Dam slope vegetation laser cleaning method based on visual identification

Abstract

The application discloses a kind of dam slope vegetation laser removal methods based on visual identification, it is related to dam maintenance and laser removal technical field, comprising: obtaining multi-view image and distance data and calibrating and reconstructing, generating dam slope surface model and calculating slope and curvature;While extracting vegetation features and spatial distribution characteristics, complete area division and construct adjacency relationship;Further establish removal priority model, sort and successively select target area to adjacent area;Different area types are used Differentiated removal strategy, preset laser removal is executed in non-edge transition area, and uncertainty band based on information entropy is constructed in edge transition area and multi-round progressive removal is implemented, and dynamic non-target object is identified into operation picture under laser emission state and laser output is suspended.By fusing multi-view data, dam slope is finely identified, the removal sequence is optimized in combination with priority, and a progressive removal strategy is used in the boundary area, improving the efficiency and safety of the operation.

Description

Technical Field

[0001] This invention relates to the field of dam maintenance and laser removal technology, specifically a method for laser removal of vegetation on dam slopes based on visual recognition. Background Technology

[0002] Vegetation growth is likely to occur on dam slopes during long-term operation. Excessive or uneven distribution of vegetation can affect dam inspection and safety assessment, and may also affect the surface structure. Therefore, it is necessary to remove vegetation from dam slopes.

[0003] In existing technologies, methods such as manual pruning, mechanical removal, or spraying chemicals are commonly used for vegetation treatment. However, these methods each have limitations in terms of efficiency, adaptability to complex slopes, and environmental impact. With the development of laser technology, laser-based non-contact removal methods are gradually being applied to dam slope vegetation treatment, and combined with image acquisition and image processing technologies, vegetation areas can be identified and located.

[0004] However, in practical applications, due to the varying slopes of dam slopes and the complex boundaries between vegetation and the base surface, existing zoning methods based on single features or simple rules are insufficient to accurately distinguish different regional characteristics. Furthermore, in boundary transition areas, the clearing process lacks an effective dynamic adjustment mechanism, making it difficult to balance clearing effectiveness with operational safety. Therefore, it is necessary to further optimize dam slope vegetation identification and clearing strategies. Summary of the Invention

[0005] Based on the shortcomings of the prior art described above, the purpose of this invention is to provide a method for laser removal of vegetation on dam slopes based on visual recognition, so as to solve the above-mentioned technical problems.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for laser removal of vegetation on dam slopes based on visual recognition, comprising: S1: Acquire multi-view image data and distance data of the dam slope, generate a dam slope surface model by combining camera calibration and visual reconstruction algorithm, and calculate the slope and curvature point by point; S2: Generate a vegetation cover probability map based on multi-view image data, and extract vegetation features and spatial distribution features through image segmentation and classification methods. Combined with slope and curvature parameters, divide the dam slope surface into regions to generate region division results and construct region adjacency relationships. The region division results include: edge transition regions and non-edge transition regions. S3: Construct a clearing priority model based on vegetation features, spatial features, and the adjacent region relationship of the current work location. Calculate and sort the priority of the adjacent regions of the current work location to obtain a priority sequence for selecting the current work location region. S4: Select regions to be processed sequentially according to the priority sequence, and determine the currently selected region as the target region; S5: Perform the clearing operation according to the target area type. When the target area is a non-edge transition area, a preset laser clearing strategy is adopted. When the target area is an edge transition area, an uncertainty band with a width that dynamically changes with the local image information entropy is constructed. A multi-round progressive clearing strategy is generated based on the uncertainty band, and the clearing is performed according to the multi-round progressive clearing strategy. During the laser clearing process, the laser clearing operation screen containing the target area and the surrounding area of ​​the laser irradiation area is collected in real time. Dynamic non-target objects entering the laser clearing operation screen are identified. When a dynamic non-target object enters the laser irradiation area or the safe expansion range of the laser irradiation area, the laser output is paused.

[0007] The present invention is further configured such that S1 includes: The multi-view image data and distance data are obtained by simultaneously scanning the dam slope surface in a grid pattern using a gimbal-controlled high-magnification zoom camera and a laser rangefinder mounted in front of the dam slope. The camera calibration includes calibration processing of the internal and external parameters of the imaging device; The visual reconstruction algorithm uses the structure-of-motion reconstruction algorithm to fuse multi-view image data and distance data to generate a digital surface model of the dam slope containing three-dimensional coordinates. Based on the three-dimensional coordinate data of each pixel in the neighborhood of the digital surface model of the dam slope, the slope of the point is calculated by local plane fitting, and the curvature of the point is calculated by local quadratic surface fitting.

[0008] The present invention is further configured such that the extraction of vegetation features and spatial distribution features in S2 includes: The vegetation features include: color features, texture features, and morphological features; A pixel-level segmentation-based image segmentation method is used to perform pixel-by-pixel classification on multi-view image data to obtain classification results. The classification results include the vegetation confidence score of each pixel belonging to a vegetation category. The vegetation confidence score of each pixel is extracted from the classification results of pixel-by-pixel classification to form a vegetation cover probability map. The classification results are mapped onto the digital surface model of the dam slope to obtain candidate vegetation areas; Feature extraction processing is performed on candidate vegetation areas. Color features are extracted by color space transformation, texture features are extracted by local texture analysis, and morphological features are extracted by connected component analysis. Based on a preset classification model, the extracted color features, texture features and morphological features are fused and judged to generate vegetation area identification results between vegetation areas and non-vegetation areas. Spatial statistical analysis is performed on the identified vegetation areas to obtain the distribution range, density, and regional connectivity of vegetation on the dam slope surface, forming corresponding spatial distribution characteristics.

[0009] The present invention is further configured such that generating the region partitioning result and constructing the region adjacency relationship in step S2 includes: Spatial correlation processing is performed between the vegetation cover probability map and the slope and curvature parameters in the dam slope digital surface model to generate a fused map; Based on the spatial distribution characteristics of vegetation cover probability in the fused map, and combined with the spatial boundary between vegetated and non-vegetated areas, areas located in the range of high and low vegetation cover probability or where the attributes of adjacent areas change transitionally are identified as edge transition areas, and the remaining areas are identified as non-edge transition areas.

[0010] The present invention is further configured such that the subdivision of the non-edge transition region includes: Continuous areas with a vegetation cover probability higher than a preset high threshold are classified as vegetation cover areas, and continuous areas with a vegetation cover probability equal to or lower than a preset low threshold are classified as bare base areas. Based on the preset dam structure location information, the drainage ditch, expansion joints and the area around the monitoring equipment are marked as key areas for removal; Record the boundary location information of each region, and construct the region adjacency relationship structure based on the spatial contact relationship and boundary connection relationship between regions.

[0011] The present invention is further configured such that S3 includes: Based on vegetation characteristics, spatial distribution characteristics, and regional adjacency structure, data organization and processing are performed on the adjacent regions of the current work location to construct a regional demand data set; For each adjacent region in the regional demand data set, vegetation features and spatial distribution features are combined and processed according to preset feature fusion rules to form a comprehensive demand index. Based on a preset priority determination strategy, the comprehensive demand indicators of each adjacent area are classified and processed to obtain the priority level of each adjacent area. The adjacent regions are sorted in descending order of priority to generate a priority sequence for selecting the current work location region, and the region sorting result of the clearing priority model is formed accordingly.

[0012] The present invention is further configured such that determining the target area in S4 and performing the clearing operation in S5 include: According to the priority sequence, each region is extracted as the target area for the current laser removal operation. During the clearing of the current target area, the laser clearing operation is captured in real time and dynamic non-target objects are identified. The dynamic non-target objects include: people, birds, animals and other moving objects that enter the laser clearing operation screen and do not belong to the vegetation to be cleared. Based on the current laser spot position, laser scanning path, and target area boundary, a laser irradiation area is generated that includes the current irradiation position and subsequent scanning path, and the laser irradiation area is extended outward by a preset range to form a safe expansion range; When a dynamic non-target object enters the laser irradiation area or the safe expansion range, the laser output is paused and the pause position is recorded. Once the dynamic non-target object leaves the laser irradiation area and the safe expansion range, the clearing process corresponding to the current target area will continue from the paused position. After clearing the current target area, mark the current target area as processed and continue to select the next area to be processed according to the priority sequence until all areas have been traversed and processed.

[0013] The present invention is further configured such that the determination of the non-edge transition region clearing strategy in S5 includes: When the target region's region type does not belong to the edge transition region: Based on vegetation features, an image classification method is used to identify the vegetation type of the target area. The vegetation type includes: weeds, shrubs, and non-cleared targets. Based on the region type of the target area and the identified vegetation type, a preset set of laser removal parameters is invoked; Configure the laser scanning path for the target area based on the laser removal parameter set; Perform a one-time removal process on the target area according to the laser scanning path.

