Intelligent network connection automobile positioning method and system based on dynamic adjustable visual label cooperation

Abstract

The invention provides an intelligent network connection automobile positioning method and system based on dynamic adjustable visual label cooperation. The system comprises a roadside subsystem and a vehicle-mounted subsystem. The roadside subsystem is used for deploying a dynamic adjustable visual label and a thundersight all-in-one machine; the dynamic adjustable visual labels are arranged in an array manner and comprise guardrails on the roadside and tunnel walls; the all-in-one machine comprises a millimeter-wave radar and a high-definition camera, and is equipped with an RSU communication unit. The vehicle-mounted subsystem is used for determining position information of a current vehicle and comprises a multispectral imaging module, a positioning fusion controller and a V2X communication module. The roadside subsystem further comprises an edge computing unit which is used for executing label state monitoring and cross-device target splicing. According to the multi-mode reasoning positioning engine, a UniVG-R1 reasoning chain is applied to vehicle positioning for the first time, the problem of complex scene understanding is solved, SegVG pixel-level segmentation is fused, and the pose calculation precision is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent networked vehicle positioning, in particular to a dynamic adjustable visual label cooperative positioning method and system for intelligent networked vehicles. In particular, it relates to a vehicle-road cooperative high-precision positioning system and method based on dynamic adjustable visual labels and multi-modal fusion algorithm. BACKGROUND

[0002] With the rapid development of intelligent networked vehicle technology, vehicle precise positioning has become a core prerequisite for realizing advanced automatic driving. Global Navigation Satellite System (GNSS) is easily affected by signal blockage in complex urban environments, resulting in positioning failure or decreased accuracy. Visual label positioning technology, as an emerging solution, assists vehicles in achieving relative positioning by deploying special visual markers in the road environment. However, the existing technology still has significant shortcomings.

[0003] On the one hand, traditional two-dimensional code labels (such as Apriltag) have a sharp decline in recognition rate under rain, fog, low light, or partial blockage conditions. Existing methods rely on the complete visibility of labels, and when they are dirty or unevenly illuminated, the corner detection algorithm is prone to failure. On the other hand, when vehicles are traveling at high speed (> 60km / h), traditional cameras have a decreased label recognition rate due to exposure blur. At the same time, existing technology lacks support for cross-camera label tracking, and when vehicles cross different radar-camera integrated machine monitoring areas, target stitching relies on a simple distance threshold method, with a high mismatch rate of up to 15%.

[0004] Patent document CN115235478A discloses an intelligent vehicle positioning method and system based on visual labels and laser SLAM, the method comprising: obtaining three-dimensional point cloud information of the road environment and extracting the vehicle's position information in the global grid map to obtain the vehicle's laser positioning information; obtaining image information of the visual label and extracting coordinate position data in the visual label to obtain the vehicle's visual label positioning information; based on the vehicle's laser positioning information and visual label positioning information, using an extended Kalman filter to fuse multi-source heterogeneous positioning information to obtain the fused vehicle position information.

[0005] However, patent document CN115235478A uses simple Kalman filtering and does not fully consider the time asynchrony and confidence difference of heterogeneous data. Although it proposes an extended Kalman filter fusion scheme, it does not solve the problem of spatio-temporal alignment of visual labels and point cloud data, resulting in large cumulative errors over long distances. SUMMARY

[0006] In view of the defects in the prior art, the purpose of the present application is to provide a dynamic adjustable visual label cooperative intelligent networked vehicle positioning method and system.

[0007] The application provides a dynamic adjustable visual label coordinated intelligent network connected automobile positioning system, which comprises a roadside subsystem and a vehicle-mounted subsystem.

[0008] The roadside subsystem is used for deploying a dynamic adjustable visual label and a radar and vision integrated machine; the dynamic adjustable visual label is arranged in an array mode at positions including a roadside guardrail and a tunnel wall; the radar and vision integrated machine comprises a millimeter wave radar and a high-definition camera and is provided with an RSU communication unit.

[0009] The vehicle-mounted subsystem is used for determining position information of a current vehicle and comprises a multispectral imaging module, a positioning fusion controller and a V2X communication module.

[0010] Preferably, the roadside subsystem further comprises an edge computing unit used for performing label state monitoring and cross-device target splicing.

