High-precision positioning support automobile driving scene lane perception and topology understanding method and model

Abstract

The application discloses a kind of based on high-precision positioning support's automobile driving scene lane perception and topological understanding method and model, comprising: propose online vectorization mapping model based on high-precision positioning, realize real-time lane perception and dynamic mapping;With the high-precision positioning information provided by low-orbit satellite as the prior of model, real-time provides vehicle absolute position, lane position, direction and speed information;Based on the encoder of Transformer, fusion BEV feature and road prior information, obtain accurate lane perception result;Design topological understanding module, construct vehicle surrounding lane network topological prediction model, according to the vehicle driving trajectory and position provided by background, real-time dynamically adjust lane topology.The background is based on a large number of vehicle driving data, reinforcement learning training is carried out, and optimization information is constantly updated to provide services for the current vehicle.The application effectively solves the problem that static high-precision map acquisition cost is high and cannot be updated in real time, and realizes high-order intelligent driving of light map.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, specifically to a method and model for lane perception and topology understanding in vehicle driving scenarios supported by high-precision positioning. Background Technology

[0002] With the rapid development of autonomous driving technology, accurate lane perception and online mapping have become key technologies for achieving safe and efficient vehicle navigation in complex traffic environments. Existing methods, when faced with complex road conditions and dynamic environmental changes, primarily rely on data from the vehicle's environmental perception devices such as LiDAR, cameras, or offline high-precision maps, neglecting the potential of high-precision positioning information as prior input. This leads to significant deficiencies in real-time performance and robustness. Particularly in densely populated urban areas with tall buildings, GNSS signals are susceptible to multipath effects, resulting in a significant decrease in positioning capability in occluded scenarios. Furthermore, topology understanding focuses on constructing the logical structure of the road network, clarifying the connections between lanes, intersection layouts, traffic flow, and other information. Accurate topology understanding enables vehicles to make reasonable plans in complex intersections, roundabouts, ramps, and other scenarios, effectively optimizing online mapping tasks.

[0003] The rapid development of low-Earth orbit (LEO) satellite technology has recently provided a revolutionary solution for high-precision positioning. Due to their low orbital altitude (500-2000 km), LEO satellites offer advantages such as low transmission latency, high signal strength, high coverage density, and strong multi-source redundancy, significantly improving positioning accuracy and robustness in complex scenarios. Furthermore, the seamless global coverage of LEO satellites is particularly suitable for remote, high-latitude, and complex environments where traditional GNSS systems struggle to provide effective service, offering fundamental support for the widespread application of autonomous driving technology. In recent years, with the accelerated deployment of global LEO satellite constellations and the integration of LEO satellite systems with traditional GNSS systems, centimeter-level and even millimeter-level high-precision positioning has become possible, opening up new technological pathways for the autonomous driving field.

[0004] Therefore, this invention proposes a high-precision positioning-supported method and model for lane perception and topology understanding in automotive driving scenarios. The method takes vehicle sensor data as input and uses high-precision positioning information provided by low-orbit satellites as prior knowledge, significantly improving the vehicle's environmental perception accuracy and navigation real-time performance, thus aiding decision-making in complex downstream environments. Furthermore, the high-precision positioning information not only provides accurate vehicle position and speed prior knowledge for online dynamic lane mapping but also helps the backend establish precise high-precision vehicle trajectory data and location information. Based on this information, the road topology can be dynamically adjusted in real time, solving the problems of high cost and inability to update traditional offline high-precision maps in real time, and enhancing the adaptability of autonomous driving systems in complex scenarios. Deeply integrating high-precision positioning information into lane perception and topology understanding tasks provides strong support for the real-time performance, robustness, and safety of lane perception and online mapping in autonomous driving, laying a technological foundation for achieving high-level autonomous driving. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and model for lane perception and topology understanding in vehicle driving scenarios based on high-precision positioning. By leveraging high-precision positioning technology from low-orbit satellites and the Transformer model, it achieves more accurate lane perception and topology mapping, applicable to both conventional and complex autonomous driving scenarios, and particularly for accurate lane perception and local path topology prediction in scenarios such as complex urban intersections, construction sections, and obstructed areas.

[0006] To achieve the above technical objectives, the present invention adopts the following technical solution:

[0007] This invention designs a lane perception and topology understanding method for vehicle driving scenarios based on high-precision positioning (hereinafter referred to as the high-precision positioning mapping method), which includes the following steps:

[0008] Step 1: Acquire multi-source input data and initialize the high-precision positioning and mapping method;

[0009] Step 2: Preprocess the acquired multi-source data separately to ensure the consistency and validity of the input data;

[0010] Step 3: Construct an online mapping model based on high-precision positioning;

[0011] Step 4: Train the built model to obtain the predicted map;

[0012] Furthermore, the multi-source input data in step 1 includes: vehicle camera data, lidar data, and high-precision positioning data. Among them, the high-precision positioning information comes from low-orbit satellites and can provide centimeter-level vehicle positioning accuracy.

[0013] Furthermore, the data processing in step 2 includes processing the vehicle camera data, LiDAR data, and high-precision positioning data separately.

[0014] The specific steps are as follows:

[0015] For vehicle-mounted camera data, images are first acquired using the vehicle's surround-view camera. Then, image denoising and enhancement operations are performed to improve image quality. Finally, ResNet-50 and FPN are used to extract image features.

[0016] For LiDAR data, the point cloud data is filtered and downsampled, and ResNet-50 and FPN are used to extract key geometric features.

[0017] For high-precision positioning data, it is transformed with the vehicle's local coordinate system to ensure spatiotemporal alignment of the data;

[0018] Furthermore, the online mapping model based on high-precision positioning in step 3 includes an information fusion module, a BEV construction module, and a topology understanding module.

[0019] The information fusion module uses the Kalman filter algorithm to fuse vehicle sensor data and high-precision positioning information provided by low-orbit satellites.

[0020] The specific steps are as follows:

[0021] The Kalman filter algorithm constructs a dynamic system model, incorporating state variables such as the vehicle's position, velocity, and acceleration, represented by the state vector x = [x, y, v]. x ,v y ,a x ,a y ] T This represents the planar coordinates, where (x, y) are the coordinates of the plane, and (v... x ,v y ) represents the velocity component, (a x ,a y ) represents the acceleration components, where x and y represent the lateral and longitudinal directions, respectively.

[0022] First, system state modeling is performed. The state transition equation, based on vehicle kinematics, can be expressed as:

[0023] x k =Fx k-1 +Bu k-1 +w k-1

[0024] Where, x kLet F represent the state vector at time k, and let F be the state transition matrix, describing the evolution of the state over time. For example, in a uniform motion model, F reflects the update of position based on velocity and the change of velocity based on acceleration. B is the control input matrix, which associates external inputs (such as control commands like accelerator pedal and steering wheel angle) with state changes. k-1 The control input vector represents acceleration information. k-1 The process noise characterizes the uncertainty of the model and the unmodeled dynamic factors. It is set as Gaussian white noise, and its covariance matrix Q reflects the noise intensity. Its magnitude is pre-calibrated according to the characteristics of different scenarios.

[0025] Then, observation modeling is performed. The observation equation connects the onboard sensor data with the system state. The observation equation can be expressed as:

[0026] z k =Hx k +v k

[0027] Among them, z k H represents the observation data at time k, where H is the observation matrix, derived from the state vector x. k Extract the observable portion. k To observe noise, we need to cover measurement errors and multipath effects.

[0028] Finally, Kalman filtering is performed alternately through two key steps: the prediction phase and the update phase.

[0029] During the prediction phase, the state estimate from the previous time step is used. Covariance estimation P k-1 The predicted state at the current time is calculated using the state transition equation. And the predicted covariance P k|k-1 :

[0030]

[0031] P k|k-1 =FP k-1 +Q

[0032] Q reflects the noise intensity.

[0033] During the update phase, new sensor observation data z is received. k Calculate the Kalman gain K k And correct the state estimate and covariance P k :

[0034] K k =P k|k-1 H T HP k|k-1 H T +R)-1

[0035]

[0036] P k =(IK k H)P k|k-1

[0037] The covariance matrix R is calibrated based on the sensor characteristics.

