Transformer substation construction site three-dimensional map key feature extraction and topological map construction method

Abstract

The invention discloses a transformer substation construction site three-dimensional map key feature extraction and topological map construction method. Three-dimensional point cloud data and image data of a transformer substation construction site are acquired; preprocessing the obtained original point cloud data, and removing noise points; point cloud registration is carried out; point cloud registration is realized based on an ICP algorithm; generating a three-dimensional map of the color point cloud; key features of the generated three-dimensional map are extracted based on a Mama algorithm, and a topological map is constructed. The method overcomes the defect that topological relations lack semantic understanding and logic rationality due to the fact that topological maps are constructed only according to point cloud geometric features in the prior art, after the key points are extracted, environmental semantic information and actual application requirements are fully considered, topological mapping is carried out according to the key points, and the topological mapping efficiency is improved. The constructed topological map not only reflects the spatial geometrical relationship, but also can reflect the connection relationship conforming to the actual scene logic, thereby providing more reliable map model support for applications such as path planning, target positioning and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional reconstruction, and particularly to a method for extracting key features of a three-dimensional map and constructing a topological map at a substation construction site. Background Art

[0002] With the rapid development of the construction of the power system, as a key node of the power grid, the construction quality and operation efficiency of substations directly affect the safety and stability of the power grid. However, the construction site environment of substations is complex, involving various power equipment, lines and buildings. Traditional construction monitoring and management methods rely on manual inspections and two-dimensional drawings, and there are the following problems: (1) Low efficiency and error accumulation: Manual monitoring relies on empirical judgment and cannot quickly and accurately obtain the overall view of the construction site and the status of equipment. (2) Information isolation and insufficient interaction: Traditional monitoring means cannot efficiently integrate multi-source data, and it is difficult to comprehensively grasp the construction progress and quality, affecting management decisions. (3) Lack of digital support: The current technology has limited digital modeling and three-dimensional perception capabilities for construction sites and is difficult to meet the needs of intelligent construction and subsequent operation and maintenance.

[0003] To address the above problems, three-dimensional mapping technology has been widely used in the effective management of substation construction sites in recent years. In existing three-dimensional mapping technologies, point cloud data is the basis for constructing a three-dimensional model. Point cloud is an important modality in computer vision and has extensive practical applications in fields such as robotics, autonomous driving, and augmented reality. Different from image processing and graph learning, point cloud analysis faces unique challenges, which stem from the inherent irregularity and sparsity of point clouds. Point cloud is a 3D unstructured data. However, the complexity of the attention mechanism is quadratic, bringing a huge computational cost, which is not friendly to devices with limited resources. In recent years, state space models (SSMs) have made progress in processing one-dimensional sequences (e.g., language and speech) and two-dimensional data (e.g., images and graphs). Extending the application of Mamba to three-dimensional point clouds has been a research hotspot in recent years. The point cloud analysis method based on Mamba adopts a two-step process. First, a specific scanning method is used to label the point cloud data as discrete tokens. Then, Mamba is used to capture the potential patterns in these tokens. For example, PointMamba proposes a hierarchical scanning strategy to encode the local and global information of 3D point clouds, and then uses ordinary Mamba as the backbone to extract features from the serialized point tokens without adding additional complex techniques. Point Cloud Mamba uses Mamba as the basic model backbone, reducing the memory usage and having better performance compared with the corresponding Transformer-based models.

[0004] In addition, in point cloud processing, multi-modal data fusion techniques can make full use of the advantages of different sensors to further improve the integrity and accuracy of point cloud data. In recent years, multi-modal large language models (MLLMs) have received extensive attention. They inherit the advanced capabilities of large language models, including powerful language expression and logical reasoning abilities. Although Transformer has always been the mainstream method in this field, Mamba has become a strong competitor by demonstrating impressive performance in aligning mixed source data and achieving linear complexity expansion of sequence length, making Mamba the optimal alternative to Transformer in multi-modal learning. For example, VL-Mamba uses Mamba's efficient architecture to solve vision-language tasks, utilizes pre-trained Mamba models for language understanding, and combines connection modules to align visual blocks with language tokens. Fusion-Mamba and Sigma attempt to fuse complementary information from different modalities (such as thermal imaging, depth, and RGB). Fusion-Mamba focuses on improving object detection, while Sigma focuses on enhancing semantic segmentation.

