A vector map encoding method, system, electronic device and medium

Abstract

Embodiments of the present application provide a vector map encoding method and system, electronic equipment and medium. In the method, the map is encoded into a structured form containing the topological relationship of lane segments and target nodes, and is given geometric information and semantic information, thereby establishing a basis for multi-scale feature fusion. Subsequently, the first graph structure data is used to generate lane line global features. The construction of node features focuses on local scale detail mining, ensuring that the microscopic geometry and semantic features of the lane line are not missed. The construction of lane line global features can accurately capture the global topological relationship of the vector map. Finally, the node features of the local scale and the lane line global features of the global scale are fused and processed, and through cross-scale feature association, the features of each target node not only retain the geometric details of the lane segment, but also integrate the global topological semantics of the entire vector map, significantly improving the accuracy of trajectory prediction in the autonomous driving system.

Description

Technical Field

[0001] This application relates to the field of autonomous driving technology, and in particular to a vectorized map encoding method, system, electronic device, and medium. Background Technology

[0002] In autonomous driving systems, high-precision vectorized maps play a crucial role in the accuracy of trajectory prediction and planning during autonomous driving. The richer the map information and the more precise the detailed features used by the autonomous driving system, the more efficient and safer the autonomous driving control will be. However, the vectorized maps currently used in autonomous driving systems cannot link multi-scale features, resulting in vectorized maps being unable to comprehensively and accurately represent the environmental features around the vehicle and specific lane line features at the feature modeling level, thus affecting the accuracy of trajectory prediction in the autonomous driving system. Summary of the Invention

[0003] To address the aforementioned issues, and in order to enable vectorized maps to comprehensively and accurately represent the environmental features surrounding vehicles and specific lane line features at the feature modeling level, this application provides a vectorized map encoding method, system, electronic device, and medium.

[0004] The embodiments of this application disclose the following technical solutions:

[0005] In a first aspect, embodiments of this application provide a vectorized map encoding method applied to an autonomous driving system for vehicles, the method comprising:

[0006] Feature processing is performed on the vectorized map from the autonomous driving system to generate first graph structure data; the first graph structure data is used to characterize the geometric and semantic information of target nodes in each lane segment within the vectorized map;

[0007] Micro-features are constructed based on the geometric and semantic information of each target node in the first graph structure data to generate node features for each target node; the node features are used to characterize the lane structure features of the target node in the same lane segment and adjacent lane segments.

[0008] as well as,

[0009] Based on the first graph structure data, global lane line features are generated for the vectorized map; the global lane line features are used to characterize the global topological relationship formed by all lane segments in the vectorized map.

[0010] Based on the node features for each target node and the global features of the lane line, feature fusion processing is performed to generate second graph structure data; the second graph structure data includes the global topological semantics of the vectorized map and the geometric details of each lane segment.

[0011] In one possible implementation, the step of constructing micro-features based on the geometric and semantic information of each target node in the first graph structure data to generate node features for each target node includes:

[0012] Based on the geometric and semantic information of each target node, determine the first association information of each target node with other nodes in the current lane segment, and the second association information of each target node with other nodes in adjacent lane segments.

[0013] Wherein, the first association information is used to characterize the positional and semantic associations between the target node and other nodes within the same lane segment, and the second association information is used to characterize the contextual influence between the target node and other nodes within adjacent lane segments;

[0014] Based on the graph attention mechanism, local feature encoding is performed according to the first association information and the second association information of each target node to generate the node features for each target node.

[0015] In one possible implementation, generating global lane alignment features for the vectorized map based on the first graph structure data includes:

[0016] Based on the first map structure data, determine the geometric information of each lane segment within the vectorized map;

[0017] Based on the geometric information of all lane segments, a lane segment topology connection map is constructed for the vectorized map;

[0018] By using a self-attention mechanism and the lane segment topology connection graph, the interaction relationships between all lane segments in the vectorized map are constructed to generate the global features of the lane lines.

[0019] In one possible implementation, the feature fusion process, based on the node features for each target node and the global features of the lane line, to generate second graph structure data includes:

[0020] In response to a trajectory prediction command from the autonomous driving system, a target prediction region associated with the trajectory prediction command is determined;

[0021] Based on the vectorized map and the first graph structure data, calculate the regional complexity score of the target prediction area;

[0022] Based on the cross-scale attention mechanism, the region complexity score is used as the basis for the execution of the cross-scale attention mechanism to perform feature fusion processing on the node features of each target node and the global features of the lane line to obtain the second graph structure data.

[0023] In one possible implementation, calculating the regional complexity score of the target prediction region based on the vectorized map includes:

[0024] Based on the vectorized map and the first map structure data, determine the number of lane branches, the number of vehicles, the number of pedestrians, and the lane segment intersection topology type of the target prediction area;

[0025] The region complexity score is calculated based on the number of lane branches, the number of vehicles, the number of pedestrians, and the lane segment intersection topology type.

[0026] In one possible implementation, the feature processing of the vectorized map from the autonomous driving system to generate first graph structure data includes:

[0027] The vectorized map is topologically encoded to determine a vector set for the vectorized map; the vector set includes multiple lane segments, and each lane segment has at least one target node; the vector set is used to define the topological relationship between the multiple target nodes and the multiple lane segments.

[0028] Geometric and semantic information are assigned to each lane segment and each target node in the vector set to generate the first graph structure data.

[0029] Secondly, embodiments of this application provide a vectorized map encoding system applied to an autonomous driving system for vehicles, the system comprising:

[0030] The feature processing module is used to perform feature processing on the vectorized map from the autonomous driving system to generate first graph structure data; the first graph structure data is used to characterize the geometric and semantic information of target nodes in each lane segment within the vectorized map;

[0031] The feature construction module is used to construct micro-features based on the geometric and semantic information of each target node in the first graph structure data, so as to generate node features for each target node; the node features are used to characterize the lane structure features of the target node in the same lane segment and adjacent lane segments.

[0032] as well as,

[0033] Based on the first graph structure data, global lane line features are generated for the vectorized map; the global lane line features are used to characterize the global topological relationship formed by all lane segments in the vectorized map.

[0034] The feature fusion module is used to perform feature fusion processing based on the node features for each target node and the global features of the lane line to generate second graph structure data; the second graph structure data includes the global topological semantics of the vectorized map and the geometric details of each lane segment.

