Autonomous driving map building model training method, autonomous driving map building method and electronic equipment

By fusing vector lane map contextual information with BEV visual features, this method solves the problem of strong dependence on high-cost sensors in traditional methods, realizes the construction of high-precision autonomous driving maps, reduces the dependence on high-precision sensors, and improves the accuracy of map construction and structured modeling capabilities.

CN120822569BActive Publication Date: 2026-11-10BEIHANG UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202511092207.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2025-08-05
Publication Date
2026-11-10
Estimated Expiration
2045-08-05

AI Technical Summary

Technical Problem

Existing deep learning-based autonomous driving map construction methods rely heavily on high-precision sensors, resulting in high hardware costs and limiting their widespread application in resource-constrained scenarios.

Method used

The system employs a vector lane map construction network, a contextual cue encoder, a BEV feature extraction network, and a query embedding generation network. By fusing contextual information from the vector lane map with BEV visual features, it achieves structured lane modeling and reduces reliance on high-precision sensors.

Benefits of technology

Build high-precision autonomous driving maps to reduce reliance on high-cost sensors and improve map building accuracy and structured modeling capabilities.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120822569B_ABST
    Figure CN120822569B_ABST
Patent Text Reader

Abstract

This invention provides a training method for an autonomous driving map construction model, an autonomous driving map construction method, and an electronic device, relating to the field of autonomous driving technology. The method involves acquiring a sample dataset, inputting vehicle trajectories and lane boundaries into a vector lane map construction network to obtain a vector lane map, inputting the vector lane map into a context cue encoder to obtain context cue embeddings, inputting multi-view images into a BEV feature extraction network to obtain BEV features, and inputting each BEV feature along with the context cue embeddings into a query embedding generation network to obtain corresponding lane group query embeddings, lane line query embeddings, and foreground segmentation query embeddings. These are then classified and regressed to obtain the classification results for lane line queries, the geometric coordinates of lane points, and the foreground segmentation map. Finally, the total loss information is calculated to optimize the parameters of the autonomous driving map construction model. This reduces the reliance on dense sensor data when constructing autonomous driving maps.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, and more specifically, to an autonomous driving map building model training method, an autonomous driving map building method, and an electronic device. Background Technology

[0002] Currently, autonomous driving perception technology is developing towards higher precision and real-time performance. Key technologies include deep learning-based visual perception, LiDAR, and data fusion and processing from radar and ultrasonic sensors. Deep neural networks have made significant progress in tasks such as object detection, tracking, and semantic segmentation.

[0003] In the construction of large-scale city-level autonomous driving maps, the technology has gradually shifted from the traditional method of relying on manual annotation to end-to-end automated solutions based on deep learning.

[0004] However, existing deep learning-based autonomous driving map construction methods are highly dependent on dense sensor data. Most existing methods rely on rich data collected by high-precision sensors (such as high-resolution LiDAR and high-definition cameras), resulting in high hardware costs and limiting their widespread application in resource-constrained scenarios. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide an autonomous driving map building model training method, an autonomous driving map building method, and an electronic device to at least partially improve the above-mentioned problems.

[0006] To achieve the above objectives, the technical solutions adopted in the embodiments of the present invention are as follows: In a first aspect, embodiments of the present invention provide a method for training an autonomous driving map construction model, wherein the autonomous driving map construction model includes a vector lane map construction network, a contextual cue encoder, a BEV feature extraction network, a query embedding generation network, and a classification, regression, and segmentation network, and the method includes: Obtain a sample dataset; the sample dataset includes vehicle-collected data and real map data of the area to be detected, the vehicle-collected data includes multiple multi-view images, multiple vehicle trajectories, and multiple lane boundaries collected by multiple vehicles; Multiple vehicle trajectories and lane boundaries collected from multiple vehicles are input into the vector lane map to construct a network, thereby obtaining the vector lane map of the area to be detected; The vector lane map is input into the context cue encoder to obtain the context cue embedding; Multiple multi-view images of each vehicle are input into the BEV feature extraction network to obtain multiple BEV features; wherein the number of BEV features is the same as the number of vehicles. Initialize lane group query, lane line query, and foreground segmentation query, and input them together with each BEV feature and the context hint embedding into the query embedding generation network to obtain the corresponding lane group query embedding, lane line query embedding, and foreground segmentation query embedding. The lane group query embedding, lane line query embedding, foreground segmentation query embedding, and BEV features are input into the classification and regression segmentation network to obtain the classification results of each lane line query, the geometric coordinates of the lane points in each lane line query, and the foreground segmentation map of the foreground segmentation query; wherein, the classification results and the geometric coordinates are used to construct an autonomous driving map of the area to be detected. Based on the total loss function, the total loss information is calculated according to the foreground segmentation map, the classification results, the geometric coordinates, and the real map data. Backpropagation is then performed based on the total loss information to optimize the parameters of the autonomous driving map construction model.

[0007] Optionally, the vector lane map construction network includes a polyline encoder, a center point encoder, and a lane decoder. The step of inputting multiple vehicle trajectories collected from multiple vehicles and multiple lane boundaries into the vector lane map construction network to obtain a vector lane map of the area to be detected includes: For any of the aforementioned vehicles, the vehicle is designated as the target vehicle, and the overlapping portions of the vehicle trajectories and lane boundaries of other vehicles with the target vehicle are designated as overlapping vehicle trajectories and overlapping lane boundaries. Clustering correction is performed on the vehicle trajectory and lane boundary of the target vehicle and the overlapping vehicle trajectory and lane boundary to obtain the corrected vehicle trajectory and lane boundary of the target vehicle. The corrected vehicle trajectory and the corrected lane boundary are input into the polyline encoder to obtain vehicle trajectory features and lane boundary features. Extract the geometric center point of the corrected vehicle trajectory and input the geometric center point into the center point encoder to obtain the center point feature; Traverse all the vehicles to obtain the vehicle trajectory features, lane boundary features, and center point features of each vehicle. All vehicle trajectory features, lane boundary features, and center point features are input into the lane decoder to generate a vector lane map of the area to be detected.