[0014] The present invention is further configured such that the construction of the uncertainty band in S5 includes: When the target region's region type is an edge transition region: Based on the dam slope surface model, local image data is extracted from multi-view image data corresponding to the target area; Calculate information entropy based on the grayscale distribution of local image data; The boundary of the target region is expanded based on the information entropy value, generating an uncertain band whose width varies with the information entropy.

[0015] The present invention is further configured such that the multi-round progressive clearing strategy generation in S5 includes: A multi-round clearing strategy is implemented in the outer region of the uncertainty zone. The first round of laser clearing uses a preset low-energy short pulse mode to clear vegetation leaves and branches. After the first round of clearing is completed, local image data is re-acquired and the vegetation cover probability map of the local area is updated. By comparing the spatial position difference of the boundary pixel set in the two vegetation cover probability maps, the average displacement of the boundary pixel set is calculated as the boundary change. The advancement step size is determined proportionally based on the boundary change, and a preset medium-energy single-pulse laser is applied to the base of the newly exposed vegetation after clearing, under the condition of meeting the preset dam surface protection constraints. Repeatedly perform the clearing, image re-acquisition, boundary update, and advance operations to gradually shrink the cleared boundary towards the dam surface; The clearing will stop when the preset termination conditions are met. The termination conditions include: the vegetation coverage ratio is lower than a threshold, the boundary change is less than a threshold, or the residual vegetation height is lower than a threshold. The temperature of the target area is monitored in real time during the clearing process, and an early warning is triggered when the temperature exceeds the preset dam surface protection temperature.

[0016] This invention provides a method for laser vegetation removal on dam slopes based on visual recognition. The method comprises: S1: Acquiring multi-view image data and distance data of the dam slope; generating a dam slope surface model after camera calibration and visual reconstruction algorithm; calculating slope and curvature point-by-point; S2: Generating a vegetation cover probability map based on the multi-view image data; extracting vegetation features and spatial distribution features using image segmentation and classification methods; combining slope and curvature parameters to divide the dam slope surface into regions, generating region division results, and constructing region adjacency relationships. The region division results include edge transition regions and non-edge transition regions; S3: Constructing a removal priority model based on vegetation features, spatial features, and the adjacency relationships of the region to which the current work position belongs; calculating and sorting the priority of the adjacency regions of the region to which the current work position belongs, and obtaining the priority for the current work position. S4: Select the areas to be processed sequentially according to the priority sequence, and determine the currently selected area as the target area; S5: Perform the clearing work according to the target area type. When the target area is a non-edge transition area, a preset laser clearing strategy is adopted. When the target area is an edge transition area, an uncertainty band with a width that dynamically changes with the local image information entropy is constructed. A multi-round progressive clearing strategy is generated based on the uncertainty band, and the clearing is performed according to the multi-round progressive clearing strategy. During the laser clearing process, the laser clearing operation screen containing the target area and the surrounding area of ​​the laser irradiation area is acquired in real time. Dynamic non-target objects entering the laser clearing operation screen are identified. When a dynamic non-target object enters the laser irradiation area or the safe expansion range of the laser irradiation area, the laser output is paused. The beneficial effects include: By fusing multi-view image data with dam slope surface models, a comprehensive identification of vegetation distribution and dam surface structural features can be achieved. Based on this, regional division can be carried out, which helps to improve the accuracy of vegetation identification and the rationality of regional division, thus providing a reliable basis for subsequent clearing operations.

[0017] By constructing a clearing priority model based on regional adjacency and selecting target regions for processing sequentially according to the priority sequence, it is beneficial to optimize the clearing operation order, improve the overall operation efficiency, and enhance the coordination of the processing process between different regions.

[0018] By constructing an uncertainty zone in the edge transition area and employing a multi-round progressive removal strategy, it is beneficial to reduce the impact on the dam surface structure while ensuring the removal effect, thereby improving the stability and safety of the removal process.

[0019] By acquiring real-time images of the laser removal operation, including the target area and the surrounding area of ​​the laser irradiation area, and dynamically identifying people, birds, animals, and other moving objects that are not part of the vegetation to be removed, the laser output is paused when a non-target object enters the laser irradiation area or the safe expansion range. This helps reduce the risk of non-target objects being irradiated by the laser when they accidentally enter the removal image, thereby improving the on-site safety of laser vegetation removal operations on dam slopes.

[0020] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 A flowchart illustrating a method for laser removal of vegetation on dam slopes based on visual recognition, as shown in an exemplary embodiment of the present invention; Figure 2 This is a schematic diagram illustrating an application scenario of a laser-based vegetation removal method for dam slopes based on visual recognition, as an exemplary embodiment of the present invention. Detailed Implementation

[0022] The embodiments of the present invention will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention and not for limiting the scope of protection of the present invention.

[0023] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0024] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0025] Example: A method for laser removal of vegetation on dam slopes based on visual recognition, the process of which is as follows: Figure 1 As shown, the application scenarios are as follows Figure 2 As shown, it includes: S1: Acquire multi-view image data and distance data of the dam slope, generate a dam slope surface model by combining camera calibration and visual reconstruction algorithm, and calculate the slope and curvature point by point; S2: Generate a vegetation cover probability map based on multi-view image data, and extract vegetation features and spatial distribution features through image segmentation and classification methods. Combined with slope and curvature parameters, divide the dam slope surface into regions to generate region division results and construct region adjacency relationships. The region division results include: edge transition regions and non-edge transition regions. S3: Construct a clearing priority model based on vegetation features, spatial features, and the adjacent region relationship of the current work location. Calculate and sort the priority of the adjacent regions of the current work location to obtain a priority sequence for selecting the current work location region. S4: Select regions to be processed sequentially according to the priority sequence, and determine the currently selected region as the target region; S5: Perform the clearing operation according to the target area type. When the target area is a non-edge transition area, a preset laser clearing strategy is adopted. When the target area is an edge transition area, an uncertainty band with a width that dynamically changes with the local image information entropy is constructed. A multi-round progressive clearing strategy is generated based on the uncertainty band, and the clearing is performed according to the multi-round progressive clearing strategy. During the laser clearing process, the laser clearing operation screen containing the target area and the surrounding area of ​​the laser irradiation area is collected in real time. Dynamic non-target objects entering the laser clearing operation screen are identified. When a dynamic non-target object enters the laser irradiation area or the safe expansion range of the laser irradiation area, the laser output is paused.

