Online high-precision map building and model building method, terminal and medium
The MapSwin model architecture solves the problems of insufficient semantic information and difficulty in handling complex road structures in online high-precision map construction, and realizes high-precision, real-time map construction capabilities, which are suitable for autonomous driving systems.
Patent Information
- Application Number
- CN202411673861.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-21
- Publication Date
- 2026-05-22
AI Technical Summary
Existing online high-precision map building technologies suffer from insufficient semantic information, an inadequate trade-off between real-time performance and accuracy, and difficulties in handling complex road structures.
The MapSwin model architecture is adopted, including a backbone network, a view transformation network, a hierarchical query embedding network, and a model optimization network. Through more powerful feature extraction, data augmentation, and model integration strategies, an online high-precision map building model is constructed.
It achieves rich multi-category semantic information extraction, improves the accuracy and robustness of map construction, and can handle complex and irregular road structures in real time, meeting the real-time requirements of autonomous driving systems.
Smart Images
Figure CN122072638A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of map building technology, specifically to an online high-precision map building and model building method, terminal, and medium. Background Technology
[0002] High-resolution maps (HD maps) are a core component of autonomous driving systems, providing detailed road information such as lane markings, road boundaries, and pedestrian crossing lines. These maps are typically built offline using SLAM (Simultaneous Localization and Mapping) based methods. However, this traditional approach suffers from several problems:
[0003] 1. Complex processing flow: The SLAM method requires a complex processing pipeline, involving multiple steps such as sensor data acquisition, feature extraction, data fusion and map generation.
[0004] 2. High maintenance costs: Offline built maps require regular updates and maintenance to reflect road changes, which places high demands on resources and time.
[0005] 3. Poor real-time performance: Offline methods cannot provide maps that are updated in real time, making it difficult to meet the needs of autonomous vehicles in dynamic environments.
[0006] With technological advancements, online high-precision map building has gradually become a research hotspot. By utilizing vehicle-mounted sensors (such as cameras and LiDAR), environmental data can be acquired in real time and local high-precision maps can be generated, thereby reducing the cost and complexity of offline map building. Although existing online high-precision map building technologies have made progress, the following problems still exist:
[0007] 1. Insufficient semantic information: Traditional lane detection methods can usually only detect lane lines and lack the ability to detect and extract semantic information from other road elements (such as road boundaries and pedestrian crossing lines).
[0008] 2. Trade-off between real-time performance and accuracy: While pursuing real-time performance, existing methods still have shortcomings in terms of accuracy and detail in map construction.
[0009] 3. Difficulty in handling complex road structures: Existing methods perform poorly in terms of handling capacity and robustness when faced with complex and irregular road structures. Summary of the Invention
[0010] To address the aforementioned shortcomings in existing technologies, this invention provides an online high-precision map construction and model building method, terminal, and medium.
[0011] According to one aspect of the present invention, a model building method for online high-precision map construction is provided, characterized in that it includes:
[0012] Based on the MapTr model architecture, a backbone network, a view transformation network, a hierarchical query embedding network, and a model optimization network are constructed sequentially to obtain the initial model; where:
[0013] The backbone network is used to extract image features;
[0014] The view transformation network is used to transform the image features from 2D image features into bird's-eye view features;
[0015] The hierarchical query embedding network is used to explicitly encode and decode map elements of the bird's-eye view features through a hierarchical query embedding mechanism.
[0016] The model optimization network is used to optimize the model parameters through loss function and data augmentation;
[0017] The initial model is trained to obtain a map building model;
[0018] By integrating multiple map building models, an online high-precision map building model is obtained.
[0019] Preferably, the backbone network is constructed using any one of ResNet-50, Swin-tiny, Swin-base, and ConvNeXt, and is used to extract features from multi-view camera images; wherein:
[0020] The multi-view camera image features are represented as: Fi={Fi}Nview i=1, where Fi is the feature of the i-th view and Nview is the total number of views.
[0021] Preferably, the view transformation network is obtained by replacing the GKT module in the MapTR model with a spatial cross-attention layer proposed by the BEVFormer neural network model, while removing the temporal self-attention layer of the MapTR model. The encoder composed of multiple spatial cross-attention layers converts multi-scale 2D image features into BEV features to generate a unified bird's-eye view representation.