[0008] Optionally, inputting the corrected vehicle trajectory and the corrected lane boundary into the polyline encoder to obtain vehicle trajectory features and lane boundary features includes: The corrected vehicle trajectory and the corrected lane boundary are input into a linear layer to extract the polyline point features of each polyline point in the corrected vehicle trajectory and the corrected lane boundary. The polyline point features of the corrected vehicle trajectory and the corrected lane boundary are aggregated into a single 1D vector by a max pooling layer to obtain temporary vehicle trajectory features and temporary lane boundary features for each corrected vehicle trajectory and the corrected lane boundary. The temporary vehicle trajectory features and temporary lane boundary features are input into a multi-layer Transformer encoder, and the temporary vehicle trajectory features and temporary lane boundary features are updated through a self-attention mechanism to obtain the updated vehicle trajectory features and lane boundary features.

[0009] Optionally, the context cue encoder includes a shared MLP subnetwork, a shared location coding subnetwork, and an independent MLP. The step of inputting the vector lane map into the context cue encoder to obtain the context cue embedding includes: The vector lane map is input into the shared MLP subnetwork, and semantic encoding is performed on each element point in the vector lane map to obtain the semantic information of each element point; The vector lane map is input into the shared location coding subnetwork to obtain the geometric information of each element point; The semantic information and geometric information are concatenated and input into the independent MLP to obtain the contextual cue embedding of the vector lane map.

[0010] Optionally, the query embedding generation network includes a query initialization module and a multi-layer decoder; the initialization of lane group queries, lane line queries, and foreground segmentation queries, along with each of the BEV features and the contextual cue embeddings, is input into the query embedding generation network to obtain the corresponding lane group query embeddings, lane line query embeddings, and foreground segmentation query embeddings, including: The query initialization module initializes a first preset number of lane group queries, a second preset number of lane line queries, and a foreground segmentation query; wherein the lane group queries and the lane line queries are randomly generated vectors, and the foreground segmentation query is a fixed vector; The lane group query, the lane line query, and the foreground segmentation query are combined into a vector set; The vector set, along with each BEV feature and the contextual cue embedding, is input into the multi-layer decoder. Through layer-by-layer iteration, the final lane group query embedding, lane line query embedding, and foreground segmentation query embedding are obtained.

[0011] Optionally, each layer of the decoder includes a self-attention module, a cross-attention module, and an inner self-attention module; the step of inputting the vector set, each BEV feature, and the contextual cue embedding into the multi-layer decoder, and iterating layer by layer to obtain the final lane group query embedding, lane line query embedding, and foreground segmentation query embedding, includes: For the processing of each layer of the decoder, each lane group query, each lane line query, and the foreground segmentation query are input into the self-attention module to perform global spatial interaction, so as to enhance the global relationship between each lane group query, each lane line query, and the foreground segmentation query. The lane group query, lane line query, foreground segmentation query, BEV feature, and contextual hint are embedded into the cross attention module, and the embedded BEV feature and contextual hint are fused to enrich the geometric features of the lane group query, lane line query, and foreground segmentation query. Each lane line query is input into the inner self-attention module, and self-attention interaction is performed on the lane points inside each lane line query to enhance the geometric smoothness of the lane line query. The decoder is updated layer by layer according to the multi-layer structure, and finally the lane group query embedding, lane line query embedding and foreground segmentation query embedding are generated.

[0012] Optionally, the classification, regression, and segmentation network includes a classification branch, a regression branch, and a foreground segmentation branch; the step of inputting the lane group query embedding, the lane line query embedding, the foreground segmentation query embedding, and the BEV features into the classification, regression, and segmentation network to obtain the classification results of each lane line query, the geometric coordinates of the lane points in each lane line query, and the foreground segmentation map of the foreground segmentation query includes: The lane group query and the lane line query are embedded into the classification branch, and the classification result of each lane line query is predicted through the first fully connected layer. The lane group query and the lane line query are embedded into the regression branch, and the geometric coordinates of the lane points in the lane line query are predicted by the second fully connected layer. The foreground segmentation query embedding and each of the BEV features are input into the foreground segmentation branch, and the foreground segmentation map of the foreground segmentation query is generated through MLP and upsampling processing.

[0013] Optionally, the step of calculating the total loss information based on the total loss function, according to the foreground segmentation map, each of the classification results, each of the geometric coordinates, and the real map data, includes: Based on the classification results, the geometric coordinates, and the real map data, the classification loss, regression loss, and point-to-point loss are calculated. Based on the foreground segmentation map, the classification results, and the geometric coordinates, the first segmentation accuracy probability of each point in the foreground segmentation map is calculated. Based on the foreground segmentation map and the real map data, the second segmentation accuracy probability of each point in the foreground segmentation map is calculated; The segmentation guidance assistance loss is calculated based on the accuracy probabilities of each of the first and second segments. Based on the total loss function, the total loss information is calculated according to the classification loss, the regression loss, the point-to-point loss, and the segmentation guidance auxiliary loss; The total loss function is:

[0014]

[0015]

[0016]

[0017]

[0018]

[0019]

[0020] in, These are the weighting coefficients. It is the L1 distance between the predicted lane point and the true value in the actual map data. It is the predicted classification probability of the lane point. As a regulating factor, For cross-entropy, The average distance from each predicted lane point to the nearest ground truth point. The average distance from each real point to the nearest predicted lane point; Let be the probability of the first segmentation being accurate. Let be the probability of accurate segmentation in the second segmentation. This is the smoothing coefficient.

