A high-precision map construction method and system based on adaptive point-element interaction

Abstract

The application relates to the field of computer vision and automatic driving high-precision map construction, and provides a high-precision map construction method and system based on adaptive point-element interaction, which comprises the following steps: acquiring vehicle multi-view image data and extracting multi-view BEV features; constructing a hybrid query comprising point query and element query; decoding the first m layers of a hybrid decoder to obtain point coordinates of map elements and element query; calculating a complexity score based on point coordinate geometric features and element query semantic features, and determining a vector point number; performing arc length parameterization resampling on the point coordinates based on the vector point number, and constructing a new hybrid query; and decoding the last n layers of the hybrid decoder to obtain a structured vector high-precision map. According to the application, the vector point number can be adaptively distributed according to the geometric complexity of elements, the precision of complex elements is improved, the redundant calculation of simple elements is reduced, and the technical problem that the existing technology cannot simultaneously consider the calculation efficiency of simple elements and the geometric precision of complex elements is solved.

Description

Technical Field

[0001] This application relates to the field of computer vision and high-precision map construction for autonomous driving, and in particular to a high-precision map construction method and system based on adaptive point-element interaction. Background Technology

[0002] High-precision maps, as a core component of autonomous driving systems, provide static environmental information with centimeter-level accuracy for perception, planning, and decision-making. In recent years, online vectorized high-precision map construction methods based on vehicle sensors have received widespread attention. These methods typically represent each map element as an ordered sequence of discrete points. However, existing vectorized high-precision map construction methods generally suffer from a common structural flaw: all map elements are represented using the same fixed number of vector points. Whether the element is a simple straight lane line or a complex curved intersection boundary, a fixed number of points is used for modeling. This strategy leads to two problems: first, for geometrically simple elements such as straight lines or gentle curves, most points in the fixed number are nearly collinear, causing unnecessary computational overhead; second, for geometrically complex elements such as curved intersections and pedestrian crossing corners, the spacing between the fixed number of points is too large, causing the smooth loss of geometric details at curves. The geometric complexity of different elements in the same scene varies significantly, and a uniform fixed number of points cannot simultaneously ensure both computational efficiency and geometric accuracy. Therefore, how to adaptively allocate the number of vector points according to the geometric complexity of the elements, so as to improve the accuracy of complex elements while reducing redundant calculations of simple elements, is a technical problem that urgently needs to be solved in the field of vectorized high-precision map construction. Summary of the Invention

[0003] This application provides a high-precision map construction method and system based on adaptive point-element interaction, which solves the technical problem in the prior art that it is impossible to balance the computational efficiency of simple elements with the geometric accuracy of complex elements.

[0004] To achieve the above objectives, this application adopts the following technical solution: Firstly, a method for constructing a high-precision map based on adaptive point-element interaction is provided, characterized by comprising: acquiring multi-view image data of a vehicle; extracting multi-view BEV features from the multi-view image data using a multi-view feature extraction model; constructing a hybrid query including point queries and element queries based on the multi-view BEV features; decoding the hybrid query through the first m layers of a hybrid decoder based on the multi-view BEV features to obtain a fixed number of point coordinates corresponding to each map element and an updated element query; calculating the complexity score of each map element based on the geometric features of the point coordinates and the semantic features of the updated element query, and determining the number of vector points and feature sampling points for each map element based on the complexity score; resampling the fixed number of point coordinates using arc length parameterization based on the number of vector points to generate point coordinates and point queries corresponding to the number of vector points, and constructing a new hybrid query based on the point coordinates and the updated element query; and decoding the new hybrid query through the last n layers of a hybrid decoder based on the multi-view BEV features to obtain a structured vector high-precision map.

[0005] In conjunction with the first aspect mentioned above, in one possible implementation, multi-view BEV features of multi-view image data are extracted through a multi-view feature extraction model, including: extracting features from multi-view image data through a backbone network to obtain multi-view image features; inputting the multi-view image features into an FPN network for multi-scale feature fusion to obtain fused multi-view image features; and converting the fused multi-view image features from the image viewpoint to the bird's-eye viewpoint using a 2D-to-BEV viewpoint transformation method based on camera calibration parameters or a learned viewpoint transformation method to obtain multi-view BEV features.

[0006] In conjunction with the first aspect mentioned above, in one possible implementation, a hybrid query including point queries and element queries is constructed based on multi-view BEV features. This includes: randomly initializing the basic hybrid query into a learnable embedding vector and decomposing it into point queries and element queries; and iteratively optimizing the hybrid query in each layer of the hybrid decoder through point-level feature extraction paths and element-level feature extraction paths to obtain the hybrid query.

[0007] In conjunction with the first aspect mentioned above, in one possible implementation, the hybrid query is iteratively optimized in each layer of the hybrid decoder through point-level feature extraction paths and element-level feature extraction paths to obtain the hybrid query. This includes: in each layer of the hybrid decoder, based on the point query and point coordinates output from the previous layer, combined with position embedding, feature sampling and fusion are performed from multi-view BEV features to obtain point-level features; based on the element query output from the previous layer, combined with the position embedding generated by the point query, element-level features are extracted from multi-view BEV features through a mask attention mechanism; based on the point-level features and element-level features, adaptive point-element interaction is performed to fuse the bidirectional information between the point-level features and element-level features, and the point query and element query are updated to obtain updated point query and element query; the updated point query and element query are concatenated and combined to generate the hybrid query of the current layer; the hybrid query of the current layer is used as the input of the next layer, and the hybrid query after layer-by-layer iterative optimization is output in the last layer.

[0008] In conjunction with the first aspect mentioned above, in one possible implementation, based on multi-view BEV features, the hybrid query is decoded through the first m layers of the hybrid decoder to obtain a fixed number of point coordinates corresponding to each map element and the updated element query. This includes: inputting the hybrid query into the first m layers of the hybrid decoder, and interactively decoding it with the multi-view BEV features in each layer, updating the point query and element query layer by layer; after decoding layer by layer through the first m layers, predicting the corresponding point coordinates based on the point query output by the m-th layer; and outputting the element query corresponding to the m-th layer to participate in the calculation of the complexity score.

[0009] In conjunction with the first aspect mentioned above, in one possible implementation, the complexity score of map elements includes curvature complexity and path complexity. Among them, curvature complexity Satisfy the following formula:

[0010] in, cuv represents the curvature complexity of the i-th map element in the m-th layer, where cuv represents curvature and i represents the map element index. Represents vector coordinate points The angle between the direction vectors, where j represents the index of the vector coordinate point, and takes the value of . , This indicates the number of coordinate points output by the m-th layer; Indicates variance; Path complexity Satisfy the following formula:

[0011] in, Let len ​​represent the path complexity of the i-th map element at the m-th level, where len represents the path and i represents the map element index. , This represents the vector coordinate point, and j represents the index of the vector coordinate point; This indicates the number of coordinate points output by the m-th layer. It is an L2 norm; Complexity score Satisfy the following formula:

[0012] in, This represents the query element of the i-th map element at the m-th level. Let represent the curvature complexity of the i-th map element in the m-th layer. This represents the path complexity of the i-th map element at the m-th layer; Indicates a splicing operation; and These are the learnable weight matrix and bias parameters of the first linear layer of the complexity scoring network, respectively; ReLU() is the activation function; and The learnable weight matrix and scalar bias of the second linear layer of the complexity scoring network; This is the Sigmoid activation function.