[0035] In one possible implementation, the feature construction module includes: a node feature construction unit, which is specifically used for:

[0036] Based on the geometric and semantic information of each target node, determine the first association information of each target node with other nodes in the current lane segment, and the second association information of each target node with other nodes in adjacent lane segments.

[0037] Wherein, the first association information is used to characterize the positional and semantic associations between the target node and other nodes within the same lane segment, and the second association information is used to characterize the contextual influence between the target node and other nodes within adjacent lane segments;

[0038] Based on the graph attention mechanism, local feature encoding is performed according to the first association information and the second association information of each target node to generate the node features for each target node.

[0039] Thirdly, embodiments of this application provide an electronic device, the device including: a processor, a memory, and a system bus;

[0040] The processor and the memory are connected via the system bus;

[0041] The memory is used to store one or more programs, the one or more programs including instructions that, when executed by the processor, cause the processor to perform any of the possible vectorized map encoding methods in the first aspect.

[0042] Fourthly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, characterized in that, when the program is executed by a processor, it implements any possible vectorized map encoding method in the first aspect.

[0043] Compared to existing technologies, this application offers the following advantages: This application provides a vectorized map encoding method, system, electronic device, and medium. In this method, not only is the map encoded into a structured form containing the topological relationships between lane segments and target nodes, but more importantly, each target node is endowed with geometric and semantic information, thus laying the foundation for multi-scale feature fusion. Subsequently, node features of each target node and global lane line features of the entire vectorized map are generated using the first map structure data. The construction of node features focuses on detailed mining at the local scale, ensuring that the microscopic geometric and semantic features of lane lines are not overlooked. The construction of global lane line features can accurately capture the global topological relationships of the vectorized map, such as lane intersection types at intersections and the merging and diverging structures of multi-lane roads. Finally, the node features at the local scale are fused with the global lane line features at the global scale. Through cross-scale feature association, the features of each target node retain the geometric details of its lane segment while incorporating the global topological semantics of the entire vectorized map. This achieves effective linkage of multi-scale features, thereby comprehensively and accurately representing the environmental features and lane line features around the vehicle. This provides a more reliable basis for the core functions of autonomous driving systems, such as trajectory prediction and path planning, and significantly improves the accuracy of trajectory prediction in autonomous driving systems. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 A flowchart illustrating a vectorized map encoding method provided in this application embodiment;

[0046] Figure 2 A flowchart illustrating a method for generating global lane lines according to an embodiment of this application;

[0047] Figure 3 A flowchart illustrating a method for generating second graph structure data according to an embodiment of this application;

[0048] Figure 4 This is a schematic diagram of the structure of a vectorized map encoding system provided in an embodiment of this application;

[0049] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with specific embodiments and accompanying drawings. It should be particularly noted that the embodiments described in this application are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0051] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0052] As described earlier, the vectorized maps currently used in autonomous driving systems cannot integrate multi-scale features. This results in vectorized maps being unable to comprehensively and accurately represent the environmental features surrounding the vehicle and specific lane line features at the feature modeling level, thus affecting the accuracy of trajectory prediction in autonomous driving systems.

[0053] Based on this, embodiments of this application provide a vectorized map encoding method, system, electronic device, and medium. In this method, not only is the map encoded into a structured form containing the topological relationships between lane segments and target nodes, but more importantly, each target node is endowed with geometric and semantic information, thereby laying the foundation for multi-scale feature fusion. Subsequently, node features of each target node and global lane line features of the entire vectorized map are generated through the first map structure data. The construction of node features focuses on the detailed mining at the local scale, ensuring that the micro-geometric and semantic features of lane lines are not overlooked; while the construction of global lane line features can accurately capture the global topological relationships of the vectorized map, such as the lane intersection types at intersections and the merging and diverging structures of multi-lane roads. Finally, the node features at the local scale are fused with the global lane line features at the global scale. Through cross-scale feature association, the features of each target node retain the geometric details of its lane segment while incorporating the global topological semantics of the entire vectorized map. This achieves effective linkage of multi-scale features, thereby comprehensively and accurately representing the environmental features and lane line features around the vehicle. This provides a more reliable basis for the core functions of autonomous driving systems, such as trajectory prediction and path planning, and significantly improves the accuracy of trajectory prediction in autonomous driving systems.

[0054] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0055] See Figure 1 The figure is a flowchart illustrating a vectorized map encoding method provided in an embodiment of this application, specifically including the following steps:

[0056] S101: Perform feature processing on the vectorized map from the autonomous driving system to generate first graph structure data; the first graph structure data is used to characterize the geometric and semantic information of target nodes in each lane segment within the vectorized map.

[0057] In autonomous driving systems, the step of generating first-order graph structure data through feature processing of vectorized maps is a crucial starting point for transforming raw map information into machine-understandable structured data. The core of this step lies in constructing a graph structure framework (i.e., first-order graph structure data) that accurately represents the spatial relationships and attribute characteristics of map elements through topological encoding and information assignment. Each lane segment, as a basic unit, is further discretized into target nodes with actual geographical significance. These nodes are typically located at key positions within the lane, such as start points, end points, curvature change points, or lane function encoding points. They form ordered topological relationships through edge connections, defining both the sequential order of nodes within the same lane segment and clarifying the spatial connections between adjacent lane segments (such as bifurcations, merging, and intersections), thus building the underlying skeleton of the road network at the structural level. This topological encoding is not a simple geometric division but an organic decomposition based on traffic function and spatial logic, transforming the physical form of the map into a computable graph structure model, which is then presented in the form of first-order graph structure data.

[0058] Specifically, the generation of the first graph structure data in this step requires the following two steps:

[0059] Step 1: Perform topological structure encoding on the vectorized map to determine the vector set for the vectorized map; the vector set includes multiple lane segments, and each lane segment has at least one target node, and the vector set is used to define the topological relationship between the multiple target nodes and the multiple lane segments.