[0026] The present invention is further configured such that S1 includes: The multi-view image data and distance data are obtained by simultaneously scanning the dam slope surface in a grid pattern using a gimbal-controlled high-magnification zoom camera and a laser rangefinder mounted in front of the dam slope. The camera calibration includes calibration processing of the internal and external parameters of the imaging device; The visual reconstruction algorithm uses the structure-of-motion reconstruction algorithm to fuse multi-view image data and distance data to generate a digital surface model of the dam slope containing three-dimensional coordinates. Based on the 3D coordinate data of the neighborhood of each pixel in the digital surface model of the dam slope, the slope of that point is calculated through local plane fitting, and the curvature of that point is calculated through local quadratic surface fitting. Specifically, for example... Figure 2As shown, the laser vegetation removal operation scenario on the dam slope includes the dam body, the dam slope surface, the operation channel, a pan-tilt-zoom camera, a laser rangefinder, a laser transmitter, and a control terminal. The pan-tilt-zoom camera, laser rangefinder, and laser transmitter are positioned at stable locations in front of the dam slope. The control terminal is communicatively connected to these components. The pan-tilt-zoom camera is used to acquire multi-view image data of the dam slope surface, the laser rangefinder is used to acquire distance data of the dam slope surface, and the laser transmitter is used to emit lasers towards the vegetation targets on the dam slope surface according to the laser removal strategy corresponding to the target area. The control terminal is used to perform area identification, removal strategy generation, laser scanning path control, and safety pause control. To achieve accurate perception before vegetation removal on the dam slope, it is first necessary to acquire the three-dimensional geometric information of the dam slope surface. A pan-tilt-zoom (PTZ) controlled data acquisition device is installed at a stable location in front of the dam slope (such as the opposite bank, the dam crest, or a dedicated support). This device integrates a high-magnification zoom camera and a laser rangefinder. The PTG can rotate the camera and rangefinder horizontally and vertically, thereby performing a grid-like scan of the dam slope surface. The acquisition process includes: the PTG moves along a preset trajectory, and the camera and laser rangefinder work synchronously; at each scanning grid point, the camera acquires a high-resolution image of that local area, while the laser rangefinder measures the straight-line distance from the device to that point on the dam slope surface; through grid-like scanning, a multi-view image sequence covering the entire dam slope and the corresponding distance data array are obtained. Camera calibration includes: before use, the camera needs to be calibrated; calibration is divided into intrinsic parameter calibration and extrinsic parameter calibration. Intrinsic parameters include focal length, principal point coordinates, lens distortion coefficient, etc., reflecting the imaging characteristics of the camera itself; extrinsic parameters include the camera's position and attitude in space (rotation angle, translation). The purpose of calibration is to eliminate image distortion caused by lens distortion and establish a mapping relationship between image pixel coordinates and world space coordinates. A common calibration method is to capture multiple images of the checkerboard calibration board from different angles, and obtain intrinsic and extrinsic parameters by detecting corner points and solving the projection equations. In this embodiment, while ensuring reconstruction accuracy, the camera focal length can be kept constant during the scanning process, and images with a uniform focal length can be used for subsequent reconstruction to reduce calibration complexity. 3D reconstruction includes: after obtaining multi-view images and distance data, using the Structure from Motion (SFM) algorithm for 3D reconstruction. This process extracts and matches feature points from the multi-view images, estimates their 3D positions by combining spatial geometric constraints, and then deduces the 3D spatial coordinates of each feature point. Distance data provided by the laser rangefinder can be used as initial values ​​or constraints to accelerate the solution process and improve reconstruction accuracy. Finally, all feature points form a sparse 3D point cloud.To obtain a continuous dam slope surface model, the point cloud needs to be meshed (e.g., through Delaunay triangulation) to generate a Digital Surface Model (DSM) containing the 3D coordinates of each mesh vertex. Each pixel (i.e., a mesh vertex) in this model corresponds to an actual location on the dam slope surface, with coordinates (X, Y, Z). Slope calculation involves: slope characterizing the inclination of the dam slope surface at a given point, typically represented by the angle between the tangent plane at that point and the horizontal plane. In the digital surface model, for each mesh vertex, the 3D coordinates of adjacent vertices within its neighborhood (e.g., a 3×3 or 5×5 window) are calculated, and these coordinates are fitted to a local plane (e.g., by finding the plane that minimizes the sum of the squared distances from all points to the plane using the least squares method). The angle between the normal vector of this plane and the vertical upward direction (the direction of gravity) is the slope at that point. A larger slope indicates a steeper surface. Curvature calculation involves defining the degree of curvature on the dam slope surface at a given point, i.e., the extent to which the surface deviates from a flat surface. In the digital surface model, each vertex and its neighborhood's 3D coordinates are used to select neighborhood data centered on that point for local fitting, resulting in a local quadratic surface (e.g., a quadratic polynomial surface). The coefficients of the quadratic surface reflect the surface's curvature in two mutually perpendicular directions. By analyzing the principal curvatures (maximum and minimum curvature) of the quadratic surface, the average curvature or Gaussian curvature at that point can be obtained. A larger curvature indicates a more uneven surface (e.g., pits formed by weathering or construction); zero curvature indicates a flat area. After these calculations, each grid point on the dam slope surface is associated with slope and curvature values, serving as the foundational data input for subsequent area division and clearing strategies, supporting subsequent processing steps.

[0027] The present invention is further configured such that the extraction of vegetation features and spatial distribution features in S2 includes: The vegetation features include: color features, texture features, and morphological features; A pixel-level segmentation-based image segmentation method is used to perform pixel-by-pixel classification on multi-view image data to obtain classification results. The classification results include the vegetation confidence score of each pixel belonging to a vegetation category. The vegetation confidence score of each pixel is extracted from the classification results of pixel-by-pixel classification to form a vegetation cover probability map. The classification results are mapped onto the digital surface model of the dam slope to obtain candidate vegetation areas; Feature extraction processing is performed on candidate vegetation areas. Color features are extracted by color space transformation, texture features are extracted by local texture analysis, and morphological features are extracted by connected component analysis. Based on a preset classification model, the extracted color features, texture features and morphological features are fused and judged to generate vegetation area identification results between vegetation areas and non-vegetation areas. Spatial statistical analysis is performed on the identified vegetation areas to obtain the distribution range, density, and connectivity of vegetation on the dam slope surface, forming corresponding spatial distribution features. Specifically, to achieve the extraction and spatial distribution analysis of vegetation features on the dam slope surface, a pixel-level segmentation-based deep learning semantic segmentation network is first used to perform pixel-by-pixel classification on multi-view image data. The semantic segmentation network can be U-Net or other network structures with pixel-level segmentation capabilities, pre-trained on the dam slope image dataset. The network outputs the vegetation confidence score for each pixel belonging to a vegetation category, with the vegetation confidence score ranging from 0 to 1, calculated by the Softmax activation function in the last layer of the network. The vegetation confidence scores of all pixels are arranged according to the original image coordinates to form a vegetation cover probability map. The binary classification results obtained by pixel-by-pixel classification (e.g., distinguishing by a preset threshold, where pixels with a vegetation confidence score greater than or equal to the threshold are classified as vegetation, and pixels with a confidence score less than the threshold are classified as non-vegetation; the preset threshold can be determined based on the model's classification performance on the validation dataset, for example by maximizing the balance between classification accuracy, F1 score, or precision and recall; for example, the threshold can be selected as 0.5 or a value within its vicinity) are mapped onto the digital surface model of the dam slope using the intrinsic and extrinsic parameters obtained from camera calibration. The set of points on the resulting 3D model is called the candidate vegetation region. Vegetation features are extracted from pixels within candidate vegetation areas. These features include color, texture, and morphological characteristics. Color features are extracted by converting the original RGB image to the HSV color space, allowing for the statistical analysis of hue and saturation distributions within the candidate vegetation area to characterize its color attributes. Texture features are extracted using the Gray-Level Co-occurrence Matrix (GLCM) method, which statistically analyzes and combines texture parameters across multiple directions within a predefined window to form a texture feature vector. Morphological features are extracted using connected component analysis, marking adjacent vegetation pixels within the candidate vegetation area as the same connected component, and extracting morphological features including area, perimeter, and shape description parameters. The color, texture, and morphological features are combined into a multi-dimensional feature vector, which is then input into a predefined classification model for fusion and discrimination. The classification model can be implemented using support vector machines, random forests, or neural network models. The parameters of the classification model can be optimized by training on labeled sample data and combining it with cross-validation to improve the ability to distinguish between different vegetation types. This model is trained using pre-labeled data samples and is used to output vegetation type labels, including weeds, shrubs, and non-cleared targets. The fusion discrimination is used to further refine the vegetation type identification results based on the pixel-level segmentation results.Non-removal targets include moss and artificial slope protection grasses. Their identification can be based on a comprehensive judgment combining geometric, color, and texture features. For example, they can be distinguished by height variation features, color distribution features, and regional morphological features. The judgment criteria can be obtained through statistical analysis of typical samples and can be adaptively adjusted according to different dam slope environments. Spatial statistical analysis is performed on the identified weed and shrub areas: the convex hull boundary coordinates and coverage area of ​​each vegetation area are calculated on the dam slope digital surface model to obtain the distribution range; the dam slope digital surface model is divided into grid cells of a preset scale, and the proportion of vegetation pixels to the total number of pixels in each grid is statistically analyzed to obtain the local vegetation coverage density; for any two vegetation areas, when their spatial distance meets the preset adjacency judgment criteria, the two areas are determined to be connected, and the connection relationship is recorded. The adjacency judgment criteria can be determined based on the minimum spatial distance or boundary contact relationship between areas. For example, when the minimum distance between areas is less than a preset distance threshold, they are determined to be connected. The distance threshold can be set according to the spatial resolution of the dam slope, thereby constructing a regional connectivity graph. The final output includes the type label, distribution range boundary coordinates, coverage area, local coverage density, and regional connectivity map for each vegetation region. These data constitute the spatial distribution characteristics.