[0011] Preferably, the multispectral imaging module comprises a visible light and near-infrared dual camera.

[0012] The positioning fusion controller is used for running a UniVG-R1 model and an adaptive Kalman filter.

[0013] The V2X communication module receives roadside label maps and auxiliary positioning information.

[0014] Preferably, the dynamic adjustable visual label has a hierarchical coding structure and comprises a base layer, a dynamic layer and an information layer.

[0015] The base layer adopts a fixed black-and-white checkerboard pattern.

[0016] The dynamic layer is an electroluminescent film.

[0017] The information layer is embedded with an NFC chip for storing label absolute coordinates and neighboring label topological relations.

[0018] The dynamic adjustable visual label ID is as follows:

[0019] Label ID = region code + type code + coordinate check code

[0020] The region code comprises 8 bits, the type code comprises 4 bits and the coordinate check code comprises 16 bits.

[0021] Preferably, the dynamic adjustable visual label is provided with an asymmetric L-shaped guide symbol at the upper right corner, is internally provided with a light sensor and a temperature sensor and regularly reports a working state through LoRa wireless communication.

[0022] The application further provides a dynamic adjustable visual label coordinated intelligent network connected automobile positioning method.

[0023] Step S1: obtaining original data;

[0024] Step S2: preprocessing the original data;

[0025] Step S3: Segment the preprocessed data using the SegVG pixel-level segmentation network to generate a pixel-level label mask;

[0026] Step S4: performing confidence evaluation on the pixel-level label mask, preprocessed data, and original data to obtain the label 6DoF pose, point cloud matching result, and corresponding confidence;

[0027] Step S5: inputting the label 6DoF pose and point cloud matching results into the spatiotemporal alignment module to perform spatiotemporal alignment of multi-source data;

[0028] Step S6: performing adaptive weighted extended Kalman filtering (EKF) on the corresponding confidence and time-space aligned data to obtain the final fusion positioning.

[0029] Preferably, the raw data includes images, point clouds and IMU;

[0030] The preprocessing includes illumination adaptive preprocessing, that is, applying a nonlinear mapping to the input image, as follows:

[0031] I_out=α·tanh(β·(I_in-γ))+δ

[0032] Where I_out and I_in represent the pixel grayscale value after preprocessing and the pixel grayscale value of the input image, respectively; α, β, γ, and δ represent the output gain coefficient, contrast adjustment factor, grayscale offset, and brightness compensation value, respectively;

[0033] The confidence evaluation includes UniVG-R1 inference, laser SLAM feature extraction and IMU pre-integration.

[0034] Preferably, the spatiotemporal alignment of multi-source data includes timestamp synchronization and unified coordinate conversion;

[0035] The timestamp synchronization uses the PTPv2 protocol to align the timestamps of the camera, IMU, and roadside data;

[0036] Coordinate unified transformation refers to defining the transformation relationship between the global coordinate system and the local coordinate system, as shown below:

[0037] [X_global]=[R]·[X_local]+[T]

[0038] Wherein, X_global represents a global coordinate system, X_local represents a local coordinate system, [R] represents a rotation matrix, and [T] represents a translation vector.

[0039] Preferably, the step S6 comprises:

[0040] Step S6.1: Construct a state vector as follows:

[0041] x = [position, velocity, attitude]T

[0042] Wherein, x represents a state quantity, position, velocity, and attitude represent position, velocity, and attitude respectively, and T represents transposition.

[0043] Step S6.2: Define observation sources, respectively, vision tag positioning results z_vision, laser SLAM point cloud matching z_lidar, and inertial navigation output z_imu.

[0044] Step S6.3: Generate a weight matrix based on the confidence evaluation network of UniVG-R1, and the formula is as follows:

[0045] W = f (ambient illumination, tag occlusion rate, vehicle speed)

[0046] Wherein, W and f represent a weight matrix and a confidence evaluation function respectively.

[0047] Step S6.4: Update the measurement equation:

[0048] z = H·x + v

[0049] v ~ N (0, W·R)

[0050] Wherein, z represents an observation, H represents a conversion matrix, v represents measurement noise, N represents a Gaussian noise distribution with mean 0 and covariance matrix WR, and R represents observation noise covariance.