[0005] Point cloud registration is a fundamental task in graphics, vision, and robotics, with the goal of estimating a rigid transformation to align two partially overlapping 3D point clouds. In recent years, research progress in this field has been mainly dominated by learning-based and correspondence-based methods.

[0006] Correspondence-based methods usually first extract corresponding points between two point clouds and then use a robust pose estimator (such as RANSAC) to recover the transformation. This type of method can be further divided into two categories: one aims to detect more repeatable key points and learn more powerful key point descriptors; the other retrieves correspondences by considering all possible matches without key point detection. However, most correspondence-based methods rely on key point detection, and it is challenging to detect repeatable key points when the overlapping area between the two point clouds is small, usually resulting in a low inlier ratio in the assumed corresponding points.

[0007] Direct registration methods have also gradually emerged. They use neural networks to estimate the transformation in an end-to-end manner and can be divided into two categories: one follows the idea of ICP, iteratively establishing soft correspondences and calculating the transformation through differentiable weighted SVD; the other first extracts global feature vectors for each point cloud and then regresses the transformation using the global feature vectors. However, direct registration methods may fail in large-scale scenarios. In addition, traditional robust estimators (such as RANSAC) have problems of slow convergence and instability in cases of high outlier ratios. Although there are deep robust estimators as alternative solutions, they require training specific networks. The Transformer-based point cloud registration network plays an important role in the fields related to point cloud processing and map construction. Such networks usually preprocess the point cloud data first, including operations such as denoising and normalization, to improve the data quality. Then, using the architectural characteristics of the Transformer, the features of the point cloud are extracted and encoded. Through the multi-head attention mechanism and others, it can capture the correlation information between different local regions of the point cloud and between points, and try to find the correspondences between different point clouds during the point cloud registration process, so as to determine the relative position and pose transformation information, laying a foundation for constructing a unified map coordinate system. In terms of extracting key information, the Transformer-based network will also design specific decoding layers or auxiliary modules to identify the feature manifestations of target objects such as power equipment and transmission lines in the point cloud data, and then extract relevant key information to provide element support for map construction.

[0008] However, there are still certain problems and limitations in the current technology, such as:

[0009] 1. Insufficient key point detection and geometric feature representation

[0010] During the point cloud registration process, the accuracy of key point detection directly affects the registration accuracy. The current methods are difficult to reliably detect repeatable key points in low-overlap regions, resulting in a low inlier ratio. This deficiency significantly limits the application of the technology in high-precision registration tasks. In addition, some methods use the Transformer framework to encode the geometric features of the point cloud. Although they show certain advantages in overall feature extraction, they are still insufficient in extracting local details, especially the features of objects with complex shapes and fine geometric structures (such as specially designed traffic signs, etc.). This lack of geometric feature representation easily leads to abnormal matches, especially in low-overlap cases, seriously affecting the reliability and accuracy of registration.

[0011] 2. High computational complexity and resource consumption

[0012] Many existing point cloud registration methods rely on robust estimators such as RANSAC to handle outliers. Although RANSAC can solve the outlier problem to a certain extent, its high computational cost and slow convergence speed increase the overall computational overhead. At the same time, the inherently high computational complexity of the Transformer architecture, especially the computational volume of its multi-head attention mechanism increases sharply with the increase of the scale of point cloud data, greatly consuming computational resources, resulting in slow processing speed and increased hardware costs. This computational burden particularly affects the efficiency in application scenarios with high real-time requirements such as autonomous driving, making it difficult to achieve real-time map updates and decision-making.

[0013] 3. Insufficient data adaptability and semantic depth fusion

[0014] The distribution of point cloud data can change significantly due to factors such as sensor performance, environmental lighting, and object occlusion. Based on this, the stability and accuracy of current point cloud registration methods decrease significantly when facing data distribution changes. For example, the reflection intensity of point clouds collected under different lighting conditions may vary, resulting in an increase in registration errors. In addition, at the semantic level, although the extraction of geometric and feature information can support the generation of topological maps to a certain extent, there is a lack of in-depth fusion with the semantic logic information of the actual application scenario. This limitation may lead to the generated topological maps being unable to accurately reflect the semantic logic in tasks such as path planning and target tracking, thereby causing unreasonable route planning or target tracking failure. Summary of the Invention

[0015] Aiming at the deficiencies of the prior art, the present invention provides a method for extracting key features of a three-dimensional map and constructing a topological map of a substation construction site, overcoming the defect that the prior art only relies on the geometric features of point clouds when constructing a topological map, resulting in a lack of semantic understanding and logical rationality of topological relationships. After extracting key points, fully considering the environmental semantic information and actual application requirements, topological mapping is carried out according to the key points, so that the constructed topological map not only reflects the spatial geometric relationship, but also can reflect the connection relationship that conforms to the actual scene logic, thereby providing a more reliable map model support for applications such as path planning and target positioning.