[0038] Through iterative processes, the Kalman filter continuously integrates satellite positioning and onboard sensor information to correct vehicle state estimates in real time.

[0039] Finally, the fused information is obtained.

[0040] The BEV building block uses a Transformer-based encoder structure to transform the fused information into features from the BEV perspective. Specifically:

[0041] First, the obtained fused information is mapped to a unified bird's-eye view coordinate system through a view projection function.

[0042] Then, a multi-layer stacked Transformer encoder structure is used to model the projected fused features. The Transformer encoder structure includes a multi-head self-attention module and a feedforward neural network.

[0043] The final BEV feature is F b ∈R C×H×W Where C is the channel dimension, and H and W represent the height and width, respectively.

[0044] The topology understanding module uses a Transformer-based decoder structure to further process BEV features and, combined with information provided by the backend, achieves dynamic topology modeling and lane map prediction. The backend (referring to cloud-based devices) uses a large amount of vehicle driving data, including high-precision trajectory data, for reinforcement learning training, continuously updating and optimizing information to dynamically update the local network topology map and provide services to the current vehicle.

[0045] The specific steps are as follows:

[0046] First, a set of learnable topology query vectors is initialized. These vectors have clear semantic meanings and represent key road elements such as lane start points, topology nodes, or control points. Each query vector has dynamic association capabilities and can accurately point to and associate the location of a specific lane segment or topology point in the complex BEV feature space, thereby providing an accurate index for subsequent information search and feature extraction.

[0047] Then, the multi-head attention mechanism in the Transformer decoder is used to search for information related to topology queries in BEV features, thereby realizing the modeling of global lane semantics and geometric features.

[0048] In this process, the decoder starts from the beginning of the lane and predicts the location of the next control point in turn based on the learned road patterns and feature relationships. This prediction process continues until the lane end is accurately located or the preset stopping conditions are met.

[0049] The control points essentially represent key locations in the lane topology, including various types such as lane intersections, curve segments, and straight segments, which together outline the precise shape and connection relationships of the lanes.

[0050] The lane topology is modeled as a series of polylines, with each lane represented by a set of ordered control points. Each control point contains two-dimensional location coordinates (x, y) and detailed topological semantic labels (such as straight sections, curves, and intersections).

[0051] In addition, the topology understanding module adjusts the attention distribution mechanism in real time and dynamically based on the accurate vehicle driving trajectory and location information provided by the background, prioritizing the focus on high-priority lanes or topology node areas near the positioning point. This allows the system to concentrate limited computing resources on the most critical and decision-making road information when dealing with complex road conditions.

[0052] If the predicted control point deviates from the actual road or lane, the topology understanding module performs timely correction operations, using the position deviation information in the positioning data and road geometric constraints to fine-tune the position of the predicted control point, ensuring the consistency and accuracy of the entire topology structure.

[0053] Step 4 involves training the built model on an NVIDIA 4090 GPU.

[0054] During training, the model's loss function consists of three parts: classification loss Γ. class Point-to-point loss piont and edge direction loss Γ dense :

[0055] Γ=β1Γ class +β2Γ piont +β3Γ dense

[0056] β1, β2, and β3 represent different weights to balance the loss term.

[0057] Based on the above method, the present invention also proposes a model for lane perception and topology understanding in a vehicle driving scenario with high-precision positioning support, which is obtained from the content of step 3 above.

[0058] The beneficial effects of this invention are:

[0059] 1) This invention proposes a method and model for lane perception and topology understanding in vehicle driving scenarios based on high-precision positioning. By introducing high-precision positioning information, the vehicle can achieve centimeter-level accuracy in perceiving its own position, lane boundaries and surrounding environmental features under various complex road conditions.

[0060] 2) Based on the introduced high-precision positioning information, this invention constructs an online vectorized mapping model, which effectively solves the problems of high cost and inability to update static high-precision maps in real time, and realizes high-level intelligent driving with lightweight maps.

[0061] 3) The online mapping model of this invention incorporates an information fusion module, which cleverly integrates low-orbit satellite positioning information and vehicle sensor data using the Kalman filter algorithm. This real-time correction of vehicle state estimation provides accurate prior information for subsequent perception and modeling, effectively compensating for the deficiencies in sensor data, correcting accumulated errors, and improving the reliability and stability of perception.

