A high-precision map real-time construction method based on point-mask double-prior query

Abstract

The application relates to a high-precision map real-time construction method based on point-mask double-prior query, and the method comprises the following steps: S1. A vehicle travels on a road, data information of a vectorized navigation map of the road traveled by the vehicle is acquired, and data information of multi-view images of the road is acquired in real time based on a vehicle-mounted camera; S2. Based on the data information of the vectorized navigation map of the road traveled by the vehicle, road key points are extracted and sine-cosine position coding is performed, and combined with road category labels, structured high-dimensional point feature data information of the road traveled by the vehicle is generated through Transform coding. The application not only introduces a vectorized navigation point (point prior) and a BEV mask structure (mask prior) into a Transform query initialization stage, significantly improves the spatial expression capability and structural integrity of map construction, and can automatically perceive spatial feature distribution in complex scenes such as curved roads and intersections, so that more robust map construction is realized.

Description

Technical Field

[0001] This invention relates to the field of high-precision map construction technology, and in particular to a method for real-time construction of high-precision maps based on point-mask dual prior queries. Background Technology

[0002] Autonomous driving based on high-precision maps relies heavily on these maps and cannot achieve timely map updates, resulting in a limited driving range for vehicles.

[0003] In the prior art, Chinese Patent (Application No.: 202410080016.1, Publication No.: CN118135515A) discloses a method, apparatus, medium, and device for generating lane line maps. The method includes: acquiring at least one frame of original image collected when a vehicle is driving in a parking lot; performing lane line region extraction processing on the at least one frame of original image to obtain at least one frame of region of interest (ROI) image, where each ROI image only indicates the lane line region in the corresponding original image; performing lane line identification and decomposition processing on each ROI image to determine lane line attribute information for at least one lane line segment; and performing fitting processing on the at least one lane line segment based on the lane line attribute information to obtain a lane line map of the parking lot. This solution mainly focuses on the construction of the parking lot map and does not use prior information, resulting in low robustness. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the present invention provides a high-precision real-time map construction method based on point-mask dual prior query. It not only introduces vectorized navigation points (point prior) and BEV mask structure (mask prior) into the Transformer query initialization stage, which significantly improves the spatial expression capability and structural integrity of map construction, but also can automatically perceive the spatial feature distribution in complex scenarios such as curves and intersections, thereby achieving more robust map construction.

[0005] To achieve the above and other related objectives, the present invention provides the following technical solution: A high-precision real-time map construction method based on point-mask dual prior queries, the method comprising: S1. As the vehicle travels on the road, it acquires data information from a vectorized navigation map of the road it is traveling on, and also acquires data information from multi-view images of the road in real time based on the onboard camera; S2. Based on the data information of the vectorized navigation map of the vehicle's driving road, extract the key points of the road and perform sine and cosine position encoding, and combine them with the road category label to generate structured high-dimensional point feature data information of the vehicle's driving road through Transformer encoding; S3. Based on the data information of the multi-view images of the road, construct the BEV spatial representation, introduce a lightweight mask decoding path, and generate the data information of the lane line shape structure mask based on the BEV features enhanced by point prior. S4. The structured high-dimensional point feature data of the vehicle's driving road and the lane line shape structure mask data are respectively injected into the query pos and query of the Transformer decoder, and semantic fusion is performed in space. The data is then input into the topology inference and lane relationship solver to construct a high-precision map, thereby obtaining the data information of the constructed high-precision map.

[0006] Furthermore, in step R4, the fusion function for performing semantic fusion in space is: , Among them, Q pos For the fused query point space encoding, Q (0) Q is the initial semantic query vector. (0) pos For the spatial query vector, P point P represents the point prior features obtained from the navigation map point encoding. mask F represents the structural mask features extracted from the BEV mask image. pos F query and F ref R is a nonlinear mapping function formed by a multilayer sensing mechanism. init R0 is the final fused representation used for downstream graph construction or topology solving. R0 is the residual initialization bias.

[0007] Furthermore, in step S3, the key point reference function for constructing the BEV space representation and introducing the lightweight mask decoding path is: , Where, k j For geometric keypoints extracted from bimodal priors, σ(·) is the Sigmoid activation function, and W outputs the normalized spatial location. r and b r These are the learnable affine transformation parameters.