[0051] Preferably, it further comprises a vehicle-road cooperative optimization step.

[0052] The vehicle-road cooperative optimization step comprises cross-device target splicing and global positioning correction.

[0053] The cross-device target splicing comprises the following sub-steps:

[0054] Step A: Extract target motion features in the overlapping area, and the motion features include acceleration vector and angular velocity change.

[0055] Step B: Calculate the similarity, and the formula is as follows:

[0056] S = w1*cos theta + w2*exp(-|Delta a|) + w3*IoU_traj

[0057] Wherein, S represents the similarity, w1, w2 and w3 represent different weight values respectively, theta represents the heading angle difference of two trajectories in the overlap area, Delta a represents the module of the acceleration vector difference of two trajectories, and IoU_traj represents the trajectory intersection over union of the overlap area.

[0058] Step C: judging whether the current similarity S is greater than a preset threshold, if yes, determining as the same target, and realizing seamless switching between devices.

[0059] Compared with the prior art, the present application has the following beneficial effects:

[0060] 1. The dynamic adjustable visual label of the present application breaks through the limitation of static label and creates a light intensity self-adaptive substrate.

[0061] 2. The multi-modal reasoning positioning engine in the present application applies UniVG-R1 reasoning chain to vehicle positioning for the first time, solves the problem of complex scene understanding, fuses SegVG pixel-level segmentation, and improves the pose solution accuracy.

[0062] 3. The pixel-level-geometry-level fusion positioning of the present application combines the absolute coordinates of the visual label and the relative motion of the laser SLAM, realizes error complementation, and improves the positioning accuracy.

[0063] 4. The present application proposes a chain measurement matrix, unifies the observation equations of vision, laser and inertial navigation, dynamically adjusts the data weight based on the GRPO algorithm, overcomes the limitation of the fixed noise model of the traditional EKF, and innovates the difficult perception strategy to allocate higher training weight to low light and high dynamic scenes. BRIEF DESCRIPTION OF DRAWINGS

[0064] Other features, objects and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments, read in conjunction with the accompanying drawings:

[0065] Figure 1 It is a working method flowchart of the present application. DETAILED DESCRIPTION

[0066] The present application will be described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any form. It should be pointed out that, for those skilled in the art, without departing from the concept of the present application, a number of changes and improvements can be made. These all belong to the protection scope of the present application.

[0067] The application is applicable to the continuous high-precision positioning needs of intelligent networked vehicles in satellite signal denial environments (such as tunnels, underground parking lots, urban canyons, etc.), and realizes centimeter-level positioning accuracy through the cooperation of visual labels deployed on the roadside and the vehicle-mounted perception system, thereby providing reliable position information for automatic driving decision control.

[0068] According to the application, a dynamic adjustable visual label cooperative intelligent networked vehicle positioning system is provided, as shown in Figure 1 The road-side subsystem is used to deploy dynamic adjustable visual labels and radar-vision integrated machines. The dynamic adjustable visual labels are arranged in an array on the roadside guardrails, tunnel walls and the like, and the interval is set according to the actual application scene, for example, 50 m in a high-speed scene or 20 m in an urban scene. The radar-vision integrated machine includes a millimeter wave radar and a high-definition camera, and is equipped with an RSU communication unit. The road-side subsystem further includes an edge computing unit for performing label state monitoring and cross-device target splicing. The vehicle-mounted subsystem is used to determine the position information of the current vehicle, and includes a multi-spectral imaging module, a positioning fusion controller and a V2X communication module. The multi-spectral imaging module includes a visible light and near-infrared (850 nm) dual camera with a resolution of 1920x1200@60fps. The positioning fusion controller is used to run a UniVG-R1 model and an adaptive Kalman filter. The V2X communication module receives the road-side label map and auxiliary positioning information.