[0060] In the processing flow of vectorized maps, topology encoding is the process of transforming the physical road network into a machine-computable graph structure. This process begins with a deep deconstruction of the original vectorized map. It does not simply replicate map pixels or coordinate data, but rather divides the continuous road surface into "lane segments" with independent traffic significance based on the functional attributes and spatial logic of the roads. The division of these lane segments follows a dual criterion: considering both physical morphology, such as changes in curvature and abrupt changes in lane width, and traffic function, such as the boundary between main roads and ramps, and the start and end points of dedicated turning lanes. For example, a road segment containing both straight and right-turn forks will be divided into a straight main road lane segment and a right-turn dedicated lane segment, connected at the fork node. Each lane segment is further discretized into several target nodes, which are typically located at key positions within the lane segment, such as extreme points of curvature radius, directional turning points, lane function encoding boundaries, or traffic element connection points. Through this segmentation, the originally continuous road is transformed into a set of triplets of "node-lane segment-connection relationship", that is, a "vector set". In essence, it uses nodes and edges in graph theory to build the underlying skeleton of the road network.

[0061] In vector sets, the definition of topological relationships encompasses both physical connections in spatial locations and implicit logical rules governing traffic flow. Target nodes within the same lane segment are connected sequentially according to their direction of travel, forming a linear sequence representing lane orientation. The edges between target nodes not only record geometric distances but also implicitly represent the continuity of driving trajectories. Target nodes within different lane segments form cross-lane connections through spatial or functional associations, such as the connection between main road lane segments and merging ramp lane segments at the merging point, or the convergence of lane segments in all directions at the intersection center node at an intersection. This topological relationship implicitly encodes traffic rules; for example, the directionality of node connections defines legal driving paths (nodes prohibiting reverse traffic will not form reverse connection edges), and the multi-directional connections of branching nodes identify possible branches in driving decisions. Through this encoding, vectorized maps transform from a visual graphical perspective into a machine-understandable logical graph structure. This structured representation lays the foundation for subsequent feature extraction and semantic analysis. When it is necessary to analyze the traffic logic of complex intersections or predict vehicle lane-changing trajectories, calculations can be directly performed based on the topological relationships of the vector set, thereby significantly improving processing efficiency and accuracy.

[0062] Step 2: Assign geometric and semantic information to each lane segment and each target node in the vector set to generate the first graph structure data.

[0063] After constructing the topology of the vector set, each lane segment and target node needs to be assigned geometric and semantic information, transforming it from a simple logical node into an intelligent entity carrying rich features. The assignment of geometric information focuses on the precise quantification of spatial attributes. For a target node, its three-dimensional coordinates (x, y, z), the Euclidean distance between adjacent nodes, the node's relative position within the lane segment (e.g., normalized coordinates with a start point of 0 and an end point of 1), and the tangent direction angle at the node's location need to be determined. For a lane segment, parameters such as its overall length, radius of curvature, azimuth angle of straight segments, and cross-sectional width are calculated.

[0064] The assignment of semantic information involves transforming traffic rules, road functions, and traffic logic into a machine-recognizable tagging system. For example, target nodes are labeled with functional attributes, such as merging nodes, forking nodes, intersection center nodes, etc., or with traffic permissions, such as nodes prohibiting U-turns, priority passage nodes, etc., as well as associated traffic signs, such as nodes corresponding to speed limit signs being labeled with speed limit values; lane segments are labeled with specific lane types, such as main road lanes, ramp lanes, emergency lanes, etc., and can also be labeled with lane direction attributes, such as one-way two-lane, two-way four-lane, etc., or lane functional restrictions, such as no-lane-changing zones, bus lane time restrictions, and other semantic tags.

[0065] The purpose of injecting geometric and semantic information is to transform traffic engineering specifications and driving rules into machine-executable logical instructions. For example, a lane segment marked "Dedicated Right Turn Lane" restricts the direction of travel for vehicles within that segment, while a lane segment marked "Construction Zone" triggers the detour strategy of the autonomous driving system. Through the dual injection of geometric and semantic information, each target node becomes a rule carrier with spatial coordinates, while each lane segment becomes a spatial unit with functional attributes. The first graph structure data, jointly formed by both, preserves the precise spatial morphology of the road (such as the curvature of curves and the direction of straight sections) and carries the semantic connotations of the traffic environment (such as which lanes allow lane changes and which nodes require slowing down and yielding). This transforms the vectorized map from a static geographic layer into a dynamic and computable intelligent data volume, providing a data source with both accuracy and depth for subsequent micro-node feature encoding and macro-lane segment relationship modeling.

[0066] S1021: Micro-features are constructed based on the geometric and semantic information of each target node in the first graph structure data to generate node features for each target node; the node features are used to characterize the lane structure features of the target node in the same lane segment and adjacent lane segments.

[0067] Steps S1021 and S1022 are two parallel execution processes. The former generates node features for each target node based on the geometric and semantic information in the target node. These node features characterize the lane structure features of the target node within its own lane segment and with adjacent lane segments. The latter generates lane line distribution (i.e., global lane line features) in the vectorized map based on the geometric and distribution information of each lane segment in the first map structure data. Next, we will first introduce the processing content of step S1021.

[0068] Regarding the execution content of step S1021, this step takes the target node as the core and generates node features that can characterize its role in the lane structure by fusing geometric information and semantic information. From a geometric perspective, not only are basic spatial data such as the three-dimensional coordinates and elevation of the target node extracted, but also its relative position within the lane segment (e.g., merging, distance between adjacent nodes, tangent direction angle, and curvature parameters). These parameters accurately characterize the morphological features of the lane where the target node is located, such as the degree of curvature and changes in direction, providing spatial scale basis for speed planning and trajectory control in autonomous driving. In the semantic dimension, the node's functional attributes (e.g., merging, branching, intersection nodes), traffic permissions (labels such as no U-turn, priority passage, etc.), and associated traffic rules (speed limits, lane function restrictions) are transformed into logical features. These features are then combined with the node's connection relationships with upstream and downstream lane segments, such as the number of connections, direction, and differences in adjacent lane types, forming structural features that reflect the node's role as a "functional hub" in the road network. For example, merging nodes need to integrate the traffic priorities of the main road and ramps, while branching nodes need to identify the path branch logic. The final generated node features are an organic unity of geometric and semantic information; that is, geometric data defines the spatial form, and semantic labels inject rule connotations, together constructing a composite feature vector containing "position-attribute-relationship".