[0008] Furthermore, in step S4, the topology inference function of the topology inference and lane relationship solver is: , Where ε is the set of predicted candidate connection edges, and W uv For learnable connection weights, p u end Let p be the coordinates of the tail end of lane segment u. v startLet V be the coordinates of the beginning of lane segment v.

[0009] Furthermore, the joint relationship total loss function L of the topology inference and lane relationship solver total for, L total =λ1·L det +λ2·L topo +λ3·L BE , Among them, L det L represents the lane detection loss. topo To predict the loss for topology connectivity, L BE This is the topology inference function for the topology inference and lane relationship solver.

[0010] To achieve the above and other related objectives, the present invention also provides a high-precision map real-time construction system based on point-mask dual prior queries, used to implement the high-precision map real-time construction method based on point-mask dual prior queries as described in any one of the claims, the system comprising: The data acquisition module is used to acquire data information of vectorized navigation maps of the road the vehicle is traveling on, and to acquire data information of multi-view images of the road in real time based on the vehicle-mounted camera; The SD map dual-modal fusion module is connected to the data acquisition module and is used to achieve efficient fusion of navigation map point sequences and mask dual-modal priors through the Query layer of the Transformer decoder. The key point referencing mechanism module is connected to the SD map dual-modal fusion module and is used to improve the spatial query and mapping capabilities under complex topology through the structure-aware key point referencing mechanism. The topology reasoning and lane relationship calculation module is connected to the key point referencing mechanism module and is used to introduce a topology-aware loss function to achieve accurate modeling of topology consistency through bidirectional endpoint alignment.

[0011] Furthermore, the SD map dual-modal fusion module includes a structured high-dimensional point feature generation unit for vehicle driving roads and a lane line shape structure mask generation unit. The structured high-dimensional point feature generation unit for vehicle driving roads is used to extract key road points based on the data information of the vectorized navigation map of the vehicle driving roads, perform sine and cosine position encoding, and combine the road category labels to generate structured high-dimensional point feature data information of the vehicle driving roads through Transformer encoding.

[0012] Furthermore, the lane line forming structure mask generation unit is used to construct a BEV spatial representation based on the data information of the multi-view image of the road, introduce a lightweight mask decoding path, and generate the data information of the lane line shape structure mask based on the BEV features enhanced by point prior.

[0013] Furthermore, the system also includes a high-precision map output module, which is connected to the topology reasoning and lane relationship calculation module, and is used to output the data information of the constructed high-precision map.

[0014] Furthermore, the keypoint reference function of the keypoint reference mechanism module is as follows: , Where, k j For geometric keypoints extracted from bimodal priors, σ(·) is the Sigmoid activation function, and W outputs the normalized spatial location. r and b r These are learnable affine transformation parameters; The topology inference function of the topology inference and lane relationship calculation module is as follows: , Where ε is the set of predicted candidate connection edges, and W uv For learnable connection weights, p u end Let p be the coordinates of the tail end of lane segment u. v start Let V be the coordinates of the beginning of lane segment v.

[0015] The present invention has the following positive effects: 1. This invention significantly improves the spatial representation and structural integrity of map construction by introducing vectorized navigation points (point priors) and BEV mask structures (mask priors) into the Transformer query initialization stage. Based on the dynamic reference of key points and the multi-anchor sampling strategy, the system can automatically perceive the spatial feature distribution in complex scenarios such as curves and intersections, thereby achieving more robust map construction.

[0016] 2. This invention introduces topology perception loss BELows, integrates structure detection, connectivity reasoning, and endpoint alignment errors, effectively ensuring the continuity and availability of lane structures during the mapping process. It also employs modular design and lightweight feature fusion strategies to ensure excellent computational efficiency, meeting the real-time perception-mapping-update linkage requirements of autonomous driving systems. It does not rely on specific sensor models or scenario assumptions, has good scalability, and can be deployed in real-world vehicle scenarios to achieve online mapping and rapid topology updates. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the system framework of the present invention. Detailed Implementation