[0069] The dynamic adjustable visual label has a hierarchical coding structure including a base layer, a dynamic layer and an information layer. The base layer adopts a fixed black and white checkerboard pattern, wherein the width ratio of “black-white-black” is 3:2:3, and a basic pose solution reference is provided. The dynamic layer is an electroluminescent film, which automatically adjusts the brightness (50-1000 cd / m 2 Adjustable) according to the ambient light intensity to ensure low-illumination recognition. The information layer embeds an NFC chip to store the absolute coordinates (WGS84 longitude and latitude) of the label and the topological relationship of adjacent labels. The ID of the dynamic adjustable visual label is as follows:

[0070] Label ID = Area Code + Type Code + Coordinate Check Code

[0071] Among them, the area code contains 8 bits, the type code contains 4 bits, and the coordinate check code contains 16 bits. The dynamically adjustable visual tag adopts Reed-Solomon error correction coding and supports ID recognition under 30% pattern occlusion. Compared with the circular tags in the prior art, the present invention adds an orientation perception mark, that is, an asymmetric L-shaped guide symbol is set in the upper right corner to accelerate image matching. In addition, the dynamically adjustable visual tag has built-in light sensors and temperature sensors, and regularly reports the working status (battery voltage, luminous intensity) through LoRa wireless communication. The roadside edge computing unit generates a tag health map based on this to guide the vehicle to filter invalid tags.

[0072] According to a method for positioning an intelligent connected vehicle using a dynamically adjustable visual tag collaboration provided by the present invention, the positioning method adopts the intelligent connected vehicle positioning system using the dynamically adjustable visual tag collaboration, and includes:

[0073] Step S1: Acquire raw data, which includes images, point clouds, and IMU.

[0074] Step S2: Preprocessing the raw data. The preprocessing includes illumination adaptive preprocessing, that is, applying nonlinear mapping to the input image. The formula is as follows:

[0075] I_out=α·tanh(β·(I_in-γ))+δ

[0076] Among them, I_out and I_in represent the preprocessed pixel grayscale value and the pixel grayscale value of the input image, respectively. α, β, γ, and δ represent the output gain coefficient, contrast adjustment factor, grayscale offset, and brightness compensation value, respectively. The parameters α and β are dynamically adjusted by the ambient light intensity to solve the problem of backlight / strong light interference.

[0077] Step S3: Use the SegVG pixel-level segmentation network to segment the preprocessed data and generate a pixel-level label mask to replace the traditional bounding box detection.

[0078] Step S4: Confidence evaluation is performed on the pixel-level label mask, preprocessed data, and raw data to obtain the label 6DoF pose, point cloud matching results, and corresponding confidence levels. The confidence evaluation includes UniVG-R1 inference, laser SLAM feature extraction, and IMU pre-integration.

[0079] The UniVG-R1 inference is based on the UniVG-R1 model to perform inference-guided positioning. UniVG-R1 inference chain example:

[0080] chain_of_thought = ["Detect all potential label regions in the image", "Evaluate the occlusion ratio and lighting quality of each region", "Select regions with confidence > 0.7 for decoding", "Fuse pose estimation results of multiple labels"]

[0081] Outputs 6DoF pose (x, y, z, roll, pitch, yaw) with an accuracy of ±0.5° (pose) and ±2cm (position). x, y, z, roll, pitch, and yaw represent the coordinates and rotation of the three axes in the local coordinate system, respectively.

[0082] Step S5: Input the tag 6DoF pose and point cloud matching results into the spatiotemporal alignment module for multi-source data spatiotemporal alignment. The multi-source data spatiotemporal alignment includes timestamp synchronization and coordinate uniform conversion. Among them, the timestamp synchronization uses the PTPv2 protocol to align the timestamps of the camera, IMU, and roadside data (error <1ms). Coordinate uniform conversion refers to defining the transformation relationship between the global coordinate system (ECEF) and the local coordinate system (center of the vehicle's front axle), as shown in the following formula:

[0083] [X_global]=[R]·[X_local]+[T]

[0084] Among them, X_global represents the global coordinate system, and X_local represents the local coordinate system. The rotation matrix [R] is calculated by the IMU heading angle, and the translation vector [T] comes from the absolute coordinates of the visual label.

[0085] Step S6: The corresponding confidence and spatiotemporally aligned data are subjected to an adaptive weighted extended Kalman filter (EKF) to obtain the final fusion positioning. The adaptive weighted extended Kalman filter (EKF) refers to the introduction of a weight allocation mechanism optimized by reinforcement learning based on the extended Kalman filter (EKF). The weight allocation mechanism includes the following steps:

[0086] Step S6.1: Construct the state vector as follows:

[0087] x=[position,velocity,attitude]^T(15 dimensions)

[0088] Among them, x represents the state quantity, position, velocity and attitude represent position, speed and attitude respectively, and T represents transpose.

