A method and system for scanning a hydropower plant model

Abstract

This invention relates to the field of water conservancy and hydropower engineering technology, and discloses a method and system for scanning a hydropower plant model. The method includes: fusing multi-sensor data to generate keyframe poses and corresponding point cloud maps for laser odometers; dividing the entire scanning process into multiple local segments and independently optimizing the laser odometer data for each segment; performing voxel filtering on the point cloud map data and filtering keyframes based on the interval threshold between keyframes to determine the number of pose nodes; correcting the point cloud map and providing initial pose values ​​by adding and deleting pose constraints; automatically detecting and adding closure constraints based on the initial pose values; integrating all pose constraints to construct a pose graph; iteratively optimizing the keyframe poses to obtain a corrected 3D point cloud map of the hydropower plant. This invention can provide a millimeter-level precision 3D point cloud data foundation for hydropower plant safety monitoring, equipment layout analysis, and intelligent operation and maintenance.

Description

Technical Field

[0001] This invention relates to the field of water conservancy and hydropower engineering technology, and in particular to a method and system for scanning a hydropower plant model. Background Technology

[0002] Hydropower plant buildings are the core infrastructure of hydropower systems, and their structural safety and stable equipment operation are directly related to the reliability of power supply. With the rapid development of the hydropower industry towards intelligence and digitalization, high-precision 3D models have become an indispensable basic data support for plant safety monitoring, equipment operation and maintenance, upgrading and transformation, and emergency management. 3D laser scanning technology, with its advantages of non-contact operation, high precision, and high efficiency, has gradually become the mainstream method for 3D modeling in industrial scenarios. The introduction of real-time positioning and mapping technologies has further enabled continuous scanning and automatic modeling in large-scale scenes, and has been widely applied in open spaces and simple industrial environments.

[0003] However, the internal environment of a hydropower station powerhouse is highly complex and unique, posing multiple technical challenges to 3D scanning modeling. The powerhouse contains not only large equipment such as bridge cranes and generator sets, but also a dense network of high-pressure pipelines, steel support frames, and various electromechanical facilities, resulting in a complex spatial structure with diverse geometric features. In particular, the hooks and booms of the bridge cranes are in continuous motion during operation, and the movement of temporary equipment and personnel during maintenance creates dynamic interference, leading to temporal misalignment and occlusion in the collected point cloud data. Traditional algorithms cannot effectively distinguish between static structures and dynamic targets, easily causing model distortion and structural defects. Simultaneously, the numerous metal components within the powerhouse have strong reflective properties; after laser irradiation, specular and secondary reflections occur, generating a large number of abnormal noise points, severely interfering with the point cloud feature extraction and registration process, and reducing modeling accuracy.

[0004] Furthermore, hydropower plant buildings typically feature large spans and multiple levels, resulting in long scanning paths. Traditional real-time localization and mapping algorithms accumulate odometry errors over long distances, and without effective global constraints, these errors can be significantly amplified. The densely packed pipes, symmetrically distributed generating units, and standardized track segments within the plant create repetitive geometric structures that cause automatic loop closure detection algorithms to frequently generate incorrect constraints, further exacerbating odometry drift and leading to issues such as layering, misalignment, and ghosting in the final point cloud model.

[0005] Existing scanning modeling solutions for complex industrial scenarios mostly focus on single-sensor fusion or algorithm parameter tuning, which has significant limitations. Single-sensor fusion can only improve pose estimation accuracy to a certain extent and cannot fundamentally solve the problems of mismatch between similar structures and dynamic interference. The adjustment of algorithm hyperparameters is highly dependent on engineering experience, and different hydropower plant structures require re-adaptation of parameters, resulting in poor universality and low efficiency. At the same time, most solutions adopt fully automated processing procedures and lack effective intervention mechanisms. When the algorithm fails due to environmental interference, it cannot quickly correct errors, which can easily lead to modeling failure.

[0006] Therefore, there is an urgent need to develop a high-precision scanning modeling method for the complex environment of hydropower plant buildings, in order to overcome the bottlenecks of existing technologies and meet the urgent needs of the digital development of the hydropower industry. Summary of the Invention

[0007] To address issues such as dynamic interference from cranes within a power plant, noise from metal surfaces, and point cloud registration failures caused by repetitive structures, this invention proposes a method and system for scanning a hydropower plant model. Through interactive manual correction, dynamic noise suppression, and graph optimization techniques, millimeter-level precision 3D modeling can be achieved.

