH) solver to M G For input, get the global access sequence [ g 1,…, g n], that is, the order of visiting subspaces.

[0039] (2) Step 2: Local boundary viewpoint selection; First, sort all centroids according to the global sequence to form a global sequence: S eqC =[ k 0,…, k n ]; Where, k 0 represents the current position of the drone, and the rest are sorted centroids.

[0040] Then, select the i The starting point of the subspace and End Viewpoint : ; ; For the last subspace, since it has no subsequent subspace, no end viewpoint is set for it.

[0041] (3) Step 3: Parallel planning of local paths; If there is a subspace R viewpoints, then the cost matrix of this TSP is M L (The cost matrix contains the cost or distance to move between different viewpoints) is defined as follows: ; in, ; Ang ( a 1, a 2)= min (| a 1- a 2|,2 π -| a 1- a 2|); in, L (.) indicates the viewpoints found by the A* algorithm. v i To Viewpoint v j The path length is set to ensure that the path has no collision with the bridge structure. v max is the maximum speed limit, ω maxare the maximum limits for pitch and yaw angles respectively. When processing the last subspace, the cost matrix M L All values in The item is set to , represents the cost between the viewpoints. For each viewpoint vp i ∈ V Use multi-threaded programming to independently and parallelize the search for local collision-free paths P C .

[0042] (4) Step 4: Local path optimization; Since the above path planning steps mainly optimize each subspace, it may cause unnecessary bends in the global path, especially at the junction of local paths. Therefore, this step adopts a local optimization strategy to bridge the global coverage path of the three-dimensional scene. P C , sub-branch set B and intersection radius r jc As input, identify all junction points and construct kd Tree TC To speed up the search.

[0043] For each intersection z (the vertex where the two sub-branches intersect), perform a radius search to obtain adjacent viewpoints v jc Then proceed K Rounds of iterations, randomly removing viewpoints in each round v 1. Select a nearby viewpoint v 2 Make subsequent or predecessor adjustments and randomly select a third viewpoint from the remaining viewpoints v 3. Finally, use the 2-opt algorithm to optimize the three paths and generate the optimized path P R , effectively reducing unnecessary detours in the path of the covered bridge three-dimensional scene and improving the efficiency of the drone flight path.

[0044] (5) Step 5: Collision-free path optimization; The purpose of this step is to optimize the path of the bridge coverage scene P R ={ v R 0 ,…, v R M} is converted into a smooth, safe and shortest flight path for the drone to pass through all viewpoints. First, the path P RDivided into M Segment trajectory and generate three-dimensional safe flight corridors (SFCs), where each convex flight corridor corresponds to a set of continuous trajectory segments. For position trajectory, SFCs are used to ensure flight safety: tp i ( t )∈ CP ( i ),∀ t ∈[0, T i ],∀1≤ i ≤ M ; in, tp i It is i Segment trajectory, T i is the duration of the trajectory, CP ( i ) indicates that the tp i Convex flight corridor.

[0045] ; ; ; ; Where, v max 、 α max and j max These represent the dynamic limits of velocity, acceleration, and jerk, respectively, to ensure that the trajectory is feasible in actual flight and does not exceed the capabilities of the equipment.

[0046] Finally, the generated drone flight path is converted into a flight route according to the requirements of DJI UAV route file format standard (WPML).

[0047] Furthermore, in step S4, the crack identification deep learning model includes an encoding layer, a feature fusion layer, and a decoding layer, specifically: The encoding layer has a dual encoder structure, including a global general feature encoder and a geometric feature encoder; the global general feature encoder is used to extract the global semantic features of the crack image; the geometric feature encoder uses dynamic snake convolution to focus on the crack area and capture local geometric structure features; The feature fusion layer includes a concatenation layer, a group normalization layer, and a ReLU activation function connected in sequence. The concatenation layer is used to perform a concatenation (Cat) operation on the features extracted by the global general feature encoder and the geometric feature encoder to integrate feature information from different sources in the channel dimension. The group normalization layer and the ReLU activation function are used to perform nonlinear transformation and feature normalization on the concatenated feature map to achieve effective feature fusion. The fused feature map contains the global general features from the pre-trained model encoder and the local geometric features extracted by the DSConv encoder, which can more comprehensively characterize the characteristic information of the crack. The decoding layer includes an upsampling module for decoding the feature map output by the feature fusion layer, gradually restoring the spatial resolution of the feature map, and finally generating a pixel-level crack semantic segmentation result.