[0028] The present invention is further configured such that generating the region partitioning result and constructing the region adjacency relationship in step S2 includes: Spatial correlation processing is performed between the vegetation cover probability map and the slope and curvature parameters in the dam slope digital surface model to generate a fused map; Based on the spatial distribution characteristics of vegetation cover probability in the fused map, and combined with the spatial boundary between vegetated and non-vegetated areas, areas located within the range of varying vegetation cover probability or where adjacent areas show transitional attribute changes are identified as edge transition areas, while the remaining areas are identified as non-edge transition areas. Specifically, the vegetation cover probability map generated in step S2 is first spatially correlated with the slope and curvature parameters calculated in step S1. Spatial correlation processing refers to the following: on the digital surface model of the dam slope, each grid point simultaneously possesses three attribute values: vegetation cover probability (range 0 to 1), slope (unit: degrees, range 0 to 90), and curvature (unit: 1 / meter, its range is adaptively determined according to the curvature calculation method and model scale). These three attribute values ​​are aligned according to the grid point coordinates to form a three-dimensional attribute vector, and the attribute vectors of all grid points constitute the fused map. Each grid point in the fused map stores a triplet (vegetation cover probability value, slope value, curvature value). The fused map is then used for regional division. First, two global thresholds are defined: a high-probability threshold and a low-probability threshold, adaptively determined based on the statistical distribution of vegetation cover probability. For example, these thresholds can be set to 0.7 and 0.3 respectively. All grid points are traversed, and the set of consecutive grid points with a vegetation cover probability greater than or equal to the high-probability threshold is marked as the vegetation core area, while the set of consecutive grid points with a vegetation cover probability less than or equal to the low-probability threshold is marked as the non-vegetation core area. Then, edge transition regions are identified: for any grid point, if its vegetation cover probability is between the low-probability and high-probability thresholds, or if the difference in vegetation cover probability between it and its adjacent grid points exceeds a preset gradient threshold (the gradient threshold is adaptively determined based on the probability change amplitude of adjacent areas; for example, it can be set to 0.2), then the grid point is marked as an edge transition point. A strip-shaped region composed of all consecutive adjacent edge transition points is defined as the edge transition region, where consecutive adjacency is determined using connected component analysis. The remaining grid points, which belong neither to the vegetation core area nor the non-vegetation core area and are not marked as edge transition points, are assigned to the corresponding region based on the principle of minimizing the difference in vegetation cover probability between their vegetation cover probability and that of their adjacent regions. Finally, the vegetated core areas were merged into vegetated cover areas, the non-vegetated core areas were merged into bare base areas, and the continuous strip-shaped areas formed by edge transition points were retained as edge transition areas. After completing the region division, no new region types were introduced; all regions were limited to vegetated cover areas, bare base areas, and edge transition areas. The boundary location information of each region was recorded: for each continuous region (vegetated cover area, bare base area, edge transition area), the boundary contour was extracted using a convex hull algorithm or an eight-neighbor boundary tracing algorithm, resulting in a series of boundary coordinate point sequences.Then, a regional adjacency structure is constructed: for any two regions, a determination is made based on grid adjacency. When two regions have a direct adjacency relationship at the grid level, they are determined to be adjacent regions, and a record is added to the adjacency table, containing the identifier of region A, the identifier of region B, and the shared boundary length (in meters, which is obtained by statistically analyzing adjacent grid boundaries). The final output of the regional division results includes the type label of each region (vegetated area, bare base area, edge transition area), the set of boundary coordinate points, and the regional adjacency table (containing all adjacent region pairs and their shared boundary lengths).

[0029] The present invention is further configured such that the subdivision of the non-edge transition region includes: Continuous areas with a vegetation cover probability higher than a preset high threshold are classified as vegetation cover areas, and continuous areas with a vegetation cover probability equal to or lower than a preset low threshold are classified as bare base areas. Based on the preset dam structure location information, the drainage ditch, expansion joints and the area around the monitoring equipment are marked as key areas for removal; Record the boundary location information of each region and construct the region adjacency structure based on the spatial contact relationship and boundary connection relationship between regions. Specifically, the further subdivision of non-edge transition regions and the recording of region boundaries and the construction of adjacency relationships are performed according to the following steps: First, divide continuous regions with vegetation cover probability higher than a preset high threshold into vegetation cover regions, and divide continuous regions with vegetation cover probability equal to or lower than a preset low threshold into bare base regions. The preset high threshold and preset low threshold are adaptively determined according to the statistical distribution of vegetation cover probability, and can be set to 0.7 and 0.3 respectively as an example. Each grid point in the vegetation cover probability map stores a probability value between 0 and 1. Traverse all grid points and use the connected component analysis algorithm (eight-neighbor connectivity) to extract all continuous regions. Before connected component extraction, first perform pixel-level initial classification of grid points based on the high threshold and low threshold, where grid points greater than or equal to the high threshold are marked as vegetation class, grid points less than or equal to the low threshold are marked as bare class, and the remaining grid points do not participate in the division of non-edge transition regions. For each continuous region, the region is labeled based on the consistency of the grid points within the region. When all grid points within the region belong to the same category, the region type is directly determined, avoiding classification bias caused by the average probability. If inconsistencies exist, the region is re-divided into multiple sub-regions according to the grid point categories. If the average probability is between 0.3 and 0.7, the region should not appear in the non-edge transition region (already classified as an edge transition region in the previous step). This step only processes grid points already marked as non-edge transition regions. The second step, based on the preset dam structure location information, marks the drainage ditches, expansion joints, and the areas surrounding the monitoring equipment as key areas for removal. The preset dam structure location information is obtained from dam design drawings or field survey data and stored in three-dimensional coordinates. For drainage ditches, the centerline coordinate sequence is obtained; for expansion joints, the centerline coordinate sequence of the joints is obtained; for monitoring equipment, the center coordinates of its installation point are obtained. On the digital surface model of the dam slope, a spatial proximity influence range is generated with the structural location as a reference. The scale of the influence range is adaptively determined according to the dam slope model resolution and operational accuracy requirements; for example, an area of ​​approximately 0.5 meters can be used. All grid points within the influence range are marked as key removal areas. For linear structures such as drainage ditches and expansion joints, discrete sampling is performed along the centerline at a preset spatial sampling interval. The sampling interval is adaptively determined according to the model resolution; for example, approximately 0.1 meters can be used, and a corresponding influence range is generated at each sampling point. If a grid point belongs to both a vegetation-covered area and a bare base area, the original mark is overwritten by the key removal area (i.e., the key removal area has the highest priority). The key removal area is used as an exemplary additional mark to guide subsequent operational control and does not participate in the basic area division result output as an independent area type. The third step is to record the boundary location information of each area.For each independent region (including vegetated areas, bare base areas, edge transition areas, and key clearing areas), the eight-neighbor boundary tracing algorithm is used to extract its outer boundary contour. Specifically, starting from the lower left grid point of the region, adjacent boundary points are traversed clockwise, and the 3D coordinates of each boundary point (obtained from the dam slope digital surface model) are recorded. All boundary points are arranged in traversal order to form a closed sequence of boundary coordinate points. For regions with holes (such as a small patch of bare base within a vegetated area), the same method is used to extract both the outer and inner boundaries, with the inner boundary recorded separately as the hole boundary. The fourth step involves constructing a regional adjacency structure based on the spatial contact and boundary connection relationships between regions. Spatial contact refers to the existence of a common line segment or common point on the boundaries of two regions in space. For any two regions, the boundary coordinate point sequences of both regions are obtained. Adjacency determination is performed based on grid topology; when two regions have directly adjacent boundary grid points at the grid level, they are considered adjacent regions. Further calculation of the shared boundary length: The shared boundary length is obtained by counting the number of adjacent boundary grid points and converting the data based on the grid scale, where the grid scale is determined according to the resolution of the dam slope digital surface model. A new record is added to the adjacency table, containing the unique identifiers of region A and region B, and the shared boundary length. If multiple unconnected shared boundary segments exist between two regions (e.g., two regions are adjacent at two points above and below each other), these are recorded as multiple adjacency relationships. The final output of the region division results includes: a type label for each region (vegetated area, bare base area, edge transition area, key clearing area marker information), a sequence of boundary coordinate points, and the adjacency table.