[0013] In conjunction with the first aspect mentioned above, in one possible implementation, based on the number of vector points, a fixed number of point coordinates are resampled using arc length parameterization to generate point coordinates and point queries corresponding to the number of vector points. A new hybrid query is then constructed based on the point coordinates and the updated element query. This includes: calculating the cumulative arc length parameter according to the Euclidean distance between adjacent points based on the fixed number of point coordinates, and normalizing the point coordinates to obtain an arc length parameter sequence; performing equidistant sampling on the arc length parameter sequence based on the number of vector points to determine the target sampling position; resampling the fixed number of point coordinates using interpolation based on the target sampling position to generate resampled point coordinates; the resampled point coordinates are the point coordinates corresponding to the number of vector points; and determining the corresponding point query based on the resampled point coordinates, combining it with the updated element query to construct a new hybrid query.

[0014] In conjunction with the first aspect mentioned above, in one possible implementation, the cumulative arc length parameter... Satisfy the following formula:

[0015]

[0016] in, Let the coordinates be the (j+1)th point of the i-th element. Let j be the coordinates of the i-th element. It is the L2 norm. This represents the straight-line distance between the j-th point and the (j+1)-th point. Let be the cumulative arc length from point 1 to point j. ,when The interpolation step is skipped, and the original point coordinates are directly used as the output.

[0017] In conjunction with the first aspect mentioned above, in one possible implementation, based on multi-view BEV features, a new hybrid query is decoded through the last n layers of the hybrid decoder to obtain a structured vector high-precision map. This includes: inputting the new hybrid query into the last n layers of the hybrid decoder and performing layer-by-layer interactive decoding in conjunction with multi-view BEV features to iteratively update the point query and element query; predicting the point coordinate information and category information of the corresponding map element based on the updated point query and element query at the nth layer; and organizing each map element in a structured manner according to the point coordinate information and the corresponding category information to generate a structured vector high-precision map.

[0018] Secondly, a high-precision map construction system based on adaptive point-element interaction is provided, comprising: a data acquisition module, a feature extraction module, a hybrid query construction module, a front-end decoding module, a complexity evaluation module, a resampling and query reconstruction module, and a back-end decoding module; wherein, the data acquisition module is used to acquire multi-view image data of the vehicle; the feature extraction module is used to extract multi-view BEV features from the multi-view image data through a multi-view feature extraction model; the hybrid query construction module is used to construct a hybrid query including point query and element query based on the multi-view BEV features; and the front-end decoding module is used to decode the hybrid query based on the multi-view BEV features through the first m layers of the hybrid decoder to obtain each map. The system consists of: a fixed number of point coordinates corresponding to each element and an updated element query; a complexity evaluation module, which calculates the complexity score of each map element based on the geometric features of the point coordinates and the semantic features of the updated element query, and determines the number of vector points and feature sampling points for each map element based on the complexity score; a resampling and query reconstruction module, which performs arc-length parameterized resampling on the fixed number of point coordinates based on the number of vector points, generates point coordinates and point queries corresponding to the number of vector points, and constructs a new hybrid query based on the point coordinates and the updated element query; and a back-end decoding module, which decodes the new hybrid query based on multi-view BEV features through the last n layers of the hybrid decoder to obtain a structured vector high-precision map.

[0019] Beneficial effects: This application provides a high-precision map construction method and system based on adaptive point-element interaction. By dynamically allocating more vector points to complex elements, it can reconstruct their geometric details (such as curved intersections and pedestrian crossing corners) more precisely, avoiding the detail smoothing and missing problems caused by fixed-point models. Simultaneously, this application achieves on-demand allocation of computing resources, avoiding redundant calculations for simple elements, thereby improving the overall computational efficiency of the model and enabling it to handle more complex scenarios with the same computational budget. Furthermore, the adaptive point allocation mechanism of this application is embedded as an independent module in the intermediate layer of the hybrid decoder, without changing the BEV feature extraction process at the input end or the prediction head structure at the output end. It can be flexibly integrated into existing vectorized high-precision map frameworks, solving the technical problem of existing technologies that cannot simultaneously achieve computational efficiency for simple elements and geometric accuracy for complex elements.

[0020] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0021] Figure 1 A system architecture diagram of a high-precision map construction system based on adaptive point-element interaction is provided in the embodiments of this application; Figure 2 A flowchart illustrating a high-precision map construction method based on adaptive point-element interaction provided in this application embodiment. Figure 1 ; Figure 3 A flowchart illustrating a high-precision map construction method based on adaptive point-element interaction provided in this application embodiment. Figure 2 ; Figure 4 A flowchart illustrating a high-precision map construction method based on adaptive point-element interaction provided in this application embodiment. Figure 3 ; Figure 5 A flowchart illustrating a high-precision map construction method based on adaptive point-element interaction provided in this application embodiment. Figure 4 ; Figure 6 A flowchart illustrating a high-precision map construction method based on adaptive point-element interaction provided in this application embodiment. Figure 5 ; Figure 7 A flowchart illustrating a high-precision map construction method based on adaptive point-element interaction provided in this application embodiment. Figure 6 ; Figure 8 A flowchart illustrating a high-precision map construction method based on adaptive point-element interaction provided in this application embodiment. Figure 7 . Detailed Implementation

[0022] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0023] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0024] The high-precision map construction method based on adaptive point-element interaction provided in this application can be applied to, for example... Figure 1 In the high-precision map construction system based on adaptive point-element interaction shown, such as Figure 1 As shown, the system includes: a data acquisition module 101, a feature extraction module 102, a hybrid query construction module 103, a front-end decoding module 104, a complexity evaluation module 105, a resampling and query reconstruction module 106, and a back-end decoding module 107. The system includes: a data acquisition module 101 for acquiring multi-view image data of the vehicle; a feature extraction module 102 for extracting multi-view BEV features from the multi-view image data using a multi-view feature extraction model; a hybrid query construction module 103 for constructing a hybrid query including point queries and element queries based on the multi-view BEV features; a front-end decoding module 104 for decoding the hybrid query based on the multi-view BEV features using the first m layers of the hybrid decoder to obtain a fixed number of point coordinates corresponding to each map element and the updated element query; and a complexity evaluation module 105 for evaluating the complexity based on the point coordinates. The system calculates the complexity score of each map element based on geometric features and the semantic features of the updated element query, and determines the number of vector points and feature sampling points for each map element based on the complexity score. The resampling and query reconstruction module 106 is used to perform arc-length parameterized resampling on the coordinates of a fixed number of points based on the number of vector points, generate the coordinates of points and point queries corresponding to the number of vector points, and construct a new hybrid query based on the coordinates of points and the updated element query. The back-end decoding module 107 is used to decode the new hybrid query based on the multi-view BEV features through the last n layers of the hybrid decoder to obtain a structured vector high-precision map.