[0089] Step S6.2: Define the observation sources, which are the visual tag positioning result z_vision, the laser SLAM point cloud matching z_lidar, and the inertial navigation output z_imu.

[0090] Step S6.3: The confidence evaluation network based on UniVG-R1 generates a weight matrix, as follows:

[0091] W=f(ambient illumination, label occlusion rate, vehicle speed)

[0092] where W and f represent the weight matrix and the confidence evaluation function, respectively. For example, the laser weight is increased in a rainy and foggy environment, and the visual label weight is increased at night.

[0093] Step S6.4: Update the measurement equation:

[0094] z=H·x+v

[0095] v~N(0,W·R)

[0096] where z represents the observation, H represents the conversion matrix, v represents the measurement noise, N represents a Gaussian noise distribution with a mean of 0 and a covariance matrix of WR, and R represents the observation noise covariance.

[0097] According to the dynamic adjustable visual label cooperative intelligent networked vehicle positioning method provided by the application, a vehicle-road cooperative optimization step is further included, that is, for long-distance positioning requirements, road-side auxiliary information is introduced, including cross-device target splicing and global positioning correction. The cross-device target splicing adopts an improved threshold method and trajectory mode matching, including the following sub-steps:

[0098] Step A: Extract target motion features (acceleration vector, angular velocity change) in the overlapping area.

[0099] Step B: Calculate the similarity, as follows:

[0100] S=w1·cosθ+w2·exp(-|Δa|)+w3·IoU_traj

[0101] where S represents the similarity, w1, w2, and w3 represent different weight values, θ represents the heading angle difference of the two trajectories in the overlapping area, Δa represents the modulus of the acceleration vector difference of the two trajectories, and IoU_traj represents the trajectory intersection over union in the overlapping area.

[0102] Step C: Determine whether the current similarity S is greater than a preset threshold value, if yes, it is determined that it is the same target, and seamless switching between devices is realized; the threshold value is preferably set to 0.85.

[0103] The global positioning correction refers to that the edge computing unit in the road-side subsystem broadcasts a positioning enhancement packet once every 500 ms.

[0104] The dynamic adjustable visual label of the application, the roadside layout has the environment response ability of the active light-emitting label. The multi-modal fusion positioning algorithm, combined with the visual processing engine of the UniVG-R1 inference model and the SegVG segmentation technology. The vehicle-road cooperation optimization mechanism, based on the global trajectory splicing and confidence evaluation of the roadside radar vision array.

[0105] Further, the specific embodiments of the application are described in detail as follows in combination with the actual application scenarios:

[0106] First, system deployment. Roadside device installation: install dynamic visual labels at an interval of 50m on both sides of the guardrail of the target road section (such as a tunnel), with a height of 1.2-1.5m and a pitch angle of 5° towards the lane. Install a radar-vision integrated machine (including RSU) every 200m to ensure that the monitoring areas of adjacent devices overlap by ≥20m. Deploy an edge server (Intel i7-12800H+RTX4090) at the entrance of the tunnel, pre-installed with a visual label map database. Vehicle-mounted system integration: install a multi-spectral camera (visible light+850nm near-infrared) on the inner side of the front windshield, with a viewing angle of 120° and a frame rate of 60fps. Install a positioning fusion controller (NVIDIA Orin platform) in the trunk, which is connected to the vehicle IMU (accuracy 0.1° / h) and wheel speed signal through the CAN bus. Install a V2X antenna (DSRC / C-V2X dual-mode) on the roof, with a communication delay of <100ms with the roadside RSU.

[0107] Then, the online positioning process, including:

[0108] Step 1: Visual label recognition (vehicle-mounted end) The vehicle enters the tunnel, and the camera captures the label image in front. Perform light compensation: when the ambient illuminance is <50lux, trigger near-infrared light compensation (850nm). Call the SegVG model to generate a pixel-level mask, which can still be recognized even if the label is 30% obscured by mud. Decode the label ID and absolute coordinates, and calculate the relative pose by combining the camera intrinsic parameters, as follows:

[0109] [x_c, y_c, z_c] = solvePnP(tag_corners, camera_matrix)

[0110] Where x_c, y_c, and z_c represent the relative pose x-axis, y-axis, and z-axis of the current vehicle, respectively, and tag_corners and camera_matrix represent the label corner and camera intrinsic matrix, respectively.

