Robust online vector high-definition map construction method and device oriented to biased prior

Abstract

The application provides a robust online vector high-precision map construction method for biased prior. The application first converts the current multi-view image into an aerial view feature, aligns the available biased vector map as the prior and the coordinate system of the aerial view, and processes the biased vector map to obtain a biased rasterized map; then compares the aerial view feature with the biased rasterized map, and outputs three types of results of reliable, error and no map for each region of the biased rasterized map. The results are used as unified control signals to respectively control the prior reliability supervision, the vector and raster dual-prior query initialization, the aerial view feature fusion guided by the confidence, and the separated auxiliary reliability supervision. Compared with the existing online vector high-precision map construction method based on prior enhancement, the application can improve the mapping robustness and precision in the biased prior scene, so as to provide more accurate online vector high-precision map for downstream applications such as automatic driving planning and other tasks.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving environmental perception and high-precision map construction technology, and particularly to a robust online vector high-precision map construction method, apparatus, electronic device, computer-readable storage medium and computer program product for biased priors. It can still robustly generate online vector high-precision maps by leveraging biased vector maps even when the prior map is outdated, misaligned or has semantic errors that lead to biased vector maps. Background Technology

[0002] High-precision maps are a crucial infrastructure for autonomous driving systems to understand road topology and geometry. Traditional high-precision maps rely on specialized surveying vehicles, high-precision positioning equipment, and manual annotation, resulting in lengthy production and review cycles. Furthermore, real-world roads are constantly changing due to construction, expansion, lane line redrawing, and temporary detours, leading to discrepancies between offline maps and actual roads. To reduce reliance on offline mapping, online vector high-precision map construction is gradually becoming an important direction for autonomous driving perception. Its goal is to reduce dependence on expensive offline maps by providing real-time vector representations of map elements such as lane lines, road boundaries, and pedestrian crossings based on onboard sensors.

[0003] Early online vector high-precision map construction typically involved first predicting the semantic results of the bird's-eye view raster, and then generating the vector map through post-processing. However, while this method performs well in conventional scenarios, it may still encounter problems such as local missed detections, broken line segments, and incomplete topological relationships in situations involving occlusion, long distances, rain, snow, nighttime conditions, and complex intersections.

[0004] To compensate for the limitations of single-frame observations, existing technologies further incorporate historical information or external map priors, using historical frames or predictions to supplement the current field of view; or using standard-precision maps, high-precision maps, historical predictions, or unified prior coding to provide beyond-line-of-sight information and long-range topological constraints for online mapping. A common assumption of these methods is that prior information is largely available, but this assumption does not always hold true in real-world deployments.

[0005] In real-world road environments, vector maps, serving as prior information, may suffer from geometric shifts, semantic errors, or topological outdated information due to positioning errors, road construction, lane line redrawing, temporary detours, or map update delays, resulting in biased vector maps. If the system does not first determine the reliability of the biased vector map but directly injects it into bird's-eye view features, query initialization, or historical memory, the erroneous priors will be amplified by the model, potentially leading to a worse vector map result than if only the current observation were relied upon.

[0006] In summary, the shortcomings of existing technologies regarding biased prior scenarios can be categorized into the following types.

[0007] First, there is a lack of region-level reliability assessment for biased vector maps. Most methods directly use biased vector maps as reliable input, failing to proactively reduce the impact of erroneous priors at road changes. Second, single prior representation methods are incomplete. Raster priors are easy to fuse with bird's-eye view features grid-by-grid, but they easily weaken instance and topological relationships; vector priors preserve instance geometry and point order structure, but are not suitable for direct participation in dense space fusion. Third, the intensity of prior fusion lacks interpretable control. Erroneous priors may be over-injected, while slight misalignment or local occlusion may cause usable priors to be completely discarded. Fourth, if reliability assessment is strongly coupled with the main mapping path, the problem of improved auxiliary classification but decreased accuracy of the main map may occur.

[0008] The root cause of these problems is that existing methods typically use biased vector maps directly as external memory without first treating them as objects to be verified. To solve this problem, the system needs to first utilize the spatial consistency between the current bird's-eye view features and the biased vector map to determine whether the prior information for each region is reliable, and then selectively use the prior information based on the judgment results in paths such as query initialization, raster query, and bird's-eye view feature fusion. Summary of the Invention

[0009] The purpose of this invention is to address the robustness issue of existing prior-enhanced online vector high-precision map construction methods in scenarios with biased priors. Specifically, existing methods typically assume that offline maps, standard-precision maps, or historical maps are reliable priors, lacking intermediate judgments on reliable areas, erroneous areas, and areas without maps. Furthermore, incorrect lane lines, incorrect road boundaries, missing map elements, or newly added map elements may propagate along both the confidence-guided feature fusion path and the dual-prior query generation path. To address this, this invention proposes a robust online vector high-precision map construction method and system oriented towards biased priors, enabling the model to verify biased vector maps and then selectively use priors based on reliability.

[0010] To address the shortcomings of existing technologies, this invention proposes a robust online vector high-precision map construction method oriented towards biased priors, which includes:

[0011] Step 1: Obtain a biased vector map of the vehicle's location and real-world images taken from multiple perspectives by the vehicle's own camera. Based on the camera's parameters, transform the biased vector map to the coordinate system of the real-world images, i.e., the vehicle's coordinate system.

[0012] Step 2: Rasterize the biased vector map to obtain a biased rasterized map;

[0013] Step 3: Process the real-world images from multiple perspectives to obtain bird's-eye view features;

[0014] Step 4: Call the confidence assessment module to compare the features of the bird's-eye view with the biased rasterized map to obtain the classification result of each raster in the biased rasterized map as reliable, erroneous, or without map.

[0015] Step 5: Based on the classification result, the dual prior query generation module is invoked. First, the biased raster map is filtered, retaining only the raster that the classification result belongs to is reliable, forming a raster prior query unit. Then, for each vector point in the biased vector map, based on the classification result of the vector point being in the corresponding raster region in the biased raster map, only the vector point that the classification result belongs to is reliable is retained, forming a vector prior query unit. Finally, the two prior query units are merged and combined with the learnable random query unit to obtain the prior query unit.

[0016] Step 6: Extract features from the biased rasterized map to obtain prior map features; then call the confidence-guided feature fusion module to first fuse the prior map features with the bird's-eye view features based on the classification results of each raster, obtaining initial fused features. Then, perform local consistency verification on the initial fused features using the classification results of each raster in the biased rasterized map, finally obtaining the map prediction features.

[0017] Step 7: Based on the prior query unit and map prediction features, the map decoder outputs the category, confidence level and vector point sequence of each map element as the vector map prediction result under the joint constraints of the real image and the biased vector map.

[0018] The aforementioned robust online vector high-precision map construction method oriented towards biased priors, wherein...

[0019] Step 1 includes: obtaining the old map, historical prediction map, cloud map or crowdsourced updated map that the vehicle can obtain at the current moment as the biased vector map;

[0020] In step 3, the bird's-eye view feature is the road representation after converting multiple real-world images from different perspectives into a top-down space.

[0021] Step 4: Obtain the classification result for each grid cell in the biased rasterized map using the following formula:

[0022]

[0023] These are the predicted confidence scores for each raster region, and their size is consistent with the size of the biased rasterized map. Each raster includes three categories of predicted probabilities. It's a prediction module. This is a characteristic of the bird's-eye view. It is a rasterized map; These are probability graphs showing that the prediction results are reliable, incorrect, and without a map.

[0024] Converted into a single-channel reliability strength map, its size is consistent with the size of the biased rasterized map:

[0025]

[0026] In the formula This is a single-channel reliability strength diagram. Represents one of the grid cells. The threshold for whether a map exists; These represent the probabilities of the three probability maps at grid u;

[0027] Step 6, the initial fusion process, specifically involves:

[0028]

[0029] In the formula These are initial fusion features. It is a feature of bird's-eye view. This is the scripture The reliability of the calculated map; These are the prior features obtained after processing a partially rasterized map, i.e., map prior features. This is point-by-point multiplication;

[0030] Step 6 also includes: performing local consistency verification on the initial fused features.