[0025] To address the technical problem in existing technologies that cannot simultaneously achieve computational efficiency for simple elements and geometric accuracy for complex elements, embodiments of this application provide a high-precision map construction method based on adaptive point-element interaction. Figure 2 A flowchart illustrating the high-precision map construction method based on adaptive point-element interaction provided in this application embodiment. Figure 1 ,like Figure 2 As shown, the method includes: S201. Acquire multi-view image data of the vehicle.

[0026] In one possible implementation, a surround-view camera array deployed around the vehicle acquires multi-view image data. For example, the vehicle can be equipped with 4-6 wide-angle cameras, covering the front, rear, left, and right views of the vehicle. The acquired image data contains rich texture and semantic information of the vehicle's surrounding environment, serving as the basic input for subsequent feature extraction.

[0027] S202. Extract multi-view BEV features from multi-view image data using a multi-view feature extraction model.

[0028] Multi-view BEV features refer to the feature representation obtained by uniformly transforming image features from multiple perspectives to a bird's-eye view. BEV features eliminate the effects of viewpoint occlusion and scale variations, intuitively representing the geometric positional relationships of road elements, and providing a unified spatial feature description for subsequent map element detection and vectorization. The multi-view feature extraction model includes a backbone network, an FPN network, and a 2D-to-BEV viewpoint transformation network.

[0029] In one possible implementation, feature extraction is performed on multi-view image data through a backbone network to obtain multi-view image features; the multi-view image features are input into an FPN network for multi-scale feature fusion to obtain fused multi-view image features; based on camera calibration parameters or a learned viewpoint transformation method, the fused multi-view image features are transformed from the image viewpoint to the bird's-eye viewpoint through a 2D-to-BEV viewpoint transformation method to obtain multi-view BEV features.

[0030] S203. Based on multi-view BEV features, construct a hybrid query that includes point queries and element queries.

[0031] In one possible implementation, a hybrid query is a key data structure used in this embodiment to iteratively generate map elements in the decoder. Specifically, a point query characterizes the geometric location features of discrete points within a map element, while an element query characterizes the overall semantic and shape features of the map element. By constructing a hybrid query, this embodiment can simultaneously focus on the local details and global attributes of map elements, providing initialization parameters for subsequent dual-path interactions.

[0032] S204. Based on multi-view BEV features, the hybrid query is decoded through the first m layers of the hybrid decoder to obtain a fixed number of point coordinates corresponding to each map element and the updated element query.

[0033] In one possible implementation, the first m layers of the hybrid decoder constitute the initial decoding stage. In this stage, regardless of the actual shape complexity of the map elements, the model forces the output of a fixed number of point coordinates (e.g., 20 points). The purpose of this design is to provide a unified geometric representation benchmark for all elements, enabling the model to quickly converge to the approximate outline and location of the map elements with low computational cost. Simultaneously, through iterative interactions across m layers, the element query also incorporates semantic information from the BEV features and is updated. It should be understood that the specific value of m can be set according to actual computing power and accuracy requirements, such as setting it to 3 or 4 layers; this embodiment does not impose such limitations.

[0034] S205. Based on the geometric features of point coordinates and the semantic features of the updated element query, calculate the complexity score of each map element, and determine the number of vector points and feature sampling points of each map element according to the complexity score.

[0035] In one possible implementation, existing technologies typically use the same fixed number of points for all elements, leading to redundant calculations for simple elements and loss of detail for complex elements. This embodiment quantifies the geometric complexity of each map element through a computational complexity score. For example, for a curved intersection boundary, the curvature of the line connecting its point coordinates is large, resulting in complex geometric features and a high computational complexity score; while for a straight lane line, its point coordinates are approximately collinear, resulting in simple geometric features and a low score. Based on the score, the model adaptively allocates more vector points (e.g., 40) to complex elements to ensure the accuracy of geometric detail reconstruction; and allocates fewer vector points (e.g., 10) to simple elements to reduce redundant calculations. Furthermore, the adaptive determination of the number of feature sampling points allows the model to sample more BEV feature points when extracting features from complex elements, thereby capturing richer local information. Through this step, on-demand allocation of computational resources is achieved.

[0036] Preferably, the complexity score of map elements includes curvature complexity and path complexity; Among them, curvature complexity Satisfy the following formula:

[0037] in, cuv represents the curvature complexity of the i-th map element in the m-th layer, where cuv represents curvature and i represents the map element index. Represents vector coordinate points The angle between the direction vectors, where j represents the index of the vector coordinate point, and takes the value of . , This indicates the number of coordinate points output by the m-th layer; Indicates variance; Preferred path complexity Satisfy the following formula:

[0038] in, Let len ​​represent the path complexity of the i-th map element at the m-th level, where len represents the path and i represents the map element index. , This represents the vector coordinate point, and j represents the index of the vector coordinate point; This indicates the number of coordinate points output by the m-th layer. It is an L2 norm; Preferred complexity score Satisfy the following formula:

[0039] in, This represents the query element of the i-th map element at the m-th level. Let represent the curvature complexity of the i-th map element in the m-th layer. This represents the path complexity of the i-th map element at the m-th layer; Indicates a splicing operation; and These are the learnable weight matrix and bias parameters of the first linear layer of the complexity scoring network, respectively; ReLU() is the activation function; and The learnable weight matrix and scalar bias of the second linear layer of the complexity scoring network; This is the Sigmoid activation function.

[0040] Preferably, the number of vector points satisfies the following formula:

[0041] in, Score the complexity. The preset maximum number of vector points, The minimum number of vector points is preset to ensure that even the simplest elements have basic geometric representation capabilities; round() is the rounding function; clamp() is the truncation function to ensure that the output is within the range of the maximum and minimum preset number of points; As a discrete structural control quantity, it determines the number of vector points of each element and does not participate in the backpropagation calculation; Preferably, the number of sampling points Satisfy the following formula:

[0042] in, Number of basic sampling points; Maximum number of expanded sampling points; complexity score Higher-order elements obtain more sampling points to extract richer local geometric features in the BEV space; round() is the rounding function.

[0043] S206. Based on the number of vector points, perform arc length parameterized resampling on the coordinates of a fixed number of points to generate the coordinates of points and point queries corresponding to the number of vector points, and construct a new hybrid query based on the coordinates of points and the updated element queries.

[0044] In one possible implementation, after determining the adaptive number of vector points, the coordinates of the fixed number of points output from the first m layers need to be adjusted. Arc-length parameterized resampling ensures that the resampled point coordinates are uniformly distributed on the element curve, avoiding shape distortion caused by changes in the number of points. For example, when the number of points needs to be increased, new coordinate points are generated between existing adjacent points through interpolation; when the number of points needs to be reduced, key points are extracted according to arc-length intervals. Simultaneously, a new point query is generated based on the resampled point coordinates and recombined with the previously updated element query to construct a new hybrid query. This process ensures that the input to the subsequent n layers of decoding includes both the geometric and semantic priors learned in the preceding stages and meets the requirements of adaptive point count.

[0045] S207. Based on multi-view BEV features, the new hybrid query is decoded through the last n layers of the hybrid decoder to obtain a structured vector high-precision map.