[0111] Step 2: Multi-source data fusion (vehicle end), synchronously receive local point cloud map of laser SLAM and 6DoF data of IMU. Run UniVG-R1 inference chain to evaluate data confidence: vision confidence = 0.92 (label complete), laser confidence = 0.85 (few features on tunnel wall), IMU confidence = 0.95. Adaptive EKF adjusts weights according to confidence: W_vision = 0.5, W_lidar = 0.3, W_imu = 0.2. Output fused pose (frequency 100Hz)

[0112] Step 3: Roadside cooperative correction (roadside end) The radar-vision integrated machine detects vehicle passing through the splicing area, extracts trajectory features as velocity vector v = [25.6 m / s, -0.3 m / s] and heading angle yaw = 1.2 rad. Match adjacent device data and calculate similarity S = 0.91 > threshold (0.85) to determine as the same target. The edge server generates a correction message and broadcasts it to the vehicle system through the RSU. The vehicle updates the global position based on the message, with error correction amount Δx = -0.2 m and Δy = +0.1 m. The real scene test performance (1 km tunnel, average speed 60 km / h) is shown in Table 1:

[0113] Table 1

[0114] index Traditional visual labels The present invention Improvement Positioning success rate 76.5％ 98.2％ +21.7％ Average position error 0.82m 0.11m -86.6％ Maximum cumulative error 2.1m 0.25m -88.1％ Processing Delays 120ms 45ms -62.5％

[0115] Finally, in the label failure scenario, when the roadside reports the label health status as "fault", the vehicle system automatically switches to pure SLAM mode, and the laser weight is increased to 0.7. In the communication interruption scenario, when the V2X link is disconnected, the local cached 500m label map data is used to assist positioning. In the extreme environment scenario, in heavy fog (visibility <50m), the radar dominant mode is activated, and the vision weight is reduced to 0.2.

[0116] In summary, the positioning accuracy of the application is improved, with a positioning error ≤0.15m (3σ) in urban tunnel scenarios, which is higher than the traditional visual label positioning accuracy. It supports high-speed driving scenarios. The roadside device density is reduced, and the existing intelligent road radar-vision integrated machine is upgraded and deployed. In rain, fog, night, shadow shielding and other working conditions, the positioning success rate is high. The application has been verified through simulation and real vehicle testing, and is especially suitable for GNSS denial environments such as highway tunnel groups, underground parking lots and urban viaducts, providing a high-reliability positioning solution for L3 and above autonomous driving.

[0117] Those skilled in the art know that, in addition to implementing the system provided by the present application and each device, module and unit thereof in the form of pure computer readable program code, the system provided by the present application and each device, module and unit thereof can also be implemented in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers and embedded microcontrollers, etc. by logically programming the method steps to achieve the same functions. Therefore, the system provided by the present application and each device, module and unit thereof can be considered as a hardware component, and the devices, modules and units included therein for achieving various functions can also be considered as structures within the hardware component; the devices, modules and units for achieving various functions can also be considered as both software modules implementing methods and structures within hardware components.

[0118] The specific embodiments of the present application are described above. It needs to be understood that the present application is not limited to the specific embodiments described above, and various changes or modifications can be made by those skilled in the art within the scope of the claims, which does not affect the essential content of the present application. The embodiments of the present application and the features in the embodiments can be combined with each other in any manner without conflict.

Claims

1. A dynamic and adjustable visual tag collaborative intelligent connected vehicle positioning system, characterized in that: include: Roadside subsystem and vehicle-mounted subsystem; The roadside subsystem is used to deploy dynamically adjustable visual tags and radar-visual integrated devices; The dynamically adjustable visual tags are arranged in arrays at locations including roadside guardrails and tunnel walls; the radar-visual integrated device includes a millimeter-wave radar and a high-definition camera, equipped with an RSU communication unit; The vehicle-mounted subsystem is used to determine the current vehicle location information, including a multispectral imaging module, a positioning fusion controller and a V2X communication module.

2. The intelligent connected vehicle positioning system with dynamically adjustable visual tