Vector high-precision map construction method and device, electronic equipment and medium

Abstract

The application discloses a kind of vectorization high-precision map construction method, device, electronic equipment and medium, it uses segmented Bezier curve to carry out parameterization representation to map element, and can guide image feature sampling according to curve structure by adaptive offset sampling, so that model can pay more attention to the feature information of the area where map element is located, to accurately capture the irregular, flexible linear characteristics of map element, both achieve high geometric fit degree, and improve detection accuracy, by pure sparse norm, completely avoid traditional dense BEV grid construction step, and the amount of calculation is significantly reduced.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving technology, and particularly relates to a vectorized high-precision map construction method, device, electronic device and medium. Background Technology

[0002] With the development of autonomous driving technology, high-definition maps (HD maps) play a crucial role in environmental perception and path planning. HD maps typically represent road structure information in vector form, including map elements such as lane lines, road boundaries, pedestrian crossings, and road segments. Compared to traditional raster maps, vectorized maps can more accurately represent the geometric relationships of road structures, thus holding significant importance in autonomous driving systems.

[0003] Existing methods for constructing vectorized high-precision maps typically rely on multi-view image data collected by vehicles, using deep learning models to detect and extract road structure information from the images. For example, some methods generate road structures by constructing a Bird's Eye View (BEV) feature representation and then performing semantic segmentation or object detection on the BEV features. However, such methods usually require constructing a dense BEV feature grid, resulting in high computational complexity.

[0004] Another approach is to use point sets to represent map elements, which lacks coherence and results in fluctuating detection results.

[0005] In addition, some studies have attempted to use Bezier curves to parametrically represent map elements. Although the geometric fit is high, the detection accuracy is far behind that of advanced point set methods. Furthermore, due to the irregular distribution characteristics of map elements, the pure sparse detection paradigm is difficult to achieve. Existing methods usually still rely on dense, scale-constrained feature representations (BEV feature grids). Summary of the Invention

[0006] Based on this, and in response to the aforementioned technical problems, a method, apparatus, electronic device, and medium for constructing vectorized high-precision maps are provided.

[0007] The technical solution adopted in this invention is as follows: As a first aspect of the present invention, a method for constructing a vectorized high-precision map is provided, comprising: Acquire multi-view perspective views and standard map data; Extract the image features of the multi-view perspective view and the map features of the standard map data; The image features, map features, query vectors representing map elements, and current piecewise Bézier curve parameters are input into a piecewise Bézier curve decoder. The piecewise Bézier curve decoder includes a decoding layer, which includes a piecewise Bézier curve attention module. The piecewise Bézier curve attention module extracts features around map elements from the image features by adaptively offset sampling the sampling point positions on the piecewise Bézier curve and interacts with the query vector. The decoding layer predicts the offset of the piecewise Bézier curve parameters based on the result of the feature interaction and updates the current piecewise Bézier curve parameters according to the offset. Vectorized representations of map elements are generated using updated piecewise Bézier curve parameters.

[0008] As a second aspect of the present invention, a vectorized high-precision map construction apparatus is provided, comprising: The first module is used to acquire multi-view perspective views and standard map data; The second module is used to extract the image features of the multi-view perspective view and the map features of the standard map data; The third module is used to input the image features, map features, query vectors representing map elements, and current piecewise Bézier curve parameters into a piecewise Bézier curve decoder. The piecewise Bézier curve decoder includes a decoding layer, which includes a piecewise Bézier curve attention module. The piecewise Bézier curve attention module extracts features around map elements from the image features by adaptively offset sampling the sampling point positions on the piecewise Bézier curve and interacts with the query vector. The decoding layer predicts the offset of the piecewise Bézier curve parameters based on the result of the feature interaction and updates the current piecewise Bézier curve parameters according to the offset. The fourth module is used to generate vectorized representations of map elements using updated piecewise Bézier curve parameters.

[0009] As a third aspect of the present invention, an electronic device is provided, including a storage module, the storage module including instructions loaded and executed by a processor, the instructions causing the processor to perform the vectorized high-precision map construction method described in the first aspect above when executed.

[0010] As a fourth aspect of the present invention, a computer-readable storage medium is provided that stores one or more programs, which, when executed by a processor, implement the vectorized high-precision map method described in the first aspect above.