[0008] The technical solution adopted in this invention is as follows: A method for scanning a hydropower plant powerhouse model, comprising: By fusing data from multiple sensors to generate keyframe poses and corresponding point cloud maps for laser odometry, the entire scanning process is divided into multiple local segments, and the laser odometry data for each segment is optimized independently. Voxel filtering is performed on the point cloud map data, and key frames are filtered according to the interval threshold between key frames to determine the number of pose nodes. The point cloud map is corrected by adding or deleting pose constraints and initial pose values ​​are provided. Based on the initial pose values, closure constraints are automatically detected and added. By integrating all pose constraints to construct a pose graph, and iteratively optimizing the pose of keyframes, a corrected 3D point cloud map of the hydropower plant is obtained.

[0009] Furthermore, the process of fusing multi-sensor data to generate keyframe poses and corresponding point cloud maps for laser odometry includes: Raw data from multiple sensors, including laser point cloud data, is collected. The relative pose between adjacent frames is calculated based on real-time localization and laser mapping algorithms. Keyframes are extracted based on the degree of pose change, and a laser odometry composed of keyframe poses is generated. The local point clouds corresponding to each keyframe are stitched together according to the pose to obtain an initial point cloud map.

[0010] Furthermore, the process of dividing the entire scanning process into multiple local segments and independently optimizing the laser odometry data for each segment includes: Based on the continuity of the scanning path or the characteristics of the odometry cumulative error, the complete scanning process is divided into multiple non-overlapping local segments; the pose optimization is performed separately for the keyframe pose sequence within each local segment to eliminate the local cumulative error within a single segment, thus obtaining the optimized initial laser odometry for each segment.

[0011] Furthermore, the voxel filtering of the point cloud map data includes: The initial point cloud map is divided into multiple three-dimensional voxel grids. The point clouds in each voxel grid are merged, and the point cloud data that can represent the geometric features of the voxel are retained. While reducing the overall size of the point cloud, the original shape and structural features of the point cloud are maintained.

[0012] Furthermore, the voxel filtering of the point cloud map data further includes: Based on the characteristics of laser reflection intensity, abnormal point clouds generated on the metal surface are eliminated; By comparing the point cloud data corresponding to adjacent keyframes, dynamic targets with displacement changes are detected. Generate a region mask corresponding to the dynamic target, remove the dynamic point cloud data within the mask range, and retain the point cloud information of the static structure of the factory.

[0013] Further, the step of filtering keyframes based on the interval threshold between keyframes to determine the number of pose nodes includes: Set the interval judgment condition between keyframes, traverse all keyframes in the initial laser odometry, and select keyframes that meet the interval judgment condition as pose nodes. Construct the initial pose graph based on the selected pose nodes.

[0014] Furthermore, the step of correcting the point cloud map and providing initial pose values ​​by adding or deleting pose constraints includes: Identify misaligned or unclosed regions in the point cloud map, select two candidate keyframes for the corresponding region, and adjust the relative pose between the two keyframes to initially align the corresponding point clouds; perform point cloud registration operation to verify the point cloud alignment effect; if the alignment effect meets the requirements, add the corresponding pose constraints; if there are incorrect pose constraints, delete them from the constraint set; after completing the addition and deletion of pose constraints, obtain the corrected initial pose value.

[0015] Furthermore, the automatic detection and addition of closure constraints based on initial pose values ​​includes: Based on the corrected initial pose values, all pose nodes are traversed, and key frame pairs that meet the conditions of spatial distance, path length, and matching degree are selected. Point cloud registration is performed on the selected key frame pairs. If the registration results meet the requirements, closure constraints are added between the corresponding two key frames. After traversal, all valid automatic closure constraints are obtained.

[0016] Furthermore, the process of integrating all pose constraints to construct a pose graph and iteratively optimizing keyframe poses includes: The selected pose nodes are used as vertices of the pose graph, and the adjacent frame pose constraints generated by laser odometry, the pose constraints obtained by addition and deletion, and the loop closure constraints obtained by automatic detection are used as edges of the pose graph to construct a complete pose graph containing all vertices and edges. A graph optimization algorithm is used to iteratively adjust the pose of each vertex to minimize the total error of all pose constraints. After the iteration is completed, the point cloud is re-stitched based on the optimized keyframe poses to obtain the corrected 3D point cloud map of the hydropower plant.