[0046] In one possible implementation, the last n layers constitute the fine-grained decoding stage. In this stage, each map element has an independent number of vector points commensurate with its complexity. Hybrid queries continue to interact with BEV features in subsequent layers, continuously optimizing the positional accuracy of point coordinates and the semantic category of the elements. Ultimately, the model outputs a structured vector high-precision map containing accurate coordinates and category labels. Because complex elements receive sufficient point support, details such as curves and corners are preserved; simple elements avoid excessive computational overhead, thus achieving a balance between map building accuracy and efficiency overall.

[0047] This application introduces an adaptive point-element interaction mechanism that combines point query and element query for collaborative modeling. Based on multi-view BEV features, it employs a two-stage hybrid decoding strategy to jointly optimize the geometric structure and semantic information of map elements. First, an initial point set is obtained through the first m layers of decoding. Then, a complexity scoring mechanism based on curvature and path features adaptively determines the number of vector points for each map element. The point set is then reconstructed through arc-length parameterized resampling. Finally, the point coordinates are refined and category information is supplemented through the last n layers of decoding, thereby generating a structured vector high-precision map. Compared to existing fixed-point or single-scale modeling methods, this application can dynamically adjust the expression precision according to the structural complexity of map elements. While ensuring the precision of geometric expression, it effectively reduces the computational overhead caused by redundant points, improving the accuracy, expressive power, and computational efficiency of map construction.

[0048] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 3 As shown, the above S202 can be specifically implemented through the following S301, S302 and S303, which are explained in detail below: S301. Feature extraction is performed on multi-view image data through a backbone network to obtain multi-view image features.

[0049] In one possible implementation, the backbone network serves as the fundamental module for feature extraction, its role being to extract high-dimensional semantic features from the raw image data. In this embodiment, the backbone network can employ a ResNet (residual network) series model, such as ResNet-50 or ResNet-101, utilizing its residual learning units to effectively address the gradient vanishing problem in deep network training, extracting robust image features. It should be understood that, in addition to the ResNet series, the backbone network can also employ deep learning models such as EfficientNet, VGG, or Swin Transformer. For example, on resource-constrained embedded automotive platforms, EfficientNet can be preferentially chosen, as it significantly reduces the number of model parameters and computational cost while maintaining feature extraction accuracy through a composite scaling method. The multi-view image data input to the backbone network is typically RGB images captured by the vehicle's surround-view camera. After processing by the backbone network, multi-view image features are output, containing rich texture, edge, and semantic information.

[0050] S302. Input the multi-view image features into the FPN network to perform multi-scale feature fusion to obtain the fused multi-view image features.

[0051] In one possible implementation, due to the significant scale differences among map elements in road scenes, features at a single scale are insufficient to simultaneously address the detection needs of both large and small objects. Feature Pyramid Networks (FPNs) fuse feature maps output from different levels of the backbone network through top-down paths and lateral connections. For example, FPNs can fuse features from three levels (C3, C4, and C5) of the backbone network output to generate a feature pyramid with multi-scale receptive fields. This fusion mechanism combines high-level features (strong semantic information, low resolution) with low-level features (rich geometric details, high resolution), thereby enhancing the model's ability to perceive map elements at different scales and providing richer feature representations for subsequent viewpoint transitions.

[0052] S303. Based on camera calibration parameters or a learning-based perspective transformation method, the fused multi-view image features are transformed from the image perspective to the bird's-eye view perspective using the 2D-to-BEV perspective transformation method to obtain multi-view BEV features.

[0053] In one possible implementation, the features of an image perspective suffer from perspective distortion, making them difficult to use directly for map construction in metric space. Therefore, they need to be converted to a bird's-eye view (BEV) perspective. This application provides two optional conversion methods: one is based on camera calibration parameters, using pre-calibrated camera intrinsics and extrinsic parameters to project image features into the BEV space through geometric transformation. This method is computationally efficient but highly dependent on calibration accuracy. The other is based on deep learning, such as using end-to-end models like BEVFormer, which automatically learns the correspondence between image features and BEV features through a spatiotemporal attention mechanism. This method is more robust to camera parameter errors. Taking BEVFormer as an example, it uses a spatial cross-attention mechanism to pre-set learnable query vectors in the BEV space, interacting with multi-view image features to aggregate BEV features. The final output multi-view BEV features are a unified spatial feature map, where each pixel corresponds to a physical location in the real world, eliminating the effects of perspective occlusion and scale changes, and providing standardized input data for subsequent hybrid decoders.

[0054] Preferably, the BEVFormer method is used to obtain multi-view BEV features. Satisfy the following formula:

[0055] in, H, W, and C represent the height, width, and number of channels of the BEV space, respectively. Representing multi-view image features, This represents the BEVFormer method.

[0056] This application embodiment fuses multi-scale features from a backbone network and an FPN network, and combines this with a 2D-to-BEV perspective transformation method to uniformly map multi-view image features to a bird's-eye view, effectively integrating spatial information from different perspectives and enhancing the global consistency and spatial expressive power of the features. Simultaneously, the introduction of multi-scale features helps to balance fine-grained structural information with high-level semantic information, thereby improving the perception of map elements at different scales. Furthermore, by using camera calibration parameters or a learned perspective transformation method, the accuracy and robustness of perspective transformation are improved, making the generated BEV features more stable and reliable, providing a high-quality feature foundation for subsequent point-element interaction modeling and high-precision map construction.

[0057] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 4 As shown, the above S203 can be specifically implemented through the following S401 and S402, which are explained in detail below: S401. Randomly initialize the basic hybrid query as a learnable embedding vector and decompose it into point queries and element queries.

[0058] In one possible implementation, during the initial stage of the hybrid decoder, the model needs to construct a set of initial query vectors as the starting point for decoding. In this embodiment, the basic hybrid query is designed as a set of learnable embedding vectors, whose initial values ​​are generated through random initialization and optimized during model training using the backpropagation algorithm. It should be understood that this random initialization method endows the model with the ability to automatically learn prior knowledge from the data. The basic hybrid query is logically decomposed into two parts: point query and element query. The point query is used to represent the local geometric features of each discrete point in the map element, such as the coordinates of each vertex of a lane line; the element query is used to represent the overall semantic features and shape attributes of the map element, such as whether the element is a lane line or a curb, and its overall directional trend. This decomposition design allows the model to focus on local details and global attributes separately, providing a structured foundation for subsequent dual-path interactions.

[0059] S402, in each layer of the hybrid decoder, the hybrid query is iteratively optimized through point-level feature extraction path and element-level feature extraction path to obtain the hybrid query.

[0060] In one possible implementation, the hybrid decoder employs a hierarchical structure, with each layer containing both point-level and element-level feature extraction paths. The feature representation of the hybrid query is continuously refined through iterative layer-by-layer processing. The advantage of this dual-path design is that the point-level path captures fine geometric details, while the element-level path maintains overall semantic consistency. The two complement each other, jointly improving the accuracy of map construction.

[0061] In one possible implementation of the embodiments of this application, combined with Figure 4 ,like Figure 5 As shown, the above S402 can be specifically implemented through the following S501 to S504, which are explained in detail below: S501. In each layer of the hybrid decoder, based on the point query and point coordinates output by the previous layer, and combined with position embedding, feature sampling and fusion are performed from the multi-view BEV features to obtain point-level features.