[0011] This invention uses piecewise Bézier curves to parametrically represent map elements, and through adaptive offset sampling, it can guide image feature sampling according to the curve structure, enabling the model to pay more attention to the feature information of the area where the map element is located, thereby accurately capturing the irregular and flexible linear features of the map element. It achieves both high geometric fit and improved detection accuracy. By using a pure sparse paradigm (i.e., without constructing any intermediate dense BEV features), it completely avoids the traditional dense BEV grid construction steps, significantly reducing the amount of computation. Attached Figure Description

[0012] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments: Figure 1 A flowchart of a vectorized high-precision map construction method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a vectorized high-precision map construction device provided in an embodiment of the present invention; Figure 3 A schematic diagram of an electronic device provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of SparseMapNet, a vectorized high-precision map construction network based on the pure sparse paradigm, according to an embodiment of the present invention. Detailed Implementation

[0013] The embodiments of the present invention will be described below with reference to the accompanying drawings. It should be noted that the embodiments described in this specification are not exhaustive and do not represent the only embodiments of the present invention. The corresponding embodiments below are only for clearly illustrating the inventive content of this patent and are not intended to limit its implementation. For those skilled in the art, different variations and modifications can be made based on the embodiments described. Any variations or modifications that fall within the technical concept and inventive content of this invention and are obvious are also within the protection scope of this invention.

[0014] like Figure 1 As shown, this application provides a method for constructing vectorized high-precision maps. This method is based on the SparseMapNet vectorized high-precision map construction network with a pure sparse paradigm. The network structure is described in [link to network diagram]. Figure 4 SparseMapNet follows the DETR-Family ensemble prediction paradigm and consists of two feature extractors and a sparse piecewise Bezier decoder. The specific process is as follows: S101. Obtain multi-view perspective views and standard map data.

[0015] Among them, multi-view perspective (Multi-PV) Figure 4 Multi-view maps (hereinafter referred to as multi-view maps) can be acquired by multiple cameras installed around the vehicle, such as front-view cameras, side-view cameras, and rear-view cameras. Standard map data can come from existing map datasets, such as open map data or other high-precision map data. In this embodiment, the standard map data comes from OpenStreetMap's labeled map (SDMap).

[0016] S102. Two feature extractors extract image features from the multi-view perspective view and map features from the standard map data (SDMap Features).

[0017] Specifically, the Multi-PV Feature Extractor extracts image features from the multi-view perspective. The Multi-PV Feature Extractor can use a combination of Backbone and Encoder to perform multi-scale feature extraction on the multi-view perspective, obtaining image features (2D feature maps of multiple views and scales). The Backbone and Encoder can be flexibly combined; the Backbone can use ResNet, VoVNet, etc., for feature extraction, while the Encoder can use FPN and Deformable Attention for feature encoding and enhancement.

[0018] The SDMap Feature Extractor extracts map features from standard map data. It can be modeled after SMERF, using both location encoding and one-hot heatmap encoding. The two encoding results are concatenated, and then a convolutional module is used to extract features to obtain the SDMap feature map.

[0019] S103. Input the image features, map features, query vectors used to represent map elements, and current piecewise Bézier curve parameters into the piecewise Bézier curve decoder.

[0020] The segmented Bézier curve decoder includes multiple decoding layers ( Figure 4The diagram illustrates the i-th decoding layer (SparsePiecewise Bezier Decoder Layer-i). Each decoding layer includes an inter-element attention module, a piecewise Bezier curve attention module, an intra-element attention module, and a standard map feature attention module (SD Map Feature Attention). The input to each decoding layer includes image features, map features, a query vector, and piecewise Bezier curve parameters. It predicts the offset of the piecewise Bezier curve parameters and updates the current piecewise Bezier curve parameters based on the offset.

[0021] During the decoding process, such as Figure 4 As shown, each attention module is executed sequentially according to a preset order.

[0022] The inter-element attention module enhances the feature representation of map elements through information interaction between query vectors. Specifically, the input to the inter-element attention module comes from the piecewise Bézier curve parameters output from the previous decoding layer and the updated query vector (the query vector of the i-th layer (E)). i (i∈[1,6])). The query vector input to the first decoding layer is obtained through random initialization, and the piecewise Bézier curve parameters (B0) input to the first decoding layer are predicted by the multilayer perceptron (MLP) based on image features, which provides a good starting point for the decoder, accelerates the model training convergence process, and ultimately achieves excellent detection performance.