[0018] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0019] Example 1: As Figure 1 As shown, a high-precision real-time map construction method based on point-mask dual prior queries is proposed, the method comprising: S1. As the vehicle travels on the road, it acquires data information from a vectorized navigation map of the road it is traveling on, and also acquires data information from multi-view images of the road in real time based on the onboard camera; S2. Based on the data information of the vectorized navigation map of the vehicle's driving road, extract the key points of the road and perform sine and cosine position encoding, and combine them with the road category label to generate structured high-dimensional point feature data information of the vehicle's driving road through Transformer encoding; S3. Based on the data information of the multi-view images of the road, construct the BEV spatial representation, introduce a lightweight mask decoding path, and generate the data information of the lane line shape structure mask based on the BEV features enhanced by point prior. S4. The structured high-dimensional point feature data of the vehicle's driving road and the lane line shape structure mask data are respectively injected into the query pos and query of the Transformer decoder, and semantic fusion is performed in space. The data is then input into the topology inference and lane relationship solver to construct a high-precision map, thereby obtaining the data information of the constructed high-precision map.

[0020] In this embodiment, in step R4, the fusion function for spatial semantic fusion is: , Among them, Q pos For the fused query point space encoding, Q (0) Q is the initial semantic query vector. (0) pos For the spatial query vector, P point P represents the point prior features obtained from the navigation map point encoding. mask F represents the structural mask features extracted from the BEV mask image.pos F query and F ref R is a nonlinear mapping function formed by a multilayer sensing mechanism. init R0 is the final fused representation used for downstream graph construction or topology solving. R0 is the residual initialization bias.

[0021] In this embodiment, in step S3, the key point reference function for constructing the BEV space representation and introducing the lightweight mask decoding path is: , Where, k j For geometric keypoints extracted from bimodal priors, σ(·) is the Sigmoid activation function, and W outputs the normalized spatial location. r and b r These are the learnable affine transformation parameters.

[0022] In this embodiment, in step S4, the topology inference function of the topology inference and lane relationship solver is: , Where ε is the set of predicted candidate connection edges, and W uv For learnable connection weights, p u end Let p be the coordinates of the tail end of lane segment u. v start Let V be the coordinates of the beginning of lane segment v.

[0023] In this embodiment, the joint relationship total loss function L of the topology inference and lane relationship solver is... total for, L total =λ1·L det +λ2·L topo +λ3·L BE , Among them, L det L represents the lane detection loss. topo To predict the loss for topology connectivity, L BE This is the topology inference function for the topology inference and lane relationship solver.

[0024] Example 2: Based on the high-precision real-time map construction method based on point-mask dual prior query in Example 1, the present invention will be further explained and described below.

[0025] like Figure 2 As shown, this invention provides a high-precision real-time map construction system based on point-mask dual prior queries, used to implement any of the high-precision real-time map construction methods based on point-mask dual prior queries, the system comprising: The data acquisition module is used to acquire data information of vectorized navigation maps of the road the vehicle is traveling on, and to acquire data information of multi-view images of the road in real time based on the vehicle-mounted camera; The SD map dual-modal fusion module is connected to the data acquisition module and is used to achieve efficient fusion of navigation map point sequences and mask dual-modal priors through the Query layer of the Transformer decoder. The key point referencing mechanism module is connected to the SD map dual-modal fusion module and is used to improve the spatial query and mapping capabilities under complex topology through the structure-aware key point referencing mechanism. The topology reasoning and lane relationship calculation module is connected to the key point referencing mechanism module and is used to introduce a topology-aware loss function to achieve accurate modeling of topology consistency through bidirectional endpoint alignment.

[0026] In this embodiment, the SD map dual-modal fusion module includes a structured high-dimensional point feature generation unit for vehicle driving roads and a lane line shape structure mask generation unit. The structured high-dimensional point feature generation unit for vehicle driving roads is used to extract key road points and perform sine and cosine position encoding based on the data information of the vectorized navigation map of the vehicle driving roads, and combine the road category labels to generate structured high-dimensional point feature data information of the vehicle driving roads through Transformer encoding.

[0027] In this embodiment, the lane line formation structure mask generation unit is used to construct a BEV spatial representation based on the data information of the multi-view image of the road, introduce a lightweight mask decoding path, and generate lane line shape structure mask data information based on the BEV features enhanced by point prior.