[0062] 4) The online mapping model of this invention includes a BEV building module, which is based on the Transformer encoder structure. It efficiently converts the fused multi-source data into a BEV perspective, giving full play to the advantages of Transformer in feature extraction and global context modeling, and providing an intuitive and comprehensive scene representation for topology understanding.

[0063] 5) The online mapping model of this invention includes a topology understanding module, which uses a polyline sequence to simplify the complexity of topology description while ensuring the stability and reliability of topology prediction. The high-precision positioning information provided by this module can assist sensors in quickly locking the target lane and optimize the accuracy of topology understanding. Attached Figure Description

[0064] Figure 1 A flowchart of a high-precision positioning-supported lane perception and topology understanding method for vehicle driving scenarios provided by the present invention;

[0065] Figure 2 The structural diagram of the online mapping model based on high-precision positioning provided by this invention;

[0066] Figure 3 This is a schematic diagram of the vectorized map provided by the present invention; Detailed Implementation

[0067] The present invention will be further described below with reference to the accompanying drawings. A more detailed description of the embodiments of the present disclosure will be provided in conjunction with the accompanying drawings. The drawings are used to provide a further understanding of the embodiments of the present disclosure and form part of the specification. They are used together with the embodiments of the present disclosure to explain the present disclosure and do not constitute a limitation thereof.

[0068] This invention proposes a high-precision positioning-supported method for lane perception and topology understanding in vehicle driving scenarios, such as... Figure 1 The diagram shown illustrates a method flowchart for a high-precision positioning-supported lane perception and topology understanding method for vehicle driving scenarios provided in this application. The method mainly includes the following steps:

[0069] Step 1: Acquire multi-source input data and initialize the high-precision positioning and mapping method;

[0070] Furthermore, the multi-source input data includes: vehicle camera data, LiDAR data, and high-precision positioning data. Among them, vehicle camera data can capture rich environmental information, such as road signs, lane lines, and dynamic obstacles; LiDAR data provides high-precision 3D point cloud information for building environmental maps and detecting surrounding objects; and high-precision positioning data comes from low-orbit satellites, which use their centimeter-level positioning accuracy to provide stable and reliable position references for the vehicle.

[0071] Step 2: Preprocess the acquired multi-source data separately to ensure the consistency and validity of the input data;

[0072] Furthermore, data preprocessing includes processing vehicle camera data, LiDAR data, and high-precision positioning data separately.

[0073] The specific steps are as follows:

[0074] For vehicle camera data, the system first uses a surround-view camera to acquire images of the vehicle's surrounding environment from all angles. Then, image denoising and enhancement are performed to optimize image quality. Finally, a ResNet-50 network and a feature pyramid network architecture are used to extract deep features from the processed images, accurately capturing key features such as the shape, texture, and location of various objects in the images. This provides strong support for target recognition and scene understanding in autonomous driving systems.

[0075] For LiDAR data, the collected data is point cloud data, which is a large amount of data and contains some redundant information. It is necessary to filter and downsample the data first. Then, the same ResNet-50 and feature pyramid network architecture are used to extract key geometric features from the preprocessed point cloud data. These geometric features enable the autonomous driving system to accurately perceive the shape, size and position of surrounding obstacles.

[0076] For high-precision positioning data, the positioning information in the global coordinate system is converted into coordinate values ​​in the vehicle's local coordinate system, thereby ensuring that the positioning data is accurately aligned with the vehicle's own position and attitude information in time and space.

[0077] Step 3: Construct an online mapping model based on high-precision positioning;

[0078] Furthermore, such as Figure 2 The structure diagram of the online mapping model based on high-precision positioning is shown in the figure. The model mainly consists of three parts: information fusion module, BEV construction module, and topology understanding module.

[0079] The information fusion module uses the Kalman filter algorithm to fuse vehicle sensor data and high-precision positioning information provided by low-orbit satellites.

