High-precision map construction method and related device

Abstract

The invention discloses a high-precision map construction method and a related device, and relates to the technical field of automatic driving. The high-precision map construction method comprises the steps that image data representing the surrounding road environment of a vehicle is obtained, a semantic topology network diagram containing nodes and associated edges is constructed based on the image data, the nodes are information units which are extracted from the surrounding road environment of the vehicle and carry semantic and spatial features, and the associated edges are associated with the nodes. The associated edge represents a node relationship of spatial and semantic dimensions between the nodes. Then determining a target anchor point meeting a space reference condition from the semantic topology network diagram, then mapping the target anchor point to a geographic coordinate system in combination with the real-time pose of the vehicle, and finally determining the spatial position of each node according to the relationship between the target anchor point and the node in the geographic coordinate system so as to generate a high-precision map. Finally, real-time and accurate road environment structured information is provided for automatic driving, and safe and efficient driving of the automatic driving vehicle is guaranteed.

Description

Technical Field

[0001] This application relates to the field of autonomous driving technology, and more specifically, to a high-precision map construction method and related apparatus. Background Technology

[0002] In the field of autonomous driving technology, the construction of high-precision maps is crucial for vehicle navigation and decision-making. Existing technologies rely on physical scale calculation methods when constructing high-precision maps. These methods typically involve calibrating sensors to determine geometric parameters such as relative positions and angles between them, then using the calibrated sensors to collect road data such as physical distances, angles, and positions in the road scene, and finally calculating the geographic coordinates of each road element based on the road data to complete map construction.

[0003] However, this physical scale calculation method has significant limitations in practical applications. It requires high accuracy in sensor calibration and environmental conditions. If there is a sensor calibration deviation, or if there is environmental interference such as obstruction or severe weather, it can easily lead to large errors in geographic coordinate calculation, which in turn affects the accuracy of map construction and makes the constructed high-precision map unable to adapt to dynamically changing road conditions. Summary of the Invention

[0004] In view of the above problems, this application proposes a high-precision map construction method and related apparatus to improve the accuracy of high-precision map construction. The specific solution is as follows:

[0005] Firstly, a high-precision map construction method is provided, including:

[0006] Acquire image data, which characterizes the surrounding road environment of the vehicle;

[0007] A semantic topology network graph is determined based on the image data. The semantic topology network graph includes nodes and associated edges. The nodes represent information units extracted from the road environment around the vehicle and carrying semantic and spatial features. The associated edges represent the node relationships between nodes. The node relationships represent the spatial and semantic dimension relationships between nodes.

[0008] The target anchor point is determined based on the semantic topology network graph, and the target anchor point is a node that satisfies the function of a spatial reference.

[0009] The target anchor point is mapped to the geographic coordinate system based on the real-time pose of the vehicle;

[0010] Based on the target anchor points in the geographic coordinate system and the node relationships, the spatial positions of each node in the semantic topology network graph in the geographic coordinate system are determined, resulting in a high-precision map.

[0011] In one possible design, in another implementation of the first aspect of the embodiments of this application, the process of determining the semantic topology network graph based on the image data includes:

[0012] The image data is input into a pre-trained image recognition model to obtain road elements corresponding to the image data output by the image recognition model. The road elements represent various entities related to driving in the road environment around the vehicle. The image recognition model is trained using different image data as training samples and the road elements corresponding to different image data as sample labels.

[0013] The road elements are classified based on their road element types, and nodes corresponding to each road element are created based on the road element types. The road element types represent the classification and definition of the functions of the road elements.

[0014] The node relationships are determined based on the node attributes of each node. The associated edges are added between nodes with node relationships. The semantic topology network graph is constructed based on each node and the associated edges. The node attributes include semantic attributes and spatial coordinates. The semantic attributes represent the traffic rules or function types carried by the node, and the spatial coordinates represent the position of the node in a vehicle coordinate system centered on the vehicle.

[0015] In one possible design, in another implementation of the first aspect of the embodiments of this application, the nodes include lane line nodes, road surface marking nodes, traffic sign nodes, and drivable area nodes. The node relationships include connection relationships, constraint relationships, parallel relationships, and inclusion relationships. The connection relationship indicates that there is a connection relationship between the lane line nodes. The constraint relationship indicates that the drivable area node is constrained by the rules of the traffic sign node. The parallel relationship indicates that there is a parallel relationship between the lane line nodes. The inclusion relationship indicates that the road surface marking node is included in the drivable area node.

[0016] The process of determining the node relationships based on the node attributes of each node includes:

[0017] Based on the node attributes of each node, it is determined whether the preset relationship matching rule is met. If the relationship matching rule is met, it is determined that there is a node relationship between the nodes. The relationship matching rule represents the judgment rule set for different node relationships. The judgment rule combines the node attributes of the node and compares them with a preset threshold to realize the judgment of the node relationship.

[0018] In one possible design, in another implementation of the first aspect of the present application, the node attributes further include detection confidence and continuous tracking frame count. The detection confidence represents the credibility of the road element corresponding to the node in the recognition process of the image recognition model. The continuous tracking frame count represents the number of times the node is continuously included in the image data and successfully matched with the node identifier. The node, candidate anchor point and target anchor point all have node attributes.

[0019] The process of determining the target anchor point based on the semantic topology network graph includes:

[0020] Based on the detection confidence and the number of consecutive tracking frames, lane intersections and / or traffic sign nodes are selected in the semantic topology network graph. The selected lane intersections and / or traffic sign nodes are used as candidate anchor points. The lane intersection is the intersection of at least two lane line nodes.

[0021] The candidate anchor points are input into a preset anchor point scoring model. The anchor point scoring model determines the node confidence score, topological centrality score, and location stability score of the candidate anchor points. The node confidence score, topological centrality score, and location stability score are then weighted and summed to obtain and output the candidate anchor point score corresponding to the candidate anchor point. The node confidence score represents the reliability of the road element corresponding to the anchor point in being identified in multiple consecutive frames of image data. The topological centrality score represents the degree of association between the anchor point and other nodes in the semantic topological network graph. The location stability score represents the degree of fluctuation of the spatial location of the anchor point in multiple consecutive frames of image data.

[0022] Candidate anchor points whose scores are greater than a preset value are selected as target anchor points.

[0023] In one possible design, in another implementation of the first aspect of the embodiments of this application, the method further includes:

[0024] Candidate anchor points whose scores are not greater than a preset value are used as backup anchor points, and pre-activated target anchor points are selected from the backup anchor points.

[0025] When the target anchor point is in a failed state, an optimal pre-activated target anchor point is selected based on the pre-activated target anchor point and the optimal pre-activated target anchor point is used as the new target anchor point to replace the failed target anchor point. The failed state indicates that the detection confidence of the target anchor point is greater than a preset value or the position offset of the target anchor point is greater than a preset value.

[0026] In one possible design, in another implementation of the first aspect of the embodiments of this application, the candidate anchor point, the backup anchor point, the pre-activated target anchor point, and the target anchor point all have node attributes and anchor point attributes. The anchor point attributes include anchor point type, spatial error value, effective duration, and historical offset record. The spatial error value represents the error value of the anchor point in the vehicle coordinate system. The effective duration represents the expected effective time of the anchor point. The historical offset record represents the positional deviation of the anchor point at different time points. The anchor point type includes lane intersections and traffic signs.

[0027] The process of selecting a pre-activated target anchor point from the backup anchor points includes:

[0028] Select backup anchor points that meet the pre-screening criteria, wherein the pre-screening criteria are that the relative distance from the target anchor point is not less than the configured distance value, and the direction of the target anchor point relative to the vehicle is different from that of the vehicle.

[0029] The location stability score corresponding to the backup anchor point that meets the pre-screening conditions is determined based on the spatial error value, the effective duration, and the historical offset record. The spatial error value and the historical offset record are negatively correlated with the location stability score, and the effective duration is positively correlated with the location stability score.

[0030] The weight of the backup anchor point that meets the pre-screening conditions is determined based on the scenario in which the vehicle is located and the anchor point type of the backup anchor point that meets the pre-screening conditions, and the scenario adaptability score is determined based on the weight.

[0031] The topological correlation score is determined based on the correlation strength between the backup anchor points that meet the pre-screening conditions and the active nodes. The correlation strength is positively correlated with the topological correlation score. The active nodes include the target anchor points and nodes with a detection confidence score higher than a preset value.

[0032] The backup score corresponding to the backup anchor point that meets the pre-screening conditions is obtained by weighted summation of the location stability, the scene adaptability score and the topology correlation score.

[0033] The pre-activated target anchor point is determined based on the backup score.

[0034] In one possible design, in another implementation of the first aspect of the embodiments of this application, the process of mapping the target anchor point to a geographic coordinate system based on the real-time pose of the vehicle includes:

[0035] The geographic coordinates of the target anchor point are determined based on the real-time pose of the vehicle and the spatial coordinates of the target anchor point. The spatial coordinates of the target anchor point represent the position of the target anchor point in the vehicle coordinate system centered on the vehicle, and the geographic coordinates are the position of the anchor point in the geographic coordinate system.

[0036] Based on the node relationships and the geographic coordinates, the positioning error and topology relationship error are determined. The positioning error and topology relationship error are weighted and summed to obtain the total error. The topology relationship error represents the spatial relationship deviation between the target anchor point and other nodes in the semantic topology network graph. The positioning error represents the measurement error of the vehicle's real-time pose.

[0037] If the total error is less than a preset error threshold, the target anchor point is mapped to the geographic coordinate system based on its geographic coordinates. If not, the geographic coordinates of the target anchor point are optimized based on the node relationship until the total error is less than the preset error threshold. Then, the target anchor point is mapped to the geographic coordinate system based on the optimized geographic coordinates.

[0038] In one possible design, in another implementation of the first aspect of the embodiments of this application, the method further includes:

[0039] When a change in the state of the high-precision map is detected, the type of change is determined. The type of change includes node changes, abnormal target anchor points, and abnormal vehicle positioning status.

[0040] If the change type is a node change, then update the node attributes and determine whether the node change affects the associated edges in the semantic topology network graph. If so, adjust the associated edges.

[0041] If the change type is that the target anchor point is abnormal, then freeze the current high-precision map, replace the target anchor point, determine and update the position of each node in the geographic coordinate system based on the position of the latest target anchor point in the geographic coordinate system, and unfreeze the current high-precision map.

[0042] If the change type is that the vehicle positioning status is abnormal, the high-precision map is optimized and updated based on the semantic topology network graph to obtain the updated high-precision map.