[0031]

[0032] In the formula This is the verification result at grid u. It is a feature splicing operation. This represents a multi-layer fully connected neural network, responsible for predicting gating weights;

[0033] Calculate the distribution preservation weights for each grid cell :

[0034]

[0035]

[0036] In the formula Let u be the prior reliable probability at grid u; Let be the prior error probability at grid u; symbol This means that only positive values ​​are retained, therefore The positive part represents the difference between the prior reliable probability and the prior false probability at grid u; The lower bound of the basic weight of grid u is used to avoid directly clearing the prior to zero; it is a pre-set threshold. Maintain weights for the distribution of grid u. This is the single-channel reliability strength diagram calculated from the three types of confidence levels mentioned above. This is the verification result at grid u; The final prior fusion weights for each grid cell;

[0037] The predicted features of the map are obtained using the following formula. :

[0038]

[0039] This is the result after weighting a biased rasterized map. The final prior fusion weights for each grid cell To merge the weighted graph; The representative represents the combination of the two, that is, the initial fused features. And weighted biased raster map Merge at the channel dimension; The fusion network is used to convert the stitched features into map prediction features that the decoder can use.

[0040] In step 7, the map decoder, based on the prior query unit and the map prediction features, refers to the long-range geometry and topology provided by the biased vector map in reliable areas, and refers to the bird's-eye view features in error areas.

[0041] The robust online vector high-precision map construction method based on biased priors further includes:

[0042] Step 8: During the training phase, three types of supervised labels are generated using the rasterized unbiased vector map and the currently obtained biased rasterized map. The label generation rules retain the following mathematical expression:

[0043]

[0044] In the formula It is a rasterized unbiased vector map. It is the currently obtained biased rasterized map; 0 represents "and", meaning that both the preceding and following conditions are met simultaneously; 0, 1, and 2 represent the three categories of no map, error, and reliable, respectively.

[0045] When training the confidence evaluation module, the three types of labels are supervised to enable the model to learn to reduce the influence of erroneous priors when there are real-world road changes, construction, misalignment, or map outdated conditions. For vector map prediction itself, training still focuses on whether the category is correct, whether the point location is accurate, and whether the direction of the line segments formed by adjacent points is correct.

[0046] In addition to reliability supervision, the final output vector map prediction results are also supervised, including map element categories, instance confidence, vector point locations, and the direction of line segments formed by adjacent points.

[0047] The robust online vector high-precision map construction method oriented towards biased priors is described in which the vector map obtained in step 7 is used to model the vehicle's surrounding environment and perform downstream autonomous driving trajectory planning tasks.

[0048] This invention also proposes a robust online vector high-precision map construction device oriented towards biased priors, comprising:

[0049] Module 1 acquires a biased vector map of the vehicle's location and real-world images taken from multiple perspectives by the vehicle's own camera. Based on the camera's parameters, the biased vector map is transformed to the coordinate system of the real-world images, i.e., the vehicle's coordinate system.

[0050] Module 2 rasterizes the biased vector map to obtain a biased rasterized map;

[0051] Module 3 processes real-world images from multiple perspectives to obtain bird's-eye view features;

[0052] Module 4 calls the confidence assessment module to compare the features of the bird's-eye view with the biased rasterized map, and obtains the classification result of each raster in the biased rasterized map as reliable, erroneous or without map;

[0053] Module 5, based on the classification result, calls the dual prior query generation module. First, it filters the biased raster map, retaining only the raster that the classification result belongs to is reliable, forming a raster prior query unit. Then, for each vector point in the biased vector map, based on the classification result of the vector point being in the corresponding raster region in the biased raster map, it retains only the vector point that the classification result belongs to is reliable, forming a vector prior query unit. Finally, the two prior query units are merged and combined with the learnable random query unit to obtain the prior query unit.

[0054] Module 6 extracts features from the biased rasterized map to obtain prior map features. Then, it calls the confidence-guided feature fusion module to first fuse the prior map features with the bird's-eye view features based on the classification results of each raster, obtaining initial fused features. Next, it performs local consistency checks on the initial fused features using the classification results of each raster in the biased rasterized map, ultimately obtaining the predicted map features.

[0055] Module 7, the map decoder, outputs the category, confidence level and vector point sequence of each map element based on the prior query unit and map prediction features, as the vector map prediction result under the joint constraints of the real image and the biased vector map.

[0056] The robust online vector high-precision map construction device oriented towards biased priors, wherein

[0057] Module 1 includes: obtaining the old map, historical prediction map, cloud map or crowdsourced updated map that the vehicle can obtain at the current moment as the biased vector map;

[0058] Module 3, the bird's-eye view feature is the road representation after converting multiple perspective real-shot images into a top-down space;

[0059] Module 4 obtains the classification result for each grid cell in the biased rasterized map using the following formula:

[0060]

[0061] These are the predicted confidence scores for each raster region, and their size is consistent with the size of the biased rasterized map. Each raster includes three categories of predicted probabilities. It's a prediction module. This is a characteristic of the bird's-eye view. It is a rasterized map; These are probability graphs showing that the prediction results are reliable, incorrect, and without a map.

[0062] Converted into a single-channel reliability strength map, its size is consistent with the size of the biased rasterized map:

[0063]

[0064] In the formula This is a single-channel reliability strength diagram. Represents one of the grid cells. The threshold for whether a map exists; These represent the probabilities of the three probability maps at grid u;

[0065] The initial fusion process for module 6 is as follows:

[0066]

[0067] In the formula These are initial fusion features. It is a feature of bird's-eye view. This is the scripture The reliability of the calculated map; These are the prior features obtained after processing a partially rasterized map, i.e., map prior features. This is point-by-point multiplication;

[0068] Module 6 also includes: performing local consistency verification on the initial fused features.

[0069]

[0070] In the formula This is the verification result at grid u. It is a feature splicing operation. This represents a multi-layer fully connected neural network, responsible for predicting gating weights;

[0071] Calculate the distribution preservation weights for each grid cell :

[0072]

[0073]

[0074] In the formula Let u be the prior reliable probability at grid u; Let be the prior error probability at grid u; symbol This means that only positive values ​​are retained, therefore The positive part represents the difference between the prior reliable probability and the prior false probability at grid u; The lower bound of the basic weight of grid u is used to avoid directly clearing the prior to zero; it is a pre-set threshold. Maintain weights for the distribution of grid u. This is the single-channel reliability strength diagram calculated from the three types of confidence levels mentioned above. This is the verification result at grid u; The final prior fusion weights for each grid cell;

[0075] The predicted features of the map are obtained using the following formula. :

[0076]

[0077] This is the result after weighting a biased rasterized map. The final prior fusion weights for each grid cell To merge the weighted graph; The representative represents the combination of the two, that is, the initial fused features. And weighted biased raster map Merge at the channel dimension; The fusion network is used to convert the stitched features into map prediction features that the decoder can use.

[0078] In module 7, the map decoder, based on the prior query unit and the map prediction features, refers to the long-range geometry and topology provided by the biased vector map in reliable areas, and refers to the bird's-eye view features in erroneous areas.

[0079] The aforementioned robust online vector high-precision map construction device oriented towards biased priors further includes:

[0080] Module 8: During the training phase, three types of supervised labels are generated using the rasterized unbiased vector map and the currently obtained biased rasterized map. The label generation rules retain the following mathematical expression:

[0081]

[0082] In the formula It is a rasterized unbiased vector map. It is the currently obtained biased rasterized map; 0 represents "and", meaning that both the preceding and following conditions are met simultaneously; 0, 1, and 2 represent the three categories of no map, error, and reliable, respectively.

[0083] When training the confidence evaluation module, the three types of labels are supervised to enable the model to learn to reduce the influence of erroneous priors when there are real-world road changes, construction, misalignment, or map outdated conditions. For vector map prediction itself, training still focuses on whether the category is correct, whether the point location is accurate, and whether the direction of the line segments formed by adjacent points is correct.

[0084] In addition to reliability supervision, the final output vector map prediction results are also supervised, including map element categories, instance confidence, vector point locations, and the direction of line segments formed by adjacent points.

[0085] The present invention also proposes an electronic device, including the aforementioned robust online vector high-precision map construction device oriented towards biased priors, wherein the electronic device may be connected to an information display device, which is used to display the vector map with user-set display parameters, attributes or through an artificial intelligence model.

[0086] The present invention also proposes a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the robust online vector high-precision map construction method oriented towards biased priors.

[0087] The present invention also proposes a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, it implements the steps of the robust online vector high-precision map construction method oriented towards biased priors.

[0088] As can be seen from the above solutions, the advantages of the present invention are:

[0089] Compared to existing prior-enhanced online vector high-precision map construction methods, this invention improves the robustness of mapping in scenarios with biased priors. First, the system no longer directly trusts external maps but instead evaluates prior confidence grid-by-grid based on the features of the current bird's-eye view. Second, the dual-prior query generation module and the confidence-guided feature fusion module selectively use priors at the query and feature layers, respectively, enabling reliable regions to provide topological constraints and long-distance structural information, while allowing erroneous regions to rely more on real-time observations. Third, distributed weighting and a separate auxiliary reliability supervision branch improve training stability and engineering deployability. Therefore, this invention can output more accurate and stable online vector high-precision maps in scenarios such as road construction, lane line redrawing, map expiration, positioning errors, occlusion, and insufficient long-distance perception, and can also feed back erroneous regions to the map maintenance process. Attached Figure Description

[0090] Figure 1 This is a flowchart illustrating the overall implementation of the present invention;

[0091] Figure 2 This is a flowchart of the input alignment and confidence evaluation process for this invention.

[0092] Figure 3 This is a flowchart of the dual prior query generation process of the present invention;

[0093] Figure 4 This is a flowchart of the confidence-guided feature fusion and decoding process of the present invention. Detailed Implementation

[0094] It should be noted that, in this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0095] In the absence of further restrictions, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0096] The processor described in this invention is the control center of an electronic device. It can be a single processor or a collective term for multiple processing elements. For example, it can be one or more central processing units (CPUs), application-specific integrated circuits (ASICs), or one or more integrated circuits configured to implement embodiments of this invention, such as one or more digital signal processors (DSPs), or one or more field-programmable gate arrays (FPGAs).

[0097] Alternatively, the processor can perform various functions of the electronic device by running or executing software programs stored in memory, and by calling data stored in memory.

[0098] In a specific implementation, as one example, the processor may include one or more CPUs. Each of these processors may be a single-core processor or a multi-core processor. Here, "processor" can refer to one or more devices, circuits, and / or processing cores for processing data (e.g., computer program instructions). Electronic devices may include servers, desktop computers, laptops, smartphones, tablets, embedded computers, etc., where the embedded computer includes vehicles and robots, etc.

[0099] The memory is used to store the software program that executes the solution of the present invention, and the execution is controlled by the processor. For specific implementation methods, please refer to the above method embodiments, which will not be repeated here.

[0100] It should be noted that the structure of the electronic device shown in the accompanying drawings is not intended to limit it. The actual electronic device may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0101] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0102] It should also be understood that the term "and / or" in this article 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 existing alone, A and B existing simultaneously, and B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0103] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0104] It should also be understood that, in various embodiments of the present invention, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0105] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0106] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0107] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0108] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0109] While researching online high-precision vector map construction using biased vector maps, the inventors discovered that performance degradation stemmed from the model's lack of an intermediate judgment regarding the reliability of prior information. Further analysis revealed that erroneous priors affect the model through two main paths: 1) being injected into the feature fusion process at the bird's-eye view feature level through confidence-guided features; 2) becoming the starting point for the decoder's search at the dual-prior query generation level. When these two paths overlap, the model may be more inclined to reproduce erroneous priors rather than correcting them based on the current image.

[0110] To address the aforementioned issues, the inventors changed the prior usage method from "directly trusting the prior" to "first verifying the prior, then selectively using the prior." This invention first converts the current multi-view image into bird's-eye view features and rasterizes the biased vector map into the same bird's-eye view raster. Then, a confidence assessment module compares the bird's-eye view features with the biased rasterized map, outputting reliability, error, and no-location status for each raster of the biased rasterized map. Figure 3 The classification result serves as a unified control signal, used to regulate the dual-priority query generation module, the confidence-guided feature fusion module, and the distribution-preserving weights. During the training phase, three types of supervised labels and a separate auxiliary reliability supervision branch are used to supervise and strengthen the confidence assessment.

[0111] In this specification, the Bird's-Eye View (BEV) space represents the space of the road surrounding the vehicle from a top-down perspective; the Bird's-Eye View grid is a discrete grid covering this space, and a grid is one of its cells. A biased vector map is a vector map that may be outdated, misaligned, or semantically incorrect, and therefore inconsistent with the current road conditions; the aligned biased vector map is the result of transforming this map to the vehicle coordinate system in step 1; the biased rasterized map is the result of rasterizing the aligned biased vector map in step 2. A map element is a single object, such as a lane line, a road boundary, or a pedestrian crossing, output in step 7; an instance is a specific map element; vector points are the sequential points that constitute the map element; topological relationships are the connections or adjacencies between map elements. Reliable, erroneous, and no map are the three categories given for each grid in step 4.

[0112] In terms of numerical representation, the probabilistic map is a numerical image obtained by arranging the probability of a certain category corresponding to each grid position in the bird's-eye view grid according to the original grid. Single channel means that each grid position corresponds to only one value, while multiple channels mean that the same grid position has multiple feature components. The prior query unit is the numerical representation input to the map decoder, not the map element itself; the instance confidence is the confidence level of the map decoder in the output of a single map element, which is different from the grid classification probability in step 4. Steps 1 to 7 constitute the main map construction path; the training phase is used to obtain model parameters, and the inference phase loads the parameters obtained from training and executes steps 1 to 7. The subsequent task is to receive the online vector high-precision map output from step 7 and then use the map; autonomous driving trajectory planning is one example, which uses this map as the map input required to determine the future driving path of the vehicle.

[0113] To achieve the above-mentioned technical effects, the present invention proposes the following key technical points:

[0114] Key point 1: Set up a confidence assessment module. This module provides a reliable, incorrect, or map-free classification result for each grid cell based on the consistency between the features of the bird's-eye view and the biased rasterized map in the bird's-eye view space. In one implementation, this module can output three probabilities from a dense prediction network, so that the system can determine whether the prior should be retained, suppressed, or ignored before using the prior.

[0115] Key Point 2: Implement a three-tiered automatic reliability label generation and supervision mechanism. During the training phase, the unbiased rasterized vector map and the biased rasterized map obtained by rasterizing the biased vector map are aligned to the same bird's-eye view raster. If neither has map elements, it is marked as "no map"; if they are occupied or semantically inconsistent, it is marked as "error"; if both exist and are semantically consistent, it is marked as "reliable". This transforms the differences in complex vector maps into raster-by-raster classification supervision, enabling the model to learn when to utilize priors and when to suppress them.

[0116] Key point 3: Setting up a dual prior query generation module. The raster branch provides continuous spatial distribution and local occupancy information based on a biased rasterized map, while the vector branch retains map instances, point order, category, geometry, and topological relationships based on an aligned biased vector map. After filtering the two priors according to the classification results, they form raster prior query units and vector query units respectively. Then, the two types of prior query units are merged and combined with learnable random query units to obtain the prior query unit, enabling the decoder to obtain both local spatial information and instance-level topological information before prediction.

[0117] Key point 4: Set up a confidence-guided feature fusion module. The system uses different fusion strengths for different grids based on the classification results: reliable grids retain priors more strongly, erroneous grids weaken priors or mainly rely on the current bird's-eye view features, and grids without maps do not introduce priors. Thus, reliable grids utilize priors to enhance structural constraints, while erroneous grids reduce prior contamination.

[0118] Key point 5: Implement local consistency gating, distribution-preserving weights, and a separate auxiliary reliability supervision branch. Local consistency gating is used to re-check whether the initial fused features and the biased rasterized map are consistent in local locations; the lower bound of the base weights is used to avoid abrupt changes in feature distribution caused by directly setting them to zero; the separate auxiliary reliability supervision branch only strengthens confidence evaluation during the training phase and does not directly perturb the main map construction path. The technical effect is to improve confidence evaluation capabilities while maintaining the stability of the main mapping path.

[0119] To make the above-mentioned features and effects of the present invention clearer and easier to understand, specific embodiments are described below in conjunction with the accompanying drawings. This specification discloses one or more embodiments incorporating the features of the present invention. The disclosed embodiments are merely illustrative. The scope of protection of the present invention is not limited to the disclosed embodiments, but is defined by the appended claims.

[0120] Example 1: Complete Online Vector High-Precision Map Construction Process:

[0121] like Figure 1 As shown, the implementation process of this invention includes input alignment and rasterization, bird's-eye view feature extraction, confidence assessment, dual prior query generation and merging, confidence-guided feature fusion, vector map decoding, training constraints, and downstream applications. The system receives multi-view images from the vehicle, camera parameters, and a biased vector map of the vehicle's location at the current moment. First, the confidence assessment module determines whether the priors are reliable in each raster. Then, based on the classification results, the priors are selectively used, and finally, the vector point sequences, categories, and confidence scores of map elements such as lane lines, road boundaries, and pedestrian crossings are output.

[0122] In this invention, step 4 corresponds to the confidence assessment module; step 5 corresponds to the dual prior query generation module, which forms a grid prior query unit and a vector prior query unit respectively, and combines them with a learnable random query to form a prior query unit; step 6 corresponds to confidence-guided feature fusion, which includes a two-step fusion process; step 7 corresponds to vector map decoding; step 8 corresponds to the automatic generation and supervision mechanism of three types of reliability labels; step 9 corresponds to the downstream application of the vector map; and embodiment two corresponds to the separate auxiliary reliability supervision branch.

[0123] For ease of understanding, the function of each module in this embodiment is summarized below using a module input / output table:

[0124] Module enter deal with Output Input, Alignment, and Rasterization Real-world images, camera parameters, biased vector map Align the biased vector map to the vehicle coordinate system, and then rasterize it. Aligned biased vector map, biased rasterized map Confidence assessment module Bird's-eye view features, biased rasterized map Grid-by-grid comparison of bird's-eye view features with biased rasterized maps Reliable, faulty, without location Figure 3 Classification results Dual Prior Query Generation Module Biased rasterized map, aligned biased vector map, classification results Grid prior query units and vector prior query units are formed separately, and then combined with learnable random queries to form prior query units. Raster prior query cell, vector query cell, and prior query cell Confidence-guided feature fusion module Bird's-eye view features, map prior features, classification results, biased rasterized map The initial fused features are then fused according to the classification results, and then local consistency is checked. Map Prediction Features Vector map decoding Prior query unit, map prediction features Predict map element categories, confidence levels, and vector point sequences. Online vector high-precision map

[0125] Table 1: Input / Output Table

[0126] like Figure 2 As shown, step 1 involves acquiring and aligning a biased vector map. The system first obtains a biased vector map of the vehicle's location, then transforms it to the vehicle's current coordinate system based on camera parameters, ensuring it corresponds to the bird's-eye view space where the projected image of the real-world shot resides. A biased vector map refers to an old map, historical prediction map, cloud map, or crowdsourced updated map that the vehicle can acquire at the current moment and use as a priori information; it may be unreliable due to road changes or positioning errors.

[0127] The core objective of this step is to place the biased vector map into a unified coordinate system corresponding to the current location, enabling comparison of information from different sources. The system transforms road points, lane line points, or boundary points to the vehicle coordinate system and the current mapping area, without changing their vector instances, point order, or topological relationships in this step. The output of Step 1 is the aligned biased vector map, i.e., the aligned map; Step 2 then rasterizes this aligned map, allowing subsequent modules to perform grid-by-grid comparisons with the features of the current bird's-eye view.

[0128] Step 2: Obtain a biased rasterized map representation. The system rasterizes the aligned map obtained in Step 1 to obtain a biased rasterized map, which is used to represent the spatial occupancy of road elements in the bird's-eye view raster and supports grid-by-grid confidence assessment and feature fusion;

[0129] Step 3: Extract current bird's-eye view features. The system converts the current multi-view images into unified bird's-eye view features. These features represent the road conditions as seen by the vehicle at that moment and are the primary basis for subsequent verification of the authenticity of biased rasterized maps. Bird's-eye view features refer to the road representation after converting the current vehicle observation into a top-down space; they reflect the road conditions as seen by the vehicle at that moment.

[0130] Multi-camera images are first processed by an image feature extraction network, and then projected onto a bird's-eye view raster based on camera parameters. The projected bird's-eye view features are in the same space as the biased rasterized map, allowing for region-by-region comparison. The output of this step is the current bird's-eye view feature. It represents the road condition seen by the vehicle from the sensors at this moment, and the subsequent confidence assessment module will use it to compare with the biased rasterized map.

[0131] Step 4: Call the confidence assessment module to determine the reliability of the prior in each grid cell. The system combines the bird's-eye view features and the biased rasterization... Figure 1 The input confidence assessment module evaluates the data and provides a rating of reliable, erroneous, or unreliable for each grid cell. Figure 3 Classification results. Reliable, erroneous, and no map correspond to: prior knowledge can be used relatively strongly, the influence of prior knowledge should be reduced, and prior knowledge should not be introduced, respectively.

[0132] This core judgment can be expressed mathematically as follows:

[0133]

[0134] It is the predicted value of the three-class confidence for each grid cell, and its size is consistent with the size of the biased rasterized map. Each grid cell includes the predicted probabilities of the three classes. For the confidence assessment module, θ represents the model parameters of this module; This is a characteristic of the bird's-eye view. This is a biased rasterized map. These are the reliable, erroneous, and mapless probability maps obtained after softmax normalization; softmax is used to make the sum of the three probabilities of the same grid cell equal to 1.

[0135] Here, "reliable" indicates that the current observation supports a biased rasterized map and can be used strongly; "incorrect" indicates that the current observation conflicts with prior knowledge, and the influence of prior knowledge should be reduced; "no map" indicates that the raster does not have the map elements that need to be constructed. Further conversion into a single-channel reliability strength map yields the following mathematical expression:

[0136]

[0137] In the formula This represents a grid position in the bird's-eye view grid. The threshold value for the no-map category is ε; ε in the formula is a positive constant set to prevent the denominator from being zero. When the no-map probability reaches or exceeds this threshold, the reliability strength is set to 0.

[0138] Therefore, step 4 essentially involves evaluating the confidence level for each grid cell. The confidence evaluation module obtains the classification results based on the bird's-eye view features and the biased rasterized map. The biased rasterized map is obtained by rasterizing the biased vector map obtained in step 1.

[0139] The three classification results output in step 4 serve as a unified control signal to control the subsequent generation of dual prior queries and confidence-guided feature fusion. During the training phase, the three types of supervisory labels are used to supervise the confidence evaluation module in learning the classification result, thus corresponding to the aforementioned key point 1.

[0140] Table 2 below shows the processing rules for the three classification results in subsequent modules:

[0141] Judgment result meaning Follow-up processing reliable The characteristics of rasterized maps and bird's-eye views are basically the same. Generate corresponding prior queries and allow stronger prior queries to participate in confidence-guided feature fusion. mistake The features of the rasterized map are inconsistent with the bird's-eye view, and it may be outdated, misaligned, or have semantic errors. No prior queries are generated, and the influence of priors is reduced during feature fusion, with bird's-eye view features being the primary focus. No map No map elements required for construction were detected in this raster. No prior queries are generated, and no biased vector map content is injected.

[0142] Table 2: Processing Rules Table

[0143] like Figure 3 As shown, in step 5, the dual-priority query generation module is invoked to generate a raster prior query. A biased rasterized map is used to represent the spatial occupancy of road elements within the bird's-eye view raster. The system only allows rasters whose classification results are considered "reliable" to participate in the raster prior query generation, avoiding directly sending erroneous priors to the decoder. For rasters that pass the selection, the system records their spatial location and map category, forming a raster prior query unit. The purpose of the raster prior query is to provide the decoder with information about the existence of map elements at their location.

[0144] Map occupancy indicates whether the corresponding raster contains map elements such as lane dividers, road boundaries, or pedestrian crossings; classification result indicates whether the biased prior of the raster can be used; spatial location and map category are used to help the map decoder locate and identify map elements. The query unit is a numerical representation input to the map decoder, not the map element itself.

[0145] Subsequently, the dual-priority query generation module generates vector prior queries. The aligned biased vector map is used to represent the instance identity, point order, geometric shape, and topological relationships of map elements. For each vector point, the system retains only vector points whose classification results are reliable, based on the classification results of the raster region corresponding to that vector point, forming a vector query unit.

[0146] Specifically, old map points classified as reliable can be used to form vector query units; old map points classified as incorrect or without a map are not included in vector query generation. For the retained vector points, the system further models the geometry of individual map elements and the topological relationships between different map elements through intra-instance and inter-instance relationships, respectively.

[0147] Intra-instance relations are used to maintain the continuity of point order within a lane line or road boundary; inter-instance relations are used to maintain the spatial relationships between different elements such as lane lines, road boundaries, and pedestrian crossings. This processing enables the vector query unit to provide not only lane line information but also geometry-related and connectivity-related information.

[0148] In step 5, after obtaining the raster prior query unit and the vector prior query unit, they are merged to form a new prior query unit. The raster prior query unit provides spatial distribution and location information, while the vector prior query unit provides the shape, point order, and connectivity of candidate map elements. The classification results from step 4 have already been used to filter out unreliable raster and vector points. Therefore, this prior query unit is not a direct copy of the old map, but rather a result of merging the two verified prior queries.

[0149] Step 5 above uses a dual prior query generation module to generate raster prior query units and vector query units respectively. Then, the two prior queries are merged and combined with a learnable random query unit to form a prior query unit, so that the decoder can obtain category, geometry, spatial location and reliability information before it starts prediction, which corresponds to the aforementioned key point 3.

[0150] like Figure 4 As shown, in step 6, the confidence-guided feature fusion module is invoked to control the fusion of map prior features and bird's-eye view features based on the classification results of step 4. Reliable rasters can make strong use of priors; faulty rasters weaken priors and rely primarily on bird's-eye view features; no map rasters introduce no priors.

[0151] Prior map features are extracted from the biased rasterized map by the map feature extraction module. When fusing prior map features with bird's-eye view features, the core control variable is the reliability weight. A higher reliability weight indicates greater prior involvement; a lower reliability weight indicates that the current observation dominates. This core fusion relationship is expressed mathematically as follows:

[0152]

[0153] In the formula For initial fusion features, It is a feature of bird's-eye view. It consists of three types of confidence levels The calculated map reliability weights corresponding to the grid locations; These are the prior features obtained after processing a partially rasterized map, i.e., map prior features, where ⊙ represents point-by-point multiplication.

[0154] This formula expresses an intuitive principle: not all priors are fused, but only those priors that remain reliable after verification by current observations are strongly fused. The first part of step 6 above provides structural constraints from reliable priors in the reliable region and suppresses contamination from erroneous priors in the low-reliability region, thus corresponding to the aforementioned key point 4.

[0155] After completing the above fusion, the latter part of step 6 continues to perform local consistency verification between the initial fused features and the biased rasterized map. Since a single confidence assessment may still be affected by occlusion or noise, the system checks again after fusion whether the two are consistent in local locations. This gating relationship is expressed mathematically as follows:

[0156]

[0157] In the formula, G(u) is the local consistency gating weight of grid u; ϕ is the processing operation of stitching the initial fused features with the biased rasterized map; σ is a multi-layer fully connected neural network used to predict the gating weight based on the stitching result.

[0158] To avoid sudden changes in feature distribution caused by simply setting them to zero, the system sets a lower bound on the base weights corresponding to the grid positions to preserve stable basic information, rather than directly using low-confidence priors as reliable priors. The distribution preservation weights for each grid are calculated. :

[0159]

[0160]

[0161] In the formula Let u be the prior reliable probability at grid u; Let be the prior error probability at grid u; symbol This means that only positive values ​​are retained, therefore The positive part represents the difference between the prior reliable probability and the prior false probability at grid u; The lower bound of the basic weight of grid u is used to avoid directly clearing the prior to zero; it is a pre-set threshold. Maintain weights for the distribution of grid u. The single-channel reliability strength calculated above using three types of confidence levels is... The verification result at grid u. The final prior fusion weights for each grid cell;

[0162] Ultimately, the system obtains map prediction features after reliability control and local verification. This is intended for use by map decoders based on the Transformer architecture. The map prediction feature construction retains the following mathematical expression:

[0163]

[0164] This represents a biased rasterized map after final weighting. More reliable locations are retained, while fewer unreliable locations are retained. The representation of stitching the two together means merging the enhanced bird's-eye view features and the weighted biased rasterized map in the channel dimension. The fusion network is used to transform the stitched features into map prediction features that can be used by the decoder. It can be a convolutional layer, a lightweight neural network, or a feature mapping module. The map decoder uses predicted map features for output. Subsequent decoders then use these features to output vector map elements such as lane lines, road boundaries, and pedestrian crossings.

[0165] In this process, local consistency gating is responsible for secondary verification, the lower bound of basic weights is responsible for maintaining feature stability, and the final fusion weights are responsible for determining the actual impact of the prior on the decoder. The difference between local consistency verification and the aforementioned confidence assessment lies in their inputs and models. Local consistency verification inputs the initial fused features and a biased rasterized map into the network to extract weights, while the aforementioned confidence assessment uses unfused bird's-eye view features.

[0166] Step 6 above performs local verification again after prior propagation, and at the same time maintains the stability of feature distribution through the lower bound of basic weights, thus corresponding to the aforementioned key point 5.

[0167] Step 7: Decode and output the predicted vector map, which is a series of line segments or polygons composed of points. The map decoder uses a prior query unit obtained by merging two prior queries and map prediction features after consistency verification to output the category, confidence score, and vector point sequence for each map element. The vector point sequence can be written as [p_1, p_2, ..., p_N], where N represents the number of control points, p_i=(x_i, y_i) represents the i-th control point in the vehicle local bird's-eye view coordinate system, and x_i and y_i represent the coordinates of the point on the two coordinate axes, respectively. Adjacent control points in the sequence are connected sequentially to form line segments, and multiple line segments constitute the corresponding vector map element. In reliable areas, the decoder can refer to the long-range geometry and topology provided by the aligned biased vector map; in error areas, the decoder relies more on the current bird's-eye view features.

[0168] The final output is not a copy of the old map, but a new vector map constrained by the combined constraints of the real-world image and the prior biased vector map.

[0169] Step 8 is performed only during the training phase, setting training constraints. During training, three types of supervisory labels are generated using the rasterized unbiased vector map and the biased rasterized map obtained by rasterizing the biased vector map. These labels are used to teach the model to identify which priors are reliable, which priors are incorrect, and which regions lack map elements. The label generation rules retain the following mathematical expression:

[0170]

[0171] In the formula, L(u) is the supervision label of grid u. For rasterized unbiased vector maps, This is a biased rasterized map obtained by rasterizing a biased vector map; the symbol ∧ means "and", and "otherwise" means the remaining cases; 0, 1, and 2 represent the three categories of no map, error, and reliability, respectively.

[0172] When training the confidence evaluation module, the system supervises the three types of labels, enabling the model to learn to reduce the influence of erroneous priors when real-world road changes, construction, misalignment, or map outdated conditions occur. For vector map prediction itself, training still focuses on whether the category is correct, whether the point location is accurate, and whether the direction of the line segments formed by adjacent points is reasonable.

[0173] In addition to reliability supervision, the system also supervises the final output vector map prediction results, including map element categories, instance confidence, vector point locations, and the direction of line segments formed by adjacent points. These supervisions collectively ensure two things: first, the accuracy of the map itself; and second, that the system knows when to trust a biased vector map. The overall training objective can be summarized as minimizing both the vector map prediction error and the confidence assessment error.

[0174] Step 8 above explains that the unbiased vector map is only used to generate supervised labels during the training phase and is not a necessary input for the inference phase. This training constraint enables the model to learn both how to determine the credibility of priors and how to output accurate vector maps, thus corresponding to the aforementioned key point 2.

[0175] Step 9: Apply the vector map output in Step 7 during the inference phase. At this stage, the system only receives the current real-world image and the biased vector map; the unbiased vector map is not required. When the prior is empty or has low confidence, the prior contribution automatically decreases, and the system mainly relies on the features of the current bird's-eye view for online mapping. The output vector map is used to model the vehicle's surrounding environment and downstream autonomous driving trajectory planning, and can also report erroneous areas to the map maintenance system.

[0176] Example 2: Separate Auxiliary Reliability Monitoring Branch

[0177] In a preferred embodiment, the method sets up a separate auxiliary reliability supervision branch outside the main map construction path in steps 1 to 7. The main map construction path still performs the confidence assessment in step 4, the dual prior query generation and merging in step 5, the confidence-guided feature fusion, local consistency verification, and distribution-preserving weight calculation in step 6; this auxiliary branch is an optional training process in step 8, providing auxiliary reliability supervision only during the training phase, and is not used as input or output for steps 1 to 7 in the inference phase.

[0178] This auxiliary branch does not change the dual prior query generation method in step 5, nor does it directly change the main fusion weights in step 6. Its input includes the bird's-eye view features after gradient stopping and the filtered single-channel prior occupancy map. The bird's-eye view feature encoder is the network part that generates the bird's-eye view features in the main map construction path; "gradient stopping" means that the training error of the auxiliary branch does not update this network part in reverse along this input. The output of the auxiliary branch also does not participate in the calculation of the main fusion weights.

[0179] Specifically, this auxiliary branch first extracts single-channel occupancy information from the biased rasterized map, retaining only locations where map elements may exist, thus reducing the interference of large background areas on confidence assessment. Only locations exceeding the effective occupancy threshold are included in the auxiliary reliability judgment; locations clearly lacking map elements are excluded from reliability / error assessment. This allows the auxiliary task to focus on "locations in the biased vector map that actually contain content," avoiding the overwhelming of genuine confidence assessments by treating locations without map backgrounds as a large number of training samples.

[0180] The separate auxiliary reliability supervision branch receives the bird's-eye view features after stopping gradients and the filtered single-channel prior occupancy map, outputting auxiliary reliable, incorrect, and no-map classification results. We also introduce a stopping gradient rule, enabling this separate auxiliary reliability supervision branch to learn to judge the reliability of the prior, but its training error does not directly and negatively affect the main bird's-eye view feature encoder. Therefore, this separate auxiliary reliability supervision branch is more like a bypass diagnostic tool during the training phase, used to improve confidence assessment capabilities rather than replacing the main map construction path.

[0181] Based on the aforementioned three types of supervisory labels, auxiliary supervision can be divided into two levels: the first level determines whether map elements exist at the current location, and the second level only determines the reliability of the prior information at locations where map elements exist. This division can reduce interference from the absence of map background and concentrate reliable / incorrect judgments on the road elements that truly need to be verified.

[0182] The output of this auxiliary branch can be further summarized into "map-containing probability" and "reliability probability within a map-containing area". The map-containing probability is used to distinguish between background and road elements; the reliability probability is used to determine whether the biased vector map is reliable at locations where road elements exist. The auxiliary reliability supervision loss includes map-containing / no-map supervision terms and reliable / erroneous supervision terms.

[0183] In this training process, map-based / map-free supervision is computed on all grid cells, while reliable / error-based supervision is computed only on grid cells containing map elements. This avoids interference from a large number of map-free background areas in the confidence assessment, improving the diagnostic value of this auxiliary branch. When using this auxiliary branch, the overall training objective adds separate auxiliary reliability supervision to the main training objective. This auxiliary supervision is only used during the training phase to improve the model's ability to identify reliable and error-based priors.

[0184] Since this auxiliary branch uses a stopping gradient input and its output does not participate in the calculation of the main fusion weights, this separate auxiliary reliability supervision branch can enhance the reliability diagnosis capability while reducing the disturbance to the main map construction path.

[0185] Alternative implementation methods:

[0186] The inputs, network structure, label categories, control parameters, and prior sources in the above embodiments can all be replaced or extended according to the deployment conditions without affecting the core idea of ​​the present invention: "first verify the biased vector map, and then selectively use the prior."

[0187] Regarding sensor input, the current observation input is not limited to multi-view images, but can also include LiDAR point clouds, millimeter-wave radar features, or multimodal combination inputs. Regardless of the sensor used, the system first converts the current observation into a bird's-eye view space before using it to verify the reliability of the biased vector map. Therefore, cameras, LiDAR, millimeter-wave radar, or multimodal combinations can all serve as sources of current observations. The key requirement is that they ultimately provide a bird's-eye view comparable to the biased vector map. picture Therefore, this invention does not limit the specific sensor type, but rather the processing relationship of "checking the biased vector map using the current observation".

[0188] Regarding reliability labels, this embodiment uses reliable, faulty, and groundless labels. Figure 3 Category labels. If expanded to more categories such as geometric misalignment, semantic error, missing map, added road elements or topology expiration, the output channel of the confidence assessment module, supervision labels, the mapping from category to reliability strength, and the screening rules of step 5 should all be adjusted synchronously; if the above adjustments are not made, formulas (1), (2) and (7) are still executed according to the three categories of labels.

[0189] Regarding prior strategies, different prior retention strategies and fusion intensities can be adopted for highways, urban intersections, construction zones, or parking areas. For example, historical maps should be used more cautiously in construction zones; in structurally stable highway scenarios, the intensity of using reliable priors can be appropriately increased.

[0190] Regarding prior sources, currently available biased vector maps are not limited to offline high-precision maps; they can also come from standard-precision maps, historical predictions, cloud-based maps, or crowdsourced map update results. Regardless of whether the biased vector map originates from an offline map, historical prediction, cloud-based map, or crowdsourced update result, it must first be converted to a vehicle coordinate system and then undergo reliability verification. Therefore, this invention does not protect a specific fixed map source, but rather a processing mechanism that verifies any available prior information before selectively using it.

[0191] The alternative implementations described above can be used individually or in combination. For example, in cost-constrained scenarios, multi-view images and lightweight segmentation networks can be employed; in high-reliability deployment scenarios, images and LiDAR can be used as joint inputs, along with multi-category reliability labels; and in cloud-based map update scenarios, erroneous areas identified online can be sent back to the map maintenance system. These changes do not alter the fundamental technical approach of this invention, which involves verifying biased vector maps through current observations and selectively utilizing prior knowledge based on reliability.

[0192] Map query is the internal representation of vector map elements such as lane lines, road boundaries, and pedestrian crossings used by the decoder; the priors, after reliability verification, only participate in the generation of dual prior queries as controlled hints.

[0193] The separate auxiliary reliability supervision branch is a bypass module used during the training phase to improve confidence assessment capabilities; its output does not directly control the main map construction path.

[0194] The above terms all revolve around the same technical approach: first, treat the biased vector map as the object to be verified, and then selectively use priors based on the reliability results.

[0195] The following are system embodiments corresponding to the above method embodiments. This embodiment can be implemented in conjunction with the above embodiments. The relevant technical details mentioned in the above embodiments are still valid in this embodiment, and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the above embodiments.

[0196] This invention also proposes a robust online vector high-precision map construction device oriented towards biased priors, comprising:

[0197] Module 1 acquires a biased vector map of the vehicle's location and real-world images taken from multiple perspectives by the vehicle's own camera. Based on the camera's parameters, the biased vector map is transformed to the coordinate system of the real-world images, i.e., the vehicle's coordinate system.

[0198] Module 2 rasterizes the biased vector map to obtain a biased rasterized map;

[0199] Module 3 processes real-world images from multiple perspectives to obtain bird's-eye view features;

[0200] Module 4 calls the confidence assessment module to compare the features of the bird's-eye view with the biased rasterized map, and obtains the classification result of each raster in the biased rasterized map as reliable, erroneous or without map;

[0201] Module 5, based on the classification result, calls the dual prior query generation module. First, it filters the biased raster map, retaining only the raster that the classification result belongs to is reliable, forming a raster prior query unit. Then, for each vector point in the biased vector map, based on the classification result of the vector point being in the corresponding raster region in the biased raster map, it retains only the vector point that the classification result belongs to is reliable, forming a vector prior query unit. Finally, the two prior query units are merged and combined with the learnable random query unit to obtain the prior query unit.

[0202] Module 6 extracts features from the biased rasterized map to obtain prior map features. Then, it calls the confidence-guided feature fusion module to first fuse the prior map features with the bird's-eye view features based on the classification results of each raster, obtaining initial fused features. Next, it performs local consistency checks on the initial fused features using the classification results of each raster in the biased rasterized map, ultimately obtaining the predicted map features.

[0203] Module 7, the map decoder, outputs the category, confidence level and vector point sequence of each map element based on the prior query unit and map prediction features, as the vector map prediction result under the joint constraints of the real image and the biased vector map.

[0204] The robust online vector high-precision map construction device oriented towards biased priors, wherein

[0205] Module 1 includes: obtaining the old map, historical prediction map, cloud map or crowdsourced updated map that the vehicle can obtain at the current moment as the biased vector map;

[0206] Module 3, the bird's-eye view feature is the road representation after converting multiple perspective real-shot images into a top-down space;

[0207] Module 4 obtains the classification result for each grid cell in the biased rasterized map using the following formula:

[0208]

[0209] These are the predicted confidence scores for each raster region, and their size is consistent with the size of the biased rasterized map. Each raster includes three categories of predicted probabilities. It's a prediction module. This is a characteristic of the bird's-eye view. It is a rasterized map; These are probability graphs showing that the prediction results are reliable, incorrect, and without a map.

[0210] Converted into a single-channel reliability strength map, its size is consistent with the size of the biased rasterized map:

[0211]

[0212] In the formula This is a single-channel reliability strength diagram. Represents one of the grid cells. The threshold for whether a map exists; These represent the probabilities of the three probability maps at grid u;

[0213] The initial fusion process for module 6 is as follows:

[0214]

[0215] In the formula These are initial fusion features. It is a feature of bird's-eye view. This is the scripture The reliability of the calculated map; These are the prior features obtained after processing a partially rasterized map, i.e., map prior features. This is point-by-point multiplication;

[0216] Module 6 also includes: performing local consistency verification on the initial fused features.

[0217]

[0218] In the formula This is the verification result at grid u. It is a feature splicing operation. This represents a multi-layer fully connected neural network, responsible for predicting gating weights;

[0219] Calculate the distribution preservation weights for each grid cell :

[0220]

[0221]

[0222] In the formula Let u be the prior reliable probability at grid u; Let be the prior error probability at grid u; symbol This means that only positive values ​​are retained, therefore The positive part represents the difference between the prior reliable probability and the prior false probability at grid u; The lower bound of the basic weight of grid u is used to avoid directly clearing the prior to zero; it is a pre-set threshold. Maintain weights for the distribution of grid u. This is the single-channel reliability strength diagram calculated from the three types of confidence levels mentioned above. This is the verification result at grid u; The final prior fusion weights for each grid cell;

[0223] The predicted features of the map are obtained using the following formula. :

[0224]

[0225] This is the result after weighting a biased rasterized map. The final prior fusion weights for each grid cell To merge the weighted graph; The representative represents the combination of the two, that is, the initial fused features. And weighted biased raster map Merge at the channel dimension; The fusion network is used to convert the stitched features into map prediction features that the decoder can use.

[0226] In module 7, the map decoder, based on the prior query unit and the map prediction features, refers to the long-range geometry and topology provided by the biased vector map in reliable areas, and refers to the bird's-eye view features in erroneous areas.

[0227] The aforementioned robust online vector high-precision map construction device oriented towards biased priors further includes:

[0228] Module 8: During the training phase, three types of supervised labels are generated using the rasterized unbiased vector map and the currently obtained biased rasterized map. The label generation rules retain the following mathematical expression:

[0229]

[0230] In the formula It is a rasterized unbiased vector map. It is the currently obtained biased rasterized map; 0 represents "and", meaning that both the preceding and following conditions are met simultaneously; 0, 1, and 2 represent the three categories of no map, error, and reliable, respectively.

[0231] When training the confidence evaluation module, the three types of labels are supervised to enable the model to learn to reduce the influence of erroneous priors when there are real-world road changes, construction, misalignment, or map outdated conditions. For vector map prediction itself, training still focuses on whether the category is correct, whether the point location is accurate, and whether the direction of the line segments formed by adjacent points is correct.

[0232] In addition to reliability supervision, the final output vector map prediction results are also supervised, including map element categories, instance confidence, vector point locations, and the direction of line segments formed by adjacent points.

[0233] The present invention also proposes an electronic device, including the aforementioned robust online vector high-precision map construction device oriented towards biased priors, wherein the electronic device may be connected to an information display device, which is used to display the vector map with user-set display parameters, attributes or through an artificial intelligence model.

[0234] The present invention also proposes a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the robust online vector high-precision map construction method oriented towards biased priors.

[0235] The present invention also proposes a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, it implements the steps of the robust online vector high-precision map construction method oriented towards biased priors.

Claims

1. A robust online vector high-precision map construction method oriented towards biased priors, characterized in that, include: Step 1: Obtain a biased vector map of the vehicle's location and real-world images taken from multiple perspectives by the vehicle's own camera. Based on the camera's parameters, transform the biased vector map to the coordinate system of the real-world images, i.e., the vehicle's coordinate system. Step 2: Rasterize the biased vector map to obtain a biased rasterized map; Step 3: Process the real-world images from multiple perspectives to obtain bird's-eye view features; Step 4: Call the confidence assessment module to compare the features of the bird's-eye view with the biased rasterized map to obtain the classification result of each raster in the biased rasterized map as reliable, erroneous, or without map. Step 5: Based on the classification result, the dual prior query generation module is invoked. First, the biased raster map is filtered, retaining only the raster that the classification result belongs to is reliable, forming a raster prior query unit. Then, for each vector point in the biased vector map, based on the classification result of the vector point being in the corresponding raster region in the biased raster map, only the vector point that the classification result belongs to is reliable is retained, forming a vector prior query unit. Finally, the two prior query units are merged and combined with the learnable random query unit to obtain the prior query unit. Step 6: Extract features from the biased rasterized map to obtain the map's prior features; Then, the confidence-guided feature fusion module is called to first fuse the map prior features with the bird's-eye view features based on the classification results of each grid to obtain the initial fused features; Then, the initial fused features are locally consistent by combining the classification results of each grid in the biased rasterized map, and finally the map prediction features are obtained. Step 7: Based on the prior query unit and map prediction features, the map decoder outputs the category, confidence level and vector point sequence of each map element as the vector map prediction result under the joint constraints of the real image and the biased vector map.