[0022] Preferably, the hierarchical query embedding network explicitly encodes each map element using a hierarchical query embedding mechanism, defining a set of instance-level queries and a set of point-level queries; wherein, the instance-level queries are used to focus on the center position of a certain map element, and the point-level queries are used to focus on the offset range of the entire map element relative to the center position; the instance-level queries and the point-level queries are combined to form a hierarchical query; each map element corresponds to a set of hierarchical queries; the head of the hierarchical query embedding network contains multiple cascaded decoder layers, each decoder layer consisting of a multi-head self-attention layer and a variable cross-attention layer; the hierarchical query sequence first focuses on the features of each element in its own sequence through the self-attention layer and updates itself, then performs variable cross-attention calculation with the BEV features through the variable attention layer and updates itself again, and finally regresses and predicts the two-dimensional normalized BEV coordinates of the reference point through a multilayer perceptron model.
[0023] Preferably, the model optimization network includes: a loss function layer and a data augmentation layer; wherein:
[0024] The loss function layer mainly consists of a comprehensive loss function composed of classification loss, point-to-point loss, edge direction loss, and point-to-line loss, which weigh different loss terms to optimize model performance. Furthermore, the comprehensive loss function L formed by the classification loss Lcls, point-to-point loss Lp2p, edge direction loss Ldir, and point-to-line loss Lp2l is expressed as: L=λLcls+αLp2p+βLdir+γLp2l;
[0025] The data augmentation layer employs geometric and photometric transformation methods to optimize the model's stability under different object orientations and visibility scenarios; it also uses a class-balanced grouping sampling method to optimize the data class balance performance.
[0026] Preferably, training the initial model to obtain the map construction model includes:
[0027] Build the dataset;
[0028] Using the AdamW optimizer and cosine annealing scheduling, the initial map building model is trained on the dataset to obtain the map building model.
[0029] Preferably, the model integration includes: static model integration and dynamic model integration, wherein multiple map construction models are integrated by removing duplicate elements using a non-maximum suppression method; wherein:
[0030] In the static integration, the threshold is set to a fixed value;
[0031] In the dynamic integration, a dynamic threshold is adjusted based on the element length, where the dynamic threshold δ is: δ = max(δmax, min(δmin, (l-lmin)*(δmax-δmin) / (lmax-lmin)+δmin)); where lmin represents the minimum length in the data, and lmax represents the maximum length in the data. δmax and δmin represent the set maximum and minimum thresholds, respectively. Furthermore, generally, δmax is between [2, 3], and δmin is between [0.5, 1].
[0032] The Chamfer distance is used as a similarity measure. The element with the highest confidence is selected and similar elements are deleted, and duplicate elements are removed. If the Chamfer distance is less than the set threshold, the two tuples are considered similar.
[0033] According to another aspect of the present invention, an online high-precision map construction method is provided, comprising:
[0034] Acquire data from multiple cameras;
[0035] An online high-precision map building model is provided. The camera data is used as the input of the online high-precision map building model, and a high-precision map is output to complete the online high-precision map building.
[0036] in:
[0037] The online high-precision map construction model is constructed using the model construction method described in any one of the above-mentioned inventions.
[0038] According to a third aspect of the present invention, a computer terminal is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it can be used to execute the model building method described in any one of the above inventions, or to execute the online high-precision map building method described in the above inventions.
[0039] According to a fourth aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, can be used to perform the model building method described in any one of the above-described inventions, or to perform the online high-precision map building method described in the above-described inventions.
[0040] By adopting the above technical solution, the present invention has at least one of the following beneficial effects compared with the prior art:
[0041] The online high-precision map construction and model building method, terminal, and medium provided by this invention demonstrate excellent performance in high-precision map construction tasks using the MapSwin model, showing a significant performance improvement compared to the benchmark model HDMapNet. By using a stronger backbone network, effective data augmentation techniques, and model ensemble strategies, the accuracy and robustness of map construction are significantly improved, resulting in performance enhancement.
[0042] The online high-precision map construction and model building method, terminal and medium provided by the present invention can process multi-camera input data in real time, dynamically build local high-precision maps, eliminate the need for offline manual intervention and meet real-time requirements.