[0062] In one possible implementation, the core idea of ​​the point-level feature extraction path is to use the predicted position of the previous layer as an anchor point to extract corresponding local features from the BEV feature map. First, the point coordinates output from the previous layer are transformed linearly to generate a location embedding, which contains the spatial location information of the point. Then, the location embedding is added to the point query from the previous layer to obtain a location-aware point query. Next, based on this location-aware point query, the model predicts the offset of a set of sampling points and their corresponding attention weights. These sampling points are typically distributed around the anchor point to capture local contextual information. Features at these sampling points are extracted from the BEV feature map using bilinear interpolation and weighted fusion according to the attention weights to finally obtain the point-level features of the current layer. This process allows the point query to dynamically focus on the region in the BEV feature map most relevant to the current predicted point, thereby improving the accuracy of geometric localization.

[0063] As an example, in the first l In the layer, with the first l-1 The coordinates of the points output by the current layer are used as the anchor point coordinates of the current layer, with the first layer as the anchor point coordinates. l-1 The point query output by the layer serves as the initial state for the current layer's point query. First, the anchor point coordinates are linearly transformed to generate the position embedding of the point query, which is then added to the point query from the previous layer to obtain the position-aware point query. Next, the position-aware point query is used to predict the sampling offset and attention weights. Using the offset and attention weights, feature fusion is performed on K sampling points on the BEV feature to obtain the fused point-level features of each sampling point. Finally, the fused point-level features of the K sampling points are fused with the point query from the previous layer to obtain the point-level features of the current layer. The specific logical reasoning process is as follows:

[0064]

[0065]

[0066]

[0067]

[0068] in, l For layer indexes in the hybrid decoder, E Indicates the number of map elements. P This represents a fixed initial number of points for each map element. Indicates the first l-1 The coordinates of the j-th point in the layer; Indicates the first l The position embedding of the query for the j-th point in the layer; Indicates the first l-1 Query the j-th point of the layer; Represents a location-aware point query; It is the offset of the sampling point; These are the attention weights, located in the range [0,1]; K represents the number of sampling points. This refers to the fusion point set features of the K sampling points of the j-th point in the current layer, due to Since it is a fractional coordinate, therefore in calculation Bilinear interpolation is then used to obtain the feature vector of the coordinate. This represents the point-level features of all anchor points in the current layer; , , , All are learnable parameters of the linear layer; poi represents the point query path.

[0069] S502. For the specific operation of the element-level feature extraction path, based on the element query output of the previous layer, combined with the position embedding generated by the point query, element-level features are extracted from the multi-view BEV features through the mask attention mechanism.

[0070] In one possible implementation, the element-level feature extraction path aims to aggregate semantic information related to the current map elements from a global perspective. First, by weighted summing of the position embeddings of the current layer's point queries, the position embeddings of the current layer's element queries are generated, enabling the element queries to perceive the overall spatial range of the current element. Then, this position embedding is added to the element queries from the previous layer to obtain the position-aware element queries. Next, a mask attention mechanism is used to calculate the correlation between the position-aware element queries and the BEV feature map. The key to the mask attention mechanism is the introduction of a binary mask, generated based on the prediction results of the previous layer, used to filter out interference from background areas or irrelevant elements. Through this masking operation, the model can focus its attention on the areas where the current element may exist, thereby extracting purer element-level features. This process effectively solves the problem of mutual interference among multiple elements in complex scenes, improving the accuracy of semantic classification.

[0071] As an example, with the first l-1 The element query output from the current layer serves as the initial state for the current layer's element query; by performing a weighted summation of the position embeddings of the current layer's point query, the position embedding of the current layer's element query is generated, and then compared with the first layer's... l-1 The element queries of each layer are summed to obtain the position-aware element query; the element-level features of the current layer are extracted from the BEV features using a masked attention mechanism. The formula reasoning process is as follows:

[0072]

[0073]

[0074] in, It indicates the fusion of the first l-1 Layer element query Embedded with the current layer position Location-aware element lookup; For the sinusoidal position encoding of BEV features, the BEV feature X is compared with... Weighted fusion yields flattened position-aware BEV features ; For the first l-1 Binary anchor point mask output by the layered pier. This represents the correlation between location-aware element queries and BEV features, normalized along the H and W dimensions using Softmax, and then compared with the first... l-1 The layer mask is multiplied element-wise to filter out irrelevant regions; This refers to merging the first l-1 The output element feature is obtained from the information of the i-th element of the layer; 1 Layer element query Given a learnable initial embedding vector, the first... 1 Layer anchor mask Initialize it as an all-1 matrix, meaning no region is masked.

[0075] S503. After extracting point-level features and element-level features, adaptive point-element interaction is performed based on the point-level features and element-level features to fuse the bidirectional information between the point-level features and element-level features, and the point query and element query are updated to obtain the updated point query and element query.

[0076] In one possible implementation, point-level features and element-level features respectively contain local geometric information and global semantic information, exhibiting a natural complementarity. An adaptive point-element interaction mechanism enables bidirectional information flow between the two. On one hand, point-level features can be passed to element queries through cross-level connections, providing precise geometric basis for element-level semantic judgment; on the other hand, element-level features can also be passed to point queries, providing global semantic guidance for point-level geometric localization. This bidirectional fusion is achieved through feature addition or concatenation operations, and further refined by a self-attention mechanism and a feedforward neural network, ultimately yielding updated point and element queries. It should be understood that this interaction mechanism is not a simple feature superposition, but rather dynamically adjusts the weights between the two through learnable parameters, enabling the model to adaptively balance local and global information according to the characteristics of different elements.

[0077] As an example, the point-level features and element-level features extracted from the current layer are refined by self-attention and feedforward neural networks respectively, and then cross-level query updates are performed to generate the final point query and element query of the current layer.

[0078] Preferred, click query and element query They respectively satisfy the following formulas:

[0079]

[0080] in, This represents the point-level features refined by self-attention and feedforward neural networks. This represents the element-level features refined by self-attention and feedforward neural networks. This represents a feedforward neural network. l For layer indexes, ele represents element path and poi represents point path.

[0081] S504. Combine the updated point query and element query to generate the mixed query of the current layer; use the mixed query of the current layer as the input of the next layer, and output the mixed query after iterative optimization at the last layer.

[0082] In one possible implementation, at the end of each decoding layer, the model concatenates the updated point query and element query along the feature dimension, recombineing them into a hybrid query. This hybrid query serves as the output of the current layer and the input of the next layer, iterating in this manner until a preset number of decoding layers is reached. Through this layer-by-layer iterative approach, the geometric and semantic information in the hybrid query is continuously fused and refined, ultimately outputting a high-quality hybrid query at the final layer of the decoder, providing strong feature support for subsequent map element prediction.

[0083] This application embodiment constructs a hybrid query mechanism that combines point queries and element queries, and introduces point-level feature extraction paths and element-level feature extraction paths into the hybrid decoder. This achieves a complete feature modeling process from initialization to layer-by-layer optimization. S401-S402 completes the construction of the hybrid query and the design of the overall iterative framework. S501-S504 further refines the feature interaction and update methods within each layer. Through adaptive point-element interaction, it achieves deep fusion of geometric and semantic information, thereby effectively improving the expressive power and stability of the query representation, avoiding the information fragmentation problem caused by single query modeling, improving the accuracy and consistency of map element representation, and providing high-quality intermediate feature support for the subsequent refined construction of high-precision maps.