[0021] Secondly, embodiments of the present invention provide an autonomous driving map construction method, including: Acquire multiple multi-view images, multiple vehicle trajectories, and multiple lane boundaries collected by multiple vehicles in the area to be constructed; Multiple multi-view images, multiple vehicle trajectories, and multiple lane boundaries are input into the autonomous driving map construction model to obtain a vector lane map of the area to be constructed, the classification results of each lane line in the vector lane map, and the geometric coordinates of the lane points in each lane line; the autonomous driving map construction model is trained by the autonomous driving map construction model training method described in any one of the first aspects. Based on the classification results and geometric coordinates, an autonomous driving map of the area to be constructed is generated.

[0022] Thirdly, embodiments of the present invention provide an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the method described in any of the above-mentioned embodiments.

[0023] The present invention provides an autonomous driving map construction model training method, an autonomous driving map construction method, and an electronic device. It integrates vector lane map context information with BEV visual features and achieves structured lane modeling through a learnable query embedding mechanism. This solves the problems of high dependence on high-cost sensors, insufficient map construction accuracy, and weak structured modeling capabilities in traditional methods, thereby reducing the dependence on dense sensor data when constructing autonomous driving maps.

[0024] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0025] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 A schematic structural block diagram of an electronic device provided in an embodiment of the present invention; Figure 2 This is a schematic structural block diagram of an autonomous driving map construction model provided in an embodiment of the present invention; Figure 3 A flowchart illustrating an autonomous driving map construction model training method provided in an embodiment of the present invention; Figure 4 This is an overall flowchart of an autonomous driving map construction model training method provided in an embodiment of the present invention; Figure 5This is another schematic structural block diagram of an autonomous driving map construction model provided in an embodiment of the present invention; Figure 6 A flowchart illustrating step S220 provided in an embodiment of the present invention; Figure 7 This is another schematic structural block diagram of an autonomous driving map construction model provided in an embodiment of the present invention; Figure 8 A flowchart illustrating step S230 provided in an embodiment of the present invention; Figure 9 This is another schematic structural block diagram of an autonomous driving map construction model provided in an embodiment of the present invention; Figure 10 A flowchart illustrating step S250 provided in an embodiment of the present invention; Figure 11 A schematic structural block diagram of a multilayer decoder provided in an embodiment of the present invention; Figure 12 A flowchart illustrating step S2532 provided in an embodiment of the present invention; Figure 13 This is another schematic structural block diagram of an autonomous driving map construction model provided in an embodiment of the present invention; Figure 14 A flowchart illustrating step S260 provided in an embodiment of the present invention; Figure 15 A flowchart illustrating step S270 provided in an embodiment of the present invention; Figure 16 This is a flowchart illustrating an autonomous driving map construction method provided in an embodiment of the present invention.

[0027] Icons: 100 - Electronic device; 101 - Memory; 102 - Communication interface; 103 - Processor; 104 - Communication bus; 300 - Autonomous driving map building model; 310 - Vector lane map building network; 320 - Context cue encoder; 330 - BEV feature extraction network; 340 - Query embedding generation network; 350 - Classification, regression, and segmentation network; 311 - Piecewise encoder; 312 - Center point encoder; 313 - Lane decoder; 321 - Shared MLP subnetwork; 322 - Shared location encoding subnetwork; 323 - Independent MLP; 341 - Query initialization module; 342 - Multilayer decoder; 3421 - Self-attention module; 3422 - Cross-attention module; 3423 - Inner self-attention module; 351 - Classification branch; 352 - Regression branch; 353 - Foreground segmentation branch. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0029] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0030] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0031] It should be noted that, in this document, relational terms such as "first" and "second" are used only 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. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0032] As described in the background section, existing deep learning-based autonomous driving map construction methods are highly dependent on dense sensor data, with most existing methods relying on rich data collected by high-precision sensors (such as high-resolution LiDAR and high-definition cameras).

[0033] Based on the above, embodiments of the present invention provide an autonomous driving map construction model training method, an autonomous driving map construction method, and an electronic device. By fusing vector lane map context information with BEV visual features and using a learnable query embedding mechanism to achieve structured lane modeling, the reliance on high-precision sensors is reduced.

[0034] To implement the process steps and functions of the various examples of this invention, please refer to [link / reference]. Figure 1 , Figure 1 This is a schematic structural block diagram of an electronic device provided in an embodiment of the present invention. The electronic device 100 includes a memory 101 and a processor 103, which are electrically connected directly or indirectly to each other to achieve data transmission or interaction. For example, these components can be electrically connected to each other through one or more communication buses 104 or signal lines. The memory 101 can be used to store software programs and modules, and the processor 103 executes the software programs and modules stored in the memory 101, thereby performing various functional applications and data processing.

[0035] Electronic device 100 can be, but is not limited to, a personal computer (PC), a server, a distributed computer, etc. It is understood that electronic device 100 is not limited to a physical server, but can also be a virtual machine on a physical server, a virtual machine built on a cloud platform, or any other computer that can provide the same functionality as the server or virtual machine. The operating system of electronic device 100 can be, but is not limited to, Windows, Linux, etc.

[0036] The memory 101 may be, but is not limited to, random access memory (RAM), read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), etc.

[0037] The communication connection between the electronic device 100 and external devices is achieved through at least one communication interface 102 (which can be wired or wireless).

[0038] Processor 103 may be an integrated circuit chip with signal processing capabilities. In implementation, the steps of this embodiment can be completed by integrated logic circuits in the hardware of processor 103 or by instructions in software form. Processor 103 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0039] Understandable. Figure 1 The structure shown is for illustrative purposes only; the electronic device 100 may also include components that are more advanced than those shown. Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown. Figure 1 The components shown can be implemented using hardware, software, or a combination thereof.