[0043] In one possible design, in another implementation of the first aspect of the embodiments of this application, the node attributes further include a node identifier and a number of consecutive tracking frames, wherein the number of consecutive tracking frames represents the number of times the same node is successfully identified and matched with the node identifier in consecutive frames of image data, and the method further includes:

[0044] The relationship strength value corresponding to the node relationship is determined based on the geometric consistency score and the number of consecutive tracking frames. The geometric consistency score characterizes the deviation between the actual angle value and / or actual distance value between nodes and the configured value. The geometric consistency score, the number of consecutive tracking frames, and the relationship strength value are positively correlated.

[0045] If the relationship strength value is less than a preset first strength value, then the node relationship is removed;

[0046] If the relationship strength value is not less than the first strength value and not greater than the preset second strength value, then the node relationship is retained and updated to a weak relationship, and the second strength value is greater than the first strength value;

[0047] If the relationship strength value is greater than the second strength value, then the node relationship is retained.

[0048] In one possible design, in another implementation of the first aspect of the embodiments of this application, the tail-integrated graph structure semantic relation parsing layer (GNN) of the image recognition model analyzes each road element output by the image recognition model to obtain the relationship probability between each road element, and the relationship probability characterizes the association strength between the road elements.

[0049] Secondly, a high-precision map building device is provided, comprising:

[0050] An image acquisition unit is used to acquire image data, which represents the surrounding road environment of the vehicle;

[0051] The topology network graph determination unit is used to determine a semantic topology network graph based on the image data. The semantic topology network graph includes nodes and associated edges. The nodes represent information units extracted from the road environment around the vehicle and carrying semantic and spatial features. The associated edges represent the node relationships between nodes. The node relationships represent the spatial and semantic dimension relationships between nodes.

[0052] The target anchor point determination unit is used to determine the target anchor point based on the semantic topology network graph, wherein the target anchor point is a node that satisfies the function of a spatial reference.

[0053] The target anchor point mapping unit is used to map the target anchor point to the geographic coordinate system based on the real-time pose of the vehicle.

[0054] The map generation unit is used to determine the spatial position of each node in the semantic topology network graph in the geographic coordinate system based on the target anchor point in the geographic coordinate system and the node relationship, so as to obtain a high-precision map.

[0055] By employing the aforementioned technical solutions, this application proposes a high-precision map construction method. This method acquires image data of the road environment surrounding a vehicle and constructs a semantic topological network graph containing nodes and associated edges. Nodes are information units extracted from the surrounding road environment, carrying semantic and spatial features. Associated edges represent the spatial and semantic relationships between nodes. By constructing this semantic topological network graph, the method overcomes the limitations of traditional high-precision map construction methods that only focus on spatial coordinates and lack semantic information. It also avoids map element misalignment problems caused by physical scale recovery errors in scenarios such as sensor calibration deviations and occlusion. Target anchor points that meet spatial reference conditions are then determined from the semantic topological network graph. Subsequently, these target anchor points are mapped to a geographic coordinate system based on the vehicle's real-time pose. Finally, the spatial positions of each node are determined according to the relationships between the target anchor points and nodes in the geographic coordinate system to generate a high-precision map. The selection of target anchor points and the node relationships between nodes ensure the accuracy of the coordinates of each node in the high-precision map, ultimately providing accurate road environment information for autonomous driving and ensuring the safe and efficient operation of autonomous vehicles. Attached Figure Description

[0056] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0057] Figure 1 A flowchart illustrating a high-precision map construction method provided in this application embodiment;

[0058] Figure 2 A schematic diagram illustrating a process for constructing a semantic topology network graph, provided as an embodiment of this application;

[0059] Figure 3 This is a schematic diagram illustrating a node inclusion relationship provided in an embodiment of this application;

[0060] Figure 4 This application provides a schematic diagram of a process for determining a target anchor point.

[0061] Figure 5 A schematic diagram of a target anchor point status monitoring process provided in an embodiment of this application;

[0062] Figure 6 A flowchart illustrating the optimization of geographic coordinates provided in this application embodiment;

[0063] Figure 7 A schematic diagram of a high-precision map building device provided in this application embodiment;

[0064] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0065] Before introducing the solution of this application, some of the technical terms involved in the embodiments of this application will be explained and described first:

[0066] Convolutional Neural Networks (CNNs): A deep learning architecture widely used in image recognition and processing. It extracts local features from an image through convolutional layers, reduces feature dimensionality using pooling layers, and then performs classification or regression through fully connected layers. CNNs can automatically learn hierarchical features in images, from simple edges to complex object shapes, effectively capturing the spatial structure of images. They perform exceptionally well in tasks such as image classification and object detection.

[0067] Transformer architecture: a deep learning model based on self-attention mechanism. It abandons the traditional recurrent neural network (RNN) structure and processes sequential data in parallel through self-attention mechanism.

[0068] IPM (Inverse Perspective Mapping) transformation: an image processing technique used to convert an image from a perspective projection to a top view. It uses geometric transformations to map elements such as lane lines and vehicles in a road image to a virtual top-view plane, thereby eliminating perspective distortion.

[0069] Graph Structure Semantic Relationship Parsing Layer (GNN): This is a component of a Graph Neural Network (GNN) used to process graph-structured data. It updates the feature representations of nodes by aggregating neighborhood information, thereby capturing the complex relationships between nodes.

[0070] The embodiments of this application are described below with reference to the accompanying drawings. The terminology used in the implementation section of this application is for explaining specific embodiments only and is not intended to limit the scope of this application.

[0071] The terms "first," "second," etc., used in this application's specification are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms are interchangeable where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of units is not necessarily limited to those units, but may include other units not explicitly listed or inherent to those processes, methods, products, or apparatuses.

[0072] This application can be applied to the field of autonomous driving. The following uses the autonomous driving scenario of a vehicle as an example to introduce the application scenario of this application.

[0073] In autonomous driving scenarios, vehicles need to acquire accurate road environment information in real time to ensure driving safety and efficiency. High-precision maps, as a core support, must have the ability to update road environment information in real time and present it accurately. Existing high-precision map construction technologies mostly rely on pre-collected static data and fixed surveying methods, which have obvious defects: poor real-time performance, long update cycles, and inability to reflect dynamic road condition changes such as temporary traffic control and sudden accidents in a timely manner. This can easily lead to vehicles driving based on outdated information, resulting in safety hazards. Traditional methods also rely on physical scale calculation methods such as IPM transformation and triangulation. These physical scale calculation methods generally involve calibrating sensors to determine geometric parameters such as relative positions and angles between sensors, then collecting road data such as physical distances, angles, and positions in the road scene using the calibrated sensors, and finally calculating the geographic coordinates of each road element based on the road data to complete map construction. However, this physical scale calculation method has significant limitations in practical applications. It requires high accuracy in sensor calibration and environmental conditions. If there is a sensor calibration deviation, or if there is environmental interference such as obstruction or severe weather, it can easily lead to large errors in geographic coordinate calculation, which in turn affects the accuracy of map construction and makes the constructed high-precision map unable to adapt to dynamically changing road conditions.

[0074] To improve the accuracy of high-precision map construction and meet the needs of complex road conditions in autonomous driving, this application acquires image data of the road environment surrounding the vehicle and constructs a semantic topological network graph containing nodes and associated edges based on the image data. Nodes are information units extracted from the road environment surrounding the vehicle, carrying semantic and spatial features, while associated edges represent the spatial and semantic relationships between nodes. By constructing this semantic topological network graph, the limitations of traditional high-precision map construction methods, which only focus on spatial coordinates and lack semantic information, are overcome. This also avoids the map feature misalignment problem caused by physical scale recovery errors in scenarios such as sensor calibration deviations and occlusion. Next, target anchor points that meet the spatial reference conditions are determined from the semantic topology network graph. Then, the target anchor points are mapped to the geographic coordinate system in combination with the real-time vehicle pose. Finally, the spatial position of each node is determined based on the relationship between the target anchor points and nodes in the geographic coordinate system to generate a high-precision map. This improves the accuracy of high-precision map construction, breaks through the dependence on traditional physical scale calculation, and realizes scale-independent high-precision map construction. Scale-independent means that it does not rely on physical scale calculation methods such as IPM inverse perspective transformation and binocular vision triangulation to calculate the actual width of lane lines and the actual distance of traffic signs to build high-precision maps, thus providing a reliable structured representation of the road environment for autonomous driving.

[0075] Next, see Figure 1 , Figure 1 This is a flowchart illustrating a high-precision map construction method provided in an embodiment of this application. The high-precision map construction method of this application can be implemented using a high-precision map construction system deployed on a terminal. As described below, the high-precision map construction method of this application embodiment specifically includes the following steps:

[0076] Step S11: Acquire image data.

[0077] Specifically, in the process of building high-precision maps for autonomous vehicles, image data representing the road environment surrounding the vehicle can be acquired first. Optionally, the image data can be continuously collected by multiple cameras installed on the vehicle, and these cameras can be distributed at different locations on the vehicle, such as front-view cameras, rear-view cameras, and side-view cameras, to cover the vehicle's all-around perspective and ensure the acquisition of complete road environment information, providing basic data support for the subsequent construction of the semantic topology network. Optionally, the timestamp accuracy alignment error between cameras can be controlled within 10 milliseconds to ensure data synchronization and accuracy.

[0078] Step S12: Determine the semantic topology network graph based on the image data.

[0079] Specifically, after acquiring image data of the road environment surrounding the vehicle, a semantic topological network graph can be constructed based on this image data. The semantic topological network graph describes the road environment surrounding the vehicle through nodes and connecting edges. Nodes represent information units extracted from the road environment surrounding the vehicle, carrying semantic and spatial features; connecting edges represent the relationships between nodes; and node relationships represent the spatial and semantic dimensions of the relationships between nodes. By constructing a semantic topological network graph, complex road environment information can be transformed into graph data, providing a foundation for subsequent high-precision map construction.

[0080] Step S13: Determine the target anchor point based on the semantic topology network graph.

[0081] Specifically, after constructing the semantic topological network graph, target anchor points can be determined based on the semantic topological network graph. Target anchor points are nodes selected from the semantic topological network graph that satisfy the criteria for serving as spatial references, which not only improves the accuracy of map construction but also enhances robustness.

[0082] Step S14: Map the target anchor point to the geographic coordinate system based on the real-time pose of the vehicle.

[0083] Specifically, the target anchor point is mapped from the vehicle's local coordinate system to the geographic coordinate system based on the vehicle's real-time pose, ensuring the accuracy of the target anchor point in the geographic coordinate system. The vehicle's local coordinate system (vehicle coordinate system) is a coordinate system with the vehicle as the origin, and the geographic coordinate system can adopt the internationally accepted geodetic coordinate system. The vehicle's real-time pose represents the vehicle's current position and attitude in the geographic coordinate system.

