A map construction method and system for high-level task requirements

Abstract

The present invention discloses a map construction method and system for high-level task requirements, and relates to the field of map construction technology. The method includes: introducing an image saliency recognition strategy to perform saliency recognition analysis on the global image of the target area to obtain the most salient block; matching the most salient image frame of the most salient block, and using the most salient image frame as the initial frame; performing three-dimensional point cloud analysis to construct an initial map; establishing a first-level neighborhood space of the initial node in the initial feature vector, and matching the first-level neighborhood node set of the first-level neighborhood space; based on a predetermined graph neural network enhancement mechanism, enhancing and updating the initial map to obtain an enhanced map. The present invention solves the technical problem in the prior art that key areas cannot be efficiently and accurately identified from complex images, and the constructed map has low precision, and achieves the technical effect of improving the efficiency and accuracy of map construction and meeting the requirements of complex application scenarios.

Description

Technical Field

[0001] The present invention relates to the technical field of map construction, and in particular to a map construction method and system oriented to high-level task requirements. Background Art

[0002] In the context of increasingly stringent requirements for map accuracy in high-level task application scenarios such as intelligent transportation and urban planning, traditional map construction methods face many difficulties. They mostly rely on a single data source, lack details in satellite remote sensing images to build maps, and have limited range and complex processing of vehicle-mounted laser scanning data. In the initial stage, there is a lack of effective significant recognition strategies, and it is difficult to quickly and accurately locate key areas due to interference from factors such as noise. The lack of a verification mechanism when building the initial map leads to large errors. In terms of map enhancement and updating, the information of neighboring nodes is not fully mined, and the requirements for high precision and semantic richness of maps in autonomous driving, urban planning, and virtual reality cannot be met, which limits the effective application of maps in high-level tasks.

[0003] The existing technology has the technical problem of being unable to efficiently and accurately identify key areas from complex images and the constructed map has low accuracy. Summary of the invention

[0004] The present application provides a map construction method and system for high-level task requirements, which is used to solve the technical problem that the existing technology cannot efficiently and accurately identify key areas from complex images, resulting in low accuracy of the constructed map.

[0005] In view of the above problems, the present application provides a map construction method and system for high-level task requirements.

[0006] In a first aspect of the present application, a map construction method for high-level task requirements is provided, the method comprising:

[0007] An image saliency recognition strategy is introduced to perform saliency recognition analysis on the global image of the target area to obtain the most salient block; the most salient image frame of the most salient block is matched in the image frame database of the target area, and the most salient image frame is used as the initial frame; a three-dimensional point cloud analysis is performed in combination with an initial feature vector of the initial frame to construct an initial map; a first-level neighborhood space of the initial node in the initial feature vector is established, and a first-level neighborhood node set of the first-level neighborhood space is matched in the image frame database; based on a predetermined graph neural network enhancement mechanism, the initial map is enhanced and updated using the first-level neighborhood node set to obtain an enhanced map.

[0008] The second aspect of the present application provides a map construction system for high-level task requirements, the system comprising:

[0009] A most significant block acquisition module, the most significant block acquisition module is used to introduce an image significant recognition strategy to perform significant recognition analysis on the global image of the target area to obtain the most significant block; an initial frame acquisition module, the initial frame acquisition module is used to match the most significant image frame of the most significant block in the image frame database of the target area, and use the most significant image frame as the initial frame; an initial map construction module, the initial map construction module is used to perform three-dimensional point cloud analysis in combination with an initial feature vector of the initial frame to construct an initial map; a first-level neighborhood space establishment module, the first-level neighborhood space establishment module is used to establish a first-level neighborhood space of an initial node in the initial feature vector, and match a first-level neighborhood node set of the first-level neighborhood space in the image frame database; an enhanced map acquisition module, the enhanced map acquisition module uses the first-level neighborhood node set to enhance and update the initial map based on a predetermined graph neural network enhancement mechanism to obtain an enhanced map.

[0010] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0011] The image saliency recognition strategy is introduced to perform saliency recognition analysis on the global image of the target area to obtain the most salient block; the most salient image frame of the most salient block is matched, and the most salient image frame is used as the initial frame; the initial feature vector of the initial frame is combined to perform three-dimensional point cloud analysis to construct an initial map; the first-level neighborhood space of the initial node in the initial feature vector is established, and the first-level neighborhood node set of the first-level neighborhood space is matched; based on the predetermined graph neural network enhancement mechanism, the first-level neighborhood node set is used to enhance and update the initial map to obtain an enhanced map. The technical effect of improving the efficiency and accuracy of map construction and meeting the needs of complex application scenarios is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0013] Figure 1 A flowchart of a map construction method for high-level task requirements provided in an embodiment of the present application;

[0014] Figure 2 A schematic diagram of the structure of a map building system for high-level task requirements provided in an embodiment of the present application.

[0015] Explanation of reference numerals: most prominent block acquisition module 10 , initial frame acquisition module 20 , initial map construction module 30 , first-level neighborhood space establishment module 40 , enhanced map acquisition module 50 . DETAILED DESCRIPTION

[0016] The present application provides a map construction method and system for high-level task requirements, which is used to solve the technical problem that the existing technology cannot efficiently and accurately identify key areas from complex images, resulting in low accuracy of the constructed map.

[0017] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0018] Embodiment 1, as Figure 1 As shown, the present application provides a map construction method for high-level task requirements, the method comprising:

[0019] Step S100: introducing an image salient recognition strategy to perform salient recognition analysis on the global image of the target area to obtain the most salient block.

[0020] Specifically, in the initial stage of map construction, an image saliency recognition strategy is introduced to perform saliency recognition analysis on the global image of the target area. First, the image saliency recognition strategy is based on the perception principle of the human visual system for salient objects, and aims to quickly locate the most attention-grabbing areas, that is, the most salient blocks, from the complex global image information. In order to achieve this goal, a variety of advanced image processing technologies and algorithms are used. For example, a calculation method based on color contrast is used to analyze the degree of difference in color between each pixel in the image and its surrounding pixels. In an image of the target area containing multiple colors, areas with high color contrast, such as bright red flowers against a green grass background, are more likely to be regarded as salient areas. At the same time, brightness differences are also an important consideration. Bright objects will stand out against a relatively dark background, such as the area where bright street lights are located in a night scene. Texture features should not be ignored either. Areas with unique textures, such as special texture patterns on the surface of buildings or special texture changes in terrain, are more significant than smooth and single areas. Through a comprehensive quantitative analysis of these color, brightness and texture features, a saliency value is assigned to each block in the image. Then, the saliency values ​​of all blocks are compared and screened to determine the block with the highest saliency value, which is the most salient block. It will serve as an important reference point for subsequent map construction, lay the foundation for the entire map construction process, and guide the subsequent steps towards building accurate and practical maps.

[0021] Step S200: matching the most significant image frame of the most significant block in the image frame database of the target area, and taking the most significant image frame as the initial frame.

[0022] Specifically, in the process of map construction, the most significant image frame of the most significant block is accurately matched in a specific image frame database and set as the initial frame. The image frame database is a set of image frames constructed by collecting block-by-block images of the target area through a special image acquisition device. These image frames contain rich information, and each image frame has geometric and semantic feature identification. The geometric feature identification covers information such as the shape, size, and position relationship of objects in the image frame, such as the outline shape of a building, the linear direction of a road, etc.; the semantic feature identification gives the image frame a deeper meaning, such as a certain area is marked as "park", "commercial area" or "residential area". When matching the most significant image frame of the most significant block in this database, the feature information of the most significant block determined in the previous step, such as unique color distribution, significant texture structure or specific shape characteristics, is first used as the key basis for the search. Then, these features are compared one by one with the geometric features and semantic feature identifications of each image frame in the database through the image matching algorithm. This image matching algorithm takes into account a variety of factors, not just simple pixel-level similarity, but also the correspondence of feature points, structural similarity of regions, and relevance of semantic information. For example, if the most prominent block has a unique circular structure and is determined to be a possible landmark building, the algorithm will search the database for image frames with similar circular structural geometric features and semantic features identified as "landmark buildings" or related to them. After the calculation and screening process, the image frame that best matches the most prominent block is finally determined, that is, the most prominent image frame. Using this most prominent image frame as the initial frame, it will provide an accurate and representative starting point for subsequent 3D point cloud analysis and map construction, ensuring that the entire map construction process can be gradually carried out based on accurate initial information, thereby improving the quality and accuracy of the final constructed map.

[0023] Step S300: performing a three-dimensional point cloud analysis in combination with the initial feature vector of the initial frame to construct an initial map.

[0024] Specifically, the key work of building the initial map is officially launched, mainly focusing on the three-dimensional point cloud analysis combined with the initial feature vector of the initial frame to achieve the goal of building the initial map. As an important starting point of the whole process, the initial frame contains a rich variety of information in its initial feature vector, which covers visual attributes such as color, brightness, texture, and geometric features such as edges and corners of objects in the image. It also contains semantic information extracted from the image, such as object categories and regional attributes in the scene. When performing three-dimensional point cloud analysis, computer vision technology and algorithms are first used to extract information related to spatial position from the initial feature vector. For example, through the principle of stereoscopic vision, the depth information of the object in three-dimensional space is calculated according to the parallax of the object in the image at different viewing angles. For each pixel or feature point in the initial frame, a corresponding depth value is assigned to it, thereby converting the two-dimensional image information into point cloud data in three-dimensional space. These point cloud data present a discrete distribution state in three-dimensional space, representing the position of the surface points of the object in the target area in space. Then, according to the spatial relationship between points in the point cloud data, point cloud processing algorithms are used to perform clustering, segmentation and fitting operations. For example, points belonging to the same object surface are clustered together, the contour boundaries of different objects are identified, and then the three-dimensional geometric model of the object is constructed through the fitting algorithm, such as the three-dimensional shape of the building, the undulation of the terrain, etc. During the construction process, the semantic information in the initial feature vector is also combined to classify and label different objects and areas. For example, the point cloud model identified as a building is labeled as the "building" category, and the terrain area is labeled as "ground", "hillside" and other categories. In this way, an initial map containing the three-dimensional spatial information of the target area, the geometric shape of the object and the semantic annotation is gradually constructed, providing a basic framework for the subsequent enhancement and improvement of the map.

[0025] Step S400: establishing a primary neighborhood space of an initial node in the initial feature vector, and matching a primary neighborhood node set of the primary neighborhood space in the image frame database.

[0026] Specifically, for the initial node in the initial feature vector, its first-level neighborhood space is established. As the key reference point in the initial map construction process, the definition of its first-level neighborhood space is of great significance. By setting certain spatial range parameters, a region containing relevant information around the initial node is determined with the initial node as the center. The points or elements in this region have a direct relationship with the initial node, such as being adjacent in geographic space, or having similarity or continuity in image features. Subsequently, the work of matching the first-level neighborhood node set of the first-level neighborhood space is carried out in the image frame database. The image frame database stores a large amount of rich and labeled image frame information, which provides sufficient data basis for finding the first-level neighborhood node set. In the matching process, the search and comparison algorithm is used to carefully compare the feature attributes of the first-level neighborhood space with the geometric features and semantic feature identifiers contained in the image frame in the database. From the perspective of geometric features, factors such as the distance relationship, angle relationship and shape similarity between nodes are considered; from the perspective of semantic features, attention is paid to whether the area type and object attributes represented by the node match the first-level neighborhood space. For example, if the initial node is located near a building, then its first-level neighborhood space covers surrounding roads, ancillary facilities and other areas. When matching in the database, we will look for nodes that are geometrically close to the building and semantically related to the surrounding environment of the building, such as nodes on nearby roads, nodes in parking lots, etc. After calculation and screening, we will eventually determine the first-level neighborhood node set that is highly matched with the first-level neighborhood space. These node sets will provide key supplementary information for the subsequent enhancement and update of the initial map based on the predetermined graph neural network enhancement mechanism, further enriching and refining the map construction content.

[0027] Step S500: Based on a predetermined graph neural network enhancement mechanism, the initial map is enhanced and updated using the first-level neighborhood node set to obtain an enhanced map.

[0028] Specifically, based on the predetermined graph neural network enhancement mechanism, the initial map is enhanced and updated using the first-level neighborhood node set to obtain an enhanced map. First, the first-level neighborhood node set is input into the predetermined graph neural network, which constructs a special architecture in which nodes represent key elements in the map and edges represent the spatial or semantic relationship between elements. The first-level neighborhood node set is used as input data to trigger a series of complex calculation processes in the network. The neurons in the graph neural network learn the characteristic patterns of the first-level neighborhood node set, including the geometric position, semantic attributes, and connection relationship between nodes, to explore the deep-level rules hidden in it. For example, it can identify the distribution rules of different types of nodes (such as building nodes, road nodes, etc.) in the node set and the mutual influence relationship between them. Based on the learned information, the graph neural network begins to enhance and update the initial map. At the geometric level, according to the connection status and trend of the road nodes in the first-level neighborhood node set, the shape, width and extension direction of the road are more accurately depicted, the outline and position relationship of the building are optimized, and the parts with inaccurate or missing geometric shapes that may exist in the initial map are filled. At the semantic level, by analyzing the semantic features of the first-level neighborhood node set, more accurate semantic labels are assigned to the regions and objects in the map, further clarifying the functional attributes of different regions, such as distinguishing between commercial areas and residential areas. After such a comprehensive and in-depth enhancement and update process, the initial map gradually evolves into a more accurate, detailed and semantically rich enhanced map, thus better meeting the requirements of high-level tasks, such as intelligent navigation, urban planning analysis and other fields that require high precision and high reliability of maps.

[0029] In a possible implementation, step S100 further includes:

[0030] Step S110: Acquire an image acquisition device, where the image acquisition device refers to a device for acquiring images of the target area;

[0031] Step S120: determining a unit acquisition length according to a predetermined width and a predetermined distance of the image acquisition device;

[0032] Step S130: performing segmentation processing on the global image based on the unit acquisition length to obtain a segmentation processing result;

[0033] Step S140: introducing the image salient recognition strategy to analyze the multiple blocks in the segmentation processing result to obtain the most salient block.

[0034] Specifically, first obtain the image acquisition device, which is the information source of the entire process and is responsible for collecting images of the target area. There are various types of image acquisition devices, such as high-definition cameras and aerial cameras carried by drones. Their performance and parameters directly affect the quality and availability of subsequent images.

[0035] The unit acquisition length is determined according to the predetermined width and predetermined distance of the image acquisition device. The predetermined width of the image acquisition device determines the lateral range that can be covered in each shot, and the predetermined distance involves the spatial relationship between the shooting point and the target area. Through these two key parameters, the width is divided proportionally according to the predetermined distance to determine the unit acquisition length. This unit acquisition length is the basic unit of measurement for subsequent precise processing of the global image. It ensures the systematic and standardized nature of image acquisition, so that images acquired at different positions and angles can be integrated and analyzed under a unified standard.

[0036] The global image is segmented based on the unit acquisition length to obtain the segmentation result. The entire global image is regarded as a continuous two-dimensional plane, and is divided into multiple regular blocks with the unit acquisition length as the side length. Each block becomes a relatively independent image unit. This segmentation method helps to reduce the complexity of image analysis and decompose the huge global image into small parts that are easy to process. For example, for a large urban area image, after segmentation according to the unit acquisition length, it can be divided into multiple block-sized blocks. Each block contains image information within a specific range in the area, such as buildings, roads, greenery and other elements.

[0037] In order to accurately determine the most significant block from multiple blocks in the segmentation processing results, an advanced and complex image saliency recognition strategy is introduced. First, in terms of color feature analysis, a quantization method based on color space models (such as RGB, HSV, etc.) is adopted. For each block, its color histogram is calculated and the distribution of different color channels is statistically analyzed. The color contrast is measured by comparing the differences in color histograms between different blocks. For example, the entropy value of the color in the block is calculated. The higher the entropy value, the more complex and diverse the color distribution is, and the more significant the block may be in terms of color features. At the same time, color moments (mean, variance, skewness, etc.) are used to describe the overall characteristics of the color. Blocks with large color mean differences from surrounding blocks and high variance (rich color changes) are given higher saliency scores. In terms of brightness feature analysis, the average brightness value of each block and the gradient change of brightness are calculated. Areas with large brightness gradients, such as edges that change sharply from dark to bright, tend to attract more visual attention. The gradient map of the brightness image is obtained by convolution operation, and then the sum or mean of the gradient values ​​in each block is counted as an indicator to measure the brightness significance of the block. Texture feature analysis uses the gray level co-occurrence matrix (GLCM) technology. The GLCM of each block in different directions (such as 0°, 45°, 90°, 135°) is calculated, and texture feature parameters such as contrast, energy, entropy, and correlation are extracted from the GLCM. For example, high-contrast textures indicate drastic texture changes, which may correspond to complex textures on the surface of buildings or undulating changes in natural terrain; high-energy textures reflect the regularity and repetitiveness of textures, such as crop arrangement textures in fields. Combining these texture feature parameters, the texture saliency value of each block is calculated through a pre-set weight formula. In shape feature analysis, edge detection algorithms (such as Canny edge detection) are used to extract the edge contours of the blocks, and then the geometric properties of the edges such as length, curvature, and closure are calculated. Blocks with longer edges, larger curvature changes (such as unique shapes such as circles and triangles) or higher closure (such as complete building outlines) are more significant in shape features. By normalizing these shape feature parameters and combining them with a specific shape saliency model, the shape-based saliency score of each block is calculated. Finally, the saliency scores of the four aspects of color, brightness, texture and shape are integrated, and a comprehensive saliency value is calculated for each block according to the preset weighted summation formula. The comprehensive saliency values ​​of all blocks are sorted, and the block with the highest comprehensive saliency value is the most salient block. It will serve as an important reference in the subsequent map construction process, guiding map construction towards a more accurate and efficient direction.

[0038] In a possible implementation, step S140 further includes:

[0039] Step S141: extracting any block from the plurality of blocks.

[0040] Step S142: performing feature quantification analysis on the arbitrary block according to the block feature quantification plan in the image saliency recognition strategy to obtain a feature value of the arbitrary block.

[0041] Step S143: using the arbitrary block feature value as a constraint to screen and obtain a maximum feature value, and reversely matching the block corresponding to the maximum feature value, which is recorded as the most significant block.

[0042] Specifically, extract any block from multiple blocks. After completing the segmentation of the global image based on the unit acquisition length, a large number of blocks will be obtained. At this time, a block is randomly selected from these block sets as the object of subsequent analysis. The selection of this arbitrary block is the beginning of the entire feature quantitative analysis, which provides a specific sample for in-depth exploration of the significant features of each block.

[0043] According to the block feature quantification plan in the image saliency recognition strategy, the feature quantification analysis of any block is performed to obtain the feature value of any block. The block feature quantification plan in the image saliency recognition strategy covers multiple key aspects of feature quantification methods. For color features, the pixel color information in the block will be converted to a specific color space (such as HSV space), and the statistics of color distribution, such as the mean and variance of the color histogram, will be calculated. These statistics are combined to form a color feature vector, and then weighted summed with other feature vectors through a pre-set weight matrix to obtain partial eigenvalue contributions. In terms of brightness features, the average value, standard deviation and brightness gradient information of the pixel brightness in the block are calculated, and these information are converted into brightness eigenvalue components through brightness-related algorithm models. Texture features use grayscale co-occurrence matrices to calculate texture parameters such as contrast, energy, and entropy. These parameters are normalized and used as components of texture eigenvalues. Shape features use edge detection algorithms to obtain the edge contours of blocks, calculate geometric metrics such as edge length, curvature, and complexity, and then convert these metrics into shape eigenvalues ​​through shape feature quantization functions. Finally, the feature values ​​of color, brightness, texture and shape are fused and calculated according to a predetermined integration rule to obtain the comprehensive feature value of the arbitrary block, that is, the feature value of the arbitrary block.

[0044] The maximum eigenvalue is obtained by using any block eigenvalue as a constraint, and the block corresponding to the maximum eigenvalue is matched in reverse, which is recorded as the most significant block. After obtaining the eigenvalue of any block of each block, these eigenvalues ​​are compared, and the maximum value is found by traversing the eigenvalues ​​of all blocks. This maximum value represents the quantitative performance of the feature with the highest degree of significance among all blocks. Then, based on the correspondence between the eigenvalue and the block, the block with the maximum eigenvalue is reversely found and determined as the most significant block. This most significant block will serve as an important identification point in the map construction process, providing a key reference for subsequent initial frame selection, map construction, and further enhancement and update operations based on the block, ensuring that the map construction work can be gradually carried out based on the most representative and significant areas, thereby improving the accuracy and effectiveness of the map.

[0045] In a possible implementation, step S142 further includes:

[0046] Step S1421: performing weighted calculation on the structural features and base tone features of the arbitrary block according to the block feature quantization plan to obtain the feature value of the arbitrary block, wherein the base tone features include red features, green features and blue features.

[0047] Specifically, first of all, for the analysis of structural features, a variety of image processing techniques are used. For example, edge detection algorithms are used to extract edge contour information of objects in the block, and indicators such as edge length, density, and complexity are calculated. Long and complex edges may indicate that there are more objects with obvious shapes in the block, such as buildings or complex terrain. At the same time, the connectivity and distribution patterns between different regions are determined by the region segmentation algorithm, such as calculating the number of regions, area size differences, and spatial relationships between regions. These quantitative indicators of structural features can reflect the layout and organization of objects in the block, which is of great significance for evaluating the prominence of the block. In terms of base color features, since base color features include red features, green features, and blue features (RGB model), statistical analysis is performed for each color channel separately. For red features, parameters such as the proportion of red pixels in the block, the brightness distribution of red pixels, and the contrast between red and other colors are calculated. For example, in a block dominated by red buildings, the proportion of red pixels will be higher, and the contrast between red and the surrounding background colors (such as green vegetation or blue sky) will also be greater. Similarly, similar calculations are performed for green features and blue features to obtain feature information of green pixels and blue pixels in terms of quantity, brightness, contrast, etc. Then, the various quantitative indicators of structural features and base tone features are weighted and summed according to the weight coefficients pre-set in the block feature quantification plan. These weight coefficients are determined based on a large number of experiments and the evaluation of the importance of different features, reflecting the relative importance of structural features and base tone features in the overall salient recognition process. For example, in some scenarios, if the structural feature is more critical for distinguishing salient blocks, the weight coefficient corresponding to the structural feature will be relatively high; while in other scenarios, if the color feature plays a dominant role, the weight coefficient of the base tone feature will be larger. Through such weighted calculations, the structural features and base tone features are organically integrated, and finally an arbitrary block feature value that can comprehensively reflect the saliency of the arbitrary block is obtained, providing an accurate and representative quantitative basis for the subsequent screening of the most salient blocks.

[0048] In a possible implementation, step S300 further includes:

[0049] Step S310: extracting initial point cloud data from the initial feature vector.

[0050] Step S320: combining the initial node in the initial feature vector with the first point cloud sample in the initial point cloud data to obtain a first map sample.

[0051] Step S330: Compare the initial point cloud data with the first point cloud sample to obtain a first verification point cloud sample.

[0052] Step S340: performing verification analysis on the first map sample according to the first verification point cloud sample to obtain a first verification result.

[0053] Step S350: When the first verification result meets the predetermined verification constraint, the first map sample is used as the initial map.

[0054] Specifically, the initial feature vector is deeply analyzed, and the part related to the point cloud data is located by identifying its specific data structure and storage format. Scan each element from the beginning of the vector, and determine its starting position with the help of the predefined point cloud data header identifier or index. Then, according to the established point cloud data format specification, the storage structure of each point cloud data point is clarified, including the arrangement order of three-dimensional coordinates and possible other attribute information. Then, pointer operations and memory offset calculations are used to read the data of each point in sequence, and the data is converted into a suitable numerical type and stored in a temporary structure. After that, an integrity check is performed to compare the number of extracted points with the expected value, and to verify the coordinates and attribute range of each point. If problems are found, they are handled according to the preset strategy. Finally, the verified point cloud data is organized into a format that is convenient for subsequent use, so as to accurately obtain the initial point cloud data and lay a solid foundation for map construction.

[0055] The first map sample is obtained by combining the initial node in the initial feature vector with the first point cloud sample in the initial point cloud data. The initial node is used as a reference point for initial map construction and is associated with the first point cloud sample. The first point cloud sample is a representative portion of point cloud data selected from the initial point cloud data, which together with the initial node determines the initial form of the map. By matching the position information of the initial node with the points in the first point cloud sample for spatial position and calculating the geometric relationship, a preliminary map framework, namely the first map sample, is constructed. This sample contains information such as the three-dimensional spatial distribution of a portion of the target area and the approximate shape of the object.

[0056] Taking the initial point cloud data as the benchmark, traverse each point in it. For each point, find the corresponding point closest to its spatial position in the first point cloud sample, and measure the degree of proximity by calculating the three-dimensional spatial distance between the two points (such as the Euclidean distance formula). When the corresponding point is found, calculate the difference in coordinate values ​​between the point and the corresponding point, including the deviation in the X, Y, and Z directions. At the same time, compare other possible attribute differences between the two, such as reflection intensity (if any). If the coordinate deviation or attribute difference between a point and its corresponding point exceeds a preset threshold (the threshold is determined based on factors such as map accuracy requirements and the error range of the acquisition equipment), the point is marked as a difference point. After traversing all the points, extract these marked difference points to form the first verification point cloud sample. This sample accurately reflects the deviation of the first point cloud sample relative to the initial point cloud data, providing a key basis for subsequent verification analysis.

[0057] For each point in the first verification point cloud sample, determine its corresponding position in the first map sample. Calculate the deviation between the coordinates of these points in the map sample and the theoretically correct coordinates in the initial point cloud data (provided by the first verification point cloud sample), and evaluate the performance of the first map sample in terms of spatial position accuracy by calculating the mean, standard deviation and other statistics of these deviation values. Analyze the difference between the distribution density of the points in the verification point cloud sample and the point cloud density of the corresponding area in the first map sample to determine whether the map sample is accurate in reflecting the density of the point cloud in the target area. At the same time, examine the degree of fit between the shape features of the objects in the verification point cloud sample (such as edge contours, curvature, etc.) and the corresponding object shapes in the first map sample, and use the shape similarity algorithm for quantitative evaluation. In addition, if the point cloud data contains attribute information (such as reflection intensity), compare the attribute differences between the corresponding points in the verification point cloud sample and the first map sample to analyze the attribute consistency. Combining the above-mentioned analysis results of position accuracy, density difference, shape fit and attribute consistency, a comprehensive first verification result is formed to accurately determine the accuracy, completeness and reliability of the first map sample, and provide a key decision basis for whether to use it as the initial map in the future.

[0058] When the first verification result obtained by the verification analysis of the first map sample is compared with the predetermined verification constraint, if the requirements are met, the first map sample is determined as the initial map. The predetermined verification constraint is a series of quantitative indicators and qualitative conditions pre-set based on factors such as the accuracy standard of map construction, application requirements, and characteristics of the acquisition equipment. For example, in terms of spatial position accuracy, it is stipulated that the mean value of the coordinate deviation shall not exceed a certain threshold, and the standard deviation shall be within a specific range to ensure that the position of the object in the map sample is sufficiently accurate; for point cloud density, it is required to be consistent with the actual point cloud distribution law of the target area, and there shall not be too sparse or too dense to cause information missing or redundancy; in terms of shape fitting, a certain similarity standard must be met to ensure that the shape of the object in the map sample can accurately reflect the actual object shape; at the same time, attribute consistency (if there is a relevant attribute) should also meet the corresponding error range. When all indicators in the first verification result are within these predetermined verification constraints, it indicates that the first map sample has met the requirements for being an initial map in terms of accuracy, completeness and reliability, and is therefore selected as the basis for map construction, providing a solid starting framework for subsequent enhancements and updates based on it, ensuring that the final constructed map can meet the needs of high-level tasks.

[0059] In a possible implementation, step S340 further includes:

[0060] Step S341: taking the average of the matching degrees between the first verification point cloud and the first point cloud of the first map sample as the first verification result, wherein the first verification point cloud refers to any one verification point cloud in the first verification point cloud sample.

[0061] Specifically, first, the concept of the first verification point cloud is clarified, and any verification point cloud selected from the first verification point cloud sample is selected. When focusing on this specific first verification point cloud, calculating its matching degree with the first point cloud in the first map sample becomes a key step. When calculating the matching degree, multiple key factors are comprehensively considered. From the perspective of spatial position, the coordinate system is used to calculate the difference between the coordinates of each point of the first verification point cloud in three-dimensional space and the coordinates of the corresponding points of the first point cloud in the first map sample. For example, for each pair of corresponding points, the spatial distance between them is calculated by the Euclidean distance formula, and the smaller the distance means the higher the degree of matching in position. If the point cloud data has other attribute information, such as reflection intensity or color information (assuming that it is recorded during the data collection process), the matching of these attributes will also be analyzed in depth. For reflection intensity, the intensity difference between the two is calculated, and the difference is converted into a part of the contribution value of the matching degree according to the pre-set rules; for color information, its weight in the matching degree is determined by comparing the distance or color similarity index in the color space. After completing the comprehensive matching degree calculation for a single first verification point cloud and the first point cloud, repeat this process for each verification point cloud in the first verification point cloud sample. When all the matching degrees are calculated, these matching degrees are summarized and their mean is calculated. This mean represents the matching degree between the first verification point cloud sample as a whole and the first map sample, which is the first verification result. It provides an intuitive and quantitative basis for the subsequent evaluation of the accuracy and reliability of the first map sample. If the mean is high, it means that the first map sample and the original initial point cloud data still maintain a high consistency and accuracy after comparison with the verification point cloud sample, and it is more likely to be identified as a qualified initial map; conversely, if the mean is low, it may be necessary to further check or correct the first map sample to ensure that the final constructed map can meet the expected quality standards.

[0062] In a possible implementation, step S500 further includes:

[0063] Step S510: extracting any node in the first-level neighborhood node set.

[0064] Step S520: improving the initial map according to any point cloud data of the any node to obtain any improved map.

[0065] Step S530: Perform node sampling on the first-level neighborhood node set based on the predetermined graph neural network enhancement mechanism to obtain a first sampling node set.

[0066] Step S540: matching a first sampling node map set of the first sampling node set in the arbitrary perfected map.

[0067] Step S550: updating the map corresponding to the initial node according to the fusion result of the first sampling node map set to obtain the enhanced map.

[0068] Specifically, firstly, we extract any node in the first-level neighborhood node set. This first-level neighborhood node set is determined by a complex matching and screening process, and each node contains rich information, including its own geometric features, semantic features, and related point cloud data. By extracting any node, we can provide basic elements for the subsequent improvement of the initial map.

[0069] The initial map is improved based on the arbitrary point cloud data of the arbitrary node, thereby obtaining an arbitrary improved map. The arbitrary point cloud data reflects the local three-dimensional spatial information of the target area corresponding to the node. When integrating it into the initial map, factors such as the position and density of the points in the point cloud data and the spatial relationship with the existing elements of the initial map will be comprehensively considered. For example, if the arbitrary point cloud data shows that there is a small building or terrain undulations that were not accurately depicted in the initial map before, then the initial map will be supplemented and corrected based on this information, so that the map can be improved in details and accuracy in the local area, thereby forming an arbitrary improved map.

[0070] With the help of the predetermined graph neural network enhancement mechanism, the node sampling work is carried out on the first-level neighborhood node set to obtain the first sampling node set. The predetermined graph neural network enhancement mechanism has been pre-built and trained, and has the ability to deeply analyze and process node features. First, the first-level neighborhood node set is input into the graph neural network, and the neurons in the network begin to extract and parse the multivariate features of each node. These features cover the geometric position information of the node, such as the coordinates in three-dimensional space, the relative position relationship with the surrounding nodes; semantic information, including the object category represented by the node, the regional functional attributes; and the structural characteristics of the node itself, such as the number and weight of the connecting edges. Based on these extracted features, the graph neural network performs sampling operations according to its internal complex algorithm model and predefined sampling strategy. The sampling strategy comprehensively considers the importance score of the node. This score is obtained by the node features through specific weighted calculation and function transformation. Nodes with high importance scores are more likely to be selected. At the same time, the correlation and diversity between nodes will be taken into account to avoid the sampling results being too concentrated or lacking representativeness. For example, for a cluster of nodes that are densely distributed in space and have similar features, some key nodes will be screened out according to certain rules to ensure that the sampled nodes can cover different areas and feature types of the first-level neighborhood node set. Through such a series of calculations and screening processes, some nodes that meet the requirements are finally determined from the first-level neighborhood node set. These nodes together form the first sampling node set, which will serve as the key data source for subsequent map updates and improvements, laying the foundation for building more accurate and rich enhanced maps.

[0071] It is clear that the first sampling node map set is a collection of local map information corresponding to each node in the first sampling node set. Each sampling node has a potential matching area in any perfect map due to its own characteristic information, such as position, geometric shape, semantic identification, etc. The matching process starts with the processing of the first sampling node map in the first sampling node map set, and its features are comprehensively compared with any perfect map. At the geometric level, by calculating the shape similarity of key contour points and shape boundaries in the sampling node map and the corresponding area in any perfect map, the shape difference between the two is quantified using algorithms such as Hausdorff distance to determine whether their geometric shapes match; in terms of semantics, the semantic categories represented by the sampling node map (such as building type, road grade, etc.) are compared with the semantic information marked in the corresponding area in any perfect map. In terms of positional relationship, the relative deviation between the central position or key feature point position of the sampling node map and the position of the relevant area in any perfect map is analyzed to ensure the rationality of its spatial position. After completing the matching of the first sampling node map, the above matching process is repeated for other sampling node maps in the first sampling node map set. In the matching process, not only the matching degree of each sampling node map with any perfect map should be considered separately, but also the mutual relationship between multiple sampling node maps in any perfect map should be comprehensively analyzed, such as whether their relative position distribution conforms to the actual geographical logic, whether the boundary connection of adjacent sampling node maps is natural and smooth, etc. Through such a meticulous and comprehensive matching operation, the accurate corresponding position and matching situation of the first sampling node map set in any perfect map are finally determined, which provides a key basis for the subsequent accurate update of the initial map based on the matching results, thereby promoting the construction process of the enhanced map.

[0072] First, the fusion results of the first sampling node atlas are comprehensively and deeply analyzed. The various geographic element information in the fusion results are carefully separated, including landmarks in the form of points, roads and rivers in the form of lines, and geometric shape information of regions and building outlines in the form of faces, such as coordinate point sets, line segment endpoints and directions, polygon vertices, etc.; semantic information, that is, the specific meaning represented by each geographic element, such as the type of road (highway, city street, etc.), the function of the region (commercial area, residential area, etc.); topological relationships, such as the connection between points and lines, the surrounding of lines and surfaces, and the adjacent relationship between surfaces; and attribute information such as the width of roads, the height and number of floors of buildings. At the same time, the information extracted from the fusion results is uniformly converted to the coordinate system used by the map corresponding to the initial node, laying the foundation for subsequent accurate updates. In the geometric shape update link of geographic elements, for point features, if the corresponding points can be found in the initial map, the difference values ​​between the midpoints of the fusion results and the midpoints of the initial map in each coordinate dimension (such as X, Y, Z coordinates, if it is a three-dimensional map) are calculated respectively. Then, the final updated coordinates are determined based on the contribution weights of different sampling nodes to the point in the fusion result. The determination of this weight can comprehensively consider the reliability of the sampling node itself, such as the accuracy of the data acquisition equipment, the environmental stability during acquisition, and the spatial distance between the sampling node and the point. Generally, the closer the distance, the higher the weight of the sampling node. If there is no corresponding point in the initial map, the new point element is added to the initial map at the appropriate position directly based on the point information in the fusion result and the distribution law of the surrounding geographical elements. For line elements, since they have a certain length and direction, the lines in the fusion result are first segmented according to certain rules, such as according to the curvature change of the line segment, the intersection with other geographical elements, etc., so that they can be matched and compared with the lines in the initial map more accurately. For each line segment, the least squares method is used to calculate the distance error between it and the corresponding line segment in the initial map at the endpoints and several intermediate points to measure the degree of fit between the two. If the degree of fit is lower than the set threshold, it indicates that the line in the initial map needs to be updated. At this time, according to the coordinate information of the line in the fusion result and the weight of each sampling node for the line segment, linear interpolation or spline interpolation algorithms are used to update and adjust the shape and position of the line in the initial map. In addition, during the update process, the continuity of the line should be checked at all times to avoid breakpoints, and the smoothness of the line should be ensured to prevent the broken line from being too abrupt. The update process of the surface element is similar to that of the line element. First, the surface in the fusion result is matched with the surface in the initial map, and the necessity of the update is determined by comparing the coordinate differences of the boundary points of the surface.Similarly, based on the weights of the boundary points, the polygon deformation algorithm is used to reconstruct the surface elements according to the updated boundary points to ensure the closure of the surface, that is, all boundary points can be connected end to end to form a complete polygon, and the geometric shape is correct, and there are no errors such as self-intersection. At the same time, the consistency of the boundaries between adjacent surfaces should be carefully checked to prevent boundary gaps or overlaps, and to ensure the integrity and accuracy of the map at the surface element level. In terms of updating semantic information, for each geographic element in the map, the semantic label information of the corresponding element in the fusion result should be obtained. Then, the semantic label of the fusion result is compared with the semantic label in the initial map based on the semantic knowledge base or a pre-defined semantic distance measurement method. If the similarity is low and the semantic label in the fusion result has a higher credibility (for example, based on more sample data or a more authoritative data source), the semantic annotation in the initial map is updated according to the semantic label in the fusion result. When a geographic element has multiple semantic labels in the fusion result, it can be selected according to the pre-set semantic label priority rules, such as giving priority to semantic labels that are more critical in specific application scenarios, or using a probability model to determine the final semantic annotation based on the frequency of each semantic label in different sampling nodes. The last step is to update the topological relationship. Based on the topological relationship between the geographic elements contained in the fusion result and the existing topological structure of the initial map, a comprehensive check is performed using relevant algorithms in graph theory. For example, check whether the connection between points and lines is correct, whether there are isolated points or unconnected line segments; whether the surrounding relationship between lines and surfaces conforms to the actual geographical logic, whether there are lines that pass through the surface without forming a correct surrounding; whether the adjacent relationship between surfaces is accurate, and whether the boundaries of adjacent surfaces match consistently. Once errors or inconsistencies are found in the topological relationship, they are corrected and improved according to the correct topological information in the fusion result to ensure that the connection of each geographic element in the map is correct and logically rigorous at the topological level. Through this complete update process covering geometric, semantic and topological relationships, the map corresponding to the initial node gradually evolves into an enhanced map, which has been significantly improved in many aspects such as the accuracy of geographic information, semantic richness, element integrity and logical coherence. It can better serve various application scenarios such as intelligent navigation, geographic information analysis, urban planning, etc., and provide users with better quality and more reliable map data support.

[0073] In a possible implementation, step S500 further includes:

[0074] Step S560: The predetermined graph neural network enhancement mechanism includes a predetermined sampling plan, and the predetermined sampling plan refers to stratified sampling.

[0075] Specifically, stratified sampling is based on the hierarchical division of the first-level neighborhood node set. These levels can be determined based on a variety of factors, such as the size of the geographical area represented by the node, the degree of abstraction of the geographical features, or the hierarchical classification of semantic information. For example, the nodes can be divided into different levels such as the macro-regional layer, the meso-block layer, and the micro-landmark layer. At the macro-regional layer, there may be nodes covering the entire city or large geographical divisions. These nodes mainly carry relatively general information such as the overall outline of the region, the distribution of major traffic arteries, and the division of large functional areas; the nodes at the meso-block layer focus on the block units within the city, and describe in detail the road network, building community distribution, and some local geographical features within the block; the nodes at the micro-landmark layer target specific landmark buildings, important transportation hubs, and other geographical elements with significant identification, and contain extremely fine geometric shapes, precise location information, and rich semantic details, such as the name, architectural style, and functional use of landmark buildings.

[0076] When implementing stratified sampling, different sampling ratios or sampling rules will be set according to different levels. For the macro-regional layer, due to its relatively large amount of data and relatively general information, a lower sampling ratio is adopted to select some representative nodes, which can roughly outline the framework structure of the entire region and provide macroscopic geographical layout information for the map. For example, in the construction of a map of a large city, only the nodes corresponding to the central area of ​​the city, major satellite towns, and the main traffic arteries connecting them may be selected from the macro-regional layer for key analysis and sampling. The meso-block layer will adopt a moderate sampling ratio according to specific needs and data distribution, ensuring that sufficient detailed information inside the block can be obtained, while not making the data volume too large to cause excessive computational complexity. For example, sampling can be carried out according to a certain block interval or based on the functional importance of the block, and those commercial and bustling blocks, densely populated blocks, or block nodes with special geographical characteristics can be selected. The micro-landmark layer will adopt a higher sampling ratio or specific sampling rules because the information contained in its nodes is highly accurate and unique, ensuring that all important landmark nodes can be included in the sampling range, because these landmark nodes play a key role in improving the accuracy and recognition of the map. For example, for a tourist city, all famous attractions, large shopping malls, important transportation hubs and other micro-landmark nodes should be sampled as much as possible.