[0030] The present invention is further configured such that S3 includes: Based on vegetation characteristics, spatial distribution characteristics, and regional adjacency structure, data organization and processing are performed on the adjacent regions of the current work location to construct a regional demand data set; For each adjacent region in the regional demand data set, vegetation features and spatial distribution features are combined and processed according to preset feature fusion rules to form a comprehensive demand index. Based on a preset priority determination strategy, the comprehensive demand indicators of each adjacent area are classified and processed to obtain the priority level of each adjacent area. The adjacent regions are sorted in descending order of priority to generate a priority sequence for selecting the current work location, thus forming the region sorting result of the clearing priority model. Specifically, the construction of the clearing priority model and the generation of the priority sequence are performed according to the following steps: First, based on vegetation characteristics, spatial distribution characteristics, and regional adjacency relationship structure, the adjacent regions of the current work location are organized and processed to construct a regional demand data set. The current work location refers to the spatial location of the laser clearing equipment, which corresponds to a grid point on the digital surface model of the dam slope. This grid point belongs to a pre-defined region (called the current region). From the regional adjacency relationship table output in step S2, all adjacent regions that share a boundary with the current region are found, and these adjacent regions are extracted to form an adjacent region list. For each region in the adjacent region list, obtain the following data from the output of step S2: region type label (vegetation-covered region, bare base region, edge transition region, key removal region marker information), vegetation cover density (unit: dimensionless, value range 0 to 1, the proportion of vegetation pixels to total pixels in the region), main vegetation type in the region (weeds, shrubs, or non-removal target), average slope of the region (unit: degrees, the arithmetic mean of the slopes of all grid points in the region), average curvature of the region (unit: 1 / meter, the arithmetic mean of the curvatures of all grid points in the region), and regional spatial location relationship parameters (used to characterize the relative spatial relationship between the region and key structures or key operation locations, which can be in the form of distance or adjacency). Organize these data according to the one-to-one correspondence between regions to form a region requirement data set. Each element contains (region identifier, region type, vegetation cover density, main vegetation type, average slope, average curvature, and regional spatial location relationship parameters). The second step involves combining vegetation features and spatial distribution features of adjacent regions within the regional demand data set according to preset feature fusion rules to form a comprehensive demand index. The preset feature fusion rules employ a weighted linear fusion method. First, each feature parameter is normalized, mapping it to a range of 0 to 1. Vegetation cover density itself is between 0 and 1 and requires no normalization. Average slope is normalized based on its range or statistical distribution (e.g., normalization by maximum slope or quantile). Average curvature is normalized based on its statistical range (e.g., normalization by maximum absolute value or quantile after taking the absolute value). Regional spatial location parameters are normalized based on their range (e.g., when using distance metrics, the normalization scale can be determined based on the operational area). The contribution of the main vegetation type is reflected through coefficients: if the main vegetation type is shrubs, the vegetation intensity coefficient is 0.8; if it is weeds, it is 0.4; if it is not a clearing target, it is 0.The influence of region type is reflected through weighting factors: the weighting factors corresponding to different region types are preset parameters that can be set according to actual operational needs. For example, vegetated areas, bare base areas, and edge transition areas can each take different weight values. The formula for calculating the comprehensive demand index is: Comprehensive demand index = Region type weighting factor × (w1 × Vegetation cover density + w2 × Vegetation intensity coefficient + w3 × Spatial location relationship parameter + w4 × Topographic feature parameter), where w1, w2, w3, and w4 are preset fusion weighting parameters that satisfy a weight sum of 1. These can be set based on historical data or experience; for example, they can be 0.4, 0.3, 0.2, and 0.1. The topographic feature parameter is obtained by combining normalized slope and normalized curvature. The comprehensive demand index for each adjacent region is calculated. In the third step, based on the preset priority determination strategy, the comprehensive demand index of each adjacent region is graded and discriminated to obtain the priority level corresponding to each adjacent region. The preset priority determination strategy divides the comprehensive demand index into multiple levels: the threshold for each level is a preset parameter that can be set according to the distribution of the comprehensive demand index. For example, a segmented interval method can be used (e.g., high, medium, and low levels). The fourth step involves sorting adjacent areas in descending order of priority level to generate a priority sequence for selecting the current work location area, thus forming the area sorting result of the priority clearing model. Sorting rules: First, sort by priority level, with high priority level first, medium priority level second, and low priority level last. Within the same priority level, sort by comprehensive demand index value from largest to smallest in descending order. If the comprehensive demand index values ​​are the same, sort by distance from the current work location (using the Euclidean distance from the nearest point on the area boundary to the current work location) from smallest to largest in ascending order. Finally, an ordered priority sequence is generated, where each element contains an area identifier and its corresponding priority level.

[0031] The present invention is further configured such that determining the target area in S4 and performing the clearing operation in S5 include: According to the priority sequence, each region is extracted as the target area for the current laser removal operation. During the clearing of the current target area, the laser clearing operation is captured in real time and dynamic non-target objects are identified. The dynamic non-target objects include: people, birds, animals and other moving objects that enter the laser clearing operation screen and do not belong to the vegetation to be cleared. Based on the current laser spot position, laser scanning path, and target area boundary, a laser irradiation area is generated that includes the current irradiation position and subsequent scanning path, and the laser irradiation area is extended outward by a preset range to form a safe expansion range; When a dynamic non-target object enters the laser irradiation area or the safe expansion range, the laser output is paused and the pause position is recorded. Once the dynamic non-target object leaves the laser irradiation area and the safe expansion range, the clearing process corresponding to the current target area will continue from the paused position. After clearing the current target area, the current target area is marked as processed, and the next area to be processed is selected according to the priority sequence until all areas have been traversed. Specifically, an ordered list is obtained from the priority sequence output in step S3. Each element in the ordered list contains a region identifier and a priority level, and the ordered list is arranged from high to low priority. An integer index variable is set, with an initial value of 0, pointing to the first element of the priority sequence. The region corresponding to the region identifier in the element pointed to by the integer index variable is extracted as the target area for the current laser clearing operation. The spatial range of the target area is defined by the boundary coordinate point sequence output in step S2. The boundary coordinate point sequence is an ordered set of three-dimensional coordinate points, each containing X, Y, and Z coordinates in meters. The laser clearing equipment controls the pan-tilt rotation angle and laser emission direction according to the boundary coordinate point sequence of the target area, so that the laser spot covers the current target area. The clearing operation is then performed on the current target area. The specific execution method of the removal work is defined in step S5: if the target area is labeled as a vegetation-covered area, a bare base area, or a key removal area, a preset laser removal strategy is adopted; if the target area is labeled as an edge transition area, an uncertainty zone is constructed and a multi-round progressive removal strategy is generated. Both the preset laser removal strategy and the multi-round progressive removal strategy include corresponding laser scanning paths, which include the scan point sequence, scan direction, scan line spacing, scan speed, and scan order. The laser scanning path corresponding to the current removal strategy is used to determine the current position of the laser spot within the current target area and the subsequent scanning path. During the removal process in the current target area, the control terminal acquires the laser removal operation image in real time using a PTZ-controlled high-magnification zoom camera. The laser removal operation image includes the current target area, the current laser spot position, the surrounding area of ​​the laser irradiation area, and the subsequent scanning path corresponding to the current removal strategy. The control terminal calls a preset dynamic non-target object recognition model to perform target detection or semantic segmentation on the laser removal operation image. The dynamic non-target object recognition model can use a YOLO series target detection model, a Mask R-CNN semantic segmentation model, or other visual recognition models with target detection capabilities. The dynamic non-target object recognition model outputs the dynamic non-target object category, recognition confidence score, bounding box coordinates, and target mask. Dynamic non-target objects include people, birds, animals, and other moving objects that enter the laser removal operation area but do not belong to the vegetation to be removed. The default recognition confidence score threshold is 0.6; when the object to be identified is a person, the default recognition confidence score threshold is 0.5. The recognition confidence score threshold can be adjusted according to the on-site false detection rate, false negative rate, and laser safety level, with a value range of 0.4 to 0.8. The control terminal generates the laser irradiation area based on the current laser spot position, the laser scanning path corresponding to the current removal strategy, and the target area boundary.The laser irradiation area includes the area currently covered by the laser spot and the scanning path area to be reached within the subsequent preset time window. The subsequent preset time window is 0.5 seconds by default, with a range of 0.2 seconds to 1 second, determined by the gimbal response time, laser shutdown response time, and scanning speed. The control terminal extends the laser irradiation area outward by a preset range to form a safe outer range. The preset range is 0.2 meters by default, or three times the diameter of the laser spot, whichever is greater. The preset range can be adjusted according to laser power, operating distance, dam slope spatial resolution, and on-site safety requirements. The control terminal performs spatial cross-referencing between the boundary coordinates or target mask of the dynamic non-target object and the laser irradiation area and the safe outer range. When the boundary coordinates or target mask of the dynamic non-target object overlaps with the laser irradiation area, or when the boundary coordinates or target mask of the dynamic non-target object enters the safe outer range, the dynamic non-target object is determined to have entered the danger zone. As an example, when the dynamic non-target object category is a person, it is determined to be in a danger zone as soon as the person's bounding box or target mask enters the safety perimeter. When the dynamic non-target object category is a bird, animal, or other moving object, the determination can be made based on the overlap ratio between the bounding box or target mask and the safety perimeter. The default threshold for the overlap ratio is 5%, with a range of 1% to 10%. When a dynamic non-target object is determined to have entered the danger zone, the control terminal immediately sends a pause command to the laser emitter, causing the laser emitter to pause laser output and record the pause position. The pause position includes the current target area identifier, the current clearing strategy type, the current scan point number, the current scan line number, the current laser spot position, the incomplete scan path, and the pause time. The pause position is used to determine the location to continue clearing after the danger has passed, avoiding repeated irradiation of already cleared locations. In the paused state, the control terminal continues to acquire laser clearing operation footage and continuously performs dynamic non-target object identification. When the dynamic non-target object leaves the laser irradiation area and the safety perimeter, and no dynamic non-target object is detected entering the danger zone again within a continuously preset safety recovery time, the control terminal determines that the danger has passed. The default safe recovery time is 3 seconds, with a range of 1 to 5 seconds. The value is determined based on the moving speed of the dynamic non-target object, the on-site wind speed, the operating distance, and the laser safety level. After the danger is cleared, the control terminal resumes the clearing process for the current target area from the paused position. If the current target area is not an edge transition area, the preset laser clearing strategy continues from the paused position. If the current target area is an edge transition area, local image data is reacquired, and a multi-round progressive clearing strategy continues based on the paused position. After clearing the current target area, a processed status is marked. The processed status mark is implemented by adding a Boolean field to the area data structure, named "Cleared Flag," with an initial value of False. After clearing is complete, the "Cleared Flag" is set to True.Simultaneously, the element corresponding to the current target region in the priority sequence is marked as completed. This is done by adding a completion status attribute to the element and setting it to true. The integer index variable is incremented by 1, pointing to the next element in the priority sequence. The process checks if the integer index variable is less than the length of the priority sequence: if so, the next element is extracted as the target region, and the clearing, security monitoring, pause control, resume control, and marking steps are repeated; if the integer index variable equals the length of the priority sequence, the entire region traversal is considered complete. In the traversal termination determination, the comparison between the integer index variable and the priority sequence length serves as the termination condition. The logic for the termination condition is as follows: when the integer index variable's value range is [0, N-1], N is the priority sequence length; when the integer index variable equals N, the traversal is considered complete. For example, when the priority sequence length is 10, the integer index variable increments from 0 to 9, corresponding to the processing of 10 regions. When the integer index variable increments to 10, it indicates that the entire region processing is complete. After the entire region traversal is complete, a clearing job completion signal is output.