2. The robust online vector high-precision map construction method oriented towards biased priors as described in claim 1, characterized in that, Step 1 includes: obtaining the old map, historical prediction map, cloud map or crowdsourced updated map that the vehicle can obtain at the current moment as the biased vector map; In step 3, the bird's-eye view feature is the road representation after converting multiple real-world images from different perspectives into a top-down space. Step 4: Obtain the classification result for each grid cell in the biased rasterized map using the following formula: These are the predicted confidence scores for each raster region, and their size is consistent with the size of the biased rasterized map. Each raster includes three categories of predicted probabilities. It's a prediction module. This is a characteristic of the bird's-eye view. It is a rasterized map; These are probability graphs showing that the prediction results are reliable, incorrect, and without a map. Converted into a single-channel reliability strength map, its size is consistent with the size of the biased rasterized map: In the formula This is a single-channel reliability strength diagram. Represents one of the grid cells. The threshold for whether a map exists; These represent the probabilities of the three probability maps at grid u; Step 6, the initial fusion process, specifically involves: In the formula These are initial fusion features. It is a feature of bird's-eye view. This is the scripture The reliability of the calculated map; These are the prior features obtained after processing a partially rasterized map, i.e., map prior features. This is point-by-point multiplication; Step 6 also includes: performing local consistency verification on the initial fused features. In the formula This is the verification result at grid u. It is a feature splicing operation. This represents a multi-layer fully connected neural network, responsible for predicting gating weights; Calculate the distribution-preserving weights for each grid cell. : In the formula Let u be the prior reliable probability at grid u; Let be the prior error probability at grid u; symbol This means that only positive values ​​are retained, therefore The positive part represents the difference between the prior reliable probability and the prior false probability at grid u; The lower bound of the basic weight of grid u is used to avoid directly clearing the prior to zero; it is a pre-set threshold. Preserve the weights for the distribution of grid u; This is the single-channel reliability strength diagram calculated from the three types of confidence levels mentioned above. This is the verification result at grid u; The final prior fusion weights for each grid cell; The predicted features of the map are obtained using the following formula. : This is the result after weighting the biased rasterized map. The final prior fusion weights for each grid cell To merge the weighted graph; The representative represents the combination of the two, that is, the initial fused features. And weighted biased raster map Merge at the channel dimension; The fusion network is used to convert the stitched features into map prediction features that the decoder can use. In step 7, the map decoder, based on the prior query unit and the map prediction features, refers to the long-range geometry and topology provided by the biased vector map in reliable areas, and refers to the bird's-eye view features in error areas.

3. The robust online vector high-precision map construction method based on biased priors as described in claim 1, characterized in that, Also includes: Step 8: During the training phase, three types of supervised labels are generated using the rasterized unbiased vector map and the currently obtained biased rasterized map. The label generation rules retain the following mathematical expression: In the formula It is a rasterized unbiased vector map. It is the currently obtained biased rasterized map; 0 represents "and", meaning that both the preceding and following conditions are met simultaneously; 0, 1, and 2 represent the three categories of no map, error, and reliable, respectively. When training the confidence evaluation module, the three types of labels are supervised to enable the model to learn to reduce the influence of erroneous priors when there are real road changes, construction, misalignment, or map expiration. For vector map prediction itself, the training still focuses on whether the category is correct, whether the position of the point is accurate, and whether the direction of the line segment formed by adjacent points is correct. In addition to reliability supervision, the final output vector map prediction results are also supervised, including map element categories, instance confidence, vector point locations, and the direction of line segments formed by adjacent points.

4. The robust online vector high-precision map construction method based on biased priors as described in claim 1, characterized in that, The vector map obtained in step 7 is used to model the vehicle's surrounding environment and perform trajectory planning tasks for downstream autonomous driving.

5. A robust online vector high-precision map construction device oriented towards biased priors, characterized in that, include: Module 1 acquires a biased vector map of the vehicle's location and real-world images taken from multiple perspectives by the vehicle's own camera. Based on the camera's parameters, the biased vector map is transformed to the coordinate system of the real-world images, i.e., the vehicle's coordinate system. Module 2 rasterizes the biased vector map to obtain a biased rasterized map; Module 3 processes real-world images from multiple perspectives to obtain bird's-eye view features; Module 4 calls the confidence assessment module to compare the features of the bird's-eye view with the biased rasterized map, and obtains the classification result of each raster in the biased rasterized map as reliable, erroneous or without map; Module 5, based on the classification result, calls the dual prior query generation module. First, it filters the biased raster map, retaining only the raster that the classification result belongs to is reliable, forming a raster prior query unit. Then, for each vector point in the biased vector map, based on the classification result of the vector point being in the corresponding raster region in the biased raster map, it retains only the vector point that the classification result belongs to is reliable, forming a vector prior query unit. Finally, the two prior query units are merged and combined with the learnable random query unit to obtain the prior query unit. Module 6 extracts features from the partially rasterized map to obtain the map's prior features; Then, the confidence-guided feature fusion module is called to first fuse the map prior features with the bird's-eye view features based on the classification results of each grid to obtain the initial fused features; Then, the initial fused features are locally consistent by combining the classification results of each grid in the biased rasterized map, and finally the map prediction features are obtained. Module 7, the map decoder, outputs the category, confidence level and vector point sequence of each map element based on the prior query unit and map prediction features, as the vector map prediction result under the joint constraints of the real image and the biased vector map.

6. The robust online vector high-precision map construction device oriented towards biased priors as described in claim 5, characterized in that, Module 1 includes: obtaining the old map, historical prediction map, cloud map or crowdsourced updated map that the vehicle can obtain at the current moment as the biased vector map; Module 3, the bird's-eye view feature is the road representation after converting multiple perspective real-shot images into a top-down space; Module 4 obtains the classification result for each grid cell in the biased rasterized map using the following formula: These are the predicted confidence scores for each raster region, and their size is consistent with the size of the biased rasterized map. Each raster includes three categories of predicted probabilities. It's a prediction module. This is a characteristic of the bird's-eye view. It is a rasterized map; These are probability graphs showing that the prediction results are reliable, incorrect, and without a map. Converted into a single-channel reliability strength map, its size is consistent with the size of the biased rasterized map: In the formula This is a single-channel reliability strength diagram. Represents one of the grid cells. The threshold for whether a map exists; These represent the probabilities of the three probability maps at grid u; The initial fusion process for module 6 is as follows: In the formula These are initial fusion features. It is a feature of bird's-eye view. This is the scripture The reliability of the calculated map; These are the prior features obtained after processing a partially rasterized map, i.e., map prior features. This is point-by-point multiplication; Module 6 also includes: performing local consistency verification on the initial fused features. In the formula This is the verification result at grid u. It is a feature splicing operation. This represents a multi-layer fully connected neural network, responsible for predicting gating weights; Calculate the distribution-preserving weights for each grid cell. : In the formula Let u be the prior reliable probability at grid u; Let be the prior error probability at grid u; symbol This means that only positive values ​​are retained, therefore The positive part represents the difference between the prior reliable probability and the prior false probability at grid u; The lower bound of the basic weight of grid u is used to avoid directly clearing the prior to zero; it is a pre-set threshold. Preserve the weights for the distribution of grid u; This is the single-channel reliability strength diagram calculated from the three types of confidence levels mentioned above. This is the verification result at grid u; The final prior fusion weights for each grid cell; The predicted features of the map are obtained using the following formula. : This is the result after weighting the biased rasterized map. The final prior fusion weights for each grid cell To merge the weighted graph; The representative represents the combination of the two, that is, the initial fused features. And weighted biased raster map Merge at the channel dimension; The fusion network is used to convert the stitched features into map prediction features that the decoder can use. In module 7, the map decoder, based on the prior query unit and the map prediction features, refers to the long-range geometry and topology provided by the biased vector map in reliable areas, and refers to the bird's-eye view features in erroneous areas.

7. The robust online vector high-precision map construction device oriented towards biased priors as described in claim 5, characterized in that, Also includes: Module 8: During the training phase, three types of supervised labels are generated using the rasterized unbiased vector map and the currently obtained biased rasterized map. The label generation rules retain the following mathematical expression: In the formula It is a rasterized unbiased vector map. It is the currently obtained biased rasterized map; 0 represents "and", meaning that both the preceding and following conditions are met simultaneously; 0, 1, and 2 represent the three categories of no map, error, and reliable, respectively. When training the confidence evaluation module, the three types of labels are supervised to enable the model to learn to reduce the influence of erroneous priors when there are real road changes, construction, misalignment, or map expiration. For vector map prediction itself, the training still focuses on whether the category is correct, whether the position of the point is accurate, and whether the direction of the line segment formed by adjacent points is correct. In addition to reliability supervision, the final output vector map prediction results are also supervised, including map element categories, instance confidence, vector point locations, and the direction of line segments formed by adjacent points.

8. An electronic device, characterized in that, The invention includes a robust online vector high-precision map construction device for biased priors as described in claims 5-7, wherein the electronic device is connected to an information display device, which displays the vector map using user-set display parameters, attributes, or through an artificial intelligence model.

9. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the robust online vector high-precision map construction method oriented towards biased priors as described in any one of claims 1-4.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the robust online vector high-precision map construction method oriented towards biased priors as described in any of claims 1-4.

Images