[0016] To achieve the above objectives, the present invention is realized through the following technical solutions:

[0017] A method for extracting key features of a three-dimensional map and constructing a topological map of a substation construction site, comprising the following steps:

[0018] Step S1: Obtain the three-dimensional point cloud data and image data of the substation construction site;

[0019] Step S2: Point cloud filtering; preprocess the original point cloud data obtained in Step S1 to remove noise points;

[0020] Step S3: Point cloud registration; realizing the registration of the point cloud based on the ICP algorithm; during the registration process, first convert the point cloud preprocessed in Step S2 into a depth image, and extract the edge and corner features in the depth image through an image processing method; then map the extracted feature points back to the point cloud space; finally, perform normalization adjustment through the ICP algorithm;

[0021] Step S4: Generate a three-dimensional map of the colored point cloud; the color image is I C , and its pixel coordinates are (u, v), and the color values are (r(u, v), g(u, v), b(u, v)); through the correspondence between the point cloud and the image established during the conversion of the point cloud to the depth image in Step S3, for the point p i =(x i , y i , z i ) in the point cloud, its corresponding pixel coordinates (u i , v i ) in the color image are known; assign the color value of this pixel in the color image to the point cloud point p i , that is, p i =(x i , y i , z i , r(u i , v i ), g(u i , v i ), b(u i , v i )) to generate a three-dimensional map of the colored point cloud;

[0022] Step S5: Extract key features from the three-dimensional map generated in Step S4 based on the Mamba algorithm and construct a topological map.

[0023] Preferably, in Step S2, the preprocessing of the original point cloud data specifically includes the following steps:

[0024] Step S21: Assume the point cloud data is:

[0025]

[0026] where n is the number of points in the point cloud, and p i is a point in the point cloud, and its coordinates are (x i , y i , z i );

[0027] Step S22: Preprocess the point cloud data and remove noise points using statistical filtering;

[0028] Determine the neighborhood range with radius r and calculate each point p i 's neighborhood point set N(p i ); For point p i , its neighborhood point set is:

[0029]

[0030] In the formula, p j is the j-th point in the neighborhood point set, is the Euclidean distance between point p j and p i , and r is the radius;

[0031] Step S23: Calculate the average distance from the points in the neighborhood point set N(p i ) to p i :

[0032]

[0033] Step S24: Set the distance threshold d t , if where is the average distance estimate value of the overall point cloud, then determine that p i is a noise point and remove it from the point cloud P.

[0034] Preferably, the step S3 specifically includes the following steps:

[0035] Step S31: Convert the point cloud preprocessed in step S2 into a depth image;

[0036] Let the internal parameter matrix of the camera that captures the panoramic image be K. For the point p i =(x i , y i , z i ) in the point cloud P, its coordinates (u i , v i ) on the image plane are calculated through the perspective projection formula:

[0037]

[0038] Thus, generate the depth image I d , and the depth value d i =z i ;

[0039] Step S32: Improve the registration accuracy through depth image information fusion;

[0040] Register the two groups of point clouds and , and the corresponding point set is Let the point cloud P be extracted through depth image features A The feature points in The edge direction in the depth image is

[0041] The corresponding point cloud P B The points in The edge direction in the depth image is If (θ th is the threshold), then search for A more suitable corresponding point;

[0042] Let the point cloud P A The feature points in The depth in the converted depth image is The corresponding point cloud P B The feature points in The depth in the converted depth image is

[0043] The objective function of the ICP algorithm is modified to:

[0044]

[0045] In the formula, the transformation matrix R is the rotation matrix, t is the translation vector, and λ is the weight coefficient;

[0046] Adopt an iterative method to continuously update the corresponding point set C and the transformation matrix T until the objective function converges.