[0043] The online high-precision map construction and model building method, terminal and medium provided by this invention provide rich multi-category semantic information, including lane lines, road boundaries, pedestrian crossing lines, etc., and can cope with complex and irregular road structures.
[0044] The online high-precision map construction and model building method, terminal and medium provided by this invention enable the processing of complex road structures. By adopting a multi-layer spatial encoder and a hierarchical query embedding mechanism, the model can handle complex and irregular road structures, thereby improving the adaptability of the autonomous driving system in various environments. Attached Figure Description
[0045] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:
[0046] Figure 1 This is a schematic diagram illustrating the working principle of a model building method for online high-precision map construction in a preferred embodiment of the present invention.
[0047] Figure 2 This is a flowchart illustrating the workflow of an online high-precision map construction method in a preferred embodiment of the present invention.
[0048] Figure 3 This is a schematic diagram of the components of an online high-precision map construction system in a preferred embodiment of the present invention. Detailed Implementation
[0049] The embodiments of the present invention are described in detail below: These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.
[0050] Existing online high-precision map building technologies typically suffer from insufficient semantic information, shortcomings in accuracy and detail despite pursuing real-time performance, and difficulties in handling complex road structures. To address these issues, one embodiment of this invention provides a model building method for online high-precision map construction. This method constructs a novel online high-precision map building model (MapSwin model) to achieve online high-precision map construction. This model improves upon the existing MapTr model by incorporating a stronger backbone network, superior view transformations, and more effective data augmentation and supervision methods to achieve higher accuracy and real-time performance, comprehensively enhancing the map building capabilities of autonomous driving systems.
[0051] Specifically, such as Figure 1 As shown, the model building method for online high-precision map construction provided in this embodiment may include the following operations:
[0052] Based on the MapTr model architecture, the backbone network, view transformation network, hierarchical query embedding network, and model optimization network are constructed sequentially to obtain the initial model; where:
[0053] The backbone network is used to extract image features;
[0054] View transformation network is used to convert image features from 2D image features into bird's-eye view (BEV) features;
[0055] Hierarchical query embedding network is used to explicitly encode and decode map elements of bird's-eye view features through a hierarchical query embedding mechanism;
[0056] Model optimization networks are used to optimize model parameters through loss functions and data augmentation.
[0057] The initial model is trained to obtain the map building model;
[0058] Multiple map building models are integrated to obtain an online high-precision map building model.
[0059] To obtain a more powerful feature extraction network and improve the ability to acquire image features, in some preferred embodiments, the backbone network may further include the following operations:
[0060] ResNet-50, Swin-tiny, Swin-base, and ConvNeXt networks are used as backbone networks to extract features from multi-view camera images; among them, ConvNeXt network can be further selected as the backbone network to further improve the feature extraction effect.
[0061] The multi-view camera image features are represented as: Fi={Fi}Nview i=1, where Fi is the feature of the i-th view and Nview is the total number of camera views.
[0062] Figure 1 In the middle, B t H represents the aerial view features at time t. t This preserves the temporal fusion features from time 0 to t. G and P are two points in the ground truth lane line and the predicted lane line, respectively. L left L right These represent the lane lines on the left and right sides of G.
[0063] To achieve view transformation, a new view transformation network is introduced to convert 2D image features into bird's-eye view (BEV) features. Based on this, in some preferred embodiments, the view transformation network may further include the following operations:
[0064] The view transformation network replaces the GKT module in the MapTR model by introducing a spatial cross-attention layer proposed by the BEVFormer neural network model, while removing the temporal self-attention layer of the MapTR model. An encoder composed of multiple spatial cross-attention layers converts multi-scale 2D image features into BEV features, generating a unified bird's-eye view representation. This view transformation network simplifies the feature transformation process and improves computational efficiency.
[0065] To achieve hierarchical query embedding, map elements are explicitly encoded and decoded using a hierarchical query embedding mechanism. Based on this, in some preferred embodiments, the hierarchical query embedding (Head) network may further include the following operations:
[0066] A hierarchical query embedding network explicitly encodes each map element using a hierarchical query embedding mechanism, defining a set of instance-level queries and a set of point-level queries. The instance-level queries focus on the center position of a map element, while the point-level queries focus on the offset range of the entire map element relative to the center position. The instance-level queries and point-level queries are combined to form a hierarchical query. Each map element corresponds to a set of hierarchical queries. The head of the hierarchical query embedding network contains multiple cascaded decoder layers, each consisting of a multi-head self-attention (MHSA) layer and a variable cross-attention layer. The hierarchical query sequence first updates itself by focusing on the features of each element in its sequence through the self-attention layer, then performs variable cross-attention calculation with the BEV features through the variable attention layer and updates itself again. Finally, it regresses and predicts the two-dimensional normalized BEV coordinates of the reference point through a multilayer perceptron model.