[0040] The following is an exemplary description of the autonomous driving map construction model training method provided by this invention. See [link to documentation]. Figure 2 The autonomous driving map construction model 300 includes a vector lane map construction network 310, a contextual cue encoder 320, a BEV feature extraction network 330, a query embedding generation network 340, and a classification, regression, and segmentation network 350. See also... Figure 3 The subject executing this method can be one of the above. Figure 1 The electronic device 100 shown, the method includes as follows Figure 3 The following steps are described: S210: Obtain the sample dataset.

[0041] The sample dataset includes vehicle-collected data and real map data of the area to be detected. The vehicle-collected data includes multiple multi-view images, multiple vehicle trajectories, and multiple lane boundaries collected by multiple vehicles.

[0042] S220: Input multiple vehicle trajectories and lane boundaries collected from multiple vehicles into a vector lane map to construct a network and obtain a vector lane map of the area to be detected.

[0043] S230: Input the vector lane map into the context cue encoder to obtain the context cue embedding.

[0044] S240: Input multiple multi-view images of each vehicle into the BEV feature extraction network to obtain multiple BEV features.

[0045] The number of BEV features is the same as the number of vehicles.

[0046] S250: Initialize lane group query, lane line query, and foreground segmentation query, and input them together with each BEV feature and contextual cue embedding into the query embedding generation network to obtain the corresponding lane group query embedding, lane line query embedding, and foreground segmentation query embedding.

[0047] S260: Input the lane group query embedding, lane line query embedding, foreground segmentation query embedding, and BEV feature into the classification and regression segmentation network to obtain the classification results of each lane line query, the geometric coordinates of the lane points in each lane line query, and the foreground segmentation map of the foreground segmentation query.

[0048] The classification results and geometric coordinates are used to construct an autonomous driving map of the area to be detected.

[0049] S270: Based on the total loss function, the total loss information is calculated according to the foreground segmentation map, classification results, geometric coordinates and real map data. Backpropagation is then performed based on the total loss information to optimize the parameters of the autonomous driving map construction model.

[0050] Before training the autonomous driving map building model, sample data is collected. For example, in an area, multiple vehicles equipped with multi-view cameras and positioning devices drive. During the journey, each vehicle collects its own trajectory and lane markings on the road surface, while simultaneously capturing multi-view images. This data will be used to build the map. For comparison, real map data for the area will also be obtained. This real map data can be manually verified and labeled data.

[0051] First, multiple vehicle trajectories and lane boundaries collected from multiple vehicles are input into a vector lane map construction network. This network then fuses the data from multiple vehicles to generate a complete vector lane map of the area to be detected. This vector lane map is then input into a contextual cue encoder to obtain contextual cue embeddings, which are used to update subsequent queries.

[0052] For multi-view images, multiple multi-view images acquired for each vehicle are input into the BEV feature extraction network to obtain the BEV features corresponding to each vehicle. Optionally, the BEV feature extraction network may include a backbone neural network and a feature pyramid network (FPN).

[0053] After obtaining the contextual cue embedding and BEV features, they are input into the query embedding generation network along with the initialized lane group query, lane line query, and foreground segmentation query. Based on the mutual constraints between lane groups and lane lines according to the contextual cue embedding and BEV features, the lane group query, lane line query, and foreground segmentation query are optimized and iterated to obtain updated lane group query embedding, lane line query embedding, and foreground segmentation query embedding. These embeddings are then input into the final classification, regression, and segmentation network to obtain the classification results of each lane line query, the geometric coordinates of lane points in each lane line query, and the foreground segmentation map of the foreground segmentation query. These data are compared with the real map data to calculate the total loss information, and backpropagation is performed based on the total loss information to optimize the parameters of the autonomous driving map construction model.

[0054] The overall training process of this method can be found in [reference needed]. Figure 4 This method generates vector lane maps based on vehicle trajectories and lane boundaries, generates BEV features from multi-view images, integrates contextual hints from the vector lane maps with BEV features, and achieves structured lane modeling through a learnable query embedding mechanism to reduce reliance on high-precision sensors.

[0055] See Figure 5 The vector lane map construction network 310 includes a polyline encoder 311, a center point encoder 312, and a lane decoder 313. To fuse trajectories and lane boundaries acquired from multiple vehicles to generate a vector lane map, one possible implementation is described in [reference needed]. Figure 6 The above step S220 may include the following steps: S221: For any vehicle, take the vehicle as the target vehicle, and take the overlapping parts of the vehicle trajectories and lane boundaries of other vehicles and the target vehicle as the overlapping vehicle trajectories and overlapping lane boundaries.

[0056] S222: Cluster and correct the target vehicle's trajectory and lane boundaries with overlapping vehicle trajectories and lane boundaries to obtain the corrected target vehicle trajectory and lane boundaries.

[0057] S223: Input the corrected vehicle trajectory and corrected lane boundary into the polyline encoder to obtain vehicle trajectory features and lane boundary features.

[0058] S224: Extract the geometric center point of the corrected vehicle trajectory and input the geometric center point into the center point encoder to obtain the center point feature.

[0059] S225: Traverse all vehicles to obtain the vehicle trajectory features, lane boundary features, and center point features of each vehicle.

[0060] S226: Input all vehicle trajectory features, lane boundary features, and center point features into the lane decoder to generate a vector lane map of the area to be detected.

[0061] The same processing is applied to the data collected from each vehicle. The processing of data from a single vehicle is described below. This vehicle is designated as the target vehicle. The overlapping portions of the vehicle trajectories and lane boundaries of other vehicles with the target vehicle are defined as overlapping vehicle trajectories and overlapping lane boundaries. By processing the overlapping portions, shared path information can be extracted, providing a data foundation for subsequent clustering correction. The data from the target vehicle and the overlapping portions are then clustered to correct their trajectories and lane boundaries, thereby improving trajectory accuracy, reducing noise and offset, and forming a unified and smooth reference path.

[0062] After correcting the vehicle's trajectory and vehicle boundaries, the corrected trajectory and lane boundaries are input into a polyline encoder to extract their feature representations.

[0063] Simultaneously, the geometric center point of the corrected trajectory is extracted and encoded using a center point encoder. The center point of the trajectory can be selected from the centroid, midpoint, reference point, etc. The center point encoder can employ structures such as MLP, CNN, or Transformer.

[0064] Repeating the above process for all vehicles will yield all vehicle trajectory features, lane boundary features, and center point features. Inputting all these features into the lane decoder will enable global mapping of the detection area, generating a vector lane map.

[0065] The extraction of trajectory features and lane boundary features can be achieved using the following methods, and step S223 above may include the following steps: S2231: Input the corrected vehicle trajectory and the corrected lane boundary into the linear layer, and extract the polyline point features of each polyline point in the corrected vehicle trajectory and the corrected lane boundary.

[0066] S2232: The features of the corrected vehicle trajectory and the corrected lane boundary are aggregated into a single 1D vector by the max pooling layer, so as to obtain the temporary vehicle trajectory features of each corrected vehicle trajectory and the temporary lane boundary features of the corrected lane boundary.

[0067] S2233: Input the temporary vehicle trajectory features and temporary lane boundary features into the multi-layer Transformer encoder, and update the temporary vehicle trajectory features and temporary lane boundary features through the self-attention mechanism to obtain the updated vehicle trajectory features and lane boundary features.

[0068] The vehicle trajectory and lane boundaries are both polylines, each composed of a series of polyline points. First, the polylines are input into a linear layer, mapping each polyline point to a high-dimensional vector, resulting in polyline point features. The goal is to transform the original polyline points into feature vectors that can be processed by the neural network. Then, the polyline point features of each polyline are max-pooled and aggregated into a single 1D vector. Through max-pooling, the polyline point-level features can be compressed into overall polyline-level features, forming temporary polyline features, i.e., temporary trajectory features and temporary lane boundary features. These temporary polyline features will be updated subsequently. They can be input into a multi-layer Transformer encoder, updating the features through a self-attention mechanism. Processing by the multi-layer Transformer encoder models the spatial context between each trajectory and the lane boundary, enhancing feature representation. The output is the final vehicle trajectory features and lane boundary features, used for subsequent lane map generation.

[0069] In one possible implementation, see Figure 7 The context-sensitive encoder 320 includes a shared MLP subnetwork 321, a shared location coding subnetwork 322, and a standalone MLP 323. See also... Figure 8 The above step S230 may include the following steps: S231: Input the vector lane map into the shared MLP subnetwork, perform semantic encoding on each element point in the vector lane map, and obtain the semantic information of each element point.

[0070] S232: Input the vector lane map into the shared location coding subnetwork to obtain the geometric information of each element point.

[0071] S233: Concatenate the semantic information with the geometric information and input them into an independent MLP to obtain the contextual cue embedding of the vector lane map.

[0072] The vector lane map is input into a shared multilayer perceptron (MLP) subnetwork to semantically encode each element point, identifying the category and function of each lane element and obtaining its semantic feature vector. The vector lane map is then input into a shared location encoding subnetwork to extract the geometric information of each element point, including its coordinates, orientation, and other geometric attributes, resulting in its geometric feature vector. The semantic and geometric information of each element point are concatenated and then input into an independent MLP for nonlinear transformation to generate the final contextual cue embedding.

[0073] This method allows for the extraction of semantic and geometric information (such as coordinates and direction) of structured elements (such as lane lines, center lines, and connection relationships) from the generated vector lane map, and their integration into a unified contextual hint embedding.

[0074] There are multiple ways to implement query embedding generative networks. To generate richer query embeddings, one possible implementation is described in [link to relevant documentation]. Figure 9 The query embedding generative network 340 includes a query initialization module 341 and a multi-layer decoder 342. See also Figure 10 The above step S250 may include the following steps: S251: Initialize the first preset number of lane group queries, the second preset number of lane line queries, and a foreground segmentation query through the query initialization module.

[0075] Among them, lane group query and lane line query are randomly generated vectors, while foreground segmentation query is a fixed vector.

[0076] S252: Combine the queries for each lane group, each lane line, and the foreground segmentation query into a vector set.

[0077] S253: Input the vector set, each BEV feature, and contextual cue embedding into the multi-layer decoder. After layer-by-layer iteration, the final lane group query embedding, lane line query embedding, and foreground segmentation query embedding are obtained.

[0078] First, a first preset number of lane group queries, a second preset number of lane line queries, and a foreground segmentation query need to be initialized. The lane group and lane line queries are randomly initialized learnable vectors, while the foreground segmentation query is a fixed vector. The lane group and lane line queries are used to model structured lane elements, while the foreground segmentation query represents the key areas of interest within the overall lane region. For example, each lane line is considered a key area of ​​interest, while others are considered non-key areas of interest. This foreground segmentation query can assist the lane decoder in improving coordinate accuracy during model training.

[0079] These initialized queries, lane line queries, and foreground segmentation queries are concatenated or combined into a unified vector set, and then fed into a multi-layer decoder along with BEV features and contextual cue embeddings. The vectors are iteratively updated layer by layer to finally obtain lane group query embeddings, lane line query embeddings, and foreground segmentation query embeddings.