[0032] The present invention is further configured such that the determination of the non-edge transition region clearing strategy in S5 includes: When the target region's region type does not belong to the edge transition region: Based on vegetation features, an image classification method is used to identify the vegetation type of the target area. The vegetation type includes: weeds, shrubs, and non-cleared targets. Based on the region type of the target area and the identified vegetation type, a preset set of laser removal parameters is invoked; Configure the laser scanning path for the target area based on the laser removal parameter set; A one-time cleanup process is performed on the target area according to the laser scanning path. Specifically, when the target area type label is a vegetation-covered area, a bare base area, or a key cleanup area, the following one-time cleanup process is performed: The color features, texture features, and morphological features of all pixels within the target area are obtained from the vegetation feature data extracted in step S2. Color features use the mean hue and mean saturation in the HSV color space. The mean hue ranges from 0 to 360 degrees, and the mean saturation ranges from 0 to 1. Texture features use the gray-level co-occurrence matrix method. The window size can be set according to the image resolution; for example, a 7×7 pixel window can be used. The contrast, correlation, energy, and homogeneity parameters are calculated in four directions: 0 degrees, 45 degrees, 90 degrees, and 135 degrees. The average value of these four directions is taken as the texture feature vector. Morphological features use connected component analysis to extract the connected component area (unit: pixels), perimeter (unit: pixels), and circularity. Circularity is equal to 4π multiplied by the area divided by the square of the perimeter. The above features are combined into a feature vector (the dimension can be set according to the selected features; for example, 15 dimensions can be used), and input into a pre-trained classification model (for example, a support vector machine classifier can be used, and the kernel function can be a radial basis function). The number of training samples and the distribution of categories can be constructed according to the actual application scenario; for example, about 2000 samples can be used (such as weeds, shrubs, and non-cleared targets distributed in a certain proportion). The classifier outputs vegetation type labels: 0 represents weeds, 1 represents shrubs, and 2 represents non-cleared targets. For the identification of non-cleared targets, it can be determined based on the output results of the classification model, or it can be combined with prior feature rules for auxiliary discrimination. The prior feature rules, for example, may include: the height change in the depth map is less than a certain threshold, a specific hue and saturation range, high texture homogeneity, and the area and shape of the connected components meeting preset conditions; wherein, the height change threshold can be set according to the scale of the dam surface structure; for example, it can be about 2 centimeters. Based on the target area type and the identified vegetation type, a preset set of laser removal parameters is invoked: For shrubs in vegetated areas, the laser power, pulse width, pulse repetition frequency, single pulse energy, and scanning speed are preset parameters, which can be determined according to vegetation density and material characteristics. For example, these can be 300 watts, 10 milliseconds, 20 Hz, 6 joules, and 0.5 m / s, respectively. For weeds in vegetated areas, the above parameters can be adjusted according to the characteristics of low-intensity vegetation. For example, these can be 150 watts, 5 milliseconds, 50 Hz, 0.75 joules, and 2 m / s, respectively. For shrubs in key removal areas, the parameters can be adjusted according to priority and removal intensity requirements. For example, these can be 350 watts, 12 milliseconds, 15 Hz, 4.2 joules, and 0.3 m / s, respectively. For weeds in key removal areas, these can be 200 watts, 6 milliseconds, 40 Hz, 0.8 joules, and 1.5 m / s, respectively. No laser is emitted for bare surfaces. No laser is emitted for non-removal targets.If the target area contains multiple vegetation types, they are processed according to their area proportions: shrub areas are cleared first, followed by weed areas. The laser scanning path corresponding to the current clearing strategy uses a preset scanning method (for example, a raster scanning method can be used), scanning line by line from left to right and from top to bottom within the convex hull of the target area boundary. The laser spot diameter can be set according to the device parameters; for example, it can be about 5 cm. The overlap rate of adjacent scanning lines can be set according to the clearing coverage requirements; for example, it can be about 20% (corresponding to a line spacing of about 4 cm). The scanning point spacing in the shrub area can be set according to the clearing accuracy requirements; for example, it can be about 2.5 cm, and the scanning point spacing in the weed area can be about 4 cm for example. The scanning path generation algorithm calculates the bounding box of the target area (minimum X, maximum X, minimum Y, maximum Y), generates grid points with the corresponding step size, and removes points outside the boundary to obtain the scanning point sequence. Lasers are emitted sequentially at each scanning point along the scanning path, with each scanning point emitting only one laser pulse (weeds) or one pulse train (shrubs). During the removal process, an infrared thermal imager is used to monitor the target area's temperature in real time at a sampling frequency that can be set according to real-time requirements. For example, approximately 10 Hz can be used. If the temperature exceeds the dam surface protection temperature threshold, the laser output is interrupted and an alarm is triggered. This dam surface protection temperature threshold can be determined based on the dam material's temperature tolerance and safety margin; for example, approximately 60 degrees Celsius can be used. After laser emission is completed at all scanning points, the target area is cleared.

[0033] The present invention is further configured such that the construction of the uncertainty band in S5 includes: When the target region's region type is an edge transition region: Based on the dam slope surface model, local image data is extracted from multi-view image data corresponding to the target area; Calculate information entropy based on the grayscale distribution of local image data; The target region boundary is expanded based on the information entropy value, generating an uncertainty band whose width varies with the information entropy. Specifically, when the target region type label is edge transition region, the following operations are performed: The spatial extent of the target region is obtained from the dam slope digital surface model, defined by the boundary coordinate point sequence output in step S2. Based on the spatial extent of the target region, local image data corresponding to the spatial extent of the target region is extracted from multi-view image data. The extraction method is as follows: using the intrinsic and extrinsic parameters obtained from camera calibration, the target region boundary coordinate point sequence is projected onto the image plane to obtain a projected polygon on the image. A viewpoint image with a high matching degree to the projected region is selected from the multi-view images, or the image region within the projected polygon is cropped after fusing multiple viewpoint images, serving as the local image data. The local image data is converted to a grayscale image. The conversion formula is: grayscale value equals 0.299 multiplied by the red component value plus 0.587 multiplied by the green component value plus 0.114 multiplied by the blue component value. The values ​​of the red, green, and blue components all range from 0 to 255. The probability distribution of each gray level (256 gray levels from 0 to 255) in a grayscale image is obtained by counting the number of pixels at each gray level and dividing by the total number of pixels in the local image. This probability distribution is denoted by P(g), where g represents the gray level. The information entropy H is calculated by multiplying the sum of P(g) from g (0 to 255) by the logarithm of P(g) to the base 2, with H equal to negative. Gray levels where P(g) is 0 are not included in the summation, or the corresponding terms of the logarithm of P(g) to the base 2 are treated as 0. The information entropy ranges from 0 to 8 bits. Information entropy is used to characterize the complexity of the texture and the clarity of the boundaries of a target area. Higher information entropy usually indicates a significant difference between vegetation and the dam surface, with clear boundaries; lower information entropy usually indicates a more concentrated gray level distribution and blurred or transitional boundaries. When setting the basic expansion width and information entropy reference value, it can be determined based on the characteristics of the dam surface material, image resolution, and vegetation distribution. The information entropy reference value can be determined based on the mean or boundary point of the entropy distribution between clearly defined and blurred regions obtained from typical sample statistics; for example, approximately 4 bits can be used. The basic expansion width can be determined based on the influence range of the laser effect and the safety margin; for example, approximately 0.1 meters can be used. When the information entropy is greater than or equal to 4 bits, the expansion width is the basic expansion width multiplied by a coefficient of 0.5, i.e., 0.05 meters. When the information entropy is less than 4 bits, the expansion width is the basic expansion width multiplied by a coefficient (1 plus (4 minus information entropy) divided by 4), i.e., the expansion width equals 0.1 multiplied by (1 plus (4 minus H) divided by 4), with a maximum expansion width of 0.2 meters. The relationship between the above expansion width and information entropy can be implemented using a linear or non-linear mapping method; the expression described is an example implementation.Each boundary point on the target area boundary is expanded outward along the boundary normal (pointing away from the center of the target area) by a certain width, resulting in an expanded boundary point. Simultaneously, it is expanded inward along the boundary normal (pointing away from the center of the target area) by half the width, resulting in a contracted boundary point. The ratio of outward expansion to inward expansion can be set according to safety protection requirements; for example, 1 and 0.5 can be used respectively. The expanded boundary points are connected to form the outer boundary of the uncertainty zone, and the contracted boundary points are connected to form the inner boundary of the uncertainty zone. The strip-shaped area between the outer and inner boundaries is the uncertainty zone. The spatial range of the uncertainty zone is defined by the sequence of coordinate points of the outer and inner boundaries. Finally, the sequence of coordinate points of the outer and inner boundaries of the uncertainty zone is output for subsequent multi-round progressive clearing strategies. In the above process, the basic expansion width of 0.1 meters, the information entropy reference value of 4 bits, the outward expansion coefficient of 1, and the inward expansion coefficient of 0.5 are all preset values, which can be adjusted on-site according to the actual dam surface material and laser performance.

[0034] The present invention is further configured such that the multi-round progressive clearing strategy generation in S5 includes: A multi-round clearing strategy is implemented in the outer region of the uncertainty zone. The first round of laser clearing uses a preset low-energy short pulse mode to clear vegetation leaves and branches. After the first round of clearing is completed, local image data is re-acquired and the vegetation cover probability map of the local area is updated. By comparing the spatial position difference of the boundary pixel set in the two vegetation cover probability maps, the average displacement of the boundary pixel set is calculated as the boundary change. The advancement step size is determined proportionally based on the boundary change, and a preset medium-energy single-pulse laser is applied to the base of the newly exposed vegetation after clearing, under the condition of meeting the preset dam surface protection constraints. Repeatedly perform the clearing, image re-acquisition, boundary update, and advance operations to gradually shrink the cleared boundary towards the dam surface; The clearing will stop when the preset termination conditions are met. The termination conditions include: the vegetation coverage ratio is lower than a threshold, the boundary change is less than a threshold, or the residual vegetation height is lower than a threshold. During the clearing process, the temperature of the target area is monitored in real time, and an early warning is triggered when the temperature exceeds the preset dam surface protection temperature. Specifically, when the target area type label is edge transition area and the uncertainty zone has been constructed, the following multi-round clearing strategy is executed. The first round of laser clearing targets the area outside the uncertainty zone, which refers to the area between the outer boundary of the uncertainty zone and the original boundary of the target area. The first round of laser clearing uses a low-energy short pulse mode, with a laser power of 100 watts, a pulse width of 3 milliseconds, a pulse repetition frequency of 100 Hz, and a single pulse energy of 0.3 joules. The laser scanning path corresponding to the first round of clearing strategy uses a raster scanning method, with a scanning point spacing of 5 cm. The clearing target is vegetation leaves and small branches, while preserving the main stem structure. The above low-energy laser parameters can be set according to the vegetation type, density, and laser performance; the above values ​​are used as an example. After the first round of clearing is completed, local image data is re-acquired. Local image data refers to multi-view images corresponding to the current processing area. A high-magnification zoom camera controlled by a pan-tilt unit is used to perform high-resolution imaging of the current processing area. Based on the newly acquired images, a semantic segmentation network is used to regenerate the vegetation cover probability map of the local area. By comparing the spatial differences of the boundary pixel set in the two vegetation cover probability maps, the boundary pixel set can be defined as an isotropic region with a vegetation cover probability close to 0.5, or the boundary pixel set extracted by setting a probability threshold range (0.45 to 0.55 for example). The displacement vector of each pixel in the boundary pixel set in the two maps is calculated. The displacement vector can be determined by nearest neighbor matching or contour distance-based matching. The arithmetic mean of all displacement vectors is taken as the boundary change, which is represented by ΔB in meters. The value of ΔB is generally between 0 and 0.5 meters. The advancement step size is determined proportionally based on the boundary change ΔB. The formula for calculating the advancement step size is step = min(ΔB × 1.2, 0.03 meters), where 1.2 is the gain coefficient and 0.03 meters is the maximum single inward distance. The advancement step size can be adaptively determined according to the boundary change trend. The above calculation formula is an example implementation. The gain coefficient and the maximum single inward distance can be set according to the equipment response characteristics, image resolution, and safety constraints. For example, they can be approximately 1.2 and approximately 0.03 meters, respectively. Under the condition of satisfying the dam surface protection constraints, a medium-energy single-pulse laser is applied to the base of the newly exposed vegetation after clearing.The dam surface protection constraints are as follows: the dam surface temperature is monitored in real time using an infrared thermal imager, and the dam surface temperature must not exceed 60 degrees Celsius. If the dam surface temperature exceeds 55 degrees Celsius, the energy of the next pulse is reduced to 80% of the original energy. If it exceeds 60 degrees Celsius, the laser output is immediately interrupted and an alarm is triggered. The temperature threshold can be determined based on the heat resistance characteristics and safety margin of the dam surface material. For example, approximately 55 degrees Celsius can be used as the warning threshold and approximately 60 degrees Celsius as the protection upper limit. The medium-energy single-pulse laser parameters are: laser power 200 watts, pulse width 6 milliseconds, single-pulse energy 1.2 joules, and the laser beam is always directed at the base of the vegetation without directly irradiating the dam surface. The medium-energy laser parameters can be set according to the vegetation structure and the required removal depth. For example, the above values ​​can be used. The removal, image re-acquisition, boundary update, and advancement operations are repeatedly performed, and the laser scanning path corresponding to the current round of removal strategy is updated according to the advancement step size of each round, so that the removal boundary gradually shrinks towards the dam surface. The specific process for each round of operation is as follows: First, apply a medium-energy single-pulse laser to the base of the newly exposed vegetation. Then, re-acquire images, update the vegetation cover probability map, calculate the new boundary change, and determine the next round's advancement step size based on the new boundary change. Clearing stops when a preset termination condition is met. The termination condition includes any one of the following three conditions: The first termination condition is that the vegetation cover ratio is less than 5%. The vegetation cover ratio refers to the proportion of pixels with a vegetation cover probability greater than 0.5 within the current uncertainty zone to the total number of pixels. The vegetation cover ratio threshold can be set according to the clearance completion requirements; for example, it can be approximately 5%. The second termination condition is that the boundary change ΔB is less than 0.005 meters. This threshold can be determined based on the image resolution and the system's minimum detectable displacement; for example, it can be approximately 0.005 meters. The third termination condition is that the residual vegetation height is less than 2 centimeters. The residual vegetation height is obtained through depth map information, taking the range of depth value changes (vertical distance from the top of the vegetation to the dam surface) corresponding to all vegetation pixels within the uncertainty zone. The height threshold can be set according to the allowable residual standard and dam surface protection requirements; for example, it can be approximately 2 centimeters. During the removal process, the temperature of the target area is monitored in real time. Temperature data is acquired by an infrared thermal imager at a sampling frequency of 10 Hz. When the temperature at any point is detected to be higher than 60 degrees Celsius, an early warning is immediately triggered and the laser output is interrupted. At the same time, the location of the over-temperature is recorded to prevent the laser from irradiating the same location in the future. The sampling frequency can be set according to the temperature change response speed requirements. For example, it can be about 10 Hz.All the preset parameters mentioned above include: low-energy short-pulse mode laser power 100 watts, pulse width 3 milliseconds, repetition frequency 100 Hz; medium-energy single-pulse laser power 200 watts, pulse width 6 milliseconds, single-pulse energy 1.2 joules; maximum single-pulse inward distance 0.03 meters; gain coefficient 1.2; dam surface protection temperature 60 degrees Celsius, warning trigger temperature 55 degrees Celsius; vegetation coverage threshold 5%; boundary change threshold 0.005 meters; residual vegetation height threshold 2 centimeters; infrared thermal imager sampling frequency 10 Hz. These parameters can be calibrated and adjusted on-site according to the actual laser model, dam surface material, and vegetation type.