[0047] Preferably, the step S5 specifically includes the following steps:

[0048] Step S51: Tensorization processing of the input of the Mamba model;

[0049] Convert the colored point cloud into a tensor X with the shape of (B, C, N), where B is the batch size, C is the number of channels, and N is the number of points;

[0050] Step S52: Training of the Mamba model;

[0051] The Mamba model contains multiple layers and parameters. Let the parameters of the model be θ; during the training process, define the loss function as L(θ), and let the true label be y = {y i} and the model output be Then

[0052]

[0053] Through the backpropagation algorithm, update the model parameters θ according to the gradient ;

[0054] Step S53: Feature extraction;

[0055] In the stage of extracting key information, the tensor of the colored point cloud is input into the trained Mamba model, and the model outputs the prediction results of key information such as lane lines and wires;

[0056] Step S54: Topological map construction;

[0057] Based on the Mamba model, use V = (V, Γ) to model each map element;

[0058] Among them, represents the point set of the map element (N v is the number of points); Γ = Υ k represents a set of equivalent permutations of the point set V; covering all possible organizational orders; where the broken line is represented as:

[0059]

[0060] The polygon is represented as:

[0061]

[0062] Use the graph G = (V, E) to represent the topological map, where the node set V is composed of the screened key points, and the edge set E is added according to the determined connection relationship; use the adjacency matrix A to represent the connection relationship, and the element a ij is defined as:

[0063]

[0064] Step S55: Topological map optimization.

[0065] Preferably, the step S55 specifically includes the following steps:

[0066] Step S551: Connection relationship reliability evaluation and adjustment;

[0067] For each edge e in the topological map ij connecting the key points k i and k j , calculate its reliability index R ij ;

[0068] Let where t represents different timestamps or measurement batches, then the position stability of the key point k i the position stability of the key point k the position stability of the key point k j the position stability of

[0069] Direction consistency wherein is the direction vector from k measured at time t i to k j ;

[0070] Comprehensively, we can obtain: R ij = αS i + βS j + γC ij

[0071] where α, β, and γ are weight coefficients;

[0072] Step S552: Map update and consistency maintenance;

[0073] When new point cloud data is obtained, repeat the above steps to obtain a new set of key points K new and a set of connection relationships E new ;

[0074] For each k new ∈K new , calculate its distance from each point in the existing set of key points K old :

[0075] and the included angle θ new-old of the direction vectors,

[0076] compare with the corresponding threshold to determine the connection relationship, integrate the new key points and connection relationships into the topological map, and update the adjacency matrix or adjacency list;

[0077] Check the connectivity of the graph through depth-first search or breadth-first search algorithms; if there are isolated nodes or unconnected subgraphs, analyze the reasons and take corresponding measures, such as adjusting the connection relationship judgment parameters or supplementing key information, to ensure that the topological map can accurately reflect the environmental state.

[0078] The present invention provides a method for extracting key features of a three-dimensional map and constructing a topological map at a substation construction site, with the following beneficial effects: By using statistical filtering and noise point removal techniques, the accuracy of point cloud data is significantly improved, laying a solid foundation for subsequent processing. By optimizing the ICP algorithm by combining the geometric features of depth images, the number of iterations is greatly reduced, the registration accuracy and speed are improved, and the problem of local optimal solutions is avoided. By constructing a feature extraction framework based on the Mamba model, key information such as power equipment and transmission lines can be extracted using Mamba in the same complex scene point cloud dataset. Compared with the prior art, the information extraction speed can be accelerated without sacrificing the precision rate. By performing key feature extraction of the three-dimensional map based on the Mamba model, when processing point cloud data of the same scale, the consumption of computing resources is reduced compared with the complexity of the Transformer-based model in the prior art, and at the same time, the processing speed is increased, thus greatly improving the timeliness of map construction. By means of equivalent permutation and reliability evaluation methods, the ambiguity of topological relationships is eliminated, and a more practical topological map is generated, providing support for intelligent construction monitoring and management. The topological map constructed by the present invention has a greater improvement in the accuracy and efficiency of path planning. The average time taken for path planning from the starting point to the ending point at the construction site is shortened, and the planned path better conforms to the actual construction specifications and the principle of the optimal construction route, with the path error reduced, greatly enhancing the operation safety and stability of construction vehicles and personnel in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the prior art.

[0080] Figure 1 The flowchart of the steps of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0081] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions in the present invention with reference to the accompanying drawings in the present invention.

[0082] Embodiment

[0083] As Figure 1 shown, a method for extracting key features of a three-dimensional map and constructing a topological map at a substation construction site includes the following steps:

[0084] Step S1: Obtain three-dimensional point cloud data and image data of the substation construction site based on a three-dimensional scanning lidar, a panoramic camera, an inertial navigation, a high-precision positioning module, etc.