[0067] To achieve data augmentation and supervision, various data augmentation methods and improved loss functions are applied to enhance the robustness and accuracy of the model. Based on this, in some preferred embodiments, the model optimization network may further include: a loss function layer and a data augmentation layer; wherein:
[0068] The loss function layer mainly consists of a comprehensive loss function composed of classification loss, point-to-point loss, edge direction loss, and point-to-line loss, which weigh different loss terms to optimize model performance. Furthermore, the comprehensive loss function L formed by classification loss Lcls, point-to-point loss Lp2p, edge direction loss Ldir, and point-to-line loss Lp2l is expressed as: L=λLcls+αLp2p+βLdir+γLp2l;
[0069] The Data Augmentation layer applies geometric and photometric transformation methods to optimize the model's stability under different object orientations and visibility scenarios, making the model more robust. The Balanced Class Group Sampling (CBGS) method is used to make the training dataset more balanced, which optimizes the balance performance of data classes and effectively solves the problem of class imbalance, especially the problem of few pedestrian crossing instances.
[0070] In some preferred embodiments, geometric transformation methods, including flipping, scaling, and rotation, make the model more robust under different object orientations and scenes. More preferably, the flipping geometric transformation method is the most effective.
[0071] In some preferred embodiments, the photometric transformation method includes: improving the robustness of the model in low-visibility scenes by adjusting photometric parameters such as brightness, contrast, and saturation.
[0072] To enable the model to acquire online high-precision map building capabilities, in some preferred embodiments, the initial model is trained to obtain a map building model, and the following operations may be further included:
[0073] Build the dataset;
[0074] Using the AdamW optimizer and cosine annealing scheduling, and leveraging the dataset, the initial map-building model is trained to obtain the map-building model.
[0075] To obtain rich multi-category semantic information, in some preferred embodiments, the dataset uses the Argoverse2 dataset, where each sample contains multiple images, and the map elements in each image include: boundaries, lane lines, and pedestrian crossing lines; the dataset segments are tens of seconds long and contain a variety of road scenes and conditions.
[0076] In some preferred embodiments, the training process also includes the following operations:
[0077] Set the batch size, total training period, learning rate, number of instance-level queries, number of point-level queries, and size of each BEV grid, and set the number of stacked transformer decoder layers.
[0078] In a specific application example, the backbone network uses ResNet50, the batch size is set to 16, the total training epochs are set to 40, and the learning rate is set to 6e-4. The number of instance-level queries and point-level queries are set to 100 and 20 respectively, the size of each BEV grid is 0.3 meters, and 6 transformer-decoder layers are stacked.
[0079] To achieve model ensemble, a model ensemble strategy is adopted to further improve the generalization ability and performance of the models. Based on this, in some preferred embodiments, model ensemble may further include the following operations:
[0080] Both static and dynamic model ensemble methods are employed, and non-maximum suppression (NMS) is used to remove duplicate elements to improve the model's generalization ability and performance; where:
[0081] In static integration, the threshold is set to a fixed value, which is simple and easy to implement; in dynamic methods, the threshold is adjusted according to the element length to better describe similarity.
[0082] In dynamic integration, a dynamic threshold is adjusted based on the element length. The dynamic threshold δ is defined as: δ = max(δmax, min(δmin, (l-lmin)*(δmax-δmin) / (lmax-lmin)+δmin)); where lmin represents the minimum length of the data, and lmax represents the maximum length. δmax and δmin represent the set maximum and minimum thresholds, respectively. Generally, δmax is between [2, 3], and δmin is between [0.5, 1].