[0028] In this embodiment, the system further includes a high-precision map output module, which is connected to the topology reasoning and lane relationship calculation module, and is used to output the data information of the constructed high-precision map.

[0029] In this embodiment, the keypoint referencing function of the keypoint referencing mechanism module is: , Where, k j For geometric keypoints extracted from bimodal priors, σ(·) is the Sigmoid activation function, and W outputs the normalized spatial location. r and b r These are learnable affine transformation parameters; The topology inference function of the topology inference and lane relationship calculation module is as follows: , Where ε is the set of predicted candidate connection edges, and W uv For learnable connection weights, p u end Let p be the coordinates of the tail end of lane segment u. v start Let V be the coordinates of the beginning of lane segment v.

[0030] In this embodiment, the present invention can be divided into a first-level module, a second-level module, and a third-level module, as shown in Table 1.

[0031] Table 1 This system achieves efficient fusion of bimodal priors—point sequences (vectorized) and masks (rasterized)—from the Query layer of the Transformer decoder. Point priors provide accurate spatial anchoring, while mask priors enhance shape and structural representation. This fusion not only improves geometric accuracy but also significantly optimizes topological consistency.

[0032] To achieve the joint expression of semantic and structural information, the system constructs the following fusion formula: , Among them, Q pos For the fused query point space encoding, Q (0) Q is the initial semantic query vector. (0) pos For the spatial query vector, P point P represents the point prior features obtained from the navigation map point encoding. mask F represents the structural mask features extracted from the BEV mask image. pos F query and F ref R is a nonlinear mapping function formed by a multilayer sensing mechanism. init R0 is the final fused representation used for downstream graph construction or topology solving. R0 is the residual initialization bias.

[0033] The goal of this module is to enhance the geometric accuracy and semantic discriminativeness of the query expression through structured guidance with point / mask dual priors, laying a precise foundation for subsequent lane line mapping and connection relationship calculation.

[0034] In this embodiment, the query modeling system enhances spatial query and mapping capabilities under complex topologies through a structure-aware keypoint referencing mechanism. This module not only supports geometric spatial queries but also incorporates topological structure for joint reasoning.

[0035] Key point reference formula: , Where, k jFor geometric keypoints extracted from bimodal priors, σ(·) is the Sigmoid activation function, and W outputs the normalized spatial location. r and b r These are the learnable affine transformation parameters.

[0036] The advantages of the module include: Strong structural sampling capability: can focus on areas with high variation such as curves and intersections; Supports spatially consistent mapping: key points drive sampling regions to guide geometric constraints; Supports topology reasoning fusion: Provides enhanced structure-aware support for topology connections and instance differentiation.

[0037] In summary, this invention not only introduces vectorized navigation points (point priors) and BEV mask structures (mask priors) into the Transformer query initialization stage, significantly improving the spatial representation capability and structural integrity of map construction, but also automatically perceives the spatial feature distribution in complex scenarios such as curves and intersections, thereby achieving more robust map construction.

[0038] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for real-time construction of high-precision maps based on point-mask dual prior queries, characterized in that, The method includes: S1. As the vehicle travels on the road, it acquires data information from a vectorized navigation map of the road it is traveling on, and also acquires data information from multi-view images of the road in real time based on the onboard camera; S2. Based on the data information of the vectorized navigation map of the vehicle's driving road, extract the key points of the road and perform sine and cosine position encoding, and combine them with the road category label to generate structured high-dimensional point feature data information of the vehicle's driving road through Transformer encoding; S3. Based on the data information of the multi-view images of the road, construct the BEV spatial representation, introduce a lightweight mask decoding path, and generate the data information of the lane line shape structure mask based on the BEV features enhanced by point prior. S4. The structured high-dimensional point feature data of the vehicle's driving road and the lane line shape structure mask data are respectively injected into the query pos and query of the Transformer decoder, and semantic fusion is performed in space. The data is then input into the topology inference and lane relationship solver to construct a high-precision map, thereby obtaining the data information of the constructed high-precision map.