[0085] Step S2: Point cloud filtering; preprocess the original point cloud data obtained in Step S1 to remove noise points; to improve data quality and processing efficiency;

[0086] Specifically, it includes:

[0087] Step S21: Let the point cloud data be:

[0088]

[0089] where n is the number of points in the point cloud, and p i is a point in the point cloud, and its coordinates are (x i , y i , z i );

[0090] Step S22: Preprocess the point cloud data and use statistical filtering to remove noise points;

[0091] Determine the neighborhood range with radius r, and calculate the neighborhood point set N(p i ) of each point p; for the point p i , its neighborhood point set is: i where p

[0092]

[0093] is the j-th point in the neighborhood point set, j is the Euclidean distance between the point p and p j , and r is the radius; i

[0094] Step S23: Calculate the average distance from the points in the neighborhood point set N(p i ) to p i :

[0095]

[0096] Step S24: Set the distance threshold d t , if where is the average distance estimate value of the overall point cloud, then determine that p i is a noise point and remove it from the point cloud P.

[0097] ​Step S3: Point cloud registration; realizing the registration of point clouds based on the ICP (Iterative Closest Point) algorithm; during the registration process, the corresponding points between two groups of point clouds need to be determined. By converting the point clouds preprocessed in step S2 into depth images, more stable and representative feature points can be extracted. Among them, features such as edges and corner points in the depth image can be extracted by simple image processing methods (such as edge detection algorithms, which are existing technologies and will not be elaborated here). These feature points correspond to the geometric feature change points of the object in the point cloud space, such as the boundary points of power equipment, transmission lines in the substation, and corner points of buildings. Mapping these feature points extracted from the depth image back to the point cloud space provides a more accurate initial guess of the corresponding points for the ICP algorithm; finally, normalization adjustment is performed through the ICP algorithm;

[0098] Specifically, it includes:

[0099] Step S31: Converting the point clouds preprocessed in step S2 into depth images;

[0100] Let the internal parameter matrix of the camera that captures the panoramic image be K. For the point p i =(x i ,y i ,z i ) in the point cloud P, its coordinates (u i ,v i ) on the image plane are calculated through the perspective projection formula:

[0101]

[0102] Thus, the depth image I d is generated, and the depth value d i =z i ;

[0103] In addition, the ICP algorithm is an iterative fine registration method, which is sensitive to the initial position. If the initial position deviation is too large, it may fall into a local optimal solution or result in a too slow convergence speed. The ICP algorithm will continuously update the corresponding points during the iteration process. The feature information provided by the depth image can assist in judging whether the currently found corresponding points are reasonable. For example, if the features (such as edge direction and gradient amplitude) of two corresponding points are found to be too different in the depth image, then the corresponding relationship can be normalized and adjusted in the ICP algorithm.

[0104] Specifically as follows:

[0105] Step S32: Depth image information fusion to improve the registration accuracy;

[0106] For two groups of point clouds and For registration, the corresponding point set is Suppose the point cloud P extracted by deep image features A The feature points in The edge direction in the depth image is

[0107] The corresponding point cloud P B Points in The edge direction in the depth image is if (θ th is the threshold), then search again More appropriate corresponding points;

[0108] Set point cloud P A The feature points in The depth in the converted depth image is The corresponding point cloud P B The feature points in The depth in the converted depth image is

[0109] The objective function of the ICP algorithm is modified as follows:

[0110]

[0111] In the formula, the transformation matrix R is the rotation matrix, t is the translation vector, and λ is the weight coefficient;

[0112] An iterative approach is usually used to continuously update the corresponding point set C and the transformation matrix T until the objective function converges.

[0113] Step S4: Generate a three-dimensional map of the color point cloud; the color image is I C , its pixel coordinates are (u, v), and its color value is (r(u, v), g(u, v), b(u, v)); through the point cloud and image correspondence established in the process of converting the point cloud into a depth image in step S3, for point p in the point cloud i =(x i ,y i ,z i ), whose corresponding pixel coordinates in the color image (u i ,v i ) is known; assign the color value of the pixel in the color image to the point cloud point p i , that is, p i =(x i ,y i ,z i ,r(u i ,v i ),g(u i ,vi ), b(u i , v i )) to generate a three-dimensional map of the colored point cloud;

[0114] Step S5: Extract key features from the three-dimensional map generated in Step S4 based on the Mamba algorithm and construct a topological map.