[0035] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for laser removal of vegetation on dam slopes based on visual recognition, characterized in that, include: S1: Acquire multi-view image data and distance data of the dam slope, generate a dam slope surface model by combining camera calibration and visual reconstruction algorithm, and calculate the slope and curvature point by point; S2: Generate a vegetation cover probability map based on multi-view image data, and extract vegetation features and spatial distribution features through image segmentation and classification methods. Combined with slope and curvature parameters, divide the dam slope surface into regions to generate region division results and construct region adjacency relationships. The region division results include: edge transition regions and non-edge transition regions. S3: Construct a clearing priority model based on vegetation features, spatial features, and the adjacent region relationship of the current work location. Calculate and sort the priority of the adjacent regions of the current work location to obtain a priority sequence for selecting the current work location region. S4: Select regions to be processed sequentially according to the priority sequence, and determine the currently selected region as the target region; S5: Perform the clearing operation according to the target area type. When the target area is a non-edge transition area, a preset laser clearing strategy is adopted. When the target area is an edge transition area, an uncertainty band with a width that dynamically changes with the local image information entropy is constructed. A multi-round progressive clearing strategy is generated based on the uncertainty band, and the clearing is performed according to the multi-round progressive clearing strategy. During the laser clearing process, the laser clearing operation screen containing the target area and the surrounding area of ​​the laser irradiation area is collected in real time. Dynamic non-target objects entering the laser clearing operation screen are identified. When a dynamic non-target object enters the laser irradiation area or the safe expansion range of the laser irradiation area, the laser output is paused.

2. The method for laser removal of vegetation on dam slopes based on visual recognition according to claim 1, characterized in that, S1 includes: The multi-view image data and distance data are obtained by simultaneously scanning the dam slope surface in a grid pattern using a gimbal-controlled high-magnification zoom camera and a laser rangefinder mounted in front of the dam slope. The camera calibration includes calibration processing of the internal and external parameters of the imaging device; The visual reconstruction algorithm uses the structure-of-motion algorithm to fuse multi-view image data and distance data to generate a digital surface model of the dam slope containing three-dimensional coordinates. Based on the three-dimensional coordinate data of each pixel in the neighborhood of the digital surface model of the dam slope, the slope of the point is calculated by local plane fitting, and the curvature of the point is calculated by local quadratic surface fitting.

3. The method for laser removal of vegetation on dam slopes based on visual recognition according to claim 1, characterized in that, The extracted vegetation features and spatial distribution features in S2 include: The vegetation features include: color features, texture features, and morphological features; A pixel-level segmentation-based image segmentation method is used to perform pixel-by-pixel classification on multi-view image data to obtain classification results. The classification results include the vegetation confidence score of each pixel belonging to a vegetation category. The vegetation confidence score of each pixel is extracted from the classification results of pixel-by-pixel classification to form a vegetation cover probability map. The classification results are mapped onto the digital surface model of the dam slope to obtain candidate vegetation areas; Feature extraction processing is performed on candidate vegetation areas. Color features are extracted by color space transformation, texture features are extracted by local texture analysis, and morphological features are extracted by connected component analysis. Based on a preset classification model, the extracted color features, texture features and morphological features are fused and judged to generate vegetation area identification results between vegetation areas and non-vegetation areas. Spatial statistical analysis is performed on the identified vegetation areas to obtain the distribution range, density, and regional connectivity of vegetation on the dam slope surface, forming corresponding spatial distribution characteristics.

4. The method for laser removal of vegetation on dam slopes based on visual recognition according to claim 3, characterized in that, The generation of region partitioning results and the construction of region adjacency relationships in S2 include: Spatial correlation processing is performed between the vegetation cover probability map and the slope and curvature parameters in the dam slope digital surface model to generate a fused map; Based on the spatial distribution characteristics of vegetation cover probability in the fused map, and combined with the spatial boundary between vegetated and non-vegetated areas, areas located in the range of high and low vegetation cover probability or where the attributes of adjacent areas change transitionally are identified as edge transition areas, and the remaining areas are identified as non-edge transition areas.

5. The method for laser removal of vegetation on dam slopes based on visual recognition according to claim 4, characterized in that, The non-edge transition region subdivision includes: Continuous areas with a vegetation cover probability higher than a preset high threshold are classified as vegetation cover areas, and continuous areas with a vegetation cover probability equal to or lower than a preset low threshold are classified as bare base areas. Based on the preset dam structure location information, the drainage ditch, expansion joints and the area around the monitoring equipment are marked as key areas for removal; Record the boundary location information of each region, and construct the region adjacency relationship structure based on the spatial contact relationship and boundary connection relationship between regions.

6. The method for laser removal of vegetation on dam slopes based on visual recognition according to claim 1, characterized in that, S3 includes: Based on vegetation characteristics, spatial distribution characteristics, and regional adjacency structure, data organization and processing are performed on the adjacent regions of the current work location to construct a regional demand data set; For each adjacent region in the regional demand data set, vegetation features and spatial distribution features are combined and processed according to preset feature fusion rules to form a comprehensive demand index. Based on a preset priority determination strategy, the comprehensive demand indicators of each adjacent area are classified and processed to obtain the priority level of each adjacent area. The adjacent regions are sorted in descending order of priority to generate a priority sequence for selecting the current work location region, and the region sorting result of the clearing priority model is formed accordingly.

7. The method for laser removal of vegetation on dam slopes based on visual recognition according to claim 1, characterized in that, The determination of the target area in S4 and the execution of the clearing operation in S5 include: According to the priority sequence, each region is extracted as the target area for the current laser removal operation. During the clearing of the current target area, the laser clearing operation is captured in real time and dynamic non-target objects are identified. The dynamic non-target objects include: people, birds, animals and other moving objects that enter the laser clearing operation and do not belong to the vegetation to be cleared. Based on the current laser spot position, laser scanning path, and target area boundary, a laser irradiation area is generated that includes the current irradiation position and subsequent scanning path, and the laser irradiation area is extended outward by a preset range to form a safe expansion range; When a dynamic non-target object enters the laser irradiation area or the safe expansion range, the laser output is paused and the pause position is recorded. Once the dynamic non-target object leaves the laser irradiation area and the safe expansion range, the clearing process corresponding to the current target area will continue from the paused position. After clearing the current target area, mark the current target area as processed and continue to select the next area to be processed according to the priority sequence until all areas have been traversed and processed.

8. The method for laser removal of vegetation on dam slopes based on visual recognition according to claim 1, characterized in that, The non-edge transition region clearing strategy in S5 includes: When the target region's region type does not belong to the edge transition region: Based on vegetation features, an image classification method is used to identify the vegetation type of the target area. The vegetation type includes: weeds, shrubs, and non-cleared targets. Based on the region type of the target area and the identified vegetation type, a preset set of laser removal parameters is invoked; Configure the laser scanning path for the target area based on the laser removal parameter set; Perform a one-time removal process on the target area according to the laser scanning path.

9. The method for laser removal of vegetation on dam slopes based on visual recognition according to claim 1, characterized in that, The uncertainty band construction in S5 includes: When the target region's region type is an edge transition region: Based on the dam slope surface model, local image data is extracted from multi-view image data corresponding to the target area; Calculate information entropy based on the grayscale distribution of local image data; The boundary of the target region is expanded based on the information entropy value, generating an uncertain band whose width varies with the information entropy.

10. The method for laser removal of vegetation on dam slopes based on visual recognition according to claim 9, characterized in that, The multi-round progressive cleanup strategy generation in S5 includes: A multi-round clearing strategy is implemented in the outer region of the uncertainty zone. The first round of laser clearing uses a preset low-energy short pulse mode to clear vegetation leaves and branches. After the first round of clearing is completed, local image data is re-acquired and the vegetation cover probability map of the local area is updated. By comparing the spatial position difference of the boundary pixel set in the two vegetation cover probability maps, the average displacement of the boundary pixel set is calculated as the boundary change. The advancement step size is determined proportionally based on the boundary change, and a preset medium-energy single-pulse laser is applied to the base of the newly exposed vegetation after clearing, under the condition of meeting the preset dam surface protection constraints. Repeatedly perform the clearing, image re-acquisition, boundary update, and advance operations to gradually shrink the cleared boundary towards the dam surface; The clearing will stop when the preset termination conditions are met. The termination conditions include: the vegetation coverage ratio is lower than a threshold, the boundary change is less than a threshold, or the residual vegetation height is lower than a threshold. The temperature of the target area is monitored in real time during the clearing process, and an early warning is triggered when the temperature exceeds the preset dam surface protection temperature.

Images