[0080] Taking the urban expressway scenario as an example, the specific operation is as follows:

[0081] The Kalman filter algorithm constructs a dynamic system model, incorporating state variables such as the vehicle's position, velocity, and acceleration, represented by the state vector x = [x, y, v]. x ,v y ,a x ,a y ] T This represents the planar coordinates, where (x, y) are the coordinates of the plane, and (v... x ,v y ) represents the velocity component, (a x ,a y () represents the acceleration component.

[0082] First, system state modeling is performed. The state transition equation, based on vehicle kinematics, can be expressed as:

[0083] x k =Fx k-1 +Bu k-1 +w k-1

[0084] Where F is the state transition matrix, describing the evolution of the state over time. For example, in a uniform motion model, F reflects the update of position based on velocity and the change of velocity based on acceleration. B is the control input matrix, which associates external inputs (such as control commands like the accelerator pedal and steering wheel angle) with state changes. k-1 The control input vector represents acceleration information. k-1 The process noise characterizes the uncertainty of the model and the unmodeled dynamic factors. It is set as Gaussian white noise, and its covariance matrix Q reflects the noise intensity. Its magnitude is pre-calibrated according to the characteristics of different scenarios.

[0085] Then, observation modeling is performed. The observation equation connects the onboard sensors (onboard camera data and LiDAR data) with the system state. The observation equation can be expressed as:

[0086] z k =Hx k +v k

[0087] Where H is the observation matrix, from the state vector x k Extract the observable portion. k To observe noise, we need to cover measurement errors and multipath effects.

[0088] Finally, Kalman filtering performs two key steps alternately: prediction and update.

[0089] During the prediction phase, the state estimate from the previous time step is used. Covariance estimation P k-1 The predicted state at the current time is calculated using the state transition equation. And the predicted covariance P k|k-1 :

[0090]

[0091] P k|k-1 =FP k-1 +Q

[0092] Q reflects the noise intensity.

[0093] During the update phase, new sensor observation data z is received. k Calculate the Kalman gain K k And correct the state estimate and covariance:

[0094] K k =P k|k-1 H T HP k|k-1 H T +R) -1

[0095]

[0096] P k =(IK k )P k|k-1

[0097] The covariance matrix R is calibrated based on the sensor characteristics.

[0098] Through iterative processes, the Kalman filter continuously integrates satellite positioning and onboard sensor information to correct vehicle state estimates in real time.

[0099] Finally, the fused information is obtained.

[0100] The BEV building module uses a Transformer-based encoder structure to transform fused information into a BEV perspective, providing the vehicle with a top-down, comprehensive, and intuitive environmental view. Specifically:

[0101] First, a view projection function is designed to map the obtained fused information to a unified top-down coordinate system, so that information from different sources can be integrated and analyzed within the same visual framework, laying the foundation for subsequent feature extraction and modeling.

[0102] Then, a multi-layer stacked Transformer encoder structure is used to model the projected fused features. The Transformer encoder structure includes a multi-head self-attention module and a feedforward neural network.

[0103] The final BEV feature is F b ∈R C×H×W Where C is the channel dimension, and H and W represent the height and width, respectively.

[0104] The topology understanding module uses a Transformer-based decoder structure to further process BEV features and, combined with information provided by the backend, achieves dynamic topology modeling and lane map prediction. The backend uses a large amount of vehicle driving data, including high-precision trajectory data, for reinforcement learning training, continuously updating and optimizing information to dynamically update the local network topology map and provide services to the current vehicle.

[0105] The specific steps are as follows:

[0106] First, a set of learnable topology query vectors is initialized. These vectors have clear semantic meanings and represent key road elements such as lane start points, topology nodes, or control points. Each query vector has dynamic association capabilities and can accurately point to and associate the location of a specific lane segment or topology point in the complex BEV feature space, thereby providing an accurate index for subsequent information search and feature extraction.

[0107] Then, the multi-head attention mechanism in the Transformer decoder is used to search for information related to topology queries in BEV features, thereby realizing the modeling of global lane semantics and geometric features.

[0108] In this process, the decoder starts from the beginning of the lane and predicts the location of the next control point in turn based on the learned road patterns and feature relationships. This prediction process continues until the lane end is accurately located or the preset stopping conditions are met.