[0017] A hydropower station powerhouse model scanning system includes: The multi-sensor data fusion and initial mapping module is configured to fuse multi-sensor data to generate key frame poses and corresponding point cloud maps for laser odometry, divide the entire scanning process into multiple local segments, and independently optimize the laser odometry data of each segment. The dynamic noise filtering and point cloud preprocessing module is configured to perform voxel filtering on the point cloud map data and filter key frames according to the interval threshold between key frames to determine the number of pose nodes. The pose constraint correction module is configured to correct the point cloud map and provide initial pose values ​​by adding or deleting pose constraints, and to automatically detect and add closure constraints based on the initial pose values. The pose graph construction and optimization module is configured to integrate all pose constraints to construct a pose graph, iteratively optimize the pose of keyframes, and obtain a corrected 3D point cloud map of the hydropower plant.

[0018] The beneficial effects of this invention are as follows: 1. This invention effectively solves the problem of long-distance cumulative error in large-scale hydropower plant scenarios by combining segment-based independent optimization with global pose graph optimization. Dividing the complete scanning process into multiple local segments for independent optimization eliminates local errors within each segment in advance, preventing errors from amplifying infinitely with scanning distance. Based on this, integrating multi-source pose constraints for global optimization achieves precise adjustment of keyframe poses, ensuring global consistency of the point cloud map and significantly improving the overall accuracy of 3D modeling.

[0019] 2. This invention designs a dedicated point cloud preprocessing workflow specifically for the environmental characteristics of hydropower plant buildings, effectively improving the quality and processing efficiency of point cloud data. By using voxel filtering, the data size is significantly compressed while preserving the core geometric features of the point cloud, reducing the computational burden. Combined with laser reflection intensity characteristics and analysis of differences between adjacent frame point clouds, it can accurately identify and remove high-reflectivity noise from metal surfaces and dynamic target point clouds such as crane hooks, obtaining clean static structural data of the plant building. This avoids the negative impact of noise and dynamic interference on subsequent registration and optimization processes from the source, improving the accuracy of point cloud registration and the integrity of the model.

[0020] 3. This invention significantly enhances the robustness and universality of the algorithm through a collaborative mechanism of pose constraint correction and automatic loop closure detection. First, significant deviations in the point cloud map are corrected by adding or deleting pose constraints, providing an accurate initial pose for automatic loop closure detection and effectively reducing the probability of mismatches caused by similar structures. Then, automatic loop closure detection is performed based on the corrected initial pose value, generating reliable loop closure constraints through multiple condition screenings, further refining the point cloud map. This process reduces the algorithm's dependence on scene-specific parameters, enabling it to adapt to scanning tasks of hydropower plant buildings of different scales and structures, thus reducing the difficulty of engineering applications.

[0021] 4. The high-precision 3D point cloud map generated by this invention can comprehensively and accurately reflect the structural details and equipment layout of hydropower plant buildings, providing a solid data foundation for the refined management of the hydropower industry. This model can be directly applied to multiple scenarios such as plant structure deformation monitoring, equipment installation calibration, pipeline collision detection, and the construction of digital operation and maintenance platforms. It helps to promptly identify safety hazards, improve equipment operation and maintenance efficiency, and reduce operation and maintenance costs, thus having significant engineering application value for promoting the intelligent transformation and safe and stable operation of hydropower plants. Attached Figure Description

[0022] Figure 1 This is a flowchart of a hydropower plant building model scanning method according to Embodiment 2 of the present invention.

[0023] Figure 2 This is a schematic diagram of the laser odometer of Embodiment 2 of the present invention. Detailed Implementation

[0024] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments are now described. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of the invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0025] Example 1 This embodiment provides a method for scanning a hydropower station powerhouse model, including: By fusing data from multiple sensors to generate keyframe poses and corresponding point cloud maps for laser odometry, the entire scanning process is divided into multiple local segments, and the laser odometry data for each segment is optimized independently. Voxel filtering is performed on the point cloud map data, and key frames are filtered according to the interval threshold between key frames to determine the number of pose nodes. The point cloud map is corrected by adding or deleting pose constraints and initial pose values ​​are provided. Based on the initial pose values, closure constraints are automatically detected and added. By integrating all pose constraints to construct a pose graph, and iteratively optimizing the pose of keyframes, a corrected 3D point cloud map of the hydropower plant is obtained.

[0026] Preferably, raw data from multiple sensors, including laser point cloud data, are collected, and the relative pose between adjacent frames is calculated based on real-time localization and laser mapping algorithms. Keyframes are extracted according to the degree of pose change, and a laser odometry composed of keyframe poses is generated. The local point clouds corresponding to each keyframe are stitched together according to the pose to obtain an initial point cloud map.

[0027] Specifically, before scanning begins, timestamp synchronization and extrinsic parameter calibration of multiple sensors are completed to ensure that data from different sensors are aligned in the spatiotemporal dimensions. During scanning, the lidar continuously emits and receives laser signals to generate a 3D point cloud, assisting the sensors in synchronously acquiring equipment motion status data. The synchronized data is input into a real-time positioning and laser mapping algorithm, which calculates relative rotation and translation pose by matching the features of adjacent frame point clouds. A keyframe extraction threshold is set based on the equipment's motion amplitude; when the pose change reaches the threshold, the current frame is marked as a keyframe. All keyframe poses are arranged in a time sequence to form a laser odometry, and the local point clouds corresponding to each keyframe are transformed to a unified coordinate system and stitched together to form an initial global point cloud map.

[0028] It should be noted that multi-sensor fusion improves the robustness of pose calculation, and extracting keyframes based on pose changes reduces redundant data while ensuring map accuracy, laying an efficient data foundation for subsequent optimization processes.

[0029] Preferably, the complete scanning process is divided into multiple non-overlapping local segments according to the continuity of the scanning path or the characteristics of the odometry cumulative error; the pose optimization of the key frame pose sequence in each local segment is performed separately to eliminate the local cumulative error in a single segment, and the optimized initial laser odometry of each segment is obtained.

[0030] Specifically, after generating the initial laser odometry, the changes in the scanning path and the odometry error growth curve are analyzed. When the scanning path shows obvious segmentation or the accumulated error reaches a preset level, the scanning process is divided into multiple continuous and non-overlapping local segments, each containing a fixed number of consecutive keyframes. For the pose sequence of the keyframes within each segment, a local optimization algorithm is used to adjust the pose parameters, eliminating the accumulated error caused by sensor noise and motion estimation bias within that segment. After optimizing all segments, the laser odometry data of each segment are stitched together in chronological order to obtain the global initial laser odometry.

[0031] It should be noted that segment-based independent optimization avoids the error from amplifying infinitely with scanning distance, significantly improving the accuracy of the initial odometry and reducing the difficulty and computational load of subsequent global optimization.

[0032] Preferably, the initial point cloud map is divided into multiple three-dimensional voxel grids, and the point clouds in each voxel grid are merged to retain the point cloud data that can represent the geometric features of the voxel; while reducing the overall size of the point cloud, the original shape and structural features of the point cloud are maintained.

[0033] Specifically, the voxel grid size is set according to the overall size of the point cloud map and the modeling accuracy requirements, and the initial point cloud is uniformly divided into several 3D cubic grids. All point cloud data is traversed, and each point is assigned to the corresponding voxel grid. For grids containing multiple points, the geometric center of the points within the grid is calculated or feature points are selected as retained points, and other redundant points are deleted; for grids containing only a single point, that point is directly retained. This process significantly compresses the total amount of point cloud data.

[0034] It should be noted that voxel filtering effectively reduces the point cloud size without losing key geometric features, significantly improving the processing efficiency of subsequent point cloud registration, constraint detection, and graph optimization.

[0035] Preferably, based on the characteristics of laser reflection intensity, abnormal point clouds generated on the metal surface are removed; by comparing the point cloud data corresponding to adjacent key frames, dynamic targets with displacement changes are detected; a region mask corresponding to the dynamic target is generated, the dynamic point cloud data within the mask range is removed, and the point cloud information of the static structure of the factory is retained.

[0036] Specifically, after voxel filtering, the laser reflection intensity information of each point cloud is extracted. Utilizing the significant difference in reflection intensity between metallic and non-metallic materials, abnormal noise points caused by high reflectivity on metallic surfaces are identified and removed. Next, the point cloud data of adjacent keyframes are compared to calculate the point cloud differences at the same spatial location at different times, locating dynamic target areas with displacement changes. A binary mask is generated based on the dynamic target areas, and all point clouds within the mask's coverage area are removed from the map, retaining only the point cloud data of static structures such as factory walls, beams, columns, and pipes.

[0037] It should be noted that this step effectively solves the problems of high reflectivity noise from metal surfaces inside the hydropower plant and interference from dynamic targets such as crane hooks and mobile devices, ensuring the purity of the static structural point cloud and significantly improving the accuracy of subsequent point cloud registration.

[0038] Preferably, an interval determination condition is set between keyframes, all keyframes in the initial laser odometry are traversed, and keyframes that meet the interval determination condition are selected as pose nodes. An initial pose graph is constructed based on the selected pose nodes.

[0039] Specifically, based on the complexity of the factory scene and the required modeling accuracy, the time interval or spatial distance interval of keyframes is set as a filtering condition. All keyframes in the initial laser odometry are traversed in chronological order, and the interval between adjacent keyframes is checked to see if it meets the preset condition. If the interval is less than a threshold, one keyframe is deleted; if the interval is greater than or equal to the threshold, the keyframe is retained. The final selected keyframes are used as vertices of the pose graph, and the initial pose graph structure is constructed in chronological order.