[0069] Specifically, step S1021 is achieved through the following two steps:

[0070] Step 1: Based on the geometric and semantic information of each target node, determine the first association information of each target node with other nodes in the current lane segment, and the second association information of each target node with other nodes in adjacent lane segments.

[0071] Wherein, the first association information is used to characterize the positional and semantic associations between the target node and other nodes within the same lane segment, and the second association information is used to characterize the contextual influence between the target node and other nodes within adjacent lane segments.

[0072] Determining the association information of a target node within its own lane segment and adjacent lane segments is the core of constructing a dynamic association model of the road network. This process is based on the geometric and semantic information of the nodes, mining the positional dependencies and logical constraints between nodes from two dimensions: within the same lane segment and between adjacent lane segments.

[0073] Specifically, the first association information is used to characterize the positional and semantic associations between the target node and other nodes within the same lane segment. It focuses on the vertical chain relationship of the target node within the lane segment, weaving the nodes within a single lane segment into a continuous feature network through dual modeling of geometric positional association and semantic logical association. In this process, geometric parameters need to be extracted based on the arrangement order of nodes within the same lane segment, such as the Euclidean distance between adjacent nodes, the cumulative travel distance of a node within the lane segment, the normalized position ratio, and the rate of change of the tangent direction between adjacent nodes. These data can present the spatial trajectory of the lane direction. For example, if the distance between a node and its preceding node suddenly decreases and the curvature increases sharply, its positional association will indicate a "sharp curve warning." Semantic association revolves around the functional consistency within the lane segment. Nodes continuously labeled "straight main road" form a semantic chain of "segment without lane change requirements," while the preceding node of a "branching node" is automatically associated with a "path branch prediction" label, and subsequent nodes form multi-path semantic extensions according to the branching direction, transforming the nodes within the same lane segment from isolated coordinate points into a logical sequence carrying driving rules.

[0074] The determination of the second set of associated information breaks down the boundaries of a single lane segment, focusing on the "contextual influence" between the target node and adjacent lane segment nodes. This influence is primarily reflected in spatial proximity. That is, by calculating parameters such as the horizontal distance and vertical height difference between the node and adjacent lane segment nodes, it is determined whether a valid association is formed (e.g., association calculation is triggered when the distance is less than the safe lane-changing distance). More importantly, there is the cross-lane segment mapping of semantic rules. For example, if the current node is the end node of the "no lane-changing" section of the main road, its adjacent ramp entrance node will form a "merging allowed" semantic association. Conversely, if the main road node is marked "construction closed", the adjacent lane node will automatically be associated with the "forced detour" rule.

[0075] Two types of related information jointly construct the vertical logical chain of the target node within its own lane segment and the horizontal semantic network across lane segments, making each target node an information hub that integrates local structural features and global interaction rules. The first type of related information ensures the consistency of driving strategies within a lane segment, while the second type of related information empowers the autonomous driving system to predict cross-lane interaction risks. Together, they provide hierarchical related feature inputs for subsequent graph neural network modeling, driving the autonomous driving system to upgrade from single-node analysis to a global understanding of the road network, thereby improving the accuracy of trajectory prediction in the autonomous driving system.

[0076] Step 2: Based on the graph attention mechanism, local feature encoding is performed according to the first association information and the second association information of each target node to generate the node features for each target node.

[0077] The core of the graph attention mechanism is to enable each target node to accurately select the most critical neighbor nodes from the first and second sets of related information, and assign weights based on decision relevance. For the first set of related information, such as the target node being the starting point of a merging zone on the main road, preceding curve nodes in the current lane and semantically continuous "merging warning" nodes (nodes used to indicate rule changes) will be given high weights due to their strong logical connection with the current node. Meanwhile, distant straight-line nodes have very low weights because they have no substantial impact on the current decision. For the second set of related information (nodes in adjacent lanes), such as merging nodes from adjacent ramps, if their spatial distance is within the safe lane-changing range and their semantic rules match the current lane, their weights will be significantly higher than those of unrelated nodes in the opposite lane. For example, the smaller the angle between the merging direction of a merging node and the current driving direction, the greater its impact on lane changing, and its weight will automatically increase. This weight allocation is not a fixed rule, but rather the mechanism automatically adjusts based on the decision value of the related information, ultimately condensing the originally scattered related nodes into the set of neighbors most important to the current node.

[0078] Subsequently, the graph attention mechanism first performs a weighted sum of the geometric, semantic, and association rule features of key neighbors (highly correlated nodes in the current lane and adjacent lanes), and then deeply integrates this sum with the node's own original features (such as coordinates and lane type). In this process, the weight of neighbors is entirely based on their influence on the current node's decision. For example, preceding curve nodes in the current lane (directly related to speed planning) and merging nodes in adjacent lanes (related to merging rules) will be given higher weights due to their strong correlation with the current node, while irrelevant or weakly correlated nodes (such as distant straight sections or opposite median strip nodes) are naturally weakened.

[0079] The fused node features have completely moved beyond isolated geometric coordinates or semantic labels, becoming local features that integrate their core attributes with the surrounding key environment. They retain their essential characteristics while also carrying the continuous logic of their own lane and the contextual influence of adjacent lanes. For autonomous driving systems, such features are a direct basis for decision-making. For example, the features of merging zone nodes automatically integrate constraints such as "slow down before the curve" and "merging is allowed on the right," allowing the system to respond quickly without additional calculations. Meanwhile, the features of intersection nodes naturally include rules such as "slow down in advance" and "priority for adjacent left turns," avoiding decision lag.

[0080] S1022: Based on the first graph structure data, generate global lane line features for the vectorized map; the global lane line features are used to characterize the global topological relationship formed by all lane segments in the vectorized map.

[0081] The process of generating global lane route features based on the first map structure data essentially transforms the entire vectorized map's lane segment network into a feature set that represents the global topological logic from a macroscopic perspective. The core is to move beyond the microscopic details of target nodes and instead build the topological relationships of the entire map using lane segments as basic units. Specifically, the global lane route features are a holistic representation from the macroscopic perspective of the vectorized map, using each lane segment as a node within a unit, integrating the topological relationships and interaction logic of all lane segments in the entire map. Essentially, it uses the "global identity of lane segments" to restore the overall structural logic of the road network. Next, with reference to the accompanying drawings of a specific embodiment, the method for generating global lane route features in step S1022 will be described.

[0082] See Figure 2 The figure is a flowchart illustrating a method for generating global lane features according to an embodiment of this application, specifically including the following steps:

[0083] S1023: Based on the first map structure data, determine the geometric information of each lane segment within the vectorized map.

[0084] The process of determining the geometric information of each lane segment based on the structural data in the first diagram essentially involves aggregating the microscopic spatial attributes of target nodes into the macroscopic geometric features of the lane segment. The aim is to string together scattered node geometric information into a complete spatial form of the lane segment. In the structural data of the first diagram, each lane segment corresponds to a set of target nodes arranged according to the driving direction, such as the starting point, curvature abrupt change point, and ending point. Each node carries information such as three-dimensional coordinates, distances to adjacent nodes, and tangent direction angles. For example, if the node sequence of a straight lane segment is N1(x1,y1,z1) to Nk(xk,yk,zk), the total length is calculated first: the Euclidean distances between adjacent nodes are accumulated to obtain the actual physical length of the lane. Then, the azimuth angle is calculated: using the coordinate difference between the starting point and the ending point, the angle between this line segment and the reference direction is calculated to determine the overall direction of the lane. If it is a curved segment, the node coordinates are used to fit an arc to calculate the radius of curvature. Even the trend of changes in the tangent direction angles of adjacent nodes can be used to determine whether it is a left or right turn. The slope is calculated by dividing the difference in z-coordinates between the starting and ending points by the total length to obtain the road's inclination. The width is taken as the average of the lane cross-sectional widths at all nodes. Because the target nodes in a lane segment are already assigned lane width information at their location, these calculations are not arbitrarily superimposed, but strictly follow the topological order and spatial relationships of the nodes, ultimately giving each lane segment clear attribute information, such as its specific length, orientation, whether it is straight or curved, curvature data, slope data, width data, etc. This information is both the foundation for subsequently constructing the lane segment topology connection graph and the key to generating global features.

[0085] S1024: Based on the geometric information of all the lane segments, construct a lane segment topology connection map for the vectorized map.

[0086] A lane segment topology connection map is constructed based on the geometric information of all lane segments. The purpose is to anchor their traffic logic relationships using the spatial attributes of each lane segment (start / end coordinates, direction, and location range), and then transform these relationships into a graph structure. The geometric information of each lane segment contains clear "spatial clues"—such as the three-dimensional coordinates of the start and end points, and the azimuth angle of the overall direction. These are key to determining connections: for example, if the end coordinates of lane segment A and the start coordinates of lane segment B almost coincide, and the directions of A and B are continuous, they form a "forward connection"; if the coordinate ranges of two lane segments intersect at the center of an intersection (such as eastbound and northbound straight lanes at a crossroads), and the intersection angle conforms to traffic rules, it is marked as an "intersection connection"; if two lane segments are parallel and adjacent (such as lane 1 and lane 2 in the same direction, with the center distance equal to the sum of the lane widths), it is an "adjacent connection." This process is not just mechanical coordinate matching; it also incorporates traffic logic. For example, merging ramps and main roads must ensure that the direction is merging rather than going against traffic; for branching main roads and side roads, it must be confirmed that the lanes after the branching are extensions of the main road. Ultimately, all lane segments become "nodes" in the topology graph, and their connections (merging, branching, crossing, and adjacency) become "edges," weaving together a lane segment topology connection graph that covers the entire vectorized map. For example, the four-way lanes at an intersection converge through crossing edges, the main road and the ramp are connected through merging edges, and lanes in the same direction form lane groups through adjacent edges, connecting isolated lane segments into an organic global network, laying the foundation for subsequent capture of lane interactions across the entire map.

[0087] S1025: Construct the interaction relationship between all lane segments in the vectorized map through the self-attention mechanism and the lane segment topology connection graph to generate the global features of the lane line.

[0088] The purpose of this step is to allow each lane segment to transcend its local perspective and integrate information from all related lane segments across the entire graph from a global viewpoint. First, the lane segment topology graph has woven all lane segments into a "relationship network," with each lane segment acting as a node in this network. Merging, branching, and intersecting traffic logic connections are the lines connecting these nodes. Next, metaphorically speaking, the self-attention mechanism is like equipping each node with a "smart radar." Each lane segment first uses its basic features (geometric and semantic information) as a query, then calculates the "correlation degree" with the features of all other lane segments in the entire graph. For example, merging ramps and main roads, intersecting straight lanes and right-turn lanes—lane segments with inherent physical or logical connections—will have a higher feature matching degree and correspondingly higher attention weights. Conversely, lane segments that are far apart and functionally unrelated (such as suburban side roads and city center main roads) will have very low weights. Simultaneously, the "lines" in the topology graph (such as two lane segments merging) serve as "prior knowledge," further reinforcing the weights of these key connections—preventing self-attention from "missing" those physically directly connected important partners.

[0089] Then, each lane segment uses these weights to sum the features of all other lane segments in the entire map, essentially fusing information from all related lane segments into its own features. For example, a right-turn lane segment at an intersection, whose original feature was simply "dedicated right-turn lane," now incorporates information about intersecting straight lanes. This means it intersects with intersecting straight lanes, and vehicles turning right must yield to straight-going and adjacent left-turn lanes ("it is alongside them, and lane changes must consider left-turning traffic"). Ultimately, each lane segment's features become a fusion of its own attributes and those of other lane segments in the entire map. All these features combined constitute the "global lane line features"—which not only record the geometric and semantic information of each lane segment itself but also contain its logical connections with other lane segments in the entire map, i.e., which lane segments it is upstream or downstream connected to, which are adjacent to, and which will affect traffic flow. These features tell the autonomous driving system what role this lane plays in the entire road network and how it interacts with other lanes. These logics can all be covered by the global features of the lane line, thus laying a data foundation for the precise control of the subsequent autonomous driving system.

[0090] The above describes the node features of the target node and the generation method and feature content of the global lane line features of the vectorized map in the embodiments of this application. Next, based on the... Figure 1 Step S103 will be described in detail.

[0091] S103: Based on the node features for each target node and the global features of the lane line, feature fusion processing is performed to generate second graph structure data; the second graph structure data includes the global topological semantics of the vectorized map and the geometric details of each lane segment.