[0048] Furthermore, this patent employs a transfer learning-based strategy, introducing the encoder of the EfficientSAM model, pre-trained on a large-scale image dataset, as a global universal feature encoder for the crack image recognition task. This method leverages the rich feature representations (e.g., common visual patterns such as edges, textures, and shapes) learned by the EfficientSAM model's encoder on this large-scale image dataset to extract global semantic features of crack images. Furthermore, this patent employs a transfer learning strategy called "freezing underlying parameters." During training, the encoder weights of the pre-trained model remain fixed, with only the final layers of the network being unfrozen to optimize for the specific task.

[0049] Furthermore, to address the challenges of complex crack morphology (such as small, curved, and multi-branched) and background interference, this patent introduces a dynamic snake-shaped convolution kernel to replace the traditional convolution kernel as a geometric structure feature extractor. By simulating the dynamic characteristics of the snake-shaped structure, it has stronger adaptability, can flexibly focus on the crack area and capture its local subtle geometric structure features. Feature extraction is performed on the small and curved crack morphology, thereby enhancing the perception of the local crack geometry.

[0050] In order to ensure that the convolution kernel conforms to the linear structure, it is necessary to gradually accumulate the offset Δ. x The axis directions are transformed as follows: ; Similarly, in y The transformation of the axis direction is: .

[0051] like Figure 2 As shown, in another embodiment of the present application, a bridge crack detection system combining drones and computer vision is provided, the system including a data acquisition module, a model building module, a route planning module, and a crack identification module; The model building module is used to build a bridge point cloud model based on the bridge point cloud data; The route planning module is used to plan a UAV flight route close to the bridge surface based on the bridge point cloud model and spatial constraints. Specifically, the module includes: based on the bridge point cloud model, a skeleton-guided spatial decomposition method is used to decompose the bridge structure into multiple subspaces; within each subspace, an initial viewpoint set is generated using a bidirectional ray method based on the bridge skeleton and spatial constraints; a gravity-like model is used to optimize the viewpoints, dynamically adjust the field of view and remove redundant viewpoints to generate an optimal viewpoint set that can effectively cover the bridge surface; for each subspace, local viewpoint paths are planned in parallel, and a global optimization algorithm is used to integrate the local paths to generate a globally optimal UAV flight route; The crack recognition module is used to integrate transfer learning and dynamic snake convolution to build and train a crack recognition deep learning model to identify cracks in bridge images collected by drones.

[0052] It should be noted here that the system provided in the above embodiment is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure can be divided into different functional modules to complete all or part of the functions described above. The system is a bridge crack detection method combining drones and computer vision applied to the above embodiment.

[0053] In another embodiment of the present application, a storage medium is provided, storing a program. When the program is executed by a processor, a bridge crack detection method combining drones and computer vision is implemented, specifically: S1. Use a drone equipped with a laser radar to obtain bridge point cloud data and build a bridge point cloud model; S2. Based on the point cloud model and spatial constraints, plan the drone's flight path close to the bridge surface, including: Based on the bridge point cloud model, a skeleton-guided spatial decomposition method is used to decompose the bridge structure into multiple subspaces. In each subspace, based on the bridge skeleton and spatial constraints, the bidirectional ray method is used to generate the initial viewpoint set; A gravity-like model is used to optimize viewpoints, dynamically adjust the field of view, and remove redundant viewpoints to generate an optimal viewpoint set that effectively covers the bridge surface. For each subspace, local viewpoint paths are planned in parallel, and through a global optimization algorithm, the local paths are integrated to generate the globally optimal UAV flight route; S3. Using a camera mounted on a drone to collect bridge images along the flight path of the drone; S4. Fusion transfer learning and dynamic snake convolution are used to build and train a deep learning model for crack recognition, and crack recognition is performed on bridge images captured by drones.

[0054] It should be understood that various parts of the present application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used to implement: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0055] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A bridge crack detection method combining drones and computer vision, characterized in that: The steps include: Use UAV equipped with LiDAR to obtain bridge point cloud data and build a bridge point cloud model; Based on the point cloud model and spatial constraints, a UAV flight route close to the bridge surface is planned. Specifically, the method includes: using the bridge point cloud model, a skeleton-guided spatial decomposition method is used to decompose the bridge structure into multiple subspaces; within each subspace, a bidirectional ray method is used to generate an initial viewpoint set based on the bridge skeleton and spatial constraints; a gravity-like model is used to optimize the viewpoints, dynamically adjust the field of view, and remove redundant viewpoints to generate an optimal viewpoint set that effectively covers the bridge surface; for each subspace, local viewpoint paths are planned in parallel, and a global optimization algorithm is used to integrate the local paths to generate a globally optimal UAV flight route; Using a camera mounted on a drone to collect bridge images along the drone's flight path; By integrating transfer learning and dynamic snake convolution, a deep learning model for crack recognition is constructed and trained to identify cracks in bridge images collected by drones.