2. The high-precision real-time map construction method based on point-mask dual prior queries according to claim 1, characterized in that, In step R4, the fusion function for spatial semantic fusion is: ， Among them, Q pos For the fused query point space encoding, Q (0) Q is the initial semantic query vector. (0) pos For the spatial query vector, P point P represents the point prior features obtained from the navigation map point encoding. mask F represents the structural mask features extracted from the BEV mask image. pos F query and F ref R is a nonlinear mapping function formed by a multilayer sensing mechanism. init R0 is the final fused representation used for downstream graph construction or topology solving. R0 is the residual initialization bias.

3. The high-precision real-time map construction method based on point-mask dual prior queries according to claim 1, characterized in that, In step S3, the key point reference function for constructing the BEV space representation and introducing the lightweight mask decoding path is: ， Where, k j For geometric keypoints extracted from bimodal priors, σ(·) is the Sigmoid activation function, and W outputs the normalized spatial location. r and b r These are the learnable affine transformation parameters.

4. The high-precision real-time map construction method based on point-mask dual prior query as described in claim 1, characterized in that, In step S4, the topology inference function of the topology inference and lane relationship solver is: ， Where ε is the set of predicted candidate connection edges, and W uv For learnable connection weights, p u end Let p be the coordinates of the tail end of lane segment u. v start Let V be the coordinates of the beginning of lane segment v.

5. The high-precision real-time map construction method based on point-mask dual prior query as described in claim 4, characterized in that: The joint relationship total loss function L of the topology inference and lane relationship solver total for, L total =λ1·L det +λ2·L topo +λ3·L BE , Among them, L det L represents the lane detection loss. topo To predict the loss for topology connectivity, L BE This is the topology inference function for the topology inference and lane relationship solver.

6. A high-precision real-time map construction system based on point-mask dual prior queries, characterized in that, The system for implementing the high-precision real-time map construction method based on point-mask dual prior queries as described in any one of claims 1-5 includes: The data acquisition module is used to acquire data information of vectorized navigation maps of the road the vehicle is traveling on, and to acquire data information of multi-view images of the road in real time based on the vehicle-mounted camera; The SD map dual-modal fusion module is connected to the data acquisition module and is used to achieve efficient fusion of navigation map point sequences and mask dual-modal priors through the Query layer of the Transformer decoder. The key point referencing mechanism module is connected to the SD map dual-modal fusion module and is used to improve the spatial query and mapping capabilities under complex topology through the structure-aware key point referencing mechanism. The topology reasoning and lane relationship calculation module is connected to the key point referencing mechanism module and is used to introduce a topology-aware loss function to achieve accurate modeling of topology consistency through bidirectional endpoint alignment.

7. The high-precision real-time map construction system based on point-mask dual prior query as described in claim 6, characterized in that: The SD map dual-modal fusion module includes a structured high-dimensional point feature generation unit for vehicle driving roads and a lane line shape structure mask generation unit. The structured high-dimensional point feature generation unit for vehicle driving roads is used to extract key points of the road based on the data information of the vectorized navigation map of the vehicle driving roads, perform sine and cosine position encoding, and combine the road category label to generate structured high-dimensional point feature data information of the vehicle driving roads through Transformer encoding.

8. The high-precision real-time map construction system based on point-mask dual prior query as described in claim 7, characterized in that: The lane line formation structure mask generation unit is used to construct a BEV spatial representation based on the data information of the multi-view image of the road, introduce a lightweight mask decoding path, and generate the data information of the lane line shape structure mask based on the BEV features enhanced by point prior.

9. The high-precision real-time map construction system based on point-mask dual prior query as described in claim 6, characterized in that, The system also includes a high-precision map output module, which is connected to the topology reasoning and lane relationship calculation module, and is used to output the data information of the constructed high-precision map.

10. The high-precision real-time map construction system based on point-mask dual prior query as described in claim 6, characterized in that: The keypoint referencing function of the keypoint referencing mechanism module is: ， Where, k j For geometric keypoints extracted from bimodal priors, σ(·) is the Sigmoid activation function, and W outputs the normalized spatial location. r and b r These are learnable affine transformation parameters; The topology inference function of the topology inference and lane relationship calculation module is as follows: ， Where ε is the set of predicted candidate connection edges, and W uv For learnable connection weights, p u end Let p be the coordinates of the tail end of lane segment u. v start Let V be the coordinates of the beginning of lane segment v.

Images