[0080] This implementation is based on the key modules of lane structured information decoding and context fusion in the Transformer architecture. By using an initialized learnable query vector, combined with BEV features and contextual cue embedding, it iterative optimization is performed in a multi-layer decoder to finally generate structured lane groups, lane lines and foreground segmentation results for the semantic representation of high-precision lane maps.

[0081] See Figure 11 Each multi-layer decoder 342 may include a self-attention module 3421, a cross-attention module 3422, and an inner self-attention module 3423. See also Figure 12 S253 includes: S2531: For the processing of each layer of the decoder, the queries of each lane group, each lane line, and the foreground segmentation query are input into the self-attention module to perform global spatial interaction, so as to enhance the global relationship of each lane group query, each lane line query, and the foreground segmentation query.

[0082] S2532: Embed the queries for each lane group, each lane line, the foreground segmentation query, each BEV feature, and contextual cues into the input cross-attention module, and fuse the embedded BEV features and contextual cues to enrich the geometric features of each lane group query, each lane line query, and the foreground segmentation query.

[0083] S2533: Input each lane line query into the internal self-attention module to perform self-attention interaction on the lane points within each lane line query, so as to enhance the geometric smoothness of the lane line query.

[0084] S2534: Update layer by layer according to the multi-layer decoder, and finally generate lane group query embedding, lane line query embedding and foreground segmentation query embedding.

[0085] In the self-attention module, all queries interact within the global space, modeling global dependencies between queries and enhancing structured semantic understanding.

[0086] The cross-attention module integrates the query set (lane group, lane line, foreground) with BEV features and contextual cues to fuse information, integrating visual features and contextual semantic information into the query representation, thereby enhancing its geometric and semantic expressive capabilities.

[0087] For each lane line query, an internal self-attention module is input separately to model the interaction relationship between its internal lane points, enhance the geometric consistency and smoothness of the internal points of the lane line, and improve the accuracy of lane line shape modeling.

[0088] In each decoder layer, a multi-level attention mechanism is used to progressively enhance the semantic representation, spatial modeling, and geometric smoothness of lane group queries, lane line queries, and foreground segmentation queries, ultimately outputting high-quality lane group query embeddings, lane line query embeddings, and foreground segmentation query embeddings.

[0089] To achieve more accurate predictions for each query embedding, multiple prediction heads can be constructed. Therefore, see [link to relevant documentation]. Figure 13 The classification, regression, and segmentation network 350 may include a classification branch 351, a regression branch 352, and a foreground segmentation branch 353. In one possible implementation, see [link to relevant documentation]. Figure 14 The above step S260 may include the following steps: S261: Embed the queries for each lane group and each lane line into the input classification branch, and predict the classification results of each lane line query through the first fully connected layer.

[0090] The classification branch can also use the Softmax classifier, which can identify the semantic category of each lane line (such as left boundary, right boundary, center line, stop line, dashed line, etc.) through classification.

[0091] S262: The query for each lane group and the query for each lane line are embedded into the input regression branch, and the geometric coordinates of the lane points in each lane line query are predicted through the second fully connected layer.

[0092] The second fully connected layer can be one or more fully connected layers. After regression through the fully connected layers, the output is the coordinates of each lane point in the lane line.

[0093] S263: The foreground segmentation query embedding and each BEV feature are input into the foreground segmentation branch, and the foreground segmentation map of the foreground segmentation query is generated through MLP and upsampling processing.

[0094] The foreground segmentation branch identifies pixel-level foreground masks belonging to lane regions in the BEV space, which are used for lane region modeling and background separation. Feature mapping is performed using MLP, and upsampling methods can include interpolation, transposed convolution, etc. The final result is a foreground segmentation map, which is a binary segmentation map. Elements with a value of 1 in the segmentation map can be considered as lane regions. This foreground segmentation map is used to calculate the segmentation guidance auxiliary loss in the subsequent process, which can help the lane decoder improve coordinate accuracy during model training.

[0095] For multi-task collaborative optimization, in one possible implementation, see [link to relevant documentation]. Figure 15 The above step S270 may include the following steps: S271: Calculate the classification loss, regression loss, and point-to-point loss based on the classification results, geometric coordinates, and real map data.

[0096] S272: Calculate the first segmentation accuracy probability of each point in the foreground segmentation map based on the foreground segmentation map, the classification results, and the geometric coordinates.

[0097] S273: Calculate the second segmentation accuracy probability of each point in the foreground segmentation map based on the foreground segmentation map and the real map data.

[0098] S274: Calculate the segmentation guidance auxiliary loss based on the accuracy probabilities of each first segmentation and each second segmentation.

[0099] S275: Based on the total loss function, the total loss information is calculated according to the classification loss, regression loss, point-to-point loss, and segmentation-guided auxiliary loss.

[0100] The total loss function is:

[0101]

[0102]

[0103]

[0104]

[0105]

[0106]

[0107] in, These are the weighting coefficients. It is the L1 distance between the predicted lane point and the ground truth value in the actual map data. It is the predicted classification probability of the lane point. As a regulating factor, For cross-entropy, The average distance from each predicted lane point to the nearest ground truth point. The average distance from each real point to the nearest predicted lane point; The probability of accurate segmentation is given. Let be the probability of accurate second segmentation. This is the smoothing coefficient.

[0108] This classification loss can solve the problem of inconsistency between classification and localization, this regression loss can optimize geometric accuracy, and this point-to-point loss can ensure the overall alignment of the point set through bidirectional constraints.