[0084] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 6 As shown, the above S205 can be specifically implemented through the following S601 to S603, which are explained in detail below: S601. Input the hybrid query into the first m layers of the hybrid decoder, and perform interactive decoding with the multi-view BEV features in each layer, updating the point query and element query layer by layer.

[0085] In one possible implementation, the first m layers of the hybrid decoder form a cascaded structure. In each layer, both the point query and the element query from the hybrid query are used as input, interacting deeply with multi-view BEV features. Local geometric details are captured through point-level feature extraction paths, global semantic information is aggregated through element-level feature extraction paths, and the two are fused through an adaptive point-element interaction mechanism. After processing at each layer, both the point query and the element query are updated, allowing the model's understanding of the location, shape, and category of map elements to deepen layer by layer. It should be understood that the number of layers *m* in the first *m* layers is a hyperparameter that can be set according to actual needs.

[0086] As an example, m can be set to 3, meaning the first 3 layers are used for initial decoding of a fixed number of points. The advantage of this design is that, through iterations of the first m layers, the model can quickly eliminate background interference, lock potential areas of map elements, and generate preliminary geometry with low computational cost, avoiding the waste of resources caused by directly performing high-complexity calculations.

[0087] S602. After decoding the first m layers one by one, predict the corresponding point coordinates based on the point query output of the m-th layer.

[0088] In one possible implementation, after the interactive update at layer m is completed, the model uses a point coordinate prediction head to decode the point query output from layer m. This prediction head typically consists of a simple linear layer used to map the high-dimensional point query features into two-dimensional coordinate values.

[0089] It's important to note that in the first m layers, regardless of the actual shape complexity of the map elements, the model forces the output of a fixed number of point coordinates. For example, the fixed number P can be set to 20, meaning that each map element output at the m-th layer contains 20 ordered coordinate points. These 20 points initially outline the geometry of the map element, such as the direction of lane lines and the boundaries of intersections. This fixed number of points provides a unified quantitative benchmark for subsequent complexity evaluation, enabling the model to fairly compare the geometric complexity of different elements.

[0090] S603. Output the element query corresponding to the m-th layer and participate in the calculation of complexity score.

[0091] In one possible implementation, in addition to point coordinates, the element query output from the m-th layer is also retained. This element query, updated iteratively through the first m layers, has incorporated semantic information from multi-view BEV features, representing the overall attributes of the map element, such as category features (lane line or curb) and association features with other elements. The updated element query, together with the predicted fixed number of point coordinates, constitutes the output of the first m-layer decoding stage. The point coordinates provide a preliminary geometric skeleton for subsequent calculations of curvature and path complexity; the element query provides semantic-level feature support to aid in the calculation of complexity scores.

[0092] Preferably, the hybrid query is input into the first m layers of the hybrid decoder. Each layer performs point-level feature extraction, element-level feature extraction, and cross-level query update according to the process of S501-S503. In the m-th layer of the hybrid decoder, the point query in the hybrid query is input into the point coordinate prediction head to predict the point coordinates of the m-th layer, satisfying the following formula:

[0093]

[0094] in, The coordinates of the predicted point in the (m-1)th layer. Let m be the coordinates of the predicted point. For the predicted coordinate residuals, , These are the weights and biases of the linear layer, which are gradually optimized through backpropagation. This represents the sigmoid function. This represents the inverse function of the Sigmoid function, used to map normalized coordinates from the interval (0,1) to the real number field.

[0095] This application embodiment decodes the hybrid query layer by layer in the first m layers of the hybrid decoder, and updates the point query and element query simultaneously during the decoding process. At the same time, a fixed number of point coordinates are directly predicted in the intermediate layers. This allows the initial geometric structure of map elements to be quickly and explicitly modeled in the shallower network. This not only improves the stability and convergence efficiency of point coordinate generation, but also provides reliable basic data for subsequent complexity evaluation and adaptive resampling. This avoids the problem of gradual degradation of geometric information in deep decoding and effectively improves the efficiency and accuracy of the overall high-precision map construction process.

[0096] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 7 As shown, the above S206 can be specifically implemented through the following S701 to S704, which are explained in detail below: S701. Based on a fixed number of point coordinates, calculate the cumulative arc length parameter according to the Euclidean distance between adjacent points, and normalize the point coordinates to obtain the arc length parameter sequence.

[0097] In one possible implementation, the point coordinates output from the first m layers can roughly outline the shape of the map elements, but the number of points is fixed and cannot be directly used for subsequent adaptive decoding. In order to maintain the geometric properties of the curve while adjusting the number of points, the cumulative arc length of the original point sequence is first calculated.

[0098] Preferred cumulative arc length parameter Satisfy the following formula:

[0099]

[0100] in, Let the coordinates be the (j+1)th point of the i-th element. Let j be the coordinates of the i-th element. It is the L2 norm. This represents the straight-line distance between the j-th point and the (j+1)-th point. Let be the cumulative arc length from point 1 to point j. ,when The interpolation step is skipped, and the original point coordinates are directly used as the output.

[0101] It should be understood that the cumulative arc length parameter reflects the actual geometric position of the original point sequence on the curve. By calculating and summing the L2 distances between adjacent points, a parameterized curve with a geometry completely consistent with the original point sequence can be obtained. This step is crucial for ensuring resampling accuracy. Compared to simple linear interpolation (such as uniform sampling along the x-axis or y-axis), arc length parameterization avoids uneven distribution of sampling points caused by curve curvature. For example, for a highly curved arc, if linear interpolation along the x-axis is used, the sampling points will become sparse at the points of severe curvature, leading to shape distortion; however, with arc length parameterization, regardless of the curve's curvature, the sampling points can be evenly distributed according to the actual arc length, thus accurately preserving the original shape.

[0102] S702. Based on the number of vector points, perform equidistant sampling on the arc length parameter sequence to determine the target sampling position.

[0103] In one possible implementation, after obtaining the normalized arc length parameter sequence, the model generates equally spaced sampling positions in the interval [0, 1] based on the adaptive vector point number.

[0104] As an example, if the number of vector points of a map element is determined to be 40, then 40 equally spaced target sampling positions are generated between the normalized arc lengths of 0 and 1. This process is equivalent to remarking 40 uniformly distributed anchor points on the original curve. Through this equally spaced sampling, it is ensured that regardless of the distribution of the original point sequence, the resampled points can uniformly cover the entire map element, avoiding local over-density or under-density phenomena.

[0105] S703. Based on the target sampling location, a fixed number of point coordinates are resampled using interpolation to generate resampled point coordinates.

[0106] Here, the resampled point coordinates are the point coordinates corresponding to the number of vector points.

[0107] In one possible implementation, after determining the target sampling locations, the model calculates the precise coordinates corresponding to these locations using a linear interpolation algorithm. Linear interpolation uses a weighted average of the coordinates of adjacent original points in the arc length parameter sequence to generate new point coordinates.