[0040] It should be noted that by reasonably selecting pose nodes, the pose graph can accurately reflect the motion trajectory of the device, while further simplifying the size of the pose graph, reducing the computational complexity of subsequent graph optimization, and improving the overall processing efficiency.

[0041] Preferably, the process involves identifying misaligned or unclosed regions in the point cloud map, selecting two candidate keyframes for the corresponding regions, adjusting the relative poses between the two keyframes to initially align the corresponding point clouds, performing point cloud registration, and verifying the point cloud alignment effect. If the alignment effect meets the requirements, corresponding pose constraints are added. If there are incorrect pose constraints, they are deleted from the constraint set. After completing the addition and deletion of pose constraints, the corrected initial pose value is obtained.

[0042] Specifically, the initial point cloud map is viewed through a visual interface to locate abnormal regions exhibiting layering, misalignment, or incomplete loop closures. Two candidate keyframes with significant common geometric features are selected within these abnormal regions. The relative rotation and translation parameters of the two keyframes are adjusted to initially align the corresponding point clouds in the interface. Point cloud registration is performed using the adjusted pose as the initial value, calculating the precise relative pose of the two keyframes. If the point cloud alignment is good after registration, this relative pose is added as a new constraint to the constraint set; simultaneously, existing constraints are checked, and invalid constraints causing map errors are deleted. After completing all constraint additions and deletions, the pose graph is initially optimized to obtain the corrected initial pose value.

[0043] It should be noted that by adding or deleting pose constraints, obvious errors in the point cloud map can be quickly corrected, providing an accurate initial pose for automatic loop closure detection. This avoids a large number of erroneous constraints caused by excessive initial errors, and significantly improves the reliability of loop closure detection.

[0044] Preferably, based on the corrected initial pose value, all pose nodes are traversed to select key frame pairs that meet the conditions of spatial distance, path length and matching degree; point cloud registration is performed on the selected key frame pairs; if the registration result meets the requirements, a closure constraint is added between the corresponding two key frames; after traversal, all valid automatic closure constraints are obtained.

[0045] Specifically, the global position of each pose node is calculated based on the corrected initial pose value. All pose nodes are traversed, and for each node, other nodes within a preset spatial distance are searched, while ensuring that the graph path length between two nodes is greater than a threshold to avoid adding redundant constraints between consecutive frames. For keyframe pairs that meet the spatial and path conditions, their corresponding local point clouds are extracted and registration is performed to evaluate the point cloud matching degree. If the matching degree meets the requirements, a closure constraint edge is added between the two keyframes. This process is repeated until all nodes are traversed, obtaining all valid automatic closure constraints.

[0046] It should be noted that automatic loop closure detection based on the corrected initial pose value significantly improves the detection accuracy, and multiple screening conditions further ensure the effectiveness of loop closure constraints, providing sufficient and reliable constraint information for global pose optimization.

[0047] Preferably, the selected pose nodes are used as vertices of the pose graph, and the pose constraints of adjacent frames generated by laser odometry, the pose constraints obtained by addition and deletion, and the loop closure constraints obtained by automatic detection are used as edges of the pose graph to construct a complete pose graph containing all vertices and edges. A graph optimization algorithm is used to iteratively adjust the pose of each vertex to minimize the total error of all pose constraints. After the iteration is completed, the point cloud is re-stitched according to the optimized keyframe pose to obtain the corrected 3D point cloud map of the hydropower plant.

[0048] Specifically, the selected pose nodes are used as vertices of the pose graph, with each vertex corresponding to the pose parameters of a keyframe. The relative poses of adjacent frames provided by laser odometry are used as odometry constraint edges, constraints obtained from addition and deletion operations are used as correction constraint edges, and automatically detected loop closure constraints are used as loop closure constraint edges, collectively forming the edges of the pose graph. After constructing the complete pose graph, the pose of each vertex is iteratively adjusted using a graph optimization algorithm, with the goal of minimizing the sum of squared errors of all constraint edges. A robust kernel function is used to handle outlier constraints during optimization, improving optimization stability. Optimization stops when the iteration converges or reaches a preset number of iterations. The local point clouds are then reassembled based on the optimized keyframe poses to obtain the final high-precision 3D point cloud map.

[0049] It should be noted that global map optimization can effectively integrate odometer information, correction information and loop closure information, completely eliminate global cumulative error, and make the generated point cloud map globally consistent, meeting the modeling requirements of hydropower plant with millimeter-level accuracy.