[0084] Step S15: Determine the spatial location of each node in the semantic topology network graph in the geographic coordinate system based on the target anchor point in the geographic coordinate system and the node relationship, and obtain a high-precision map.

[0085] Specifically, based on the vehicle's real-time pose, after mapping the target anchor point to the geographic coordinate system, the spatial position of each node in the semantic topology network graph is determined in the geographic coordinate system using the target anchor point and the node relationships in the semantic topology network graph, thus completing the construction of a high-precision map. The target anchor point provides a reference, the node relationships provide relative position information, and the geographic coordinates of all nodes in the geographic coordinate system are determined to generate a high-precision map, providing navigation basis for autonomous driving.

[0086] This embodiment acquires image data of the road environment surrounding the vehicle and constructs a semantic topological network graph containing nodes and associated edges based on the image data. Nodes are information units extracted from the road environment surrounding the vehicle, carrying semantic and spatial features, while associated edges represent the spatial and semantic relationships between nodes. By constructing this semantic topological network graph, the limitations of traditional high-precision map construction methods, which only focus on spatial coordinates and lack semantic information, are overcome. This avoids map element misalignment problems caused by physical scale calculation methods in scenarios such as sensor calibration deviations and occlusion. Target anchor points that meet spatial reference conditions are then determined from the semantic topological network graph. Subsequently, these target anchor points are mapped to the geographic coordinate system based on the vehicle's real-time pose. Finally, the spatial position of each node in the geographic coordinate system is determined based on the relationship between the target anchor points and nodes to generate a high-precision map. The selection of target anchor points and the node relationships between nodes ensure the accuracy of the coordinates of each node in the high-precision map, ultimately providing accurate road environment information for autonomous driving and ensuring the safe and efficient operation of autonomous vehicles.

[0087] Furthermore, in some embodiments of this application, the process of constructing a semantic topology network graph using image data can be described in more detail, see [link to relevant documentation]. Figure 2 , Figure 2 This is a schematic diagram of a process for constructing a semantic topology network graph provided in an embodiment of this application. The following will be combined with... Figure 2 This section will introduce the relevant content.

[0088] Step S210: Obtain road elements.

[0089] Alternatively, in some embodiments of this application, a pre-trained image recognition model can be used to recognize image data, which will be described in detail below.

[0090] Alternatively, image data can be input into a pre-trained image recognition model to obtain road elements corresponding to the image data, which represent various entities related to driving in the road environment surrounding the vehicle. Optionally, the image recognition model can be trained using different image data as training samples and the corresponding road elements as sample labels. Alternatively, the image recognition model can be a convolutional neural network (CNN) or a Transformer architecture.

[0091] Furthermore, in an alternative approach, the image recognition model also integrates a GNN layer (Graph Structure Semantic Relationship Parsing Layer). The GNN layer analyzes each road element output by the image recognition model to obtain the relationship probability between each road element. The relationship probability represents the strength of the association between road elements.

[0092] In one optional approach, relation probabilities can lay the foundation for constructing node relationships in a semantic topology network graph. Optionally, if the relation probability between nodes is greater than a preset relation probability value, it can be prioritized to determine whether a node pair has a relationship. If it is determined that a node pair has a relationship, an associated edge is added to the node pair. By using relation probabilities, blind calculations between all pairs of nodes are avoided, thus improving the efficiency of constructing the semantic topology network graph.

[0093] Step S211: Perform element detection on the road elements output by the image recognition model.

[0094] Specifically, element detection is performed on the road elements output by the image recognition model to obtain the road element type corresponding to each road element. The road element type represents the classification and definition of the function of the road element. Optionally, the road element type can include lane lines, road markings, traffic signs, and drivable areas. Optionally, the attributes of the road element can include any one or more of the following: output category, location, width, curvature, and pixel-level segmentation mask, where the output category is related to the road element type.

[0095] Step S212: Based on the road element types obtained from element detection, classify each road element according to its type.

[0096] Step S213: Create nodes corresponding to each road element based on the road element type.

[0097] Optionally, nodes can be created based on the road element type. The node corresponding to the road element is a lane line node, the node corresponding to the road element is a traffic sign node, the node corresponding to the road element is a road surface marking node, and the node corresponding to the road element is a drivable area node.

[0098] In one alternative approach, the node attributes may include any one of the following: node identifier, road element type, spatial coordinates, semantic attributes, detection confidence, and number of consecutive tracking frames. The node identifier is a unique identifier for the node. The road element type includes lane lines, traffic signs, road markings, and drivable areas. The spatial coordinates represent the position of the node in a vehicle coordinate system centered on the vehicle. The semantic attributes represent the traffic rules or functional types carried by the node. The detection confidence represents the credibility of the road element corresponding to the node in the recognition process of the image recognition model. The number of consecutive tracking frames represents the number of times the node is continuously included in the image data and successfully matched with the node identifier.

[0099] In one alternative approach, different types of nodes correspond to different semantic attributes. The semantic attributes of lane line nodes may include solid lines, dashed lines, and double solid lines, etc. The semantic attributes of road marking nodes may include arrows and zebra crossings, etc. The semantic attributes of traffic sign nodes may include speed limit signs and stop signs, etc. The semantic attributes of drivable area nodes may include drivable areas and non-drivable areas.

[0100] Step S214: Determine the node relationships based on the node attributes of each node.

[0101] In one alternative approach, node relationships may include connection relationships, constraint relationships, parallel relationships, and inclusion relationships. Here, a connection relationship indicates that there is a connection between lane line nodes, a constraint relationship indicates that the drivable area nodes are constrained by the rules of traffic sign nodes, a parallel relationship indicates that there is a parallel relationship between lane line nodes, and an inclusion relationship indicates that the road marking node is included in the drivable area nodes.

[0102] In one alternative approach, in some embodiments of this application, it can be determined whether a preset relationship matching rule is met based on the node attributes of each node. If the relationship matching rule is met, it is determined that there is a node relationship between the nodes. The relationship matching rule represents the judgment rule set for different node relationships. The judgment rule combines the node attributes of the nodes and realizes the judgment of the node relationship by comparing it with a preset threshold.

[0103] Optionally, the relationship matching rules include different judgment rules for four types of relationships: connection, constraint, parallel, and inclusion. One optional approach is that the relationship matching rules may include a first judgment rule for connection, a second judgment rule for constraint, a third judgment rule for parallel, and a fourth judgment rule for inclusion. Specifically, the first judgment rule for connection is: all nodes are lane line nodes, and the endpoint distance between lane line nodes is less than a preset endpoint distance, and the directional difference between lane line nodes is less than a first preset angle. The second judgment rule for constraint is: nodes are respectively a drivable area node and a traffic sign node, and the viewing angle of the traffic sign node within the drivable area node is less than a preset viewing angle range, and the distance between the traffic sign node and the drivable area node is less than a first preset distance. The third judgment rule for parallel is: all nodes are lane line nodes, and the directional difference between lane line nodes is less than a second preset angle, and the minimum distance between lane line nodes is less than a preset distance range. The fourth judgment rule for inclusion is: nodes are respectively a road marking node and a drivable area node, and the coverage rate of the road marking node within the drivable area node is greater than or equal to a preset coverage rate.

[0104] Optionally, when determining whether a connection exists between nodes, the semantic attributes in the node attributes can be used to determine whether all nodes are lane line nodes. If so, the spatial coordinates in the node attributes can be used to determine whether the endpoint distance between the lane line nodes is less than a preset endpoint distance and whether the directional difference is less than a first preset angle. If both conditions are met, a connection exists between the nodes. Optionally, the preset endpoint distance can be 0.3m, and the first preset angle can be 10°. That is, if the endpoint distance between two lane line nodes is less than 0.3m and the directional difference is less than 10°, a connection can be determined between the two lane line nodes.

[0105] Optionally, when determining whether a constraint relationship exists between nodes, the semantic attributes in the node attributes can be used to determine whether the nodes are respectively a drivable area node and a traffic sign node. If so, the spatial coordinates in the node attributes can be used to determine whether the viewing angle of the traffic sign node from the drivable area node is within a preset viewing angle range and whether the distance between the traffic sign node and the drivable area node is less than a first preset distance. If both conditions are met, then a constraint relationship exists between the nodes. Optionally, the preset viewing angle range can be 15°, and the first preset distance can be 30m. That is, if the traffic sign node is within ±15° of the viewing angle directly in front of the drivable area node and the distance is less than 30m, then it can be determined that a constraint relationship exists between the traffic sign node and the drivable area node.

[0106] Optionally, when determining whether a parallel relationship exists between nodes, the semantic attributes in the node attributes can be used to determine whether all nodes are lane line nodes. If so, the spatial coordinates in the node attributes can be used to determine whether the directional difference between lane line nodes is less than a second preset angle and whether the minimum distance between lane line nodes is within a preset distance range. If both conditions are met, then a parallel relationship exists between the nodes. Optionally, the second preset angle can be 5°, and the preset distance range can be the standard lane width range (3.5±0.5m). That is, if the directional difference between lane line nodes is less than 5° and the minimum distance between lane line elements is within the standard lane width range (3.5±0.5m), then a parallel relationship can be determined between the lane line nodes.

[0107] Optionally, when determining whether an inclusion relationship exists between nodes, the semantic attributes in the node attributes can be used to determine whether the nodes are respectively road marking nodes and drivable area nodes. If so, the spatial coordinates in the node attributes can be used to determine whether the coverage rate of the road marking node in the drivable area nodes is greater than or equal to a preset coverage rate. If this is satisfied, an inclusion relationship exists between the nodes. Optionally, the preset coverage rate can be 95%, meaning that if the coverage rate of the road marking node in the drivable area nodes is greater than or equal to 95%, it can be determined that there is an inclusion relationship between the road marking node and the drivable area node. See also an optional method. Figure 3 , Figure 3 This is a schematic diagram illustrating the inclusion relationship between nodes provided in an embodiment of this application. Figure 3 It can be seen that the road marking nodes are included in the drivable area nodes.

[0108] This embodiment constructs a relationship matching rule system based on node attributes (semantic attributes and spatial coordinates), and formulates differentiated judgment logic for four types of node relationships: connection, constraint, parallelism, and inclusion. This enables automatic determination of the association relationship between road elements, covering the association logic of road elements in autonomous driving scenarios, and providing flexible adjustment space for rule adaptation in different scenarios. For example, different road standards can be adapted by modifying preset thresholds. Ultimately, this lays the foundation for relationship determination in the efficient construction of semantic topology network graphs and realizes the mapping of high-precision maps to the semantic logic of the road environment.