[0109] The segmentation-guided loss calculation does not directly apply to the foreground segmentation map. Instead, it first compares the foreground segmentation map with the predicted and ground truth data to determine which elements in the foreground segmentation map are correctly segmented. For example, if the foreground segmentation map has 100 points, including 60 ones and 40 zeros, comparing it with the predicted data reveals points in the foreground segmentation map that match the predicted data. For instance, if a point in the predicted data represents a lane or lane line, and the corresponding point in the foreground segmentation map is a 1, then the prediction for that point in the segmentation map is correct, and the first segmentation accuracy probability is calculated. Similarly, the second segmentation accuracy probability corresponding to the ground truth data is calculated. Finally, based on the first and second segmentation accuracy probabilities, the segmentation-guided loss is calculated using cross-entropy and Dice loss.

[0110] Furthermore, this embodiment of the invention also provides a method for constructing autonomous driving maps, see [link to relevant documentation]. Figure 16 The autonomous driving map construction method includes the following steps: S410: Acquire multiple multi-view images, multiple vehicle trajectories, and multiple lane boundaries of multiple vehicles in the area to be constructed.

[0111] S420: Input multiple multi-view images, multiple vehicle trajectories, and multiple lane boundaries into the autonomous driving map construction model to obtain a vector lane map of the area to be constructed, the classification results of each lane line in the vector lane map, and the geometric coordinates of the lane points in each lane line. The autonomous driving map construction model is trained using the above-described autonomous driving map construction model training method.

[0112] S430: Construct an autonomous driving map of the area to be constructed based on the classification results and geometric coordinates.

[0113] In summary, the autonomous driving map construction model training method, autonomous driving map construction method, and electronic device provided by this invention achieve structured lane modeling by fusing vector lane map contextual information with BEV visual features and introducing a learnable query embedding mechanism. This effectively reduces reliance on high-cost, high-precision sensors (such as LiDAR and high-definition cameras) during autonomous driving map construction, lowering hardware costs and improving the economy and applicability of map construction. Furthermore, this method significantly improves the accuracy and robustness of lane modeling by collaboratively optimizing model parameters through multi-task loss functions (including classification loss, regression loss, point-to-point loss, and segmentation guidance assistance loss), exhibiting stronger structured modeling capabilities and geometric consistency, especially in complex traffic scenarios.

[0114] In the embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative; for example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0115] In addition, the functional modules in the various embodiments of the present invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0116] If the functionality is implemented as a software module and sold or used as an independent product, it 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 computer-readable 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 of 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.

[0117] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0118] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for training an autonomous driving map construction model, characterized in that, The autonomous driving map construction model includes a vector lane map construction network, a contextual cue encoder, a BEV feature extraction network, a query embedding generation network, and a classification, regression, and segmentation network. The method includes: Obtain a sample dataset; the sample dataset includes vehicle-collected data and real map data of the area to be detected, the vehicle-collected data includes multiple multi-view images, multiple vehicle trajectories, and multiple lane boundaries collected by multiple vehicles; Multiple vehicle trajectories and lane boundaries collected from multiple vehicles are input into the vector lane map to construct a network, thereby obtaining the vector lane map of the area to be detected; The vector lane map is input into the context cue encoder to obtain the context cue embedding; Multiple multi-view images of each vehicle are input into the BEV feature extraction network to obtain multiple BEV features; wherein the number of BEV features is the same as the number of vehicles. Initialize lane group query, lane line query, and foreground segmentation query, and input them together with each BEV feature and the context hint embedding into the query embedding generation network to obtain the corresponding lane group query embedding, lane line query embedding, and foreground segmentation query embedding. The lane group query embedding, lane line query embedding, foreground segmentation query embedding, and BEV features are input into the classification regression segmentation network to obtain the classification results of each lane line query, the geometric coordinates of the lane points in each lane line query, and the foreground segmentation map of the foreground segmentation query; wherein, the classification results and the geometric coordinates are used to construct an autonomous driving map of the area to be detected. Based on the total loss function, the total loss information is calculated according to the foreground segmentation map, the classification results, the geometric coordinates, and the real map data. Backpropagation is then performed based on the total loss information to optimize the parameters of the autonomous driving map construction model.

2. The method according to claim 1, characterized in that, The vector lane map construction network includes a polyline encoder, a center point encoder, and a lane decoder. The process of inputting multiple vehicle trajectories and lane boundaries collected from multiple vehicles into the vector lane map construction network to obtain a vector lane map of the area to be detected includes: For any of the aforementioned vehicles, the vehicle is designated as the target vehicle, and the overlapping portions of the vehicle trajectories and lane boundaries of other vehicles with the target vehicle are designated as overlapping vehicle trajectories and overlapping lane boundaries. Clustering correction is performed on the vehicle trajectory and lane boundary of the target vehicle and the overlapping vehicle trajectory and lane boundary to obtain the corrected vehicle trajectory and lane boundary of the target vehicle. The corrected vehicle trajectory and the corrected lane boundary are input into the polyline encoder to obtain vehicle trajectory features and lane boundary features. Extract the geometric center point of the corrected vehicle trajectory and input the geometric center point into the center point encoder to obtain the center point feature; Traverse all the vehicles to obtain the vehicle trajectory features, lane boundary features, and center point features of each vehicle. All vehicle trajectory features, lane boundary features, and center point features are input into the lane decoder to generate a vector lane map of the area to be detected.

3. The method according to claim 2, characterized in that, The step of inputting the corrected vehicle trajectory and the corrected lane boundary into the polyline encoder to obtain vehicle trajectory features and lane boundary features includes: The corrected vehicle trajectory and the corrected lane boundary are input into a linear layer to extract the polyline point features of each polyline point in the corrected vehicle trajectory and the corrected lane boundary. The polyline point features of the corrected vehicle trajectory and the corrected lane boundary are aggregated into a single 1D vector by a max pooling layer to obtain temporary vehicle trajectory features and temporary lane boundary features for each corrected vehicle trajectory and the corrected lane boundary. The temporary vehicle trajectory features and temporary lane boundary features are input into a multi-layer Transformer encoder, and the temporary vehicle trajectory features and temporary lane boundary features are updated through a self-attention mechanism to obtain the updated vehicle trajectory features and lane boundary features.