[0050] Accordingly, this embodiment also provides a hydropower plant building model scanning system, including a multi-sensor data fusion and initial mapping module, a dynamic noise filtering and point cloud preprocessing module, a pose constraint correction module, and a pose graph construction and optimization module.

[0051] Specifically, the multi-sensor data fusion and initial mapping module is responsible for receiving raw data from multiple sensors, completing time synchronization and spatial calibration, generating initial laser odometry and point cloud maps, and performing segment-by-segment independent optimization of the scanning process. The dynamic noise filtering and point cloud preprocessing module is responsible for voxel filtering, keyframe selection, and pose node determination of the point cloud, while also removing metallic reflection noise and dynamic target point clouds. The pose constraint correction module is responsible for correcting point cloud map deviations by adding or deleting pose constraints, providing accurate initial pose values, and automatically detecting and adding effective closure constraints. The pose graph construction and optimization module is responsible for integrating all constraints to construct a complete pose graph, performing global pose optimization, and generating the final high-precision 3D point cloud map based on the optimization results.

[0052] It should be noted that the modules can communicate through standardized interfaces, ensuring the accuracy of data transmission and the efficiency of the processing flow, and can reliably complete high-precision 3D scanning and modeling tasks in complex industrial environments.

[0053] Example 2 See Figure 1 This embodiment provides a method for scanning a hydropower station powerhouse model, including: Step 1: Multi-sensor data fusion and initial mapping.

[0054] Inputting sensor data, existing real-time localization and laser mapping algorithms are used to generate keyframe poses and corresponding point cloud maps for laser odometry. The laser odometry is composed of keyframe poses, and the point cloud corresponding to each keyframe is a collection of local point clouds near that keyframe. The point cloud map is constructed by stitching together the point clouds corresponding to the keyframes according to their poses. See the laser odometry schematic diagram. Figure 2 .

[0055] The pose of keyframes in laser odometry is denoted as... , Indicates the first Each keyframe pose includes a rotation and a translation component. It is a 3x3 rotation matrix. It is a 1*3 translation vector. and The reference coordinate system is based on the initial starting point, and the pose constraints formed between consecutive frames in the laser odometry are denoted as... Meanwhile, the entire scanning process is divided into multiple local segments, and the odometry data of each segment is optimized independently, thereby reducing the initial positioning deviation caused by the cumulative error of the sensor and avoiding the problem of the error gradually amplifying during long-distance scanning.

[0056] Step 2: Dynamic noise filtering and point cloud preprocessing.

[0057] Voxel filtering is applied to the point cloud data to reduce its size. Simultaneously, keyframes are selected based on an interval threshold to determine the number of pose nodes in the initial pose graph. This step primarily reduces subsequent computation. Voxel filtering preserves the shape characteristics of the point cloud while reducing its size. Dynamic noise removal is also performed, including reflection intensity filtering, dynamic mask generation, and semantic enhancement downsampling, detailed below.

[0058] Reflection intensity filtering: Removing abnormal points from metal surfaces based on laser reflection intensity thresholds. I(p) <I threshold (Filter out areas with reflectivity higher than the threshold).

[0059] Dynamic mask generation: Detect dynamic targets (such as hooks) by the difference in point clouds between adjacent frames, generate dynamic region masks and remove them.

[0060] Semantic enhancement downsampling: Fine-grained voxel filtering (0.05m voxel size) is used for crane tracks and densely piped areas, while coarse-grained filtering (0.1m voxel size) is used for other areas to balance accuracy and efficiency.

[0061] Step 3: Interactive constraint correction.

[0062] Interactive constraint correction includes the process of adding and deleting pose constraints, the process of automatic loop closure detection, and the process of semantic-assisted registration.

[0063] Adding or deleting pose constraints: Adding or deleting loop closure constraints to perform coarse correction of the point cloud map. Manually correcting obvious errors in the point cloud map, adding loop closure constraints to areas where no loops have formed, and deleting loop closure constraints to areas where erroneous loops have formed. This coarse correction of the point cloud map provides good initial pose values ​​for the subsequent automatic loop closure detection process.

[0064] For areas where loop closures should form but don't, firstly, two candidate keyframes are explicitly selected via the GUI. The poses of these two keyframes are adjusted to roughly align the corresponding point clouds, and these poses serve as initial values ​​for matching. Then, the GICP scan-matching method is executed to perform point cloud registration. Finally, the alignment of the point clouds is checked. If successful, the estimated relative poses are added as loop closure constraint edges to the pose graph for optimization. The added pose constraints are denoted as... Simultaneously, when erroneous loops exist in the point cloud map, constraint edges connected to the keyframes of the erroneous loops can be explicitly selected for deletion. This optimization process uses the g2o (General Graphic Optimization) library. Each time a correction is performed, the pose graph calls g2o's Levenberg-Marquardt optimizer for optimization. After optimization by this hypergraph optimizer, the latest corrected keyframe poses and corresponding point cloud maps can be obtained.