2. The bridge crack detection method combining drone and computer vision according to claim 1 is characterized in that: The construction of the bridge point cloud model specifically includes: Design a target bridge LiDAR scanning route, use a drone equipped with LiDAR to execute the route mission, and scan and obtain bridge point cloud data; Perform PPK calculation on the collected bridge point cloud data to reconstruct the bridge point cloud model; The bridge point cloud model is subjected to denoising, registration and streamlining processing to obtain processed bridge point cloud data.

3. The bridge crack detection method combining drone and computer vision according to claim 1 is characterized in that: Based on the bridge point cloud model, the skeleton-guided spatial decomposition method is used to decompose the bridge structure into multiple subspaces, specifically including: 1) Extract the bridge skeleton: Calculate the normal vector of each point in the bridge point cloud data to obtain the oriented point cloud P O ; The random sampling method was used to P O Downsampled to P D , and is calculated based on the iterative algorithm of plane cutting P D Each point p The ROSA point is denoted as r p =( x p ,v p ),in x p and v p Respectively represent the position and direction of the ROSA point; From the initial direction v p 0 Starting from, use iterative optimization to gradually adjust v p : ； In the formula, the update direction of each iteration is v p , and using the neighborhood N p i Normal covariance matrix of the interior points To guide optimization; After determining the optimal direction, calculate the position of the ROSA point according to the following formula to optimize the target x p : ； In the formula, each neighborhood point p k With normals n ( p k ), used for calculation of ROSA points; Finally, all ROSA points are smoothed by one-dimensional moving least squares method to generate skeleton lines that express the main structural features of the bridge point cloud and store them in an undirected graph. G In the vertex set V s Represents all ROSA points, edges E s Describe the connection relationship between ROSA points; 2) Decomposition of the bridge skeleton: The extracted bridge skeleton is decomposed into several geometric branches, each of which is represented by an undirected graph G The decomposed branches are composed of a set of edges in B Represents, and meets the following constraints: Define joints: transform the undirected graph G Vertices with degree greater than 2 are defined as joints J , and each branch starts from a joint and ends at another joint or leaf vertex; Branch direction consistency: The edges in the same branch must have similar directions, and the similar directions are determined by a preset threshold. θ Control; if the direction change in a branch is greater than θ , it will be recursively decomposed into smaller branches with more consistent directions until all branches meet the direction consistency requirements; 3) Allocate the bridge point cloud space and generate point cloud subspaces: Plane definition: Discretize the edge of each branch to obtain a series of directed points, whose direction is the direction of the edge where the directed point is located. Each directed point determines a plane Γ; Point cloud assignment: For each plane Γ, from the oriented point cloud P O Search for a point near the plane; if there is a point p ∈ P O If the point is near multiple planes at the same time, p Assigned to the plane closest to its directed point; Subspace generation: assign points that are searched on the same branching plane A Contained in the same subspace, thus completing all P O The distribution of midpoints; wherein, the edge of each branch is discretized to obtain a series of directed points, whose direction is the direction of the edge where the directed point is located.