[0115] Specifically, it includes:

[0116] Step S51: Tensorization processing of the input of the Mamba model;

[0117] Convert the colored point cloud into a tensor X with the shape of (B, C, N), where B is the batch size, C is the number of channels, and N is the number of points;

[0118] Step S52: Training of the Mamba model;

[0119] The Mamba model contains multiple layers and parameters. Let the parameters of the model be θ; during the training process, define the loss function as L(θ). For example, the classification task of the power equipment contour and the transmission line can use the cross-entropy loss. Let the true label be y = {y i}, and the model output is Then

[0120]

[0121] Through the backpropagation algorithm, update the model parameters θ according to the gradient ;

[0122] Step S53: Feature extraction;

[0123] In the stage of extracting key information, input the tensor of the colored point cloud into the trained Mamba model, and the model outputs the prediction results of key information such as lane lines and wires; for example, the probability distribution of the key points of the power equipment is output as P dev (k) (k is the key point index). If P dev (k) > P th (P th is the probability threshold), then regard this point as the key point of the power equipment.

[0124] Step S54: Topological map construction;

[0125] It is crucial to determine the link relationships between key points and expand the graph structure of the topological map. The open shape elements are discretized into a set of points on a broken line, initially realizing a representation of map elements. However, at this time, the arrangement order of the point set is not clearly defined and is not unique, resulting in ambiguity. When constructing the map, since the starting points and directions of power equipment and transmission lines are ambiguous, each endpoint of the surface of the power equipment line can be regarded as a starting point, and the point set can be organized in multiple directions. For transmission lines, each point of the formed polygon can be regarded as a starting point, and the polygon can be connected in two opposite directions (clockwise and counterclockwise). To bridge this gap, the present invention models each map element based on the Mamba model with V = (V, Γ);

[0126] where, V represents the set of points of the map element (N v is the number of points); Γ = Υ k represents a set of equivalent permutations of the point set V; covering all possible organization orders; where the broken line is represented as:

[0127]

[0128] The polygon is represented as:

[0129]

[0130] By introducing equivalent permutations, the map elements are modeled in a unified manner, solving the ambiguity problem. To achieve a more comprehensive environmental perception and understanding, the graph structure of the topological map is further constructed.

[0131] The topological map is represented by the graph G = (V, E), where the node set V consists of the screened key points, and the edge set E is added according to the determined connection relationships; the adjacency matrix A is used to represent the connection relationships, and the element a ij is defined as:

[0132]

[0133] Step S55: Topological map optimization;

[0134] (1) Reliability evaluation and adjustment of connection relationships;

[0135] For each edge e ij in the topological map that connects the key points k i and k j , calculate its reliability index R ij ;

[0136] Let where t represents different timestamps or measurement batches, then the position stability i of the key point k Key point k j Position stability Direction consistency where k is measured at time t i to k j direction vector;

[0137] Comprehensively, we can get: R ij = αS i + βS j + γC ij

[0138] where α, β, γ are weight coefficients;

[0139] Assume the reliability threshold R th , if R ij < R th , then re-evaluate the connection relationship between k i and k j , and it is necessary to adjust the connection judgment criteria or re-perform key point extraction and connection relationship determination.

[0140] (2) Map update and consistency maintenance;

[0141] When new point cloud data is obtained, repeat the above steps to obtain a new set of key points K new and connection relationship set E new ; for each k new ∈ K new , calculate its distance from each point in the existing set of key points K old :

[0142] and the included angle θ of the direction vector new-old (similar to the previous calculation method), compare with the corresponding threshold to determine the connection relationship, integrate the new key points and connection relationships into the topological map, and update the adjacency matrix or adjacency list. To check the consistency of the updated topological map, check the connectivity of the graph through depth-first search (DFS) or breadth-first search (BFS) algorithms. If there are isolated nodes or unconnected subgraphs, analyze the reasons and take corresponding measures, such as adjusting the connection relationship judgment parameters or supplementing key information, to ensure that the topological map can accurately reflect the environmental state.

[0143] Further optimize and implement applications from the extracted features and constructed topological structures, including path planning and incremental mapping. When path planning is required, it is usually necessary to combine graph search algorithms to find the best path from the starting point to the target point in the topological structure, such as the A* algorithm or the Dijkstra algorithm. The Dijkstra algorithm is a greedy algorithm that does not consider heuristics and only performs shortest path search based on the actual costs between nodes. Incremental mapping is the ability to dynamically update the map when new sensor data is available, which is suitable for real-time navigation and dynamic environment perception. During the construction of a 3D map, new poses (positions and orientations) and corresponding measurements are incrementally added to the map, and optimization is used to maintain the accuracy of the map. Or, by comparing the current field of view and the historical field of view, feature matching techniques are used to detect visited locations and perform loop closure detection to reduce cumulative errors and improve map accuracy.