[0109] Step S215: Determine the relationship strength value corresponding to the node relationship.

[0110] An alternative approach is to determine the relationship strength value corresponding to node relationships based on geometric consistency scores and the number of consecutive tracking frames. The geometric consistency score characterizes the deviation between the actual angle and / or actual distance values ​​between nodes and the configured values; a smaller deviation results in a higher geometric consistency score. The number of consecutive tracking frames characterizes the number of times the same node successfully identifies and matches node identifiers across multiple consecutive frames of image data. Geometric consistency scores, the number of consecutive tracking frames, and the relationship strength value are positively correlated.

[0111] An alternative approach is to determine the relationship strength value corresponding to the node relationship based on the geometric consistency score and the number of consecutive tracking frames. The relationship strength formula can be: Relationship strength value = 0.6 × Geometric consistency score + 0.4 × Number of consecutive tracking frames.

[0112] In one optional approach, the geometric consistency score can be calculated by acquiring image data corresponding to a preset number of frames, calculating the deviation between the actual angle value and / or actual distance value of two nodes in each frame and the configured values, performing normalization processing, and using the average value of the normalized deviation values ​​for each frame as the geometric consistency score. Alternatively, the configured values ​​may include configured distance and angle.

[0113] Step S216: Determine whether the relationship strength value is less than the preset first strength value. If yes, proceed to step S218, which is to remove the node relationship. If no, proceed to step S217, which is to retain the node relationship.

[0114] Alternatively, if the relationship strength value is less than a preset first strength value, the node relationship is removed; if the relationship strength value is not less than the first strength value, the node relationship is retained. Alternatively, the first strength value can be 0.4.

[0115] Step S217: Preserve node relationships.

[0116] In one alternative approach, if the relationship strength value is not less than the first strength value, the node relationship is retained, and step S219 can be executed to further determine whether the relationship strength value is greater than a preset second strength value.

[0117] Step S218: Remove node relationships.

[0118] Step S219: Determine whether the relationship strength value is greater than the preset second strength value. If not, proceed to step S220; if yes, proceed to step S221.

[0119] In one optional approach, if the relationship strength value is not less than the first strength value and not greater than a preset second strength value, the node relationship is retained and updated to a weak relationship; if the relationship strength value is greater than the second strength value, the node relationship is retained and may be further updated to a strong relationship. In another optional approach, the second strength value can be 0.7.

[0120] Step S220: Preserve the node relationships and update them to weak relationships.

[0121] An optional approach is to retain the node relationship and update it to a weak relationship if the relationship strength value is not less than the first strength value and not greater than a preset second strength value. During the construction of the semantic topology network graph, weak relationships can, on the one hand, avoid topology jitter caused by frequent additions or removals of relationships when the relationship strength value is insufficient; on the other hand, weak relationships are continuously tracked. If the relationship strength value of a weak relationship is greater than the second strength value in a subsequent frame, it is upgraded to a strong relationship; if it is less than the first strength value, it is removed. Furthermore, in scenarios of brief sensor occlusion or misidentification, weak relationships can maintain the basic integrity of the topology and prevent map abrupt changes.

[0122] Step S221: Preserve node relationships.

[0123] Alternatively, if the relationship strength value is greater than the second strength value, the node relationship is retained and can be further updated to a strong relationship.

[0124] An alternative approach is to retain the node relationship and update it as a strong relationship if the relationship strength value is greater than 0.7; retain the node relationship and update it as a weak relationship if the node relationship value is neither greater than 0.7 nor less than 0.4; and remove the node relationship if the node relationship value is less than 0.4.

[0125] S222. Construct a semantic topology network graph based on the relationships between nodes.

[0126] A semantic topology network graph can be constructed based on nodes and node relationships. Associative edges are added between nodes with existing relationships, and the semantic topology network graph is constructed based on each node and its associated edges. Alternatively, after constructing the semantic topology network graph, the relationship strength values ​​corresponding to each node relationship can be continuously tracked, and node relationships can be updated based on the relationship strength values. Associative edges can also be updated based on node relationships to update the semantic topology network graph in real time.

[0127] In this embodiment, the semantic topology network graph construction process uses a pre-trained image recognition model combined with a GNN layer to obtain road elements and relationship probabilities. Based on differentiated relationship matching rules, it determines four types of node relationships: connections, constraints, etc., and then manages each node relationship hierarchically according to relationship strength. This approach avoids blind node calculations, improving construction efficiency, and dynamically maintains strong and weak relationships to prevent topology fluctuations. It also adapts to different road standards, achieving a mapping between the semantic logic of the road environment and spatial coordinates, resulting in a high-precision map and laying the foundation for real-time updates of the high-precision map.

[0128] Furthermore, in some embodiments of this application, the process of determining and replacing target anchor points based on semantic topology network graphs can be described in detail, see reference. Figure 4 , Figure 4 This application provides a flowchart illustrating the process of determining a target anchor point, in conjunction with... Figure 4 This section will be described in detail.

[0129] S31. Determine candidate anchor points based on the semantic topology network graph.

[0130] Optionally, lane intersections and / or traffic sign nodes are selected from the semantic topology network graph based on detection confidence and the number of consecutive tracking frames. These selected lane intersections and / or traffic sign nodes are then used as candidate anchor points. A lane intersection is the intersection of at least two lane line nodes. Since lane intersections and traffic sign nodes possess high stability and strong topological correlation (i.e., their spatial locations are relatively stable and they are closely associated with other nodes), the selection range of candidate anchor points includes lane intersections and traffic sign nodes; that is, the types of candidate anchor points include lane intersections and traffic signs. However, road marking nodes and drivable area nodes, due to their high dynamic nature (e.g., temporary road arrows), are not currently included in the selection range of candidate anchor points.

[0131] The selection criteria for candidate anchor points can be that the number of consecutive tracking frames for lane intersections and / or traffic sign nodes is not less than a preset value, and the average detection confidence score is not less than a preset value. Specifically, lane intersections with a consecutive tracking frame count not greater than the preset value and an average detection confidence score not less than the preset value can be used as candidate anchor points, and traffic sign nodes with a consecutive tracking frame count not greater than the preset value and an average detection confidence score not less than the preset value can be used as candidate anchor points. The average detection confidence score can be the average of the detection confidence scores across the preset number of frames. Alternatively, lane intersections and / or traffic sign nodes with a consecutive tracking frame count of not less than 10 frames and an average detection confidence score of not less than 0.85 can be selected as candidate anchor points.

[0132] S32. Determine the candidate anchor point score based on the anchor point scoring model.

[0133] In one optional approach, candidate anchor points are input into an anchor point scoring model. The anchor point scoring model determines the node confidence score, topological centrality score, and location stability score of the candidate anchor points. The node confidence score, topological centrality score, and location stability score are then weighted and summed to obtain and output the candidate anchor point score corresponding to the candidate anchor point. The node confidence score represents the reliability of the road element corresponding to the candidate anchor point being identified in multiple consecutive frames of image data. The topological centrality score represents the degree of association between the anchor point and other nodes in the semantic topological network graph. The location stability score represents the degree of fluctuation of the spatial location of the candidate anchor point in multiple consecutive frames of image data.

[0134] The anchor point scoring model can obtain the candidate anchor point score by weighting and summing the node confidence score, topological centrality score, and location stability score according to the configured weights. Optionally, the candidate anchor point score = 40% × node confidence score + 30% × topological centrality score + 30% × location stability score.

[0135] One possible implementation method, where each score is calculated, is as follows:

[0136] The node confidence score is calculated as follows: obtain the detection confidence of the candidate anchor point in the current frame (current frame confidence), and obtain the average of the average confidence of the candidate anchor point in multiple consecutive historical frames (historical average confidence). Add the current frame confidence and the historical average confidence, and take the average of the sum to obtain the node confidence score, that is, node confidence score = (current frame confidence + historical average confidence) / 2.

[0137] The calculation method of topological centrality score is as follows: obtain the relationship distance between the candidate anchor point and each associated node, calculate the average relationship distance, add 1 to the average relationship distance to get a value, and finally divide 1 by this value to get the topological centrality score, that is, topological centrality score = 1 / (1 + average relationship distance).

[0138] The location stability score is calculated as follows: First, obtain the location data of the candidate anchor point in multiple consecutive historical frames and calculate the standard deviation of this location data. Then, obtain the maximum permissible deviation of the candidate anchor point, which is a pre-defined maximum acceptable fluctuation range for the node position. Next, divide the standard deviation of the location data from the multiple consecutive historical frames by the maximum permissible deviation to obtain the relative proportion of location fluctuation. Finally, subtract this relative proportion from 1. The final result is the location stability score of the candidate anchor point. That is, location stability score = 1 - (standard deviation of location data from the multiple consecutive historical frames / maximum permissible deviation).

[0139] S33. Determine whether the candidate anchor point score is greater than the preset value. If yes, proceed to step S34, that is, use the candidate anchor point as the target anchor point; otherwise, proceed to step S35, that is, use the candidate anchor point as the backup anchor point.

[0140] Optionally, it can be determined whether the candidate anchor score of the candidate anchor is greater than a preset value. Candidate anchors with scores greater than the preset value are used as target anchors, and candidate anchors with scores not greater than the preset value are used as backup anchors. The backup anchors are used as new target anchors when the state of the target anchor is abnormal.

[0141] S34. Use the candidate anchor point as the target anchor point.

[0142] Candidate anchor points whose scores are greater than a preset value are selected as target anchor points.

[0143] S35. Use the candidate anchor point as a backup anchor point.

[0144] Candidate anchor points whose scores are no greater than a preset value are used as backup anchor points, so as to lay the foundation for using backup anchor points as new target anchor points when the state of the target anchor point is abnormal.

[0145] This embodiment, based on a semantic topological network graph, comprehensively considers detection confidence and the number of consecutive tracking frames to select lane intersections and traffic sign nodes as candidate anchor points, ensuring the stability and topological value of the candidate anchor points. Next, an anchor point scoring model is used to comprehensively score the candidate anchor points from three dimensions: node confidence, topological centrality, and positional stability, scientifically quantifying the reliability of the anchor points and providing an objective basis for subsequent selection. By setting a scoring threshold, high-scoring candidate anchor points are determined as target anchor points, while low-scoring candidate anchor points are used as backup anchor points. This hierarchical mechanism effectively balances the stability and usability of anchor points.