[0108] As an example, if the target sampling position is located between the original j-th point and the (j+1)-th point, the coordinates of the two points are linearly weighted according to the ratio of the arc length distance from this position to these two points to obtain the coordinates of the new point. This interpolation method is simple and efficient in calculation, and combined with arc length parameterization, it can ensure that the coordinates of the generated point fall accurately on the trajectory of the original curve. When the number of vector points is greater than the original fixed number of points (e.g., increasing from 20 points to 40 points), this process is equivalent to denser sampling of the curve, which can supplement more geometric details; when the number of vector points is less than the original fixed number of points (e.g., decreasing from 20 points to 10 points), this process is equivalent to sparse sampling of the curve, which can remove redundant points and improve computational efficiency.

[0109] Preferred point coordinates after resampling And point query They respectively satisfy the following formulas:

[0110]

[0111]

[0112] in, For interpolation weights, Let be the interpolated coordinates of the k-th point of the i-th element. Let j be the coordinates of the j-th original point of the i-th element. For the k-th adaptive point, inherit the geometric and semantic information learned from the previous m layers for subsequent point queries. For the i-th element, query the j-th original point output at the m-th level. Indicates the first k The parameter positions corresponding to each adaptive point and Represents the first point in the original point sequence. j The original point and the first j+ The position of the normalized arc length parameter corresponding to a single original point, and satisfying the following conditions: , and Jointly determine the first k In the interpolation interval where the adaptive points are located, when the number of vector points is greater than the original fixed number of points, the sampling on the curve is densified to generate more point coordinates and point queries that match the real shape of the elements. This enables the subsequent decoding stage to model local details such as curves, corners, and boundary changes with higher geometric resolution, thereby improving the fitting accuracy of real map elements. When the number of vector points is less than the original fixed number of points, the sampling on the curve is sparse to reduce redundant calculations for simple elements such as straight lines and gentle curves. When the number of vector points is equal to the original fixed number of points, it degenerates into uniform reparameterization along the arc length, without changing the number of points but optimizing the uniformity of the distribution of each point on the curve.

[0113] S704. Based on the resampled point coordinates, determine the corresponding point query and combine it with the updated element query to construct a new hybrid query.

[0114] In one possible implementation, the model simultaneously generates corresponding point query features while generating the resampled point coordinates. The generation of point query features is similar to coordinate interpolation; linear interpolation is performed on the original point query features to obtain the feature representation of the new point position. These newly generated point query features inherit and integrate the geometric and semantic information learned from the first m layers. The resampled point queries are then concatenated with the updated element queries output from step S204 to construct a new hybrid query. This new hybrid query not only includes adaptively adjusted point count information but also retains the semantic features accumulated in previous stages, providing high-quality initialization input for the subsequent n layers of refined decoding.

[0115] It should be noted that the decoder in this application can be configured according to actual needs, such as using a transformer-based decoder, and this application does not limit this.

[0116] This application embodiment resamples a fixed number of point coordinates based on vector points using arc-length parameterization, and performs equidistant sampling and interpolation reconstruction in the arc-length space. This allows the resampled point coordinates to be more evenly distributed on the real geometric paths of map elements, effectively avoiding the problem of local information redundancy or missing information caused by uneven distribution of original points. At the same time, by reconstructing the point query corresponding to the point coordinates and recombining it with the element query, the new hybrid query can better match the expression needs of map elements with different complexities, further improving the precision and consistency of geometric modeling, providing a more reasonable and adaptive input representation for the subsequent decoding stage, thereby improving the overall accuracy and robustness of high-precision map construction.

[0117] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 8 As shown, the above S207 can be specifically implemented through the following S801 to S803, which are explained in detail below: S801. Input the new hybrid query into the last n layers of the hybrid decoder, and perform layer-by-layer interactive decoding by combining multi-view BEV features to iteratively update the point query and element query.

[0118] In one possible implementation, the new hybrid query is constructed after arc-length parameterized resampling, which includes adaptively adjusted point queries and updated element queries. In each of the subsequent n layers, the hybrid query continues to interact deeply with multi-view BEV features. It should be understood that, unlike the decoding process with a fixed number of points in the first m layers of embodiments S501-S504, the decoding process in the subsequent n layers has significant differences: the number of point queries corresponding to each map element is no longer uniform, but dynamically determined according to its actual complexity.

[0119] As an example, for a curved intersection with a high complexity score, the number of point queries might be 40. During feature sampling, the model extracts local geometric features for each of these 40 points, increasing the number of sampling points accordingly. Conversely, for a straight lane with a low complexity score, the number of point queries might only be 10, and the model only needs to process the feature interactions of these 10 points. This adaptive processing mechanism allows computational resources to be precisely concentrated on geometrically complex areas, ensuring the reconstruction accuracy of complex elements while avoiding computational redundancy for simple elements. In each iteration, point queries continuously optimize their coordinate positions, while element queries further aggregate global semantic information. Both achieve bidirectional feature fusion through an adaptive point-element interaction mechanism until the preset number of decoding layers is reached.

[0120] S802. Based on the updated point query and element query at the nth layer, predict the point coordinate information and category information of the corresponding map element.

[0121] In one possible implementation, in the final layer of the hybrid decoder, the model uses a point coordinate prediction head to decode the updated point query and predict the point coordinate information of the map element; simultaneously, it uses a classification prediction head to decode the updated element query and predict the category information of the map element. Point coordinate information characterizes the geometric shape and spatial location of the map element, while category information identifies the specific type of the element, such as lane lines, pedestrian crossings, or road boundaries. Because the point query continuously optimizes the local geometric features of the map element during the subsequent n layers of decoding, and the element query continuously aggregates the overall semantic and shape features of the map element, accurate prediction of both the geometric and semantic information of the map element can be achieved.

[0122] S803. Based on the point coordinate information and the corresponding category information, each map element is organized in a structured manner to generate a structured vector high-precision map.

[0123] In one possible implementation, the final output structured vector high-precision map is a structured result containing rich semantic and geometric information. For each detected map element, its data structure contains at least two parts: the first part is a sequence of geometric coordinates, i.e., an adaptive number of point coordinates obtained after n layers of refinement by the decoder, which accurately depict the geometric shape of the element; the second part is a semantic category label, used to indicate the attribute of the element. By jointly organizing the point coordinate information and category information, a structured vector high-precision map with both geometric structure and semantic attributes can be generated.

[0124] This application embodiment decodes new hybrid queries layer by layer in the last n layers of the hybrid decoder and continuously iterates and optimizes point queries and element queries by combining multi-view BEV features. This allows the resampled geometric and semantic information to be further integrated and refined in the deep network. At the same time, based on the updated point and element queries, the point coordinate information and category information of map elements are predicted respectively, realizing the unified modeling of geometric representation and semantic structure. This enables the structured organization of each map element, improves the completeness of the high-precision map representation, and ultimately improves the accuracy and practicality of the high-precision map construction results.

[0125] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0126] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of this application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and modifications of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and modifications.