[0144] The present invention utilizes statistical filtering and noise point removal techniques to significantly improve the accuracy of point cloud data, laying a solid foundation for subsequent processing. By optimizing the ICP algorithm in combination with the geometric features of depth images, the number of iterations is greatly reduced, the registration accuracy and speed are improved, and the problem of local optimal solutions is avoided. By constructing a feature extraction framework based on the Mamba model, key information such as power equipment and transmission lines can be extracted using Mamba in the same complex scene point cloud dataset. Compared with existing technologies (such as MapTR), Mamba can accelerate the information extraction speed without sacrificing the precision rate, especially in the processing of large-scale traffic environment data, where the processing time is shortened. In terms of power construction, accurate extraction of key substation information helps to timely discover potential safety hazards at the line construction site, improve the efficiency of power construction safety management and control, reduce power outages caused by construction safety, and thus ensure the normal progress of social production and life.

[0145] In addition, the present invention performs key feature extraction of 3D maps based on the Mamba model. When processing point cloud data of the same scale, its computational resource consumption is reduced compared to the complexity of some traditional Transformer-based models. At the same time, the processing speed is increased. For example, when processing the point cloud data of a large and complex substation to construct a map, the existing technology may require more than 10GB of memory and several hours of computing time, while adopting this solution reduces the memory and inference time, greatly improving the timeliness of map construction. In actual application scenarios, such as the real-time map construction and update of a substation construction site, lower resource consumption and higher processing speed mean that the construction environment can update the 3D map faster, obtain accurate map information in a timely manner, and thus make more reasonable safety decisions, improving the safety of substation construction.

[0146] In addition, through the equivalent arrangement and reliability evaluation method, the present invention eliminates the ambiguity of topological relationships, generates a more practical topological map, and provides support for intelligent construction monitoring and management. The topological map constructed by the present invention has greatly improved accuracy and efficiency in path planning. The average time consumed for path planning from the starting point to the ending point at the construction site is shortened, and the planned path better conforms to the actual construction specifications and the principle of the optimal construction route, with reduced path errors, greatly enhancing the running safety and stability of construction vehicles and personnel in complex environments.

[0147] The map is updated in real time with the construction site through the incremental mapping and loop closure detection functions, effectively enhancing the adaptability of navigation and construction planning. In scenarios such as substation construction sites and emergency rescue, an accurate and efficient topological map can enable construction personnel to obtain the best construction path faster, shorten the troubleshooting time, and improve the inspection efficiency, which has important social value.

[0148] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for extracting key features of a three-dimensional map of a substation construction site and constructing a topological map, characterized in that: The following steps are involved: Step S1: Acquire three-dimensional point cloud data and image data of the substation construction site; Step S2: point cloud filtering: preprocessing the original point cloud data obtained in step S1 to remove noise points; Step S3: point cloud registration; point cloud registration is implemented based on the ICP algorithm; in the registration process, the point cloud preprocessed in step S2 is first converted into a depth image, and the edge and corner features in the depth image are extracted by an image processing method; the extracted feature points are then mapped back to the point cloud space; and finally, normalization adjustment is performed by the ICP algorithm; Step S4: Generate a three-dimensional map of the color point cloud; the color image is I C , its pixel coordinates are (u, v), and its color value is (r(u, v), g(u, v), b(u, v)); through the point cloud and image correspondence established in the process of converting the point cloud into a depth image in step S3, for point p in the point cloud i =(x i ,y i ,z i ), whose corresponding pixel coordinates in the color image (u i ,v i ) is known; assign the color value of the pixel in the color image to the point cloud point p i , that is, p i =(x i ,y i ,z i ,r(u i ,v i ),g(u i ,v i ),b(u i ,v i )) to generate a three-dimensional map of the colored point cloud; Step S5: extract key features of the three-dimensional map generated in step S4 based on the Mamba algorithm and construct a topological map.