[0065] Automatic loop closure detection: After adding or deleting loop closure constraints, a relatively accurate initial pose value is obtained by reducing a large pose error. Then, automatic loop closure detection is performed according to the set threshold, and more loop closure constraint edges are automatically added for fine point cloud map correction.

[0066] The automatic loop closure detection process based on scan matching still uses the GICP method to find keyframe pairs that meet the following conditions. And add a closure constraint between them: (1) (2) (3) in, Indicates the first The location estimated for each keyframe This represents the minimum graph path length between two keyframes. This is the matching score of the GICP scan results. , , These are the set threshold parameters used to determine whether to create a loop closure constraint between two keyframes. Equation 1 indicates that the positional distance between two keyframes needs to be less than a given threshold to ensure that the distance between the two keyframes is close enough. Equation 2 indicates that the minimum graph path length between two keyframes is greater than a given threshold to prevent the creation of loop closure constraints between consecutive frames from causing excessive loop closure constraints in local areas, thus increasing the computational burden. Equation 3 indicates that the scan matching score is less than a given threshold to ensure that the two candidate loop closure keyframes have a good matching degree.

[0067] In addition, to prevent interference from noise and outliers during the matching process, the Huber robust kernel function is used in the point cloud registration process. Unlike the traditional squared error function, it decays linearly when the error is small, but decays at a slower rate when the error is large. This characteristic can effectively reduce the impact of outliers or noise on the final registration result and improve the robustness and accuracy of the registration.

[0068] During this process, keyframes are automatically compared with their neighboring keyframes. If they match well, a closure constraint is added between them. This process is repeated for all keyframes. The closure constraint added in this process is denoted as . .

[0069] Step 4: Optimize the pose side graph.

[0070] During the process of adding and deleting pose constraints and automatic loop closure detection, a pose optimization process is performed. The pose of key frames in the odometry is iteratively optimized. The above steps are repeated to obtain the corrected point cloud map.

[0071] After adding or deleting pose constraints and establishing loop closure constraints that meet the threshold conditions through automatic loop closure detection, the pose of the keyframes is then optimized. The objective function is defined as follows: (4) (5) In Equation 4, It is the keyframe pose sequence to be optimized, and the objective function is... It consists of the sum of three error terms, namely the pose constraints generated by the instantaneous localization and laser mapping algorithms. Added pose constraints Pose constraints added during automatic loop closure detection process In Equation 5, This indicates that for a set of constraints The sum of error functions, error function Representation and Constraints Related parameter blocks The error, It is an information matrix, and the parameter set is optimized using graph optimization methods. This minimizes the value of equation 4.

[0072] Example 3 This embodiment provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement a hydropower station powerhouse model scanning method according to Embodiment 1 or 2. The computer program can be in the form of source code, object code, executable file, or some intermediate form.

[0073] Example 4 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a hydropower plant model scanning method of Embodiment 1 or 2. The computer program can be in the form of source code, object code, executable file, or some intermediate form. The storage medium includes any entity or device capable of carrying computer program code, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0074] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

[0075] It should be noted that, for the sake of simplicity, the foregoing method embodiments are described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

Claims

1. A method for scanning a hydropower station powerhouse model, characterized in that, include: By fusing data from multiple sensors to generate keyframe poses and corresponding point cloud maps for laser odometry, the entire scanning process is divided into multiple local segments, and the laser odometry data for each segment is optimized independently. Voxel filtering is performed on the point cloud map data, and key frames are filtered according to the interval threshold between key frames to determine the number of pose nodes. The point cloud map is corrected by adding or deleting pose constraints and initial pose values ​​are provided. Based on the initial pose values, closure constraints are automatically detected and added. By integrating all pose constraints to construct a pose graph, and iteratively optimizing the pose of keyframes, a corrected 3D point cloud map of the hydropower plant is obtained.

2. The method for scanning a hydropower station powerhouse model according to claim 1, characterized in that, The process of generating keyframe poses and corresponding point cloud maps for laser odometry by fusing multi-sensor data includes: Raw data from multiple sensors, including laser point cloud data, is collected. The relative pose between adjacent frames is calculated based on real-time localization and laser mapping algorithms. Keyframes are extracted based on the degree of pose change, and a laser odometry composed of keyframe poses is generated. The local point clouds corresponding to each keyframe are stitched together according to the pose to obtain an initial point cloud map.