[0092] Finally, the node features of the target nodes (detailed attributes such as 3D coordinates, curvature abrupt changes, and lane width of each node) are fused with the global features of the lane lines to generate a second graph structure data that simultaneously covers global topological semantics and lane segment geometric details. Specifically, the node features of each target node are anchored to the global context of its lane segment, and the detailed attributes of the node are imbued with the global identity of its lane segment (e.g., whether the lane segment is a merging section connecting the main road and the ramp, and its position in the overall graph topology), giving the originally isolated node data a "position in the global network." At the same time, the global features of each lane segment absorb the node features of all the target nodes it contains, making the macro-topological relationship of the lane segment concrete in terms of the coordinates of specific nodes (e.g., the 3D position of a node corresponding to a merging point), and the geometric attributes of the lane segment are sequentially pieced together from the details of the nodes.

[0093] This fusion allows the second map structure data to possess two core pieces of information simultaneously: firstly, global topological semantics, namely the overall logic of lane segment connections and functional distinctions across the entire map. These semantics are no longer abstract descriptions but rather concrete relationships bound to specific node locations. Secondly, the geometric details of each lane segment include micro-attributes such as specific curvature abrupt changes, width variations, and the coordinates of key nodes. These details are no longer scattered point data but structured information incorporated into the overall map topological framework. Ultimately, the second map structure data achieves a unification of global context and local details. Global topological semantics provide the overall logic of the entire map, while geometric details support precise descriptions of the micro-attributes of lane segments. Together, they constitute a complete and detailed vectorized map structure.

[0094] Next, the method for generating the structure data of the second diagram in step S103 will be described with reference to the specific implementation drawings. See [link to relevant documentation]. Figure 3 The figure is a flowchart illustrating a method for generating second graph structure data according to an embodiment of this application, specifically including the following steps:

[0095] S1031: In response to a trajectory prediction command from the autonomous driving system, determine a target prediction area associated with the trajectory prediction command.

[0096] When an autonomous driving system issues a trajectory prediction command, it first needs to define the target prediction area directly related to the command based on the vehicle's real-time status (including precise 3D position, driving direction, current lane ID, and intent cues such as turn signal activation). In practice, the system uses the vehicle's current position as a reference point and determines the immediate impact range based on the prediction time scale. For example, if predicting the trajectory within 3 seconds, the range extends approximately 50 meters forward, covering 1-2 adjacent lanes on the left and right. Then, combining the topological relationships of the vectorized map, it supplements areas with critical influence on the trajectory. If there is a fork between a main road and a ramp ahead, the starting section of the ramp must be included; if it's before an intersection, it must include straight lanes, left-turn or right-turn lanes in the intersecting directions. The road structure in these areas directly affects the vehicle's possible future trajectory, therefore, they must be included in the target prediction area to ensure that subsequent predictions cover all necessary factors, while excluding irrelevant suburban roads or unrelated lanes to avoid computational redundancy.

[0097] S1032: Based on the vectorized map and the first graph structure data, calculate the regional complexity score of the target prediction area.

[0098] Next, based on the geometric details of the vectorized map, such as the number of lanes, width variations, and curve curvature, and the topological semantic information of the first map's structural data, such as merging / bifaction relationships between lanes and functional classifications, a comprehensive calculation of the complexity score for the target prediction area is required. The scoring process necessitates weighted integration of various influencing factors: for example, the more lanes, the denser the merging and bifurcation points, and the more frequent the curve curvature changes within the area, the higher the final area complexity score, indicating a greater challenge to trajectory prediction posed by the road structure in that area.

[0099] Specifically, the calculation of the region complexity score is achieved through the following two steps:

[0100] Step 1: Based on the vectorized map and the first map structure data, determine the number of lane branches, the number of vehicles, the number of pedestrians, and the lane segment intersection topology type of the target prediction area;

[0101] Step 2: Calculate the region complexity score based on the number of lane branches, the number of vehicles, the number of pedestrians, and the lane segment intersection topology type.

[0102] When calculating the regional complexity score of a target prediction area based on a vectorized map, it is first necessary to combine the static structural information of the vectorized map with the topological semantics and dynamic correlation data of the first map structure data to extract and integrate the core elements affecting regional complexity. Specifically, by using the coordinates of lane connection points and bifurcation points in the vectorized map and the semantic annotation of lane topological relationships in the first map structure data, the number of lane branches in the target prediction area is counted, that is, the total number of all lane merging or bifurcation points in the area, and the number of lanes involved in each branch point. This indicator directly reflects the degree of bifurcation of the road structure. The more branches, the richer the possible future trajectories of vehicles, and the more complex the scenarios that need to be covered in the prediction.

[0103] Next, based on the geometric layout of lane intersections in the vectorized map and the definition of intersection types (such as cross intersections, T-junctions, multi-lane merging, or continuous bifurcation) in the first map structure data, the lane segment intersection topology type is determined. Different types correspond to different traffic flow interaction patterns (for example, cross intersections have traffic conflict points in four directions, while T-junctions only have three), and therefore are assigned different complexity weights. Simultaneously, relying on real-time perception information integrated in the first map structure data, the current number of vehicles and pedestrians in the target area is statistically analyzed. These two indicators reflect the real-time traffic congestion level in the area; a higher number means that vehicles need to consider obstacle avoidance scenarios and interaction behaviors more frequently when predicting their trajectories.

[0104] Subsequently, the four extracted elements—number of lane branches, number of vehicles, number of pedestrians, and lane segment intersection topology type—are weighted and integrated according to preset weights. For example, a base score is first assigned to each intersection topology type (e.g., 5 points for a cross intersection and 3 points for a T intersection). Then, the number of lane branches is multiplied by the corresponding coefficient (e.g., 1 point for each additional branch), and the number of vehicles and pedestrians are multiplied by their respective coefficients (e.g., 0.5 points for each additional vehicle and 0.3 points for each additional pedestrian). Finally, all scores are summed to obtain the area complexity score.