4. The bridge crack detection method combining drone and computer vision according to claim 1 is characterized in that: In each subspace, based on the bridge skeleton and spatial constraints, a bidirectional ray method is used to generate an initial viewpoint set, specifically including: 1) Use volume maps to represent the interior space. Each voxel is initialized to an "idle" state. If the voxel contains point cloud data P O , the status is updated to "occupied"; 2) For each plane γ i ∈Γ, from the skeleton point O i Towards its corresponding target point A i Emit a ray. Before encountering an "occupied" voxel, all voxels that the ray passes through are marked as "interior space", indicating that it is before an obstacle or scene surface, while the part after passing through the "occupied" voxel belongs to "free space"; 3) Using a bidirectional ray casting method, traverse from both ends of the ray simultaneously, optimize viewpoint generation through spatial constraints, and set the constraint area S , filter out safe viewpoint positions: v i ∈ S i ,{ S i ∈ S , v i ∈ V}； Where, S i is the sub-region of the spatial constraint, v i For the viewpoints to be screened, eliminate the viewpoints located in the dangerous area; The viewpoint posture is iteratively updated through a gravity-like model. The viewpoint posture is defined as follows: vp =[ p , θ , ϕ , id ]； Where, p represents the three-dimensional position of the viewpoint, θ is the elevation angle of the viewpoint, ϕ is the yaw angle of the viewpoint, parameter id represents the bridge subspace to which the viewpoint belongs; Viewpoint sampling ray r vs The definition is as follows: sr =&[ x sr , y sr , z sr ]； dr =&[ nx dr , ny dr , nz dr ]； Where, sr Sampling rays for the viewpoint r vs The three-dimensional starting point coordinates, dr is a three-dimensional direction vector; Distance along the ray D For sampling, the viewpoint coordinates and angle calculation formulas are as follows: p = sr + D * dr ； θ = arcsin ((- nz dr ) ⁄ || dr ||2)； ϕ = arctan (- ny dr ⁄ - nx dr )； The initial viewpoint set obtained is recorded as vp ini .

5. The bridge crack detection method combining drone and computer vision according to claim 1 is characterized in that: The method of optimizing viewpoints using a gravity-like model, dynamically adjusting the field of view and removing redundant viewpoints, and generating an optimal viewpoint set that can effectively cover the bridge surface includes: Calculate the coverage of each viewpoint through bidirectional ray casting and optimize the viewpoint allocation strategy: 1) If a voxel is observed by multiple viewpoints, it is preferentially assigned to the viewpoint that covers the most voxels to achieve optimal allocation; 2) For viewpoints that do not cover any voxels, remove them from the initial viewpoint set vp ini Deleted to reduce redundancy; 3) Build kd Tree for multidimensional data query: Initial viewpoint set vp ini Build kd Tree T ini , and set the "active" or "dormant" state for each viewpoint, and the initial state is "active"; Adjust the viewpoint positions based on the gravity-like model, merging viewpoints with a coverage range smaller than the preset value into viewpoints with a coverage range larger than the preset value. The specific process is as follows: Viewpoints are processed in order of decreasing number of covered voxels, and each viewpoint vp q exist kd Tree T ini Query its neighborhood viewpoint set VP q , query radius r q By maximum visible distance d v and the field of view FOV, the field of view FOV includes the horizontal viewing angle f h and vertical viewing angle f w : r q = d v *tan( min ( f h , f w )⁄2); Viewpoint VP q The posture of the VP a Affected, and the position is updated according to the following formula: ； Where, c q and c a Respectively represent viewpoints vp q and the active viewpoint set VP a The number of voxels covered by each viewpoint in vp q The updated position of p q ;After the update, the pitch and yaw angles of the viewpoint are also adjusted accordingly; The subspace to which the viewpoint belongs is given by kd The nearest initial viewpoint in the tree determines the neighboring viewpoints after the update. VP a The state is set to "sleep" and no longer participates in subsequent updates; the process is repeated until all voxels are covered by the viewpoint; Finally, all viewpoints that are still in the "active" state are assigned to the corresponding subspace according to their numbers, and the final viewpoint set is obtained: V ={ VP 1,…, VP N }。 6. The bridge crack detection method combining drone and computer vision according to claim 1 is characterized in that: For each subspace, local viewpoint paths are planned in parallel, and the global optimization algorithm is used to integrate the local paths to generate the globally optimal UAV flight route. Specifically, the following steps are performed: 1) Global path sequence planning: Calculate the centroid of the viewpoint in each subspace and simplify it to a single point; Build the Euclidean distance matrix M G , record the distance between the center of mass of each subspace and the current position of the drone; Using the LKH solver, M G For input, get the global access sequence [ g 1,…, g n ], that is, the order of visiting subspaces; 2) Local boundary viewpoint selection: Sort all centroids according to the global sequence to form a global sequence: S eqC =[ k 0,…, k n ]； Where, k 0 represents the current position of the drone, and the rest are sorted centroids; Select the i The starting point of the subspace and End Viewpoint : ； ； For the last subspace, no end viewpoint is set; 3) Parallel planning of local paths: If there is a subspace R viewpoints, the cost matrix M L The definition is as follows: ； in: ； Ang ( a 1, a 2 )= min (| a 1- a 2 |,2 π -| a 1- a 2 |)； in, L (.) indicates the viewpoints found by the A* algorithm. v i To Viewpoint v j Path length; v max is the maximum speed limit, ω max are the maximum limits of pitch angle and yaw angle respectively; when processing the last subspace, the cost matrix M L All values in The item is set to , represents the cost between the viewpoints; for each viewpoint vp i ∈ V Use multi-threaded programming to independently and parallelize the search for local collision-free paths P C ; 4) Local path optimization: Adopt local optimization strategy to bridge the global coverage path of 3D scene P C , sub-branch set B and intersection radius r jc As input, identify all junction points and construct kd Tree TC To speed up the search; For each intersection z , perform radius search to obtain adjacent viewpoints v jc ; conduct K Rounds of iterations, randomly removing viewpoints in each round v 1. Select a nearby viewpoint v 2 Make subsequent or predecessor adjustments and randomly select a third viewpoint from the remaining viewpoints v 3; Finally, the 2-opt algorithm is used to optimize the three paths and generate the optimized path P R ; 5) Collision-free path optimization: The path P R Divided into M Segment trajectory, and generate a three-dimensional safe flight corridor for ensuring flight safety, wherein each convex flight corridor corresponds to a set of continuous trajectory segments; tp i ( t )∈ CP ( i ),∀ t ∈[0, T i ],∀1≤ i ≤ M ； in, tp i It is i Segment trajectory, T i is the duration of the trajectory, CP ( i ) indicates that the tp i Convex flight corridor; ； ； ； ； Where, v max 、 α max and j max represent the dynamic limits of velocity, acceleration and jerk respectively; Finally, the generated UAV flight path is converted into a UAV flight route.