4. The method according to claim 1, characterized in that, The context cue encoder includes a shared MLP subnetwork, a shared location coding subnetwork, and an independent MLP. The step of inputting the vector lane map into the context cue encoder to obtain the context cue embedding includes: The vector lane map is input into the shared MLP subnetwork, and semantic encoding is performed on each element point in the vector lane map to obtain the semantic information of each element point; The vector lane map is input into the shared location coding subnetwork to obtain the geometric information of each element point; The semantic information and geometric information are concatenated and input into the independent MLP to obtain the contextual cue embedding of the vector lane map.

5. The method according to claim 1, characterized in that, The query embedding generation network includes a query initialization module and a multi-layer decoder. The initialization of lane group queries, lane line queries, and foreground segmentation queries, along with the BEV features and the contextual cue embeddings, is input into the query embedding generation network to obtain the corresponding lane group query embeddings, lane line query embeddings, and foreground segmentation query embeddings, including: The query initialization module initializes a first preset number of lane group queries, a second preset number of lane line queries, and a foreground segmentation query; wherein the lane group queries and the lane line queries are randomly generated vectors, and the foreground segmentation query is a fixed vector; The lane group query, the lane line query, and the foreground segmentation query are combined into a vector set; The vector set, along with each BEV feature and the contextual cue embedding, is input into the multi-layer decoder. Through layer-by-layer iteration, the final lane group query embedding, lane line query embedding, and foreground segmentation query embedding are obtained.

6. The method according to claim 5, characterized in that, Each layer of the decoder includes a self-attention module, a cross-attention module, and an inner self-attention module; the vector set, along with each BEV feature and the contextual cue embedding, are input into the multi-layer decoder, and through layer-by-layer iteration, the final lane group query embedding, lane line query embedding, and foreground segmentation query embedding are obtained, including: For the processing of each layer of the decoder, each lane group query, each lane line query, and the foreground segmentation query are input into the self-attention module to perform global spatial interaction, so as to enhance the global relationship between each lane group query, each lane line query, and the foreground segmentation query. The lane group query, lane line query, foreground segmentation query, BEV feature, and contextual hint are embedded into the cross attention module, and the embedded BEV feature and contextual hint are fused to enrich the geometric features of the lane group query, lane line query, and foreground segmentation query. Each lane line query is input into the inner self-attention module, and self-attention interaction is performed on the lane points inside each lane line query to enhance the geometric smoothness of the lane line query. The decoder is updated layer by layer to finally generate lane group query embedding, lane line query embedding, and foreground segmentation query embedding.

7. The method according to claim 1, characterized in that, The classification, regression, and segmentation network includes a classification branch, a regression branch, and a foreground segmentation branch. The step of inputting the lane group query embedding, the lane line query embedding, the foreground segmentation query embedding, and the BEV features into the classification, regression, and segmentation network to obtain the classification results of each lane line query, the geometric coordinates of the lane points in each lane line query, and the foreground segmentation map of the foreground segmentation query includes: The lane group query and the lane line query are embedded into the classification branch, and the classification result of each lane line query is predicted through the first fully connected layer. The lane group query and the lane line query are embedded into the regression branch, and the geometric coordinates of the lane points in the lane line query are predicted by the second fully connected layer. The foreground segmentation query embedding and each of the BEV features are input into the foreground segmentation branch, and the foreground segmentation map of the foreground segmentation query is generated through MLP and upsampling processing.

8. The method according to claim 1, characterized in that, The total loss information, calculated based on the total loss function and according to the foreground segmentation map, the classification results, the geometric coordinates, and the real map data, includes: Based on the classification results, the geometric coordinates, and the real map data, the classification loss, regression loss, and point-to-point loss are calculated. Based on the foreground segmentation map, the classification results, and the geometric coordinates, the first segmentation accuracy probability of each point in the foreground segmentation map is calculated. Based on the foreground segmentation map and the real map data, the second segmentation accuracy probability of each point in the foreground segmentation map is calculated; The segmentation guidance assistance loss is calculated based on the accuracy probabilities of each of the first and second segments. Based on the total loss function, the total loss information is calculated according to the classification loss, the regression loss, the point-to-point loss, and the segmentation guidance auxiliary loss; The total loss function is: in, These are the weighting coefficients. It is the L1 distance between the predicted lane point and the ground truth value in the actual map data. It is the predicted classification probability of the lane point. As a regulating factor, For cross-entropy, The average distance from each predicted lane point to the nearest ground truth point. The average distance from each real point to the nearest predicted lane point; Let be the probability of the first segmentation being accurate. Let be the probability of accurate segmentation in the second segmentation. This is the smoothing coefficient.

9. A method for constructing autonomous driving maps, characterized in that, include: Acquire multiple multi-view images, multiple vehicle trajectories, and multiple lane boundaries collected by multiple vehicles in the area to be constructed; Multiple multi-view images, multiple vehicle trajectories, and multiple lane boundaries are input into the autonomous driving map construction model to obtain a vector lane map of the area to be constructed, the classification results of each lane line in the vector lane map, and the geometric coordinates of the lane points in each lane line; the autonomous driving map construction model is trained by the autonomous driving map construction model training method according to any one of claims 1 to 8. Based on the classification results and geometric coordinates, an autonomous driving map of the area to be constructed is generated.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the method according to any one of claims 1 to 9.

Citation Information

Patent Citations

  • Multi-modal end-to-end automatic driving method and system based on unified aerial view representation

    CN119049000A

  • High-precision positioning and supporting vehicle driving scene lane perception and topology understanding method and model

    CN120088421A