[0109] The control points essentially represent key locations in the lane topology, including various types such as lane intersections, curve segments, and straight segments, which together outline the precise shape and connection relationships of the lanes.

[0110] The lane topology is modeled as a series of polylines, with each lane represented by a set of ordered control points. Each control point contains two-dimensional location coordinates (x, y) and detailed topological semantic labels (such as straight sections, curves, and intersections).

[0111] In addition, the topology understanding module adjusts the attention distribution mechanism in real time and dynamically based on the accurate vehicle driving trajectory and location information provided by the background, prioritizing the focus on high-priority lanes or topology node areas near the positioning point. This allows the system to concentrate limited computing resources on the most critical and decision-making road information when dealing with complex road conditions.

[0112] If the predicted control point deviates from the actual road or lane, the topology understanding module performs timely correction operations, using the position deviation information in the positioning data and road geometric constraints to fine-tune the position of the predicted control point, ensuring the consistency and accuracy of the entire topology structure.

[0113] Step 4: Train the built online mapping model to obtain the predicted map;

[0114] The built model was trained on an NVIDIA 4090 GPU.

[0115] During training, the model's loss function consists of three parts: classification loss Γ. class Point-to-point loss piont and edge direction loss Γ dense :

[0116] Γ=β1Γ class +β2Γ piont +β3Γ dense

[0117] β1, β2, and β3 represent different weights to balance the loss term.

[0118] During testing, three static map elements—lane dividers, pedestrian crossings, and road boundaries—were used as evaluation objects. An average accuracy metric based on chamfer distance was used for evaluation; a prediction was considered correct when the distance was less than a specified threshold. An example prediction is shown below. Figure 3 The vectorized map shown is a schematic diagram.

[0119] Based on the above method, embodiments of the present invention also include a model that can be deployed in an autonomous vehicle. The model is obtained in step 3. During deployment, the code of the model is stored in the vehicle's storage device, and the vehicle's processor can execute the function of the model when executing the code in the storage device.

[0120] The detailed descriptions listed above are merely specific descriptions of feasible embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. All equivalent methods or modifications that do not depart from the technology of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for lane perception and topology understanding in vehicle driving scenarios supported by high-precision positioning, characterized in that, include: Step 1: Obtain multi-source input data; The multi-source input data includes: vehicle camera data, lidar data, and high-precision positioning data, wherein the high-precision positioning data is sourced from low-orbit satellites; Step 2: Preprocess the acquired multi-source data to ensure the consistency and validity of the input data; Step 3: Construct an online mapping model based on high-precision positioning; The online mapping model based on high-precision positioning includes: an information fusion module, a BEV construction module, and a topology understanding module; The information fusion module obtains fused information of satellite positioning and vehicle sensor information based on vehicle status information; The BEV construction module converts the information fused by Jiang into BEV perspective features. The topology understanding module uses a Transformer-based decoder structure to further process the BEV perspective features and combines them with information provided by the background to achieve dynamic lane topology modeling and lane map prediction. The information fusion module uses a Kalman filter algorithm to fuse vehicle sensor data and high-precision positioning information provided by low-orbit satellites; specifically: A dynamic system model is constructed based on the Kalman filter algorithm, comprehensively considering the vehicle's position, velocity, and acceleration state variables, and using a state vector. It means that, among them For planar coordinates, For velocity components, For acceleration components; First, system state modeling is performed; the state transition equation, based on vehicle kinematics, can be expressed as: in, Let be the state transition matrix, describing the evolution of the state over time. In a uniform motion model, This demonstrates that position is updated based on velocity, and velocity is updated based on changes in acceleration; To control the input matrix, external inputs are correlated with state changes; To control the input vector, which represents acceleration information, The process noise, representing the uncertainty of the model and unmodeled dynamic factors, is defined as Gaussian white noise, and its covariance matrix is... It reflects the noise intensity, and its magnitude is pre-calibrated according to the characteristics of different scenarios; Then, observation modeling is performed; the observation equation connects the onboard sensor data with the system state, and the observation data... Represented as: in, The observation matrix is ​​derived from the state vector. Extract the observable portion; To observe noise, including measurement errors and multipath effects; Finally, Kalman filtering performs two key steps alternately: prediction and update. During the prediction phase, the state estimate from the previous time step is used. Covariance Estimation The predicted state at the current time is calculated using the state transition equation. and predicted covariance : in, Reflects noise intensity; During the update phase, new sensor observation data is received. Calculate Kalman gain And correct the state estimate and covariance: Wherein, the covariance matrix Calibrate according to sensor characteristics; Through iterative processes, the Kalman filter continuously fuses satellite positioning and vehicle sensor information, corrects vehicle state estimation in real time, and finally obtains fused information. Step 4: Train the built model to obtain the predicted topology map, and realize lane perception and topology understanding.