[0023] First, based on the current piecewise Bézier curve parameters (i.e., the input piecewise Bézier curve parameters), several sampling points are sampled on the curve (Piecewise Bézier Curve Sampling), and the corresponding two-dimensional position coordinates are obtained through 3D to 2D image plane projection (Points Perspective Projection). After positional encoding (PE) of the position coordinates, they are concatenated with the query vector and used as input for the inter-element attention module. Figure 4 In the input, the piecewise Bézier curve parameters are the Bézier curve parameters of the i-th layer (B i (i∈[1,6]).

[0024] Through this processing, the module can perform feature interaction between query vectors. By mixing and enhancing the content and location information of each query vector, information interaction between query vectors is achieved, thereby improving the ability of query vectors to express the overall features of map elements and providing updated query vectors for the subsequent piecewise Bézier curve attention module.

[0025] After the feature interactions of the inter-element attention module are completed, the updated query vector is fed into the piecewise Bézier curve attention module. This module is used to extract only the features surrounding the map elements from the image features for interaction.

[0026] Specifically, the piecewise Bézier curve attention module fully leverages the sparsity paradigm. First, it generates several sample points on the curve based on the current piecewise Bézier curve parameters. Then, it performs grid sampling on the image using these points to obtain grid image features. Next, it adaptively offsets the sample points to obtain the offset sampling positions. In this embodiment, based on the grid image features, it calculates the offsets of the sample points and the corresponding attention weights using a multilayer perceptron (MLP), such as using nn.Linear to calculate the offsets. These offsets are then superimposed onto the sample point positions to obtain the offset sampling positions. Finally, based on the offset sampling positions, it performs feature sampling on the image feature map and performs weighted aggregation on the sampled features according to the attention weights to obtain local feature representations (samplefeats) around the map elements, which are then used to update the query vector.

[0027] The aforementioned sampling point offset and feature aggregation mechanism can draw inspiration from the DeformableAttention concept in Deformable DETR. This involves adaptively offsetting a small number of key sampling points to extract sparse features and fuse attention from image features. Unlike DeformableAttention, which samples on an image feature grid, this invention applies this sparse attention mechanism to sampling points on piecewise Bézier curves. By offsetting and aggregating the curve sampling points, it can capture environmental information surrounding map elements in the vicinity of the curve, primarily serving the subsequent prediction of the offset of the piecewise Bézier curve parameters.

[0028] Through the above process, the piecewise Bézier curve attention module can guide image feature sampling according to the curve structure, enabling the model to pay more attention to the feature information of the area where map elements (curves) are located while maintaining a large receptive field. This allows for the accurate capture of irregular and flexible linear features of slender curve structures such as lane lines, enhancing the modeling ability of curve shapes and avoiding the loss of local features despite a large receptive field.

[0029] The updated query vector is then fed into the intra-element attention module. The intra-element attention module is used to augment and sparsify the query vectors of the same map element. Its purpose is to filter redundant features, making the query vector features more closely match the geometric representation of lane lines or other map elements, and improving the efficiency of sparse features in map element representation.

[0030] Subsequently, the updated query vector is input into the standard map feature attention module. The standard map feature attention module is used to introduce global SDMap features, which interact with the current query vector to enhance the representation and perception of lane lines in the distance, compensating for the shortcomings of pure visual perception in distant areas.

[0031] After passing through the attention modules described above, the decoding layer predicts the offset of the piecewise Bézier curve parameters based on the updated query vector. Specifically, a regression network outputs the corresponding parameter offset based on the query vector, and this offset is superimposed on the current piecewise Bézier curve parameters to obtain the updated piecewise Bézier curve parameters. This purely sparse decoding mechanism based on attention fundamentally avoids the computational bottleneck of constructing a dense BEV feature grid, and is the core of this scheme for achieving long-range, high real-time performance.

[0032] In the multi-layer decoding structure, subsequent decoding layers use the piecewise Bézier curve parameters updated by the previous decoding layer as input and repeat the above process. The computation is performed in a purely sparse manner (i.e., without constructing any intermediate dense BEV features). The output of each decoding layer is not the final result, but rather a refined offset of the Bézier curve parameters. Through a recursive optimization approach of updating the piecewise Bézier curve parameters layer by layer, the true geometric shape of the map elements is gradually approximated. By updating multiple decoding layers layer by layer, the prediction accuracy of the piecewise Bézier curve parameters can be gradually improved, thereby obtaining a more accurate vectorized representation of map elements.