[0146] Furthermore, in some embodiments of this application, a pre-activated target anchor point can be selected based on a backup anchor point to ensure that when the state of the target anchor point is abnormal, a new target anchor point replaces the target anchor point of the state agenda. Some of these contents are described below.

[0147] In one optional approach, since candidate anchors, backup anchors, pre-activated target anchors, optimal pre-activated target anchors, and target anchors are all selected based on nodes, each node, candidate anchor, backup anchor, pre-activated target anchor, optimal pre-activated target anchor, and target anchor all possess node attributes. Furthermore, candidate anchors, backup anchors, pre-activated target anchors, optimal pre-activated target anchors, and target anchors also possess anchor attributes. These anchor attributes include anchor identifier, binding time, anchor type, spatial error value, effective duration, historical offset record, and associated nodes. The binding time represents the timestamp of the anchor binding to the vehicle; the anchor type includes lane intersections and traffic signs; the spatial error value represents the error value of the anchor in the vehicle coordinate system; the effective duration represents the expected effective time of the anchor; the historical offset record represents the positional deviation of the anchor at different time points; and the associated nodes represent other nodes in the semantic topology network graph that are directly associated with the anchor.

[0148] Alternatively, the spatial error value can be determined by the image recognition model outputting the spatial error value of the road element corresponding to the anchor point when outputting the road element corresponding to the anchor point. That is, the spatial error value can be output by the image recognition model. Or, the spatial error value can be determined based on the position of the anchor point in several consecutive historical frames. That is, the standard deviation of the position data of the anchor point in several consecutive historical frames is used as the spatial error value. The greater the position jump of the anchor point, the greater the spatial error value.

[0149] Alternatively, the process of selecting a pre-activated target anchor point from the backup anchor points can be as follows:

[0150] First, select backup anchor points that meet the pre-screening criteria. These criteria include a relative distance from the target anchor point that is not less than a configured distance value, and a different direction relative to the vehicle. Alternatively, select an anchor point with a relative distance from the target anchor point that is not less than a preset value (5m) to avoid simultaneous failure due to obstruction in the same area. This also ensures that the backup anchor points and the target anchor point are located in different orientations relative to the vehicle. For example, if the target anchor point is in front, prioritize backup anchor points located to the side or rear, where the direction is different from the main anchor point.

[0151] Next, the alternative anchor points that meet the pre-screening criteria can be scored. One alternative approach is to obtain the alternative score corresponding to the alternative anchor point that meets the pre-screening criteria by weighted summing of the location stability score, topological relevance score, and scene adaptability score.

[0152] An alternative approach to assess the stability of backup anchor points that meet pre-screening criteria is to determine their positional stability score based on spatial error values, effective duration, and historical offset records. Spatial error values ​​and historical offset records are negatively correlated with the positional stability score, while effective duration is positively correlated. Specifically, a smaller spatial error value indicates higher spatial accuracy of the anchor point, resulting in a higher positional stability score; conversely, a larger spatial error value leads to a lower positional stability score. A longer effective duration means the anchor point remains stable over a longer period, resulting in a higher positional stability score. Conversely, a shorter effective duration results in a lower positional stability score. Historical offset records reflect the positional changes of the anchor point at different points in time; a smaller offset indicates a more stable anchor point position, resulting in a higher positional stability score; a larger offset results in a lower positional stability score. By comprehensively considering these three factors, the positional stability of each backup anchor point that meets the pre-screening criteria can be accurately assessed, thus providing a basis for subsequently selecting pre-activated target anchor points.

[0153] The topological correlation score is determined based on the correlation strength between the backup anchor points that meet the pre-screening conditions and the active nodes. The active nodes include the target anchor points and nodes with high detection confidence. The nodes with high detection confidence are those with a detection confidence greater than a preset value. The correlation strength is positively correlated with the topological correlation score, that is, the higher the correlation strength, the higher the topological correlation score.

[0154] The scene adaptability score is determined by assigning weights to backup anchor points that meet the pre-screening criteria based on the vehicle's current scene and anchor point type. The scene adaptability score characterizes the degree of matching between the backup anchor points that meet the pre-screening criteria and the current scene. Optionally, the current scene type can be determined using real-time image data and positioning information collected by the vehicle's cameras. Scene types can be divided into two main categories: highway scenes and urban intersection scenes. Highway scenes are characterized by straight roads, clear and continuous lane lines, and minimal dynamic interference (such as pedestrians and non-motorized vehicles). Urban intersection scenes are characterized by complex road intersections and dense traffic signs (such as no left turn and yield to pedestrian signs).

[0155] Based on the scene characteristics of the vehicle, the weight of the anchor type of the backup anchor point that meets the pre-screening conditions is determined. This weight can be directly used or converted into a scene adaptability score for the backup anchor point that meets the pre-screening conditions. One optional approach, in high-speed scenarios, is to increase the weight of lane intersection anchor points (e.g., set the weight ratio to 60%) and decrease the weight of traffic sign anchor points (e.g., set the weight ratio to 40%), as lane intersections are more suitable for the long-distance continuous driving requirements of high-speed scenarios due to their stability. That is, if the backup anchor point that meets the pre-screening conditions is a lane intersection anchor point, the scene adaptability score can be determined based on the weight of lane intersection anchor points in high-speed scenarios; if the backup anchor point that meets the pre-screening conditions is a traffic sign anchor point, the scene adaptability score can be determined based on the weight of traffic sign anchor points in high-speed scenarios. In urban intersection scenarios, the weight of traffic sign anchors can be increased (e.g., set to 60%), while the weight of lane intersection anchors can be decreased (e.g., set to 40%), because the semantic information of traffic signs better supports complex driving decisions at intersections. Specifically, if a backup anchor that meets the pre-screening criteria is a lane intersection anchor, the scenario suitability score can be determined based on the weight of lane intersection anchors in urban intersection scenarios; conversely, if a backup anchor that meets the pre-screening criteria is a traffic sign anchor, the scenario suitability score can be determined based on the weight of traffic sign anchors in urban intersection scenarios.

[0156] An optional approach is to weight and sum the location stability score, topology correlation score, and scene adaptability score according to the configured weights to obtain the backup score corresponding to the backup anchor point that meets the pre-screening conditions. Optionally, the backup score = 40% × location stability score + 30% × topology correlation score + 30% scene adaptability score.

[0157] Finally, select the backup scores that meet the requirements as pre-activated target anchors. The requirements can be either backup anchors that meet the pre-screening criteria and whose ranking is higher than the preset value, or backup anchors that meet the pre-screening criteria and whose scores are higher than the preset value.

[0158] Furthermore, when the target anchor point is in a failed state, an optimal pre-activated target anchor point is selected based on the pre-activated target anchor points and used as the new target anchor point to replace the failed target anchor point. Optionally, when the failure state occurs when the detection confidence of the target anchor point is greater than a preset value (25%) or the positional offset of the target anchor point is greater than a preset value (1.5m), the failed target anchor point can be switched to the optimal pre-activated target anchor point within 0.1 seconds. Optionally, the optimal pre-activated target anchor point can be the highest-ranked pre-activated target anchor point or the pre-activated target anchor point with the highest backup score.

[0159] An alternative approach, see Figure 5 , Figure 5 This is a flowchart illustrating a target anchor point status monitoring method provided in an embodiment of this application. Figure 5 It can be seen that monitoring the status of the target anchor point can first determine whether the detection confidence of the target anchor point is greater than a preset value. If so, the optimal pre-activated target anchor point is activated. If not, it can determine whether the position offset of the target anchor point is greater than a preset value. If so, the optimal pre-activated target anchor point is activated. If not, the current target anchor point is maintained.

[0160] In this embodiment, when the target anchor point's state is abnormal, the optimal pre-activated target anchor point can be quickly selected from the pre-activated target anchor points for replacement. The entire switching process can be completed within 0.1 seconds, greatly improving robustness and response speed. Furthermore, the pre-activated target anchor point is selected based on backup anchor points, and the selection of pre-activated target anchor points integrates three scoring dimensions: position stability score, topological correlation score, and scene adaptability score, ensuring a high degree of matching and rapid adaptation of the new target anchor point to the current scene. This not only improves the accuracy and real-time performance of high-precision map construction but also enhances stability and safety in complex dynamic environments, providing more reliable road environment information for autonomous vehicles and ensuring their safe and efficient driving.

[0161] Furthermore, in some embodiments of this application, the process of mapping the target anchor point to the geographic coordinate system based on the real-time pose of the vehicle can also be described, and this part will be described in detail below.

[0162] One alternative approach is to determine the geographic coordinates of the target anchor point based on the vehicle's real-time pose and the spatial coordinates of the target anchor point. The spatial coordinates of the target anchor point represent its position in a vehicle-centric coordinate system, while the geographic coordinates are the anchor point's coordinates within the geographic coordinate system. The vehicle's real-time pose represents its current position and attitude within the geographic coordinate system. The real-time pose can be obtained by fusing GNSS (Global Navigation Satellite System) and INS (Inertial Navigation System). Optionally, the vehicle's real-time pose may include its coordinates and heading angle in the geographic coordinate system. The target anchor point is then mapped to the geographic coordinate system based on its geographic coordinates.

[0163] Alternatively, the formula for determining the geographic coordinates of the target anchor point can be:

[0164] Ag_lat = Pv_lat + (y_local / 111111);

[0165] Ag_lon = Pv_lon + (x_local / (111111×cos(Pv_lat)));

[0166] Where Ag_lat represents the latitude of the anchor point in the geographic coordinate system, Ag_lon represents the longitude of the anchor point in the geographic coordinate system, Pv_lat represents the latitude of the vehicle in the geographic coordinate system, Pv_lon represents the longitude of the vehicle in the geographic coordinate system, y_local represents the ordinate of the anchor point in the vehicle coordinate system, x_local represents the abscissa of the anchor point in the vehicle coordinate system, 111111 is a constant representing the number of meters corresponding to each degree of latitude, and cos(Pv_lat) represents the latitudinal projection coefficient. Since the Earth is a sphere, the ground distance corresponding to each degree of longitude is different at different latitudes; therefore, it needs to be multiplied by the latitudinal projection coefficient (cos(Pv_lat)). The latitudinal projection coefficient takes into account the influence of latitude on longitude distance, ensuring that the calculated longitude variation is accurate. Alternatively, the geographic coordinates of each node can also be determined using the above formula.

[0167] Alternatively, the geographic coordinates of the target anchor point may contain errors, so the geographic coordinates of the target anchor point can be further optimized. See also Figure 6 , Figure 6 This is a flowchart illustrating the optimization of geographic coordinates provided in an embodiment of this application. The process of optimizing the geographic coordinates of anchor points is described below.