Claims

1. A high-precision map construction method based on adaptive point-element interaction, characterized in that, include: Acquire multi-view image data of the vehicle; Multi-view BEV features of the multi-view image data are extracted using a multi-view feature extraction model; Based on the multi-view BEV features, a hybrid query including point query and element query is constructed; Based on the multi-view BEV features, the hybrid query is decoded through the first m layers of the hybrid decoder to obtain a fixed number of point coordinates corresponding to each map element and the updated element query. Based on the geometric features of the point coordinates and the semantic features of the updated element query, the complexity score of each map element is calculated, and the number of vector points and feature sampling points of each map element are determined according to the complexity score. Based on the number of vector points, the coordinates of the fixed number of points are resampled using arc length parameterization to generate the coordinates of points and point queries corresponding to the number of vector points, and a new hybrid query is constructed based on the coordinates of points and the updated element query. Based on the multi-view BEV features, the new hybrid query is decoded through the last n layers of the hybrid decoder to obtain a structured vector high-precision map.

2. The method according to claim 1, characterized in that, The step of extracting multi-view BEV features from the multi-view image data using a multi-view feature extraction model includes: Feature extraction is performed on multi-view image data using a backbone network to obtain multi-view image features; The multi-view image features are input into the FPN network for multi-scale feature fusion to obtain the fused multi-view image features. Based on camera calibration parameters or a learning-based perspective transformation method, the fused multi-view image features are transformed from the image perspective to the bird's-eye view perspective using a 2D-to-BEV perspective transformation method to obtain multi-view BEV features.

3. The method according to claim 1, characterized in that, Based on the multi-view BEV features, a hybrid query including point query and element query is constructed, including: The basic hybrid query is randomly initialized as a learnable embedding vector and decomposed into point queries and element queries; In each layer of the hybrid decoder, the hybrid query is iteratively optimized through point-level feature extraction paths and element-level feature extraction paths to obtain the hybrid query.

4. The method according to claim 3, characterized in that, In each layer of the hybrid decoder, the hybrid query is iteratively optimized using point-level feature extraction paths and element-level feature extraction paths to obtain the hybrid query, including: In each layer of the hybrid decoder, based on the point query and point coordinates output by the previous layer, combined with position embedding, feature sampling and fusion are performed from the multi-view BEV features to obtain point-level features; Based on the element query output of the previous layer, combined with the position embedding generated by the point query, element-level features are extracted from the multi-view BEV features through the mask attention mechanism. Based on the point-level features and element-level features, adaptive point-element interaction is performed to fuse the bidirectional information between the point-level features and element-level features, and to update the point query and element query to obtain the updated point query and element query. The updated point query and element query are concatenated and combined to generate a hybrid query for the current layer; The hybrid query of the current layer is used as the input of the next layer, and the hybrid query after layer-by-layer iteration optimization is output in the last layer.

5. The method according to claim 1, characterized in that, Based on the multi-view BEV features, the hybrid query is decoded through the first m layers of the hybrid decoder to obtain a fixed number of point coordinates corresponding to each map element and the updated element query, including: The hybrid query is input into the first m layers of the hybrid decoder, and interactively decoded with multi-view BEV features in each layer, updating the point query and element query layer by layer; After decoding through the first m layers, the corresponding point coordinates are predicted based on the point query output of the m-th layer. Output the element query corresponding to the m-th layer, and participate in the complexity score calculation of map elements.

6. The method according to claim 1, characterized in that, The complexity score of the map element includes curvature complexity and path complexity; Wherein, the curvature complexity Satisfy the following formula: in, cuv represents the curvature complexity of the i-th map element in the m-th layer, where cuv represents curvature and i represents the map element index. Represents vector coordinate points The angle between the direction vectors, where j represents the index of the vector coordinate point, and takes the value of . , This indicates the number of coordinate points output by the m-th layer; Indicates variance; The path complexity Satisfy the following formula: in, Let len ​​represent the path complexity of the i-th map element at the m-th level, where len represents the path and i represents the map element index. , This represents the vector coordinate point, and j represents the index of the vector coordinate point; This indicates the number of coordinate points output by the m-th layer. It is an L2 norm; The complexity score Satisfy the following formula: in, This represents the query element of the i-th map element at the m-th level. Let represent the curvature complexity of the i-th map element in the m-th layer. This represents the path complexity of the i-th map element at the m-th layer; Indicates a splicing operation; and These are the learnable weight matrix and bias parameters of the first linear layer of the complexity scoring network, respectively; ReLU() is the activation function; and The learnable weight matrix and scalar bias of the second linear layer of the complexity scoring network; This is the Sigmoid activation function.

7. The method according to claim 1, characterized in that, Based on the number of vector points, the fixed number of point coordinates are resampled using arc length parameterization to generate point coordinates and point queries corresponding to the number of vector points. A new hybrid query is then constructed based on the point coordinates and the updated element query, including: Based on a fixed number of point coordinates, the cumulative arc length parameter is calculated according to the Euclidean distance between adjacent points, and the point coordinates are normalized to obtain the arc length parameter sequence. Based on the number of vector points, perform equally spaced sampling on the arc length parameter sequence to determine the target sampling position; Based on the target sampling position, the fixed number of point coordinates are resampled using interpolation to generate resampled point coordinates; the resampled point coordinates are the point coordinates corresponding to the number of vector points. Based on the resampled point coordinates, the corresponding point query is determined and combined with the updated element query to construct a new hybrid query.

8. The method according to claim 7, characterized in that, The cumulative arc length parameter Satisfy the following formula: in, Let the coordinates be the (j+1)th point of the i-th element. Let j be the coordinates of the i-th element. It is the L2 norm. This represents the straight-line distance between the j-th point and the (j+1)-th point. Let be the cumulative arc length from point 1 to point j. ,when The interpolation step is skipped, and the original point coordinates are directly used as the output.

9. The method according to claim 1, characterized in that, The process of decoding the new hybrid query based on the multi-view BEV features through the last n layers of the hybrid decoder to obtain a structured vector high-precision map includes: The new hybrid query is input into the last n layers of the hybrid decoder and combined with multi-view BEV features for layer-by-layer interactive decoding, and the point query and element query are iteratively updated. Based on the updated point and element queries at the nth layer, predict the point coordinates and category information of the corresponding map elements; Based on the point coordinates and corresponding category information, each map element is organized in a structured manner to generate a structured vector high-precision map.

10. A high-precision map construction system based on adaptive point-element interaction, characterized in that, include: Data acquisition module, feature extraction module, hybrid query construction module, front-end decoding module, complexity evaluation module, resampling and query reconstruction module, and back-end decoding module; The data acquisition module is used to acquire multi-view image data of the vehicle. The feature extraction module is used to extract multi-view BEV features of the multi-view image data through a multi-view feature extraction model; The hybrid query construction module is used to construct a hybrid query including point query and element query based on the multi-view BEV features; The front-end decoding module is used to decode the hybrid query based on the multi-view BEV features through the first m layers of the hybrid decoder to obtain a fixed number of point coordinates corresponding to each map element and the updated element query. The complexity evaluation module is used to calculate the complexity score of each map element based on the geometric features of the point coordinates and the semantic features of the updated element query, and to determine the number of vector points and feature sampling points of each map element according to the complexity score. The resampling and query reconstruction module is used to perform arc-length parameterized resampling of the coordinates of a fixed number of points based on the number of vector points, generate the coordinates of points and the query corresponding to the number of vector points, and construct a new hybrid query based on the coordinates of points and the updated element query. The back-end decoding module is used to decode the new hybrid query based on the multi-view BEV features through the last n layers of the hybrid decoder to obtain a structured vector high-precision map.

Images