[0033] Finally, after completing the calculations for all decoding layers, the final updated piecewise Bézier curve parameters are obtained.

[0034] In this embodiment, the piecewise Bézier curve parameters include the number of segments, the endpoints of each Bézier curve segment, and the offset of the coordinates of intermediate control points. This invention uses piecewise Bézier curves to vectorize map elements. Different curve orders are used to address the structural complexity of different element categories. Road elements are divided into three main categories: road dividers and lane dividers, pedestrian crossings, and road segments and lanes. For these three categories, dynamic piecewise Bézier curves of orders 3, 2, and 4 are used respectively. The number of segments for each curve is adaptively predicted by the network based on the actual length and shape complexity of the element, thereby achieving flexible modeling of elements at different scales.

[0035] Each decoding layer contains a Piecewise Bézier CurveHead, which includes both classification and regression components. Map elements are categorized into three classes, each corresponding to a number of segments for class encoding. The regression function combines the class information to adaptively regress the coordinates of intermediate control points.

[0036] In this embodiment, chamfer distance is used. and Distance is used to measure the distance between the sampled point sets of the predicted piecewise Bézier curve, and is used in the loss function and label matching. In addition, the loss function also includes Focal's classification loss. and the Dice loss after expanding the region of the piecewise Bézier curve. ,in .

[0037] S104. Generate a vectorized representation of map elements using the updated piecewise Bézier curve parameters.

[0038] As can be seen from the above, the vectorized high-precision map construction method provided in this application uses piecewise Bézier curves to parametrically represent map elements. Its inherent continuous geometric constraints make the output results inherently smooth and stable compared to point set representation, effectively reducing jitter. Moreover, through adaptive offset sampling, it can guide image feature sampling according to the curve structure, enabling the model to pay more attention to the feature information of the area where the map element is located, thereby accurately capturing the irregular and flexible linear features of the map element. It achieves both high geometric fit and improved detection accuracy. Through the pure sparse paradigm, it completely avoids the traditional dense BEV grid construction steps, and the perception line distance is no longer limited by the grid size, significantly reducing the amount of computation and achieving a unity of long line distance perception and high real-time performance.

[0039] By introducing SDMap as prior information for feature fusion, the network is provided with powerful context and long-distance guidance, which significantly improves the detection accuracy and reliability of map elements in distant and occluded areas.

[0040] While piecewise Bézier curves can closely fit the shapes of map elements, networks struggle to learn to adapt them effectively, making training difficult. This invention utilizes image features for efficient initialization, providing a reliable adaptive query location reference for the query. It optimizes the position of the parameter curves layer by layer, making it easier for the network to learn the required features, successfully applying the sparse paradigm to the task of constructing vectorized high-precision maps.

[0041] In high-precision map building tasks, achieving high accuracy for distant map elements is often challenging. This invention combines readily available SDMaps with monitoring data to provide a supervisory reference, ensuring the detection accuracy of distant map elements.

[0042] The vectorized high-precision map building apparatus of one or more embodiments of the present invention will be described in detail below. Those skilled in the art will understand that these apparatuses can be configured using commercially available hardware components through the steps taught in this solution. Figure 2 The present invention illustrates a vectorized high-precision map construction device, which includes a first module 11, a second module 12, a third module 13, and a fourth module 14.

[0043] The first module 11 is used for S101, acquiring multi-view perspective views and standard map data.

[0044] The multi-view perspective (Multi-PV) can be acquired by multiple cameras installed around the vehicle, such as front-view cameras, side-view cameras, and rear-view cameras. The standard map data can come from existing map datasets, such as open map data or other high-precision map data. In this embodiment, the standard map data comes from OpenStreetMap's labeled map (SDMap).

[0045] The second module 12 is used in S102 to extract image features of the multi-view perspective view and map features of the standard map data by two feature extractors respectively.

[0046] Specifically, the Multi-PV Feature Extractor extracts image features from the multi-view perspective. The Multi-PV Feature Extractor can use a combination of Backbone and Encoder to perform multi-scale feature extraction on the multi-view perspective, obtaining image features (2D feature maps of multiple views and scales). The Backbone and Encoder can be flexibly combined; the Backbone can use ResNet, VoVNet, etc., for feature extraction, while the Encoder can use FPN, Deformable Attention, etc., for feature encoding and enhancement.