[0077] Through this predetermined sampling plan of stratified sampling, it is possible to effectively screen and collect data for the first-level neighborhood node set at different levels, so that the first sampling node set obtained by sampling contains both macro-geographic layout information and meso-block details and micro-landmark features, thereby providing a comprehensive, rich and hierarchical data foundation for subsequent map updates and enhancement operations based on these sampling nodes, which helps to build more accurate, detailed and enhanced maps that meet the needs of different applications, and can play an important role in urban planning, intelligent navigation and other application areas of geographic information systems.

[0078] In a possible implementation, step S550 further includes:

[0079] Step S551: establishing a secondary neighborhood space of the arbitrary node, and matching a secondary neighborhood node set of the secondary neighborhood space in the image frame database.

[0080] Step S552: Based on the predetermined graph neural network enhancement mechanism, the enhanced map is iteratively enhanced and updated using the secondary neighborhood node set.

[0081] Specifically, a secondary neighborhood space of any node is established. This secondary neighborhood space is a regional concept further expanded on the basis of the existing map construction and node relationship system. It takes any node as the core and determines the surrounding related regional range according to a specific spatial distance range or topological connection rule. After determining the secondary neighborhood space, a matching operation is performed in the image frame database to obtain the secondary neighborhood node set. The image frame database stores a large amount of image data and the node information associated with it. By comparing and screening the features of the secondary neighborhood space with the information in the database, those nodes located in the neighborhood space and meeting specific conditions (such as similar geometric features, semantic attributes or topological relationships) are found, thereby forming a secondary neighborhood node set. These nodes not only contain their own point cloud data, geometric shapes, semantic annotations and other information, but also reflect the spatial association and interaction with the core arbitrary node in a wider area.

[0082] Based on the predetermined graph neural network enhancement mechanism, the enhanced map is iteratively enhanced and updated using the secondary neighborhood node set. The predetermined graph neural network enhancement mechanism provides a powerful algorithm framework and processing logic for this update process. The secondary neighborhood node set is input into the graph neural network, and the network conducts in-depth analysis of the multivariate features of each node based on its internal node feature extraction module, including information on location, shape, semantics, etc. Then, the mutual influence relationship and information transmission path between nodes are calculated through the edge connection information in the network and the pre-trained weight parameters. During the iterative update process, each area or geographic element in the enhanced map is gradually optimized based on the information provided by the secondary neighborhood node set. For example, in terms of geometric shape, if a node in the secondary neighborhood node set carries more accurate outline information about a building, then this information will be transmitted to the representation of the corresponding building in the enhanced map through the network's information fusion and propagation mechanism, and its shape will be corrected and refined; in terms of semantic annotation, if the nodes in the secondary neighborhood node set provide new clues or more accurate definitions about the functional attributes of a region, then the semantic annotation of the region in the enhanced map will be updated and improved; at the topological relationship level, based on the connection relationship between the secondary neighborhood node sets and between them and the original nodes of the enhanced map, the topological structure of the map is checked and optimized to ensure that the connection between geographic elements is more reasonable and accurate. Each iteration will further improve the information richness, accuracy and logic of the enhanced map. After multiple iterations, a deeply optimized and improved enhanced map will be obtained, which can better meet the needs of various complex geographic information applications, such as high-precision navigation, urban planning simulation and analysis, etc.

[0083] Embodiment 2 is based on the same inventive concept as the map construction method for high-level task requirements in the above embodiment. Figure 2 As shown, the present application provides a map construction system for high-level task requirements, and the embodiments of the present application and the method embodiments are based on the same inventive concept. The system includes:

[0084] The most significant block acquisition module 10 is used to introduce an image significant recognition strategy to perform significant recognition analysis on the global image of the target area to obtain the most significant block.

[0085] The initial frame acquisition module 20 is used to match the most significant image frame of the most significant block in the image frame database of the target area, and use the most significant image frame as the initial frame.

[0086] The initial map construction module 30 is used to perform three-dimensional point cloud analysis in combination with the initial feature vector of the initial frame to construct an initial map.

[0087] The first-level neighborhood space establishment module 40 is used to establish the first-level neighborhood space of the initial node in the initial feature vector, and match the first-level neighborhood node set of the first-level neighborhood space in the image frame database.

[0088] The enhanced map acquisition module 50 enhances and updates the initial map using the first-level neighborhood node set based on a predetermined graph neural network enhancement mechanism to obtain an enhanced map.

[0089] Furthermore, the most prominent block acquisition module 10 further includes:

[0090] An image acquisition device acquisition unit, wherein the image acquisition device acquisition unit is used to acquire an image acquisition device, wherein the image acquisition device refers to a device that acquires images of the target area.

[0091] A unit acquisition length determination unit is used to determine the unit acquisition length according to a predetermined width and a predetermined distance of the image acquisition device.

[0092] A segmentation processing result acquisition unit is configured to perform segmentation processing on the global image based on the unit acquisition length to obtain a segmentation processing result.

[0093] A block analysis unit is used to introduce the image salient recognition strategy to analyze multiple blocks in the segmentation processing result to obtain the most salient block.

[0094] Furthermore, the block analysis unit also includes:

[0095] A block extraction unit is used to extract any block from the plurality of blocks.

[0096] The block feature value acquisition unit is used to perform feature quantification analysis on the arbitrary block according to the block feature quantification plan in the image saliency recognition strategy to obtain the arbitrary block feature value.

[0097] The maximum eigenvalue matching unit is used to obtain the maximum eigenvalue by using the eigenvalue of any block as a constraint, and reversely match the block corresponding to the maximum eigenvalue, which is recorded as the most significant block.

[0098] Furthermore, the block feature value acquisition unit further includes:

[0099] A weighted calculation unit, the weighted calculation unit is used to perform weighted calculation on the structural features and base color features of the arbitrary block according to the block feature quantization plan to obtain the feature value of the arbitrary block, wherein the base color feature includes red feature, green feature and blue feature.

[0100] Furthermore, the initial map construction module 30 also includes:

[0101] An initial point cloud data extraction unit is used to extract initial point cloud data from the initial feature vector.

[0102] The first map sample acquisition unit is used to combine the initial node in the initial feature vector with the first point cloud sample in the initial point cloud data to obtain a first map sample.

[0103] The first verification point cloud sample acquisition unit is used to compare the initial point cloud data with the first point cloud sample to obtain a first verification point cloud sample.

[0104] The first verification result acquisition unit is used to perform verification analysis on the first map sample according to the first verification point cloud sample to obtain a first verification result.

[0105] An initial map acquisition unit is configured to use the first map sample as the initial map when the first verification result meets a predetermined verification constraint.

[0106] Furthermore, the first verification result obtaining unit further includes:

[0107] A point cloud matching degree mean acquisition unit, wherein the point cloud matching degree mean acquisition unit is used to obtain the mean of the matching degrees between the first verification point cloud and the first point cloud of the first map sample as the first verification result, wherein the first verification point cloud refers to any one verification point cloud in the first verification point cloud sample.

[0108] Furthermore, the enhanced map acquisition module 50 further includes:

[0109] An arbitrary node extraction unit is used to extract an arbitrary node from the first-level neighborhood node set.

[0110] An arbitrary improved map acquisition unit is used to improve the initial map according to the arbitrary point cloud data of the arbitrary node to obtain an arbitrary improved map.

[0111] A first sampling node set acquisition unit is configured to perform node sampling on the first-level neighborhood node set based on the predetermined graph neural network enhancement mechanism to obtain a first sampling node set.

[0112] An arbitrary perfected map matching unit, wherein the arbitrary perfected map is used to match a first sampling node map set of the first sampling node set in the arbitrary perfected map.

[0113] A map updating unit, wherein the map updating unit is used to update the map corresponding to the initial node according to the fusion result of the first sampling node map set to obtain the enhanced map.

[0114] Furthermore, the enhanced map acquisition module 50 further includes:

[0115] The predetermined graph neural network enhancement mechanism includes a predetermined sampling plan, and the predetermined sampling plan refers to stratified sampling.

[0116] Furthermore, the enhanced map acquisition module 50 further includes:

[0117] A secondary neighborhood space establishing unit is used to establish a secondary neighborhood space of the arbitrary node and match a secondary neighborhood node set of the secondary neighborhood space in the image frame database.

[0118] An iterative enhancement update unit, wherein the iterative enhancement update unit performs iterative enhancement update on the enhanced map based on the predetermined graph neural network enhancement mechanism and utilizing the secondary neighborhood node set.

[0119] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description and does not represent the advantages and disadvantages of the embodiments. And the above-mentioned specific embodiments of this specification are described. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0120] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

[0121] This specification and the drawings are merely exemplary illustrations of the present application and are deemed to cover any and all modifications, variations, combinations or equivalents within the scope of the present application. Obviously, a person skilled in the art may make various modifications and variations to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalents, the present application intends to include these modifications and variations.

Claims

1. A map construction method for high-level task requirements, characterized in that: include: The image saliency recognition strategy is introduced to perform saliency recognition analysis on the global image of the target area to obtain the most salient block; Matching the most significant image frame of the most significant block in the image frame database of the target area, and using the most significant image frame as the initial frame; Performing a three-dimensional point cloud analysis based on the initial feature vector of the initial frame to construct an initial map; Establishing a primary neighborhood space of an initial node in the initial feature vector, and matching a primary neighborhood node set of the primary neighborhood space in the image frame database; Based on a predetermined graph neural network enhancement mechanism, the initial map is enhanced and updated using the first-level neighborhood node set to obtain an enhanced map.

2. A map construction method for high-level task requirements according to claim 1, characterized in that: The image saliency recognition strategy is introduced to perform saliency recognition analysis on the global image of the target area to obtain the most salient blocks, including: Acquire an image acquisition device, wherein the image acquisition device refers to a device that acquires images of the target area; Determining a unit acquisition length according to a predetermined width and a predetermined distance of the image acquisition device; Performing segmentation processing on the global image based on the unit acquisition length to obtain a segmentation processing result; The image salient recognition strategy is introduced to analyze multiple blocks in the segmentation processing result to obtain the most salient block.

3. A map construction method for high-level task requirements according to claim 2, characterized in that: The image salient recognition strategy is introduced to analyze multiple blocks in the segmentation processing result to obtain the most salient block, including: Extracting any block from the plurality of blocks; Performing feature quantification analysis on the arbitrary block according to the block feature quantification plan in the image saliency recognition strategy to obtain a feature value of the arbitrary block; The maximum eigenvalue is obtained by using the eigenvalue of any block as a constraint, and the block corresponding to the maximum eigenvalue is reversely matched and recorded as the most significant block.

4. A map construction method for high-level task requirements according to claim 3, characterized in that: The structural features and base color features of the arbitrary block are weightedly calculated according to the block feature quantization plan to obtain the feature value of the arbitrary block, wherein the base color features include red features, green features and blue features.

5. A map construction method for high-level task requirements according to claim 1, characterized in that: The initial feature vector of the initial frame is combined to perform a three-dimensional point cloud analysis to construct an initial map, including: Extracting initial point cloud data from the initial feature vector; Combining the initial node in the initial feature vector with the first point cloud sample in the initial point cloud data to obtain a first map sample; Comparing the initial point cloud data with the first point cloud sample to obtain a first verification point cloud sample; Performing verification analysis on the first map sample according to the first verification point cloud sample to obtain a first verification result; When the first verification result meets the predetermined verification constraint, the first map sample is used as the initial map.

6. A map construction method for high-level task requirements according to claim 5, characterized in that: An average of the matching degrees between the first verification point cloud and the first point cloud of the first map sample is taken as the first verification result, wherein the first verification point cloud refers to any one verification point cloud in the first verification point cloud sample.

7. A map construction method for high-level task requirements according to claim 1, characterized in that: Based on a predetermined graph neural network enhancement mechanism, the initial map is enhanced and updated using the first-level neighborhood node set to obtain an enhanced map, including: Extracting any node from the first-level neighborhood node set; Perfecting the initial map according to any point cloud data of any node to obtain any perfected map; Perform node sampling on the first-level neighborhood node set based on the predetermined graph neural network enhancement mechanism to obtain a first sampling node set; a first sampling node map set matching the first sampling node set in the arbitrary perfected map; The map corresponding to the initial node is updated according to the fusion result of the first sampling node map set to obtain the enhanced map.

8. A map construction method for high-level task requirements according to claim 7, characterized in that: The predetermined graph neural network enhancement mechanism includes a predetermined sampling plan, and the predetermined sampling plan refers to stratified sampling.

9. A map construction method for high-level task requirements according to claim 7, characterized in that: Also includes: Establishing a secondary neighborhood space of the arbitrary node, and matching a secondary neighborhood node set of the secondary neighborhood space in the image frame database; Based on the predetermined graph neural network enhancement mechanism, the enhanced map is iteratively enhanced and updated using the secondary neighborhood node set.

10. A map construction system for high-level task requirements, characterized in that: include: A most significant block acquisition module, wherein the most significant block acquisition module is used to introduce an image significant recognition strategy to perform significant recognition analysis on the global image of the target area to obtain the most significant block; An initial frame acquisition module, the initial frame acquisition module is used to match the most significant image frame of the most significant block in the image frame database of the target area, and use the most significant image frame as the initial frame; An initial map construction module, the initial map construction module is used to perform three-dimensional point cloud analysis in combination with the initial feature vector of the initial frame to construct an initial map; A primary neighborhood space establishment module, the primary neighborhood space establishment module is used to establish a primary neighborhood space of an initial node in the initial feature vector, and match a primary neighborhood node set of the primary neighborhood space in the image frame database; The enhanced map acquisition module uses the first-level neighborhood node set to enhance and update the initial map based on a predetermined graph neural network enhancement mechanism to obtain an enhanced map.

Images