[0168] S41. Determine the positioning error and topology error based on the node relationships and the geographic coordinates.

[0169] Specifically, the positioning error and topology relationship error can be determined. The positioning error is the measurement error of the vehicle's real-time pose, and the topology relationship error is the deviation of the spatial relationship between the target anchor point and other nodes in the semantic topology network graph.

[0170] Alternatively, the positioning error can be determined by the following method.

[0171] The essence of positioning error is the quantification of the difference between the vehicle's real-time pose and the reference pose. The reference pose can be either the vehicle's historical pose or its predicted pose. The historical pose can be the vehicle pose determined after optimization from the previous frame or several previous frames. The predicted pose can be derived from this, specifically by combining the historical pose from the previous moment with real-time acquired motion parameters such as vehicle speed and angular velocity to calculate the theoretically correct position of the vehicle at the current moment. Furthermore, the GNSS covariance matrix can be introduced as a supplementary basis for positioning error. The GNSS covariance matrix reflects the impact of factors such as satellite signal strength and obstruction on positioning accuracy. By combining the GNSS covariance matrix with the pose deviation, the confidence level of the positioning error can be more comprehensively characterized, avoiding the limitations of single deviation calculations that ignore environmental interference.

[0172] One alternative method is to calculate the positioning error using the following formula:

[0173] Where Pv represents the real-time pose of the vehicle. Given the vehicle's reference pose, the positioning error is obtained by calculating the square of the second-order norm of the difference vector between the two. The smaller the positioning error, the stronger the consistency between the real-time pose and the reference pose, and the more stable the current state of the positioning system.

[0174] Alternatively, the determination of topological relationship error can be achieved through the following method.

[0175] The topological relationship error of a target anchor point can be determined by its node relationships with other nodes in the semantic topological network graph and its geographic coordinates. Taking a lane intersection (the intersection of two lane line nodes) as an example, if the lane line nodes in the target anchor point have node relationships with other nodes, the topological relationship error can be calculated based on these relationships. For instance, if the lane line nodes of the target anchor point are connected to adjacent lane line nodes, the actual spatial distance between them is determined based on the geographic coordinates of the target anchor point's lane line nodes and the adjacent lane line nodes. It is then determined whether the actual spatial distance satisfies the constraint rules for connection (endpoint distance less than 0.3m and directional difference less than 10°), and the difference between the actual spatial distance and the value in the constraint rules is calculated as the topological relationship error.

[0176] S42. The total error is obtained by weighted summation of the positioning error and the topological relationship error.

[0177] The total error of the target anchor point is obtained by weighted summing of the positioning error and the topology error. Alternatively, the total error can be calculated as: Total Error = Positioning Error + 0.6 × Topology Error.

[0178] S43. Determine whether the total error is less than the preset error threshold. If yes, proceed to step S44, that is, map the target anchor point to the geographic coordinate system based on the geographic coordinates. If no, proceed to step S45, that is, optimize the geographic coordinates of the target anchor point based on the node relationship until the total error is less than the preset error threshold, and then map the target anchor point to the geographic coordinate system based on the optimized geographic coordinates of the target anchor point.

[0179] Specifically, if the total error is less than a preset threshold, the geographic coordinates of the target anchor point are output, and the target anchor point is mapped to the geographic coordinate system based on the geographic coordinates of the target anchor point; otherwise, the geographic coordinates of the target anchor point are optimized based on the node relationship until the total error is less than the preset error threshold, and then the target anchor point is mapped to the geographic coordinate system based on the optimized geographic coordinates of the target anchor point.

[0180] Step S44: Map the target anchor point to the geographic coordinate system based on its geographic coordinates.

[0181] If the total error is less than the preset threshold, the geographic coordinates of the target anchor point are output, and the target anchor point is mapped to the geographic coordinate system based on the geographic coordinates of the target anchor point.

[0182] Step S45: Optimize the geographic coordinates of the target anchor point based on the node relationship until the total error is less than the preset error threshold, and then map the target anchor point to the geographic coordinate system based on the optimized geographic coordinates of the target anchor point.

[0183] Specifically, the geographic coordinates of the target anchor point are optimized based on node relationship and topology relationship errors. For example, the geographic coordinates of the target anchor point are adjusted based on node relationships to reduce topology relationship errors. The real-time pose of the vehicle is then re-determined based on the optimized geographic coordinates of the target anchor point. The positioning error and topology relationship error are then re-determined based on the new real-time pose of the vehicle, and the total error is re-determined based on the new positioning error and topology relationship error. It is then determined whether the new total error is less than a preset error threshold. If so, the target anchor point is mapped to the geographic coordinate system based on the optimized geographic coordinates. If not, the geographic coordinates of the target anchor point are optimized until the total error is less than the preset error threshold, at which point the target anchor point is mapped to the geographic coordinate system.

[0184] This embodiment utilizes the vehicle's real-time pose information, including coordinates and heading angles obtained through GNSS and INS fusion, and combines this with the anchor point's position in the vehicle coordinate system to calculate the anchor point's latitude and longitude in the geographic coordinate system. To address potential errors, the geographic coordinates of the target anchor point are further calculated and optimized by combining topological relationship errors and positioning errors in the semantic topology network graph, ensuring they meet preset accuracy requirements. The total error is obtained by weighted summation of positioning and topological relationship errors, and the geographic coordinates of the target anchor point are adjusted accordingly. This not only reduces the impact of GNSS noise but also ensures high accuracy of the target anchor point's geographic coordinates through iterative calculation until the error meets the requirements. This provides real-time, accurate structured road environment information for autonomous driving, guaranteeing the safe and efficient operation of autonomous vehicles.

[0185] Furthermore, the completed high-precision map can be updated, and this part will be described in detail below.

[0186] It can track the status of high-precision maps in real time. When changes in the status of high-precision maps are detected, the type of change is determined. The types of changes are mainly divided into three categories: node changes, abnormal target anchor points, and abnormal vehicle positioning status.

[0187] Different update processes can be executed for different types of changes:

[0188] An optional approach is to use a partial update mechanism if the change type is a node change. The triggering condition for a node change can be a change in non-anchor elements. This change can include any one or more of the following: adding temporary obstacles, the disappearance of existing non-anchor elements, and changes in node attributes. Adding temporary obstacles could be road debris, temporary traffic cones, etc.; the disappearance of existing non-anchor elements could be the removal of construction warning signs; and changes in node attributes could be traffic lights changing from green to red, or variable lane indicator arrows changing direction. Since node changes only affect a local area of ​​the map and do not require adjustment of anchor point references, a partial update mechanism can be used. In the processing flow of the partial update mechanism, the node attributes of the nodes that have changed in the semantic topology network graph are updated first. It is then determined whether the node change affects the associated edges in the semantic topology network graph. If it does, the associated edges are adjusted; otherwise, the updated data is directly submitted. The entire process of the partial update mechanism is completed in milliseconds, which can quickly reflect real-time changes in the road environment while minimizing computational overhead and avoiding delays caused by full map updates, thus continuously maintaining the map's accurate representation of real-time road conditions.

[0189] Another optional approach is to use an anchor point rebinding mechanism if the change type involves an anomaly in the target anchor point. Therefore, when a target anchor point is detected as ineffective (e.g., a drop in confidence exceeding 25% or a positional shift exceeding 1.5 meters), or when a vehicle makes a sharp turn, the anchor point rebinding mechanism is triggered. The anchor point rebinding mechanism is a corrective mechanism for anomalies in target anchor points. As the core of the map's spatial reference, the stability of the target anchor point directly determines the map's accuracy. In the anchor point rebinding mechanism's processing flow, the current high-precision map state can be frozen first to prevent data corruption during the update process. Then, the optimal pre-activated target anchor point can be selected. To avoid map coordinate jumps caused by switching between old and new anchor points, a smooth transition formula is used to calculate the latest geographic coordinates of the new target anchor point (the optimal pre-activated target anchor point). Next, the positions of all associated nodes are updated based on the latest geographic coordinates to ensure consistency between the topological relationship and the geographic mapping. Finally, the high-precision map state is unfrozen, restoring normal use. One alternative approach involves a smooth transition formula: the geographic coordinates of the new target anchor point = the geographic coordinates of the old target anchor point × 0.3 + the geographic coordinates of the optimal pre-activated target anchor point × 0.7. This retains 30% of the historical data from the old target anchor point to maintain continuity, and incorporates 70% of the real-time data from the optimal pre-activated target anchor point to ensure accuracy. The entire process triggered by the anchor point rebinding mechanism can be completed within seconds, quickly correcting map deviations caused by anchor point anomalies and preventing erroneous decisions by autonomous driving due to map jumps through a smooth transition, thus ensuring driving stability.

[0190] Another optional approach is to use a global optimization mechanism if the change type is an anomaly in vehicle positioning status. This global optimization mechanism for anomalies in vehicle positioning status is a minute-level calibration mechanism to address long-term accumulated errors. It primarily targets map accuracy degradation issues during long-distance vehicle travel or in special scenarios. Triggering conditions can include a cumulative vehicle displacement exceeding 500 meters, GNSS signal loss exceeding 60 seconds, or the detection of a loopback. Loopback detection can occur when the vehicle travels to a previously visited location (e.g., in a parking lot scenario), and a location matching error of less than 3 meters is found through visual perception and comparison with historical data. The global optimization mechanism is a minute-level calibration mechanism to address long-term accumulated errors. It primarily targets map issues arising from long-distance vehicle travel or special scenarios. The global optimization process first loads historical topology data stored during vehicle travel, then establishes spatiotemporal constraints, comparing the current semantic topology network graph with historical data. It utilizes location matching relationships obtained through loop closure detection and road standards as constraints. A preset global graph optimization algorithm calibrates the geographic coordinates of all nodes, achieving loop closure correction of historical data. After optimization, it checks if convergence conditions are met. If not, it relaxes constraints and continues global graph optimization. If so, it automatically updates a high-precision map conforming to the ASAMOpenDRIVE 1.7 standard, incrementing the high-precision map version number according to the major version.minor version.revision number rule, for example, updating from v1.2.3 to v1.3.0. This facilitates identification of map data update status and provides support for subsequent data backtracking and version management.

[0191] Alternatively, the three update mechanisms mentioned above can operate independently or work together. Through update strategies at different levels and time scales, they address different problems, ultimately ensuring that high-precision maps maintain real-time performance, accuracy, and consistency in complex and ever-changing autonomous driving scenarios, providing a reliable navigation foundation for autonomous vehicles' path planning, obstacle avoidance, and traffic rule compliance functions.