[0047] The SDMap Feature Extractor extracts map features from standard map data. It can be modeled after SMERF, using both location encoding and one-hot heatmap encoding. The two encoding results are concatenated, and then a convolutional module is used to extract features to obtain the SDMap feature map.

[0048] The third module 13 is used in S103 to input image features, map features, query vectors (Query) used to represent map elements, and current piecewise Bézier curve parameters into the piecewise Bézier curve decoder.

[0049] The segmented Bézier curve decoder includes multiple decoding layers ( Figure 4 The diagram illustrates the i-th decoding layer (SparsePiecewise Bezier Decoder Layer-i). Each decoding layer includes an inter-element attention module, a piecewise Bezier curve attention module, an intra-element attention module, and a standard map feature attention module (SD Map Feature Attention). The input to each decoding layer includes image features, map features, a query vector, and piecewise Bezier curve parameters. It predicts the offset of the piecewise Bezier curve parameters and updates the current piecewise Bezier curve parameters based on the offset.

[0050] During the decoding process, such as Figure 4 As shown, each attention module is executed sequentially according to a preset order.

[0051] The inter-element attention module enhances the feature representation of map elements through information interaction between query vectors. Specifically, the input to the inter-element attention module comes from the piecewise Bézier curve parameters output from the previous decoding layer and the updated query vector. The query vector input to the first decoding layer is randomly initialized, and the piecewise Bézier curve parameters input to the first decoding layer are predicted by a multilayer perceptron (MLP) based on image features, providing a good starting point for the decoder, accelerating the model training convergence process, and ultimately achieving excellent detection performance.

[0052] First, based on the current piecewise Bézier curve parameters (i.e., the input piecewise Bézier curve parameters), several sampling points are sampled on the curve (Piecewise Bézier Curve Sampling), and the corresponding two-dimensional position coordinates are obtained through 3D to 2D image plane projection (Points Perspective Projection). After positional encoding (PE) of the position coordinates, they are concatenated with the query vector and used as input for the inter-element attention module.

[0053] Through this processing, the module can perform feature interaction between query vectors. By mixing and enhancing the content and location information of each query vector, information interaction between query vectors is achieved, thereby improving the ability of query vectors to express the overall features of map elements and providing updated query vectors for the subsequent piecewise Bézier curve attention module.

[0054] After the feature interactions of the inter-element attention module are completed, the updated query vector is fed into the piecewise Bézier curve attention module. This module is used to extract only the features surrounding the map elements from the image features for interaction.

[0055] Specifically, the piecewise Bézier curve attention module fully leverages the sparsity paradigm. First, it generates several sample points on the curve based on the current piecewise Bézier curve parameters. Then, it performs grid sampling on the image using these points to obtain grid image features. Next, it adaptively offsets the sample points to obtain the offset sampling positions. In this embodiment, based on the grid image features, it calculates the offsets of the sample points and the corresponding attention weights using a multilayer perceptron (MLP), such as using nn.Linear to calculate the offsets. These offsets are then superimposed onto the sample point positions to obtain the offset sampling positions. Finally, based on the offset sampling positions, it performs feature sampling on the image feature map and performs weighted aggregation on the sampled features according to the attention weights to obtain local feature representations (samplefeats) around the map elements, which are then used to update the query vector.

[0056] The aforementioned sampling point offset and feature aggregation mechanism can draw inspiration from the DeformableAttention concept in Deformable DETR. This involves adaptively offsetting a small number of key sampling points to extract sparse features and fuse attention from image features. Unlike DeformableAttention, which samples on an image feature grid, this invention applies this sparse attention mechanism to sampling points on piecewise Bézier curves. By offsetting and aggregating the curve sampling points, it can capture environmental information surrounding map elements in the vicinity of the curve, primarily serving the subsequent prediction of the offset of the piecewise Bézier curve parameters.

[0057] Through the above process, the piecewise Bézier curve attention module can guide image feature sampling according to the curve structure, enabling the model to pay more attention to the feature information of the area where map elements (curves) are located while maintaining a large receptive field. This allows for the accurate capture of irregular and flexible linear features of slender curve structures such as lane lines, enhancing the modeling ability of curve shapes and avoiding the loss of local features despite a large receptive field.