2. The high-precision positioning-supported lane perception and topology understanding method for vehicle driving scenarios according to claim 1, characterized in that, The data processing in step 2 includes processing the vehicle-mounted camera data, LiDAR data, and high-precision positioning data separately; the specific operations are as follows: For vehicle camera data, images are first acquired using the vehicle's surround-view camera; then image denoising and enhancement operations are performed to improve image quality; finally, ResNet-50 and FPN are used to extract image features. For LiDAR data, point cloud data is filtered and downsampled, and key geometric features are extracted using ResNet-50 and FPN. For high-precision positioning data, it is converted to the vehicle's local coordinate system to ensure spatiotemporal alignment of the data.

3. The high-precision positioning-supported lane perception and topology understanding method for vehicle driving scenarios according to claim 1, characterized in that, The BEV building module uses a Transformer-based encoder structure to convert fused information into feature data from the BEV perspective; specifically as follows: First, the obtained fused information is mapped to a unified top-view coordinate system through a view projection function; Then, the projected fused features are modeled using a multi-layer stacked Transformer encoder structure, which includes a multi-head self-attention module and a feedforward neural network. The final features obtained from the BEV perspective are Where C is the channel dimension, and H and W represent the height and width, respectively.

4. The high-precision positioning-supported lane perception and topology understanding method for vehicle driving scenarios according to claim 1, characterized in that, The specific process by which the topology understanding module implements dynamic topology relationship modeling and lane map prediction is as follows: First, initialize a set of learnable topology query vectors, which have clear semantic meanings and represent lane start points, topology nodes, or control points, respectively. Then, the multi-head attention mechanism in the Transformer decoder is used to search for information related to the topology query vector in the BEV features to achieve global lane semantic and geometric feature modeling; specifically: the decoder starts from the lane start point and predicts the position of the next control point in turn based on the learned road patterns and feature associations, and continues this prediction process until the lane end point is accurately located or the preset stopping conditions are met. Among them, control points represent key locations in the lane topology, including various types such as lane intersections, curve segments, and straight segments, which together outline the precise shape and connection relationships of the lanes; The lane topology is modeled as a polyline sequence, with each lane represented by a set of ordered control points, each containing two-dimensional position coordinates. And detailed topological semantic tags, including straight road segments, curves, and intersections.

5. The high-precision positioning-supported lane perception and topology understanding method for vehicle driving scenarios according to claim 1 or 4, characterized in that, The topology understanding module adjusts the attention distribution mechanism in real time and dynamically based on the accurate vehicle driving trajectory and location information provided by the background, prioritizing the focus on high-priority lanes or topology node areas near the positioning point. When dealing with complex road conditions, it can concentrate limited computing resources on the most critical and decision-making road information. If the predicted control points deviate from the actual road or lane, the model will perform timely correction operations, using the position deviation information in the positioning data and road geometric constraints to fine-tune the predicted control point positions, ensuring the consistency and accuracy of the entire topology.

6. The high-precision positioning-supported lane perception and topology understanding method for vehicle driving scenarios according to claim 1, characterized in that, The model training in step 4 is performed on an NVIDIA 4090 GPU. The model's loss function consists of three parts: classification loss and classification loss. Point-to-point loss and edge direction loss : in , and Indicates the weight.

7. A model for lane perception and topology understanding in vehicle driving scenarios supported by high-precision positioning, characterized in that, The model is obtained according to step S3 of claim 1, and the code of the model is stored in the vehicle's storage device. When the vehicle's processor executes the code in the storage device, it can perform the function of the model.

Images