3. The method for scanning a hydropower station powerhouse model according to claim 1, characterized in that, The process of dividing the entire scanning process into multiple local segments and independently optimizing the laser odometry data for each segment includes: Based on the continuity of the scanning path or the characteristics of the odometry cumulative error, the complete scanning process is divided into multiple non-overlapping local segments; the pose optimization is performed separately for the keyframe pose sequence within each local segment to eliminate the local cumulative error within a single segment, thus obtaining the optimized initial laser odometry for each segment.

4. The method for scanning a hydropower station powerhouse model according to claim 1, characterized in that, The voxel filtering of the point cloud map data includes: The initial point cloud map is divided into multiple three-dimensional voxel grids. The point clouds in each voxel grid are merged, and the point cloud data that can represent the geometric features of the voxel are retained. While reducing the overall size of the point cloud, the original shape and structural features of the point cloud are maintained.

5. The method for scanning a hydropower station powerhouse model according to claim 4, characterized in that, The voxel filtering of the point cloud map data also includes: Based on the characteristics of laser reflection intensity, abnormal point clouds generated on the metal surface are eliminated; By comparing the point cloud data corresponding to adjacent keyframes, dynamic targets with displacement changes are detected. Generate a region mask corresponding to the dynamic target, remove the dynamic point cloud data within the mask range, and retain the point cloud information of the static structure of the factory.

6. The method for scanning a hydropower station powerhouse model according to claim 1, characterized in that, The step of filtering keyframes based on the interval threshold between keyframes to determine the number of pose nodes includes: Set the interval judgment condition between keyframes, traverse all keyframes in the initial laser odometry, filter out the keyframes that meet the interval judgment condition as pose nodes, and construct the initial pose graph based on the filtered pose nodes.

7. The method for scanning a hydropower station powerhouse model according to claim 1, characterized in that, The step of correcting the point cloud map and providing initial pose values ​​by adding or deleting pose constraints includes: Identify misaligned or unclosed regions in the point cloud map, select two candidate keyframes for the corresponding region, and adjust the relative pose between the two keyframes to initially align the corresponding point clouds; perform point cloud registration operation to verify the point cloud alignment effect; if the alignment effect meets the requirements, add the corresponding pose constraints; if there are incorrect pose constraints, delete them from the constraint set; after completing the addition and deletion of pose constraints, obtain the corrected initial pose value.

8. The method for scanning a hydropower station powerhouse model according to claim 1, characterized in that, The automatic detection and addition of closure constraints based on initial pose values ​​includes: Based on the corrected initial pose values, all pose nodes are traversed, and key frame pairs that meet the conditions of spatial distance, path length, and matching degree are selected. Point cloud registration is performed on the selected key frame pairs. If the registration results meet the requirements, closure constraints are added between the corresponding two key frames. After traversal, all valid automatic closure constraints are obtained.

9. The method for scanning a hydropower station powerhouse model according to claim 1, characterized in that, The process of integrating all pose constraints to construct a pose graph and iteratively optimizing keyframe poses includes: The selected pose nodes are used as vertices of the pose graph, and the adjacent frame pose constraints generated by laser odometry, the pose constraints obtained by addition and deletion, and the loop closure constraints obtained by automatic detection are used as edges of the pose graph to construct a complete pose graph containing all vertices and edges. A graph optimization algorithm is used to iteratively adjust the pose of each vertex to minimize the total error of all pose constraints. After the iteration is completed, the point cloud is re-stitched based on the optimized keyframe poses to obtain the corrected 3D point cloud map of the hydropower plant.

10. A hydropower station powerhouse model scanning system, characterized in that, include: The multi-sensor data fusion and initial mapping module is configured to fuse multi-sensor data to generate key frame poses and corresponding point cloud maps for laser odometry, divide the entire scanning process into multiple local segments, and independently optimize the laser odometry data of each segment. The dynamic noise filtering and point cloud preprocessing module is configured to perform voxel filtering on the point cloud map data and filter key frames according to the interval threshold between key frames to determine the number of pose nodes. The pose constraint correction module is configured to correct the point cloud map and provide initial pose values ​​by adding or deleting pose constraints, and to automatically detect and add closure constraints based on the initial pose values. The pose graph construction and optimization module is configured to integrate all pose constraints to construct a pose graph, iteratively optimize the pose of keyframes, and obtain a corrected 3D point cloud map of the hydropower plant.

Images