[0058] The updated query vector is then fed into the intra-element attention module. The intra-element attention module is used to augment and sparsify the query vectors of the same map element. Its purpose is to filter redundant features, making the query vector features more closely match the geometric representation of lane lines or other map elements, and improving the efficiency of sparse features in map element representation.

[0059] Subsequently, the updated query vector is input into the standard map feature attention module. The standard map feature attention module is used to introduce global SDMap features, which interact with the current query vector to enhance the representation and perception of lane lines in the distance, compensating for the shortcomings of pure visual perception in distant areas.

[0060] After passing through the attention modules described above, the decoding layer predicts the offset of the piecewise Bézier curve parameters based on the updated query vector. Specifically, a regression network outputs the corresponding parameter offset based on the query vector, and this offset is superimposed on the current piecewise Bézier curve parameters to obtain the updated piecewise Bézier curve parameters. This purely sparse decoding mechanism based on attention fundamentally avoids the computational bottleneck of constructing a dense BEV feature grid, and is the core of this scheme for achieving long-range, high real-time performance.

[0061] In the multi-layer decoding structure, subsequent decoding layers use the piecewise Bézier curve parameters updated by the previous decoding layer as input and repeat the above process. The computation is performed in a purely sparse manner (i.e., without constructing any intermediate dense BEV features). The output of each decoding layer is not the final result, but rather a refined offset of the Bézier curve parameters. Through a recursive optimization approach of updating the piecewise Bézier curve parameters layer by layer, the true geometric shape of the map elements is gradually approximated. By updating multiple decoding layers layer by layer, the prediction accuracy of the piecewise Bézier curve parameters can be gradually improved, thereby obtaining a more accurate vectorized representation of map elements.

[0062] Finally, after completing the calculations for all decoding layers, the final updated piecewise Bézier curve parameters are obtained.

[0063] In this embodiment, the piecewise Bézier curve parameters include the number of segments, the endpoints of each Bézier curve segment, and the offset of the coordinates of intermediate control points. This invention uses piecewise Bézier curves to vectorize map elements. Different curve orders are used to address the structural complexity of different element categories. Road elements are divided into three main categories: road dividers and lane dividers, pedestrian crossings, and road segments and lanes. For these three categories, dynamic piecewise Bézier curves of orders 3, 2, and 4 are used respectively. The number of segments for each curve is adaptively predicted by the network based on the actual length and shape complexity of the element, thereby achieving flexible modeling of elements at different scales.

[0064] Each decoding layer contains a Piecewise Bézier CurveHead, which includes both classification and regression components. Map elements are categorized into three classes, each corresponding to a number of segments for class encoding. The regression function combines the class information to adaptively regress the coordinates of intermediate control points.

[0065] In this embodiment, chamfer distance is used. and Distance is used to measure the distance between the sampled point sets of the predicted piecewise Bézier curve, and is used in the loss function and label matching. In addition, the loss function also includes Focal's classification loss. and the Dice loss after expanding the region of the piecewise Bézier curve. ,in .

[0066] Module 4, 14, is used in S104 to generate a vectorized representation of map elements using updated piecewise Bézier curve parameters.

[0067] In summary, the vectorized high-precision map construction apparatus provided in the above embodiments can execute the vectorized high-precision map construction methods provided in the foregoing embodiments.

[0068] Similar to the above concept, Figure 3 A schematic block diagram of the structure of an electronic device provided by an embodiment of the present invention is shown.

[0069] For example, the electronic device includes a storage module 21 and a processor 22. The storage module 21 includes instructions loaded and executed by the processor 22, which, when executed, cause the processor 22 to perform the steps described in the above section of this specification, "A Method for Constructing a Vectorized High-Precision Map," according to various exemplary embodiments of the present invention.

[0070] It should be understood that processor 22 can be a Central Processing Unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, the general-purpose processor can be a microprocessor or any conventional processor.

[0071] This invention also provides a computer-readable storage medium that stores one or more programs, which, when executed by a processor, implement the steps described in the above section on a vectorized high-precision map construction method according to various exemplary embodiments of the invention.

[0072] Those skilled in the art will understand that all or some of the steps, systems, and apparatuses disclosed above, and their functional modules / units, can be implemented as software, firmware, hardware, or suitable combinations thereof. In hardware implementations, the division between functional modules / units mentioned above does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed collaboratively by several physical components. Some or all physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit (ASIC). Such software can be distributed on a computer-readable storage medium, which may include computer-readable storage media (or non-transitory media) and communication media (or transient media).