[0083] In this preferred embodiment, a static ensemble method is further used to integrate two well-trained map building models. Non-maximum suppression (NMS) is used to remove duplicate elements, significantly improving model performance. During the ensemble process, Chamfer distance is used as a similarity metric; the element with the highest confidence is selected, and similar elements are removed. If the Chamfer distance is less than a set threshold, the two tuples are considered similar.
[0084] The online high-precision map construction model obtained by the method provided in the above embodiments of the present invention has been verified by the End-to-End Autonomous Driving competition, and its performance is significantly better than the benchmark model HDMapNet.
[0085] One embodiment of the present invention provides an online high-precision map construction method.
[0086] Specifically, such as Figure 2 As shown, the online high-precision map construction method provided in this embodiment may include the following operations:
[0087] S1, acquire data from multiple cameras;
[0088] S2 provides an online high-precision map building model. It takes camera data as input to the online high-precision map building model and outputs a high-precision map to complete the online high-precision map building.
[0089] in:
[0090] The online high-precision map construction model is constructed using any of the model construction methods described in the above embodiments of the present invention.
[0091] One embodiment of the present invention provides an online high-precision map construction system.
[0092] Specifically, such as Figure 3 As shown, the online high-precision map building system provided in this embodiment may include the following modules:
[0093] Data acquisition module, which is used to acquire data from multiple cameras;
[0094] The map building module provides an online high-precision map building model. It takes camera data as input to the model and outputs a high-precision map, thus completing the online high-precision map building.
[0095] in:
[0096] The online high-precision map construction model is constructed using any of the model construction methods described in the above embodiments of the present invention.
[0097] An embodiment of the present invention also provides a computer terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it can be used to execute the model building method of any of the above embodiments of the present invention, or the online high-precision map building method.
[0098] Optionally, the memory is used to store programs; the memory may include volatile memory, such as random-access memory (RAM), such as static random-access memory (SRAM), double data rate synchronous dynamic random-access memory (DDR SDRAM), etc.; the memory may also include non-volatile memory, such as flash memory. The memory is used to store computer programs (such as application programs, functional modules, etc. that implement the above methods), computer instructions, etc., and the aforementioned computer programs, computer instructions, etc., can be partitioned and stored in one or more memories. Furthermore, the aforementioned computer programs, computer instructions, data, etc., can be accessed by the processor.
[0099] The aforementioned computer programs, computer instructions, etc., can be stored in partitions within one or more memory locations. Furthermore, the aforementioned computer programs, computer instructions, data, etc., can be accessed by a processor.
[0100] A processor is used to execute computer programs stored in memory to implement the various steps of the methods or various modules of the systems involved in the above embodiments. For details, please refer to the relevant descriptions in the preceding method and system embodiments.
[0101] The processor and memory can be separate structures or integrated structures. When the processor and memory are separate structures, they can be coupled together via a bus.
[0102] An embodiment of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can be used to perform any of the model building methods or online high-precision map building methods described in the above embodiments of the present invention.
[0103] The online high-precision map construction and model building method, system, terminal, and medium provided in the above embodiments of the present invention achieve efficient and accurate high-precision map construction through a stronger backbone network, effective view transformation and data augmentation techniques, and model integration strategies. The present invention exhibits excellent performance in multi-category semantic information extraction and is suitable for real-time map construction tasks in autonomous driving systems.
[0104] Any matters not covered in the above embodiments of the present invention are well-known in the art.
[0105] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention.
Claims
1. A model building method for online high-precision map construction, characterized in that, include: Based on the MapTr model architecture, a backbone network, a view transformation network, a hierarchical query embedding network, and a model optimization network are constructed sequentially to obtain the initial model; where: The backbone network is used to extract image features; The view transformation network is used to transform the image features from 2D image features into bird's-eye view features; The hierarchical query embedding network is used to explicitly encode and decode map elements of the bird's-eye view features through a hierarchical query embedding mechanism. The model optimization network is used to optimize the model parameters through loss function and data augmentation; The initial model is trained to obtain a map building model; By integrating multiple map building models, an online high-precision map building model is obtained.
2. The model building method for online high-precision map construction according to claim 1, characterized in that, The backbone network is constructed using any one of ResNet-50, Swin-tiny, Swin-base, and ConvNeXt networks, and is used to extract features from multi-view camera images; wherein: The multi-view camera image features are represented as: F = {F i } -1＜i＜Nview In the formula, F i Let N be the feature of the i-th viewpoint, and Nview be the total number of viewpoints.