2. A method for extracting key features of a three-dimensional map of a substation construction site and constructing a topological map according to claim 1, characterized in that: In step S2, preprocessing the original point cloud data specifically includes the following steps: Step S21: Set the point cloud data to: Where n is the number of points in the point cloud, p i is a point in the point cloud, and its coordinates are (x i ,y i ,z i ); Step S22: pre-processing the point cloud data and removing noise points using statistical filtering; Determine the neighborhood range with radius r and calculate each point p i The neighborhood point set N(p i ); for point p i , its neighborhood point set is: In the formula, p j is the jth point in the neighborhood point set, For point p j With p i The Euclidean distance between them, r is the radius; Step S23: Calculate the neighborhood point set N(p i ) from the inner point to p i The average distance: Step S24: Setting the distance threshold d t ,like in is the average distance estimate of the entire point cloud, then p i are noise points and are removed from the point cloud P.

3. A method for extracting key features of a three-dimensional map of a substation construction site and constructing a topological map according to claim 1, characterized in that: The step S3 specifically comprises the following steps: Step S31: converting the point cloud preprocessed in step S2 into a depth image; Assume that the intrinsic parameter matrix of the camera that takes the panoramic image is K. For point p in point cloud P, i =(x i ,y i ,z i ), whose coordinates on the image plane (u i ,v i ), calculated by the perspective projection formula: This generates a depth image I d , depth value d i =z i ; Step S32: Fusion of deep image information improves registration accuracy; For two sets of point clouds and For registration, the corresponding point set is Suppose the point cloud P extracted by deep image features A The feature points in The edge direction in the depth image is The corresponding point cloud P B Points in The edge direction in the depth image is if (θ th is the threshold), then search again More appropriate corresponding points; Set point cloud P A The feature points in The depth in the converted depth image is The corresponding point cloud P B The feature points in The depth in the converted depth image is The objective function of the ICP algorithm is modified as follows: In the formula, the transformation matrix R is the rotation matrix, t is the translation vector, and λ is the weight coefficient; The corresponding point set C and the transformation matrix T are continuously updated in an iterative manner until the objective function converges.

4. A method for extracting key features of a three-dimensional map of a substation construction site and constructing a topological map according to claim 1, characterized in that: The step S5 specifically comprises the following steps: Step S51: tensor quantization of Mamba model input; Convert the colored point cloud into a tensor X of shape (B,C,N), where B is the batch size, C is the number of channels, and N is the number of points; Step S52: Mamba model training; The Mamba model contains multiple layers and parameters. Let the model parameters be θ. During the training process, the loss function is defined as L(θ) and the true label is y = {y i }, the model output is but Through the back propagation algorithm, according to the gradient To update the model parameters θ; Step S53: feature extraction; In the key information extraction stage, the colored point cloud tensor is input into the trained Mamba model, and the model outputs the prediction results of key information such as lane lines and wires; Step S54: constructing a topological map; Based on the Mamba model, each map element is modeled using V = (V, Γ); in, The point set representing the map element (N v is the number of points); Γ=Υ k Represents a set of equivalent arrangements of the point set V; covers all possible organization orders; where the polyline is represented as: The polygon is represented as: The topological map is represented by a graph G = (V, E), where the node set V is composed of the selected key points, and the edge set E is added according to the determined connection relationship; the adjacency matrix A is used to represent the connection relationship, and the element a ij Defined as: Step S55: topology map optimization.

5. A method for extracting key features of a three-dimensional map of a substation construction site and constructing a topological map according to claim 4, characterized in that: The step S55 specifically The following steps are involved: Step S551: connection relationship reliability assessment and adjustment; For each edge e in the topological map ij Connect key points k i and k j , calculate its reliability index R ij ; set up Where t represents different timestamps or measurement batches, then the key point k i Position stability Key point k j Position stability Directional consistency in is k measured at time t i to k j The direction vector of Comprehensively obtained: R ij =αS i +βS j +γC ij Among them, α, β, and γ are weight coefficients; Step S552: map update and consistency maintenance; When new point cloud data is obtained, repeat the above steps to obtain a new key point set K new and the connection relationship set E new ; For each k new ∈K new , calculate its difference with the existing key point set K old The distance between points in: The angle between the direction vector and new-old , Compare with the corresponding threshold to determine the connection relationship, integrate the new key points and connection relationship into the topological map, and update the adjacency matrix or adjacency table; Check the connectivity of the graph through depth-first search or breadth-first search algorithm; if there are isolated nodes or disconnected subgraphs, analyze the reasons and take corresponding measures, such as adjusting the connection relationship judgment parameters or supplementing key information to ensure that the topological map can accurately reflect the environmental status.

Images