[0073] As is known to those skilled in the art, the term computer-readable storage medium includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer-readable storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, it is known to those skilled in the art that communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0074] For example, the computer-readable storage medium may be an internal storage unit of the electronic device described in the foregoing embodiments, such as a hard disk or memory of the electronic device. The computer-readable storage medium may also be an external storage device of the electronic device, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, etc., provided on the electronic device.

[0075] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for constructing a vectorized high-precision map, characterized in that, include: Acquire multi-view perspective views and standard map data; Extract the image features of the multi-view perspective view and the map features of the standard map data; The image features, map features, query vectors representing map elements, and current piecewise Bézier curve parameters are input into a piecewise Bézier curve decoder. The piecewise Bézier curve decoder includes a decoding layer comprising an inter-element attention module, a piecewise Bézier curve attention module, an intra-element attention module, and a standard map feature attention module. The inter-element attention module enhances the feature representation between map elements through information interaction between the query vectors before the piecewise Bézier curve attention module. The piecewise Bézier curve attention module extracts map features from the image features by adaptively offset sampling the sampling point positions on the piecewise Bézier curve. The features around the map element are interacted with the query vector. The intra-element attention module is used to perform information enhancement and sparsification processing on the query vector of the same map element after the piecewise Bézier curve attention module. The standard map feature attention module is used to interact with the map features and the query vector after the intra-element attention module. The decoding layer predicts the offset of the piecewise Bézier curve parameters based on the result of the feature interaction and updates the current piecewise Bézier curve parameters according to the offset. The inputs and outputs of the inter-element attention module, the intra-element attention module and the standard map feature attention module are respectively residual connected. Vectorized representations of map elements are generated using the updated piecewise Bézier curve parameters; The number of decoding layers is multiple, and the multiple decoding layers update the segmented Bézier curve parameters layer by layer in sequence.

2. The vectorized high-precision map construction method according to claim 1, characterized in that, The parameters of the piecewise Bézier curve include the number of segments, the endpoints of each Bézier curve segment, and the offset of the coordinates of the intermediate control points.

3. The vectorized high-precision map construction method according to claim 1, characterized in that, The query vector input to the first decoding layer is obtained through random initialization, and the piecewise Bézier curve parameters input to the first decoding layer are predicted by the multilayer perceptron based on the image features.

4. A vectorized high-precision map construction device, characterized in that, include: The first module is used to acquire multi-view perspective views and standard map data; The second module is used to extract the image features of the multi-view perspective view and the map features of the standard map data; The third module is used to input the image features, map features, query vectors representing map elements, and current piecewise Bézier curve parameters into the piecewise Bézier curve decoder. The piecewise Bézier curve decoder includes a decoding layer, which comprises an inter-element attention module, a piecewise Bézier curve attention module, an intra-element attention module, and a standard map feature attention module. The inter-element attention module enhances the feature representation between map elements through information interaction between the query vectors before the piecewise Bézier curve attention module. The piecewise Bézier curve attention module adaptively offsets the sampling point positions on the piecewise Bézier curve from the image features... The features around the map element are extracted and interacted with the query vector. The intra-element attention module is used to perform information enhancement and sparsification processing on the query vector of the same map element after the piecewise Bézier curve attention module. The standard map feature attention module is used to interact with the map features and the query vector after the intra-element attention module. The decoding layer predicts the offset of the piecewise Bézier curve parameters based on the result of the feature interaction and updates the current piecewise Bézier curve parameters according to the offset. The inputs and outputs of the inter-element attention module, the intra-element attention module and the standard map feature attention module are respectively residual connected. The fourth module is used to generate vectorized representations of map elements using the updated piecewise Bézier curve parameters; The number of decoding layers is multiple, and the multiple decoding layers update the segmented Bézier curve parameters layer by layer in sequence.

5. An electronic device, characterized in that, The system includes a storage module comprising instructions loaded and executed by a processor, which, when executed, cause the processor to perform a vectorized high-precision map construction method according to any one of claims 1-3.

6. A computer-readable storage medium storing one or more programs, characterized in that, When the one or more programs are executed by the processor, they implement the vectorized high-precision map construction method according to any one of claims 1-3.

Images