[0192] The above describes a high-precision map construction method provided by the embodiments of this application. The following describes the apparatus for performing the above high-precision map construction method.

[0193] Please see Figure 7 , Figure 7 This is a schematic diagram of a high-precision map building device provided in an embodiment of this application. Figure 7 As shown, the high-precision map building device includes:

[0194] Image acquisition unit 11 is used to acquire image data, which represents the surrounding road environment of the vehicle;

[0195] The topology network graph determination unit 12 is used to determine a semantic topology network graph based on the image data. The semantic topology network graph includes nodes and associated edges. The nodes represent information units extracted from the road environment around the vehicle and carrying semantic and spatial features. The associated edges represent the node relationships between nodes. The node relationships represent the spatial and semantic dimension relationships between nodes.

[0196] The target anchor point determination unit 13 is used to determine the target anchor point based on the semantic topology network graph, wherein the target anchor point is a node that satisfies the function of a spatial reference.

[0197] The target anchor point mapping unit 14 is used to map the target anchor point to the geographic coordinate system based on the real-time pose of the vehicle.

[0198] The map generation unit 15 is used to determine the spatial position of each node in the semantic topology network graph in the geographic coordinate system based on the target anchor point in the geographic coordinate system and the node relationship, so as to obtain a high-precision map.

[0199] In one possible implementation, the process by which the topology network graph determination unit 12 determines the semantic topology network graph based on the image data includes:

[0200] The image data is input into a pre-trained image recognition model to obtain road elements corresponding to the image data output by the image recognition model. The road elements represent various entities related to driving in the road environment around the vehicle. The image recognition model is trained using different image data as training samples and the road elements corresponding to different image data as sample labels.

[0201] The road elements are classified based on their road element types, and nodes corresponding to each road element are created based on the road element types. The road element types represent the classification and definition of the functions of the road elements.

[0202] The node relationships are determined based on the node attributes of each node. The associated edges are added between nodes with node relationships. The semantic topology network graph is constructed based on each node and the associated edges. The node attributes include semantic attributes and spatial coordinates. The semantic attributes represent the traffic rules or function types carried by the node, and the spatial coordinates represent the position of the node in a vehicle coordinate system centered on the vehicle.

[0203] In one possible implementation, the nodes include lane line nodes, road marking nodes, traffic sign nodes, and drivable area nodes. The node relationships include connection relationships, constraint relationships, parallel relationships, and inclusion relationships. The connection relationship indicates that there is a connection between the lane line nodes. The constraint relationship indicates that the drivable area node is constrained by the rules of the traffic sign node. The parallel relationship indicates that there is a parallel relationship between the lane line nodes. The inclusion relationship indicates that the road marking node is included in the drivable area node.

[0204] The process by which the topology network graph determination unit 12 determines the node relationships based on the node attributes of each node includes:

[0205] Based on the node attributes of each node, it is determined whether the preset relationship matching rule is met. If the relationship matching rule is met, it is determined that there is a node relationship between the nodes. The relationship matching rule represents the judgment rule set for different node relationships. The judgment rule combines the node attributes of the node and compares them with a preset threshold to realize the judgment of the node relationship.

[0206] In one possible implementation, the node attributes further include detection confidence and consecutive tracking frame count. The detection confidence represents the credibility of the road element corresponding to the node in the recognition process of the image recognition model. The consecutive tracking frame count represents the number of times the node is continuously included in the image data and successfully matched with the node identifier. The node, candidate anchor point and target anchor point all have node attributes.

[0207] The process by which the target anchor point determination unit 13 determines the target anchor point based on the semantic topology network graph includes:

[0208] Based on the detection confidence and the number of consecutive tracking frames, lane intersections and / or traffic sign nodes are selected in the semantic topology network graph. The selected lane intersections and / or traffic sign nodes are used as candidate anchor points. The lane intersection is the intersection of at least two lane line nodes.

[0209] The candidate anchor points are input into a preset anchor point scoring model. The anchor point scoring model determines the node confidence score, topological centrality score, and location stability score of the candidate anchor points. The node confidence score, topological centrality score, and location stability score are then weighted and summed to obtain and output the candidate anchor point score corresponding to the candidate anchor point. The node confidence score represents the reliability of the road element corresponding to the anchor point in being identified in multiple consecutive frames of image data. The topological centrality score represents the degree of association between the anchor point and other nodes in the semantic topological network graph. The location stability score represents the degree of fluctuation of the spatial location of the anchor point in multiple consecutive frames of image data.

[0210] Candidate anchor points whose scores are greater than a preset value are selected as target anchor points.

[0211] In one possible implementation, the high-precision map building apparatus further includes:

[0212] The pre-activated target anchor point determination unit is used to select candidate anchor points whose scores are not greater than a preset value as backup anchor points, and select pre-activated target anchor points from the backup anchor points.

[0213] The target anchor point replacement unit is used to select the optimal pre-activated target anchor point based on the pre-activated target anchor point and replace the target anchor point in the failed state when the target anchor point is in a failed state. The failed state indicates that the detection confidence of the target anchor point is greater than a preset value or the position offset of the target anchor point is greater than a preset value.

[0214] In one possible implementation, the candidate anchor point, the backup anchor point, the pre-activated target anchor point, and the target anchor point all have node attributes and anchor point attributes. The anchor point attributes include anchor point type, spatial error value, effective duration, and historical offset record. The spatial error value represents the error value of the anchor point in the vehicle coordinate system. The effective duration represents the expected effective time of the anchor point. The historical offset record represents the positional deviation of the anchor point at different time points. The anchor point type includes lane intersections and traffic signs.

[0215] The process by which the pre-activated target anchor point determination unit selects a pre-activated target anchor point from the spare anchor points includes:

[0216] Select backup anchor points that meet the pre-screening criteria, wherein the pre-screening criteria are that the relative distance from the target anchor point is not less than the configured distance value, and the direction of the target anchor point relative to the vehicle is different from that of the vehicle.

[0217] The location stability score corresponding to the backup anchor point that meets the pre-screening conditions is determined based on the spatial error value, the effective duration, and the historical offset record. The spatial error value and the historical offset record are negatively correlated with the location stability score, and the effective duration is positively correlated with the location stability score.

[0218] The weight of the backup anchor point that meets the pre-screening conditions is determined based on the scenario in which the vehicle is located and the anchor point type of the backup anchor point that meets the pre-screening conditions, and the scenario adaptability score is determined based on the weight.

[0219] The topological correlation score is determined based on the correlation strength between the backup anchor points that meet the pre-screening conditions and the active nodes. The correlation strength is positively correlated with the topological correlation score. The active nodes include the target anchor points and nodes with a detection confidence score higher than a preset value.

[0220] The backup score corresponding to the backup anchor point that meets the pre-screening conditions is obtained by weighted summation of the location stability, the scene adaptability score and the topology correlation score.

[0221] The pre-activated target anchor point is determined based on the backup score.

[0222] In one possible implementation, the process by which the target anchor point mapping unit 14 maps the target anchor point to a geographic coordinate system based on the real-time pose of the vehicle includes:

[0223] The geographic coordinates of the target anchor point are determined based on the real-time pose of the vehicle and the spatial coordinates of the target anchor point. The spatial coordinates of the target anchor point represent the position of the target anchor point in the vehicle coordinate system centered on the vehicle, and the geographic coordinates are the position of the anchor point in the geographic coordinate system.

[0224] Based on the node relationships and the geographic coordinates, the positioning error and topology relationship error are determined. The positioning error and topology relationship error are weighted and summed to obtain the total error. The topology relationship error represents the spatial relationship deviation between the target anchor point and other nodes in the semantic topology network graph. The positioning error represents the measurement error of the vehicle's real-time pose.

[0225] If the total error is less than a preset error threshold, the target anchor point is mapped to the geographic coordinate system based on its geographic coordinates. If not, the geographic coordinates of the target anchor point are optimized based on the node relationship until the total error is less than the preset error threshold. Then, the target anchor point is mapped to the geographic coordinate system based on the optimized geographic coordinates.

[0226] In one possible implementation, the high-precision map building apparatus further includes:

[0227] A change type determination unit is used to determine the change type when a change in the state of the high-precision map is detected. The change type includes node change, target anchor point anomaly, and vehicle positioning anomaly. If the change type is node change, the node attributes of the node are updated, and it is determined whether the node change affects the associated edges in the semantic topology network graph. If so, the associated edges are adjusted. If the change type is target anchor point anomaly, the current high-precision map is frozen, the target anchor point is replaced, the position of each node in the geographic coordinate system is determined and updated based on the position of the latest target anchor point in the geographic coordinate system, and the current high-precision map is unfrozen. If the change type is vehicle positioning anomaly, the high-precision map is optimized and updated based on the semantic topology network graph to obtain an updated high-precision map.

[0228] In one possible implementation, the node attributes further include a node identifier and a number of consecutive tracking frames, wherein the number of consecutive tracking frames represents the number of times the same node is successfully identified and matched with the node identifier in multiple consecutive frames of image data, and the high-precision map construction device further includes:

[0229] A relationship strength value determination unit is used to determine the relationship strength value corresponding to the node relationship based on the geometric consistency score and the number of consecutive tracking frames. The geometric consistency score represents the deviation between the actual angle value and / or actual distance value between nodes and the configured value. The geometric consistency score, the number of consecutive tracking frames, and the relationship strength value are positively correlated. If the relationship strength value is less than a preset first strength value, the node relationship is removed. If the relationship strength value is not less than the first strength value and not greater than a preset second strength value, the node relationship is retained and updated to a weak relationship, where the second strength value is greater than the first strength value. If the relationship strength value is greater than the second strength value, the node relationship is retained.

[0230] In one possible implementation, the image recognition model has a tail-integrated graph structure semantic relation parsing layer (GNN layer). The GNN layer analyzes each road element output by the image recognition model to obtain the relationship probability between each road element. The relationship probability characterizes the association strength between the road elements.

[0231] This application also provides an electronic device in its embodiments. (See reference...) Figure 8 The diagram illustrates a structural schematic suitable for implementing the electronic device in the embodiments of this application. The electronic device in the embodiments of this application may include, but is not limited to, fixed terminals such as mobile phones, laptops, PDAs (personal digital assistants), PADs (tablet computers), desktop computers, etc. Figure 8 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0232] like Figure 8 As shown, the electronic device may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage device 608 into a random access memory (RAM) 603. When the electronic device is powered on, the RAM 603 also stores various programs and data required for the operation of the electronic device. The processing unit 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0233] Typically, the following devices can be connected to I / O interface 605: input devices 606 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 607 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 608 including, for example, memory cards, hard drives, etc.; and communication devices 609. Communication device 609 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 8 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or have alternatively.