3. The model building method for online high-precision map construction according to claim 1, characterized in that, The view transformation network is obtained by replacing the GKT module in the MapTR model with a spatial cross-attention layer proposed by the BEVFormer neural network model, while removing the temporal self-attention layer of the MapTR model. The encoder composed of multiple spatial cross-attention layers converts multi-scale 2D image features into BEV features to generate a unified bird's-eye view representation.
4. The model building method for online high-precision map construction according to claim 1, characterized in that, The hierarchical query embedding network explicitly encodes each map element using a hierarchical query embedding mechanism, defining a set of instance-level queries and a set of point-level queries. The instance-level queries focus on the center position of a map element, while the point-level queries focus on the offset range of the entire map element relative to its center position. The instance-level and point-level queries are combined to form a hierarchical query. Each map element corresponds to a set of hierarchical queries. The head of the hierarchical query embedding network contains multiple cascaded decoder layers, each consisting of a multi-head self-attention layer and a variable cross-attention layer. The hierarchical query sequence first uses the self-attention layer to focus on the features of each element in its own sequence and updates itself. Then, it uses the variable attention layer to perform variable cross-attention calculation with the BEV features and updates itself again. Finally, it uses a multilayer perceptron model to regress and predict the two-dimensional normalized BEV coordinates of the reference point.
5. The model building method for online high-precision map construction according to claim 1, characterized in that, The model optimization network includes: a loss function layer and a data augmentation layer; wherein: The loss function layer mainly consists of a comprehensive loss function composed of classification loss, point-to-point loss, edge direction loss, and point-to-line loss, which weigh different loss terms to optimize model performance. The comprehensive loss function L formed by classification loss Lcls, point-to-point loss Lp2p, edge direction loss Ldir, and point-to-line loss Lp2l is expressed as: L=λLcls+αLp2p+βLdir+γLp2l; The data augmentation layer employs geometric and photometric transformation methods to optimize the model's stability under different object orientations and visibility scenarios; it also uses a class-balanced grouping sampling method to optimize the data class balance performance.
6. The model building method for online high-precision map construction according to claim 1, characterized in that, The process of training the initial model to obtain the map construction model includes: Build the dataset; Using the AdamW optimizer and cosine annealing scheduling, the initial map building model is trained on the dataset to obtain the map building model. It also includes any one or more of the following: - The dataset uses the Argoverse2 dataset, each sample contains multiple images, and the map elements in each image include: boundaries, lane lines, and pedestrian crossing lines; the dataset segments are tens of seconds long and contain a variety of road scenes and conditions; -During the training process, the following settings are configured respectively: Batch size, total training duration, and learning rate; Number of instance-level queries and point-level queries; Each BEV mesh size and the number of stacked decoder layers.
7. The model building method for online high-precision map construction according to claim 1, characterized in that, The model integration includes: static model integration and dynamic model integration, which involves removing duplicate elements using a non-maximum suppression method and integrating multiple map construction models; wherein: In the static integration, the threshold is set to a fixed value; In the dynamic integration, the dynamic threshold is adjusted according to the element length, wherein the dynamic threshold δ is: δ=max(δmax,min(δmin,(l-lmin)*(δmax-δmin) / (lmax-lmin)+δmin)); where lmin represents the minimum length in the data, lmax represents the maximum length in the data; δmax and δmin represent the set maximum threshold and minimum threshold respectively, δmax is between [2,3], and δmin is between [0.5,1]. The Chamfer distance is used as a similarity measure. The element with the highest confidence is selected and similar elements are deleted, and duplicate elements are removed. If the Chamfer distance is less than the set threshold, the two tuples are considered similar.
8. A method for constructing high-precision maps online, characterized in that, include: Acquire data from multiple cameras; An online high-precision map building model is provided. The camera data is used as the input of the online high-precision map building model, and a high-precision map is output to complete the online high-precision map building. in: The online high-precision map construction model is constructed using the model construction method described in any one of claims 1-7.
9. A computer terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it can be used to perform the model building method of any one of claims 1-7, or to perform the online high-precision map building method of claim 8.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program can be used to perform the model building method of any one of claims 1-7, or to perform the online high-precision map building method of claim 8.