[0105] The core of introducing regional complexity scoring is to provide dynamic weight guidance for cross-scale fusion of micro-node-level and macro-lane segment-level features. It shifts the fusion from a fixed ratio to on-demand adaptation. In complex scenarios, it strengthens the connection between macro-lane topology and micro-nodes, ensuring features retain both local lane line geometric details and global road structure semantics. In simple scenarios, it weakens the injection of macro-features, focusing on the basic geometric information of micro-nodes to reduce redundancy. This adaptive fusion enables vectorized map encoding in complex scenarios to accurately capture multi-lane interaction constraints, improving the accuracy of trajectory prediction in the Houxi autonomous driving system. It also maintains efficient inference in simple scenarios, ultimately achieving a precise match between the fusion effect and scenario requirements, supporting a balance between robustness and efficiency in autonomous driving control.

[0106] S1033: Based on the cross-scale attention mechanism, the region complexity score is used as the execution basis of the cross-scale attention mechanism to perform feature fusion processing on the node features of each target node and the global features of the lane line to obtain the second graph structure data.

[0107] The core of feature fusion based on a cross-scale attention mechanism is to use the region complexity score as a dynamic adjustment signal to guide the integration of node features with global lane features. When processing each target node, the cross-scale attention mechanism first calculates the correlation between the node's features and the global lane features. However, the weighting of this correlation is not fixed; the region complexity score directly affects the direction of the weighting. For example, if the complexity score of the target prediction region is high, the mechanism increases the weight of the global lane features, allowing the node features to more fully integrate global topological information, as the global structure has a more significant impact on the node's decision-making. If the score is low, the mechanism reduces the weight of the global features, retaining more of the node's own local features and avoiding redundant information interference. In the specific fusion operation, the cross-scale attention mechanism adjusts the parameters of the attention weight matrix according to the score, so that each target node automatically focuses on the global information that is more critical to the current scene when interacting with global features. For example, under high complexity, it focuses on the lane connection rules of intersections, while under low complexity, it focuses on the geometric continuity of the node itself. Ultimately, the features of each target node are updated to comprehensive features that integrate local details and adaptive global information. All nodes and their topological relationships together constitute the second graph structure data. At this point, the graph structure retains the local accuracy at the node level and has the semantic association of the global lane network, providing more comprehensive structural support for trajectory prediction in the subsequent autonomous driving control process.

[0108] This application provides a vectorized map encoding method. This method not only encodes the map into a structured form containing the topological relationships between lane segments and target nodes, but more importantly, it assigns geometric and semantic information to each target node, thus laying the foundation for multi-scale feature fusion. Subsequently, node features of each target node and global lane line features of the entire vectorized map are generated using the first map structure data. The construction of node features focuses on detailed mining at the local scale, ensuring that the micro-geometric and semantic features of lane lines are not overlooked. The construction of global lane line features accurately captures the global topological relationships of the vectorized map, such as lane intersection types at intersections and the merging and diverging structures of multi-lane roads. Finally, the node features at the local scale and the global lane line features at the global scale are fused. Through cross-scale feature association, the features of each target node retain both the geometric details of its lane segment and the global topological semantics of the entire vectorized map, achieving effective linkage of multi-scale features. This comprehensively and accurately represents the environmental features and lane line features around the vehicle, providing a more reliable basis for core functions such as trajectory prediction and path planning in autonomous driving systems, and significantly improving the accuracy of trajectory prediction in autonomous driving systems.

[0109] The following describes a vectorized map encoding system provided by an embodiment of this application. The vectorized map encoding system described below can be referred to in correspondence with the vectorized map encoding method described above.

[0110] See Figure 4 The figure is a schematic diagram of the structure of a vectorized map coding system provided in an embodiment of this application, which specifically includes the following modules:

[0111] The feature processing module 100 is used to perform feature processing on the vectorized map from the autonomous driving system to generate first graph structure data; the first graph structure data is used to characterize the geometric and semantic information of target nodes in each lane segment within the vectorized map;

[0112] The feature construction module 200 is used to construct micro-features based on the geometric and semantic information of each target node in the first graph structure data, so as to generate node features for each target node; the node features are used to characterize the lane structure features of the target node in the same lane segment and adjacent lane segments.

[0113] as well as,

[0114] Based on the first graph structure data, global lane line features are generated for the vectorized map; the global lane line features are used to characterize the global topological relationship formed by all lane segments in the vectorized map.

[0115] The feature fusion module 300 is used to perform feature fusion processing based on the node features for each target node and the global features of the lane line to generate second graph structure data; the second graph structure data includes the global topological semantics of the vectorized map and the geometric details of each lane segment.

[0116] In one possible implementation, the feature construction module includes: a node feature construction unit, which is specifically used for:

[0117] Based on the geometric and semantic information of each target node, determine the first association information of each target node with other nodes in the current lane segment, and the second association information of each target node with other nodes in adjacent lane segments.

[0118] Wherein, the first association information is used to characterize the positional and semantic associations between the target node and other nodes within the same lane segment, and the second association information is used to characterize the contextual influence between the target node and other nodes within adjacent lane segments;

[0119] Based on the graph attention mechanism, local feature encoding is performed according to the first association information and the second association information of each target node to generate the node features for each target node.

[0120] See Figure 5 The figure is a schematic diagram of the structure of an electronic device provided in an embodiment of this application, including:

[0121] Memory 11 is used to store computer programs;

[0122] The processor 12 is configured to implement the steps of the vectorized map encoding method described in any of the above method embodiments when executing the computer program.

[0123] In this embodiment, the device can be an in-vehicle computer, a PC (Personal Computer), or a terminal device such as a smartphone, tablet computer, handheld computer, or portable computer.

[0124] The device may include a memory 11, a processor 12, and a bus 13.

[0125] The memory 11 includes at least one type of readable storage medium, such as flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 11 may be an internal storage unit of the device, such as the hard disk of the device. In other embodiments, the memory 11 may be an external storage device of the device, such as a plug-in hard disk, SmartMedia Card (SMC), Secure Digital (SD) card, Flash Card, etc. Furthermore, the memory 11 may include both internal and external storage units of the device. The memory 11 can be used not only to store application software and various types of data installed on the device, such as program code executing vectorized map encoding methods, but also to temporarily store data that has been output or will be output. In some embodiments, the processor 12 may be a central processing unit (CPU).