[0234] It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the device embodiment drawings provided in this application, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.

[0235] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware, or it can be implemented by special-purpose hardware including application-specific integrated circuits, special-purpose CPUs, special-purpose memory, special-purpose components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for this application, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, training equipment, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0236] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0237] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, training device, or data center to another website, computer, training device, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

[0238] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

[0239] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

Claims

1. A high-definition map construction method, characterized by comprising: The method comprises: acquiring image data representing a surrounding road environment of a vehicle; determining a semantic topological network graph based on the image data, the semantic topological network graph comprising nodes and associated edges, the nodes representing information units carrying semantic and spatial features extracted from the surrounding road environment of the vehicle, the associated edges representing node relationships between the nodes, the node relationships representing spatial and semantic relationships between the nodes; determining a target anchor point based on the semantic topological network graph, the target anchor point being a node satisfying a spatial reference; mapping the target anchor point into a geographic coordinate system based on a real-time pose of the vehicle; determining spatial positions of the nodes in the geographic coordinate system based on the target anchor point in the geographic coordinate system and the node relationships, to obtain a high-definition map.

2. The method of claim 1, wherein, The process of determining the semantic topological network graph based on the image data comprises: inputting the image data into a pre-trained image recognition model to obtain road elements corresponding to the image data output by the image recognition model, the road elements representing various types of entity objects related to driving in the surrounding road environment of the vehicle, the image recognition model being trained with different image data as training samples and road elements corresponding to different image data as sample labels; performing a classification operation on each of the road elements based on a road element type, and creating a node corresponding to each of the road elements based on the road element type, the road element type representing a classification definition of a function of the road element; determining the node relationships based on node attributes of the nodes, adding the associated edges between nodes having node relationships, and constructing the semantic topological network graph based on each node and the associated edges, the node attributes including semantic attributes and spatial coordinates, wherein the semantic attributes represent traffic rules or function types carried by the nodes, and the spatial coordinates represent positions of the nodes in a vehicle coordinate system centered on the vehicle.

3. The method of claim 2, wherein, The nodes comprise lane line nodes, road surface marking nodes, traffic sign nodes, and drivable area nodes, the node relationships comprise connection relationships, constraint relationships, parallel relationships, and inclusion relationships, the connection relationships represent that the lane line nodes have connection relationships, the constraint relationships represent that the drivable area nodes are constrained by the traffic sign nodes, the parallel relationships represent that the lane line nodes have parallel relationships, and the inclusion relationships represent that the road surface marking nodes are included in the drivable area nodes; The process of determining the node relationships based on the node attributes of the nodes comprises: determining whether a preset relationship matching rule is satisfied based on the node attributes of the nodes, and determining that there is a node relationship between the nodes if the relationship matching rule is satisfied, the relationship matching rule representing a judgment rule set for different node relationships, the judgment rule being combined with the node attributes of the nodes and realized by comparison with a preset threshold to determine the node relationship.

4. The method of claim 3, wherein, The node attribute further includes a detection confidence and a continuous tracking frame number, the detection confidence representing a reliability of a road element corresponding to the node in an identification process of the image recognition model, and the continuous tracking frame number representing a number of times that the node is successfully included in image data and matched with a node identifier in succession, wherein the node, the candidate anchor point, and the target anchor point all have node attributes; The process of determining the target anchor point based on the semantic topological network graph includes: selecting a lane intersection and / or a traffic sign node in the semantic topological network graph based on the detection confidence and the continuous tracking frame number, and taking the selected lane intersection and / or traffic sign node as a candidate anchor point, the lane intersection being an intersection of at least two lane line nodes; inputting the candidate anchor point into a preset anchor point scoring model, determining a node confidence score, a topological centrality score, and a position stability score of the candidate anchor point through the anchor point scoring model, and performing weighted summation on the node confidence score, the topological centrality score, and the position stability score to obtain and output a candidate anchor point score corresponding to the candidate anchor point; wherein the node confidence score represents a score corresponding to a reliable degree of identification of a road element corresponding to the anchor point in continuous multiple frames of image data, the topological centrality score represents a score corresponding to an association degree of the anchor point with other nodes in the semantic topological network graph, and the position stability score represents a score corresponding to a fluctuation degree of a spatial position of the anchor point in continuous multiple frames of image data; taking a candidate anchor point with a candidate anchor point score greater than a preset value as the target anchor point.

5. The method of claim 4, wherein, The method further includes: taking a candidate anchor point with a candidate anchor point score not greater than a preset value as a backup anchor point, and selecting a pre-activated target anchor point from the backup anchor point; when the target anchor point is in a failure state, selecting an optimal pre-activated target anchor point based on the pre-activated target anchor point and taking the optimal pre-activated target anchor point as a new target anchor point to replace the target anchor point in the failure state, the failure state representing that the detection confidence of the target anchor point is greater than a preset value or a position offset of the target anchor point is greater than a preset value.

6. The method of claim 5, wherein, The candidate anchor point, the backup anchor point, the pre-activated target anchor point, and the target anchor point all have node attributes and anchor point attributes, the anchor point attributes including an anchor point type, a spatial error value, a valid survival time, and a historical offset record, the spatial error value representing an error value of the anchor point in the vehicle coordinate system, the valid survival time representing a predicted valid time of the anchor point, the historical offset record representing a position deviation of the anchor point at different time points, and the anchor point type including a lane intersection and a traffic sign; The process of selecting a pre-activated target anchor point from the backup anchor point includes: selecting a backup anchor point satisfying a pre-screening condition, the pre-screening condition being that a relative distance to the target anchor point is not less than a configured distance value and a direction relative to the vehicle is different from that of the target anchor point; and determine a position stability score corresponding to the backup anchor point meeting the pre-screening condition based on the spatial error value, the valid survival time and the historical offset record, the spatial error value, the historical offset record and the position stability score being negatively correlated, and the valid survival time and the position stability score being positively correlated; determine a weight of the backup anchor point meeting the pre-screening condition based on a scenario in which the vehicle is located and an anchor point type of the backup anchor point meeting the pre-screening condition, and determine a scenario adaptability score based on the weight; determine a topological correlation score based on an association strength of the backup anchor point meeting the pre-screening condition and an active node, the association strength and the topological correlation score being positively correlated, and the active node including a target anchor point and a node with a detection confidence higher than a preset value; weight and sum the position stability score, the scenario adaptability score and the topological correlation score to obtain a backup score corresponding to the backup anchor point meeting the pre-screening condition; determine the pre-activated target anchor point based on the backup score.

7. The method of claim 1, wherein, mapping the target anchor point into a geographic coordinate system based on a real-time pose of the vehicle, including: determining a geographic coordinate of the target anchor point based on the real-time pose of the vehicle and a spatial coordinate of the target anchor point, the spatial coordinate of the target anchor point representing a position of the target anchor point in a vehicle coordinate system with the vehicle as a center, and the geographic coordinate representing a position of the anchor point in the geographic coordinate system; determining a positioning error and a topological relationship error based on the node relationship and the geographic coordinate, weighting and summing the positioning error and the topological relationship error to obtain a total error, the topological relationship error representing a spatial relationship deviation of the target anchor point and other nodes in a semantic topological network graph, and the positioning error representing a measurement error of the real-time pose of the vehicle; determining whether the total error is less than a preset error threshold, and if so, mapping the target anchor point into the geographic coordinate system based on the geographic coordinate of the target anchor point, and if not, optimizing the geographic coordinate of the target anchor point based on the node relationship until the total error is less than the preset error threshold, and then mapping the target anchor point into the geographic coordinate system based on the optimized geographic coordinate of the target anchor point.

8. The method according to any one of claims 1-7, characterized in that, The method further includes: determining a change type when a change in the high-definition map is monitored, the change type including a node change, an abnormality of a target anchor point and an abnormality of a vehicle positioning state; if the change type is the node change, updating a node attribute of the node, and determining whether the node change affects an associated edge in the semantic topological network graph, and if so, adjusting the associated edge; if the change type is the abnormality of the target anchor point, freezing the current high-definition map, replacing the target anchor point, determining and updating positions of each node in the geographic coordinate system based on a position of the latest target anchor point in the geographic coordinate system, and un-freezing the current high-definition map; if the change type is the abnormality of the vehicle positioning state, optimizing and updating the high-definition map based on the semantic topological network graph to obtain an updated high-definition map.

9. The method of claim 2 or 3, wherein, The node attribute further includes a node identifier and a continuous tracking frame number, the continuous tracking frame number representing a number of times that the same node is continuously identified and matched with the node identifier in continuous multiple frame image data, and the method further includes: determining a relationship strength value corresponding to the node relationship based on a geometric consistency score and the continuous tracking frame number, the geometric consistency score representing a deviation of an actual angle value and / or an actual distance value between nodes from a configured value, wherein the geometric consistency score, the continuous tracking frame number, and the relationship strength value are positively correlated; if the relationship strength value is less than a preset first strength value, removing the node relationship; if the relationship strength value is not less than the first strength value and not greater than a preset second strength value, retaining the node relationship and updating it to a weak relationship, the second strength value being greater than the first strength value; if the relationship strength value is greater than the second strength value, retaining the node relationship.

10. The method of claim 2, wherein, The image recognition model includes a tail integrated graph structure semantic relationship analysis layer GNN layer, the GNN layer analyzes each road element output by the image recognition model to obtain a relationship probability between the road elements, the relationship probability representing an association strength between the road elements. 11.A high-definition map construction apparatus, comprising: The method includes: an image acquisition unit configured to acquire image data representing a surrounding road environment of a vehicle; a topological network graph determination unit configured to determine a semantic topological network graph based on the image data, the semantic topological network graph including nodes and associated edges, the nodes representing information units carrying semantic and spatial features extracted from the surrounding road environment of the vehicle, and the associated edges representing node relationships between the nodes, the node relationships representing relationships between the nodes in spatial and semantic dimensions; a target anchor point determination unit configured to determine a target anchor point based on the semantic topological network graph, the target anchor point being a node that meets a criterion for serving as a spatial reference; a target anchor point mapping unit configured to map the target anchor point to a geographic coordinate system based on a real-time pose of the vehicle; a map generation unit configured to determine spatial positions of each node in the semantic topological network graph in the geographic coordinate system based on the target anchor point in the geographic coordinate system and the node relationships, and to obtain a high-definition map.

Images