7. The bridge crack detection method combining drone and computer vision according to claim 1 is characterized in that: The crack identification deep learning model includes an encoding layer, a feature fusion layer, and a decoding layer, specifically: The encoding layer has a dual encoder structure, including a global general feature encoder and a geometric feature encoder; the global general feature encoder is used to extract the global semantic features of the crack image; the geometric feature encoder uses dynamic snake convolution to focus on the crack area and capture local geometric structure features; The feature fusion layer includes a splicing layer, a group normalization layer, and a ReLU activation function connected in sequence. The splicing layer is used to perform a splicing operation on the features extracted by the global general feature encoder and the geometric feature encoder; the group normalization layer and the ReLU activation function are used to perform nonlinear transformation and feature normalization on the spliced feature map; The decoding layer includes an upsampling module for decoding the feature map output by the feature fusion layer, gradually restoring the spatial resolution of the feature map, and finally generating a pixel-level crack semantic segmentation result.

8. The bridge crack detection method combining drone and computer vision according to claim 7 is characterized in that: The global universal feature encoder uses the encoder of the EfficientSAM model pre-trained on the image dataset. The encoder of the EfficientSAM model freezes the underlying parameters, and during the training process, the weights remain fixed and only the last layer of the network is unfrozen.

9. The bridge crack detection method combining drone and computer vision according to claim 7 is characterized in that: The dynamic snake convolution is specifically: Gradually accumulate the offset Δ of the convolution kernel, x The axis directions are transformed as follows: ； Similarly, in y The transformation of the axis direction is: 。 10. A bridge crack detection system combining drones and computer vision, characterized in that: A bridge crack detection method combining drone and computer vision, applied to any one of claims 1-9, comprising a data acquisition module, a model building module, a route planning module, and a crack identification module; The data acquisition module uses a drone equipped with a laser radar to obtain bridge point cloud data and collects bridge images according to the drone's flight path; The model building module is used to build a bridge point cloud model based on the bridge point cloud data; The route planning module is used to plan a UAV flight route close to the bridge surface based on the bridge point cloud model and spatial constraints. Specifically, the module includes: based on the bridge point cloud model, a skeleton-guided spatial decomposition method is used to decompose the bridge structure into multiple subspaces; within each subspace, an initial viewpoint set is generated using a bidirectional ray method based on the bridge skeleton and spatial constraints; a gravity-like model is used to optimize the viewpoints, dynamically adjust the field of view and remove redundant viewpoints to generate an optimal viewpoint set that can effectively cover the bridge surface; for each subspace, local viewpoint paths are planned in parallel, and a global optimization algorithm is used to integrate the local paths to generate a globally optimal UAV flight route; The crack recognition module is used to integrate transfer learning and dynamic snake convolution to build and train a crack recognition deep learning model to identify cracks in bridge images collected by drones.

Images