[0126] In some embodiments, processor 12 may be a central processing unit (CPU), controller, microcontroller, microprocessor or other data processing chip, used to run program code stored in memory 11 or process data, such as program code for executing vectorized map encoding methods.

[0127] This bus 13 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 5 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0128] Furthermore, the device may also include a network interface 14, which may optionally include a wired interface and / or a wireless interface (such as a Wi-Fi interface, a Bluetooth interface, etc.), typically used to establish communication connections between the device and other electronic devices.

[0129] Optionally, the device may further include a user interface 15, which may include a display, an input unit such as a keyboard, and optionally, a standard wired interface or a wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen, etc. The display may also be appropriately referred to as a screen or display unit, used to display information processed in the device and to display a visual user interface.

[0130] Figure 5 Only devices with components 11-15 are shown; those skilled in the art will understand that... Figure 5 The structure shown does not constitute a limitation on the device and may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0131] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides a computer-readable storage medium storing computer instructions for causing the computer to execute the vectorized map encoding method as described in any of the above embodiments.

[0132] The computer-readable media in this application embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0133] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the vectorized map encoding method as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0134] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for the system, method, electronic device, and medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiments. The system, method, electronic device, and medium embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components indicated 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 the solution in this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0135] The above description is merely one specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A vectorized map encoding method, characterized in that, An autonomous driving system applied to vehicles, the method comprising: Feature processing is performed on the vectorized map from the autonomous driving system to generate first graph structure data; the first graph structure data is used to characterize the geometric and semantic information of target nodes in each lane segment within the vectorized map; Micro-features are constructed based on the geometric and semantic information of each target node in the first graph structure data to generate node features for each target node; the node features are used to characterize the lane structure features of the target node in the same lane segment and adjacent lane segments. as well as, Based on the first graph structure data, global lane line features are generated for the vectorized map; the global lane line features are used to characterize the global topological relationship formed by all lane segments in the vectorized map. Based on the node features for each target node and the global features of the lane line, feature fusion processing is performed to generate second graph structure data; the second graph structure data includes the global topological semantics of the vectorized map and the geometric details of each lane segment; The step of constructing micro-features based on the geometric and semantic information of each target node in the first graph structure data to generate node features for each target node includes: Based on the geometric and semantic information of each target node, determine the first association information of each target node with other nodes in the current lane segment, and the second association information of each target node with other nodes in adjacent lane segments. Wherein, the first association information is used to characterize the positional and semantic associations between the target node and other nodes within the same lane segment, and the second association information is used to characterize the contextual influence between the target node and other nodes in adjacent lane segments; Based on the graph attention mechanism, local feature encoding is performed according to the first association information and the second association information of each target node to generate the node features for each target node.

2. The method according to claim 1, characterized in that, The step of generating global lane line features for the vectorized map based on the first map structure data includes: Based on the first map structure data, determine the geometric information of each lane segment within the vectorized map; Based on the geometric information of all lane segments, a lane segment topology connection graph is constructed for the vectorized map; By using a self-attention mechanism and the lane segment topology connection graph, the interaction relationships between all lane segments in the vectorized map are constructed to generate the global features of the lane lines.

3. The method according to claim 1, characterized in that, The step of performing feature fusion processing based on the node features of each target node and the global features of the lane line to generate second graph structure data includes: In response to a trajectory prediction command from the autonomous driving system, a target prediction region associated with the trajectory prediction command is determined; Based on the vectorized map and the first graph structure data, calculate the regional complexity score of the target prediction area; Based on the cross-scale attention mechanism, the region complexity score is used as the basis for the execution of the cross-scale attention mechanism to perform feature fusion processing on the node features of each target node and the global features of the lane line to obtain the second graph structure data.

4. The method according to claim 3, characterized in that, Calculate the regional complexity score of the target prediction area based on the vectorized map, including: Based on the vectorized map and the first map structure data, determine the number of lane branches, the number of vehicles, the number of pedestrians, and the lane segment intersection topology type of the target prediction area; The region complexity score is calculated based on the number of lane branches, the number of vehicles, the number of pedestrians, and the lane segment intersection topology type.

5. The method according to claim 1, characterized in that, The step of performing feature processing on the vectorized map from the autonomous driving system to generate the first graph structure data includes: The vectorized map is topologically encoded to determine a vector set for the vectorized map; the vector set includes multiple lane segments, and each lane segment has at least one target node; the vector set is used to define the topological relationship between the multiple target nodes and the multiple lane segments. Geometric and semantic information are assigned to each lane segment and each target node in the vector set to generate the first graph structure data.

6. A vectorized map coding system, characterized in that, An autonomous driving system for vehicles, the system comprising: The feature processing module is used to perform feature processing on the vectorized map from the autonomous driving system to generate first graph structure data; the first graph structure data is used to characterize the geometric and semantic information of target nodes in each lane segment within the vectorized map; The feature construction module is used to construct micro-features based on the geometric and semantic information of each target node in the first graph structure data, so as to generate node features for each target node; the node features are used to characterize the lane structure features of the target node in the same lane segment and adjacent lane segments. as well as, Based on the first graph structure data, global lane line features are generated for the vectorized map; the global lane line features are used to characterize the global topological relationship formed by all lane segments in the vectorized map. The feature fusion module is used to perform feature fusion processing based on the node features for each target node and the global features of the lane line to generate second graph structure data; the second graph structure data includes the global topological semantics of the vectorized map and the geometric details of each lane segment. The feature construction module includes a node feature construction unit, which is specifically used for: Based on the geometric and semantic information of each target node, determine the first association information of each target node with other nodes in the current lane segment, and the second association information of each target node with other nodes in adjacent lane segments. Wherein, the first association information is used to characterize the positional and semantic associations between the target node and other nodes within the same lane segment, and the second association information is used to characterize the contextual influence between the target node and other nodes in adjacent lane segments; Based on the graph attention mechanism, local feature encoding is performed according to the first association information and the second association information of each target node to generate the node features for each target node.

7. An electronic device, characterized in that, The device includes: a processor, a memory, and a system bus; The processor and the memory are connected via the system bus; The memory is used to store one or more programs, the one or more programs including instructions that, when executed by the processor, cause the processor to perform the vectorized map encoding method according to any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the vectorized map encoding method as described in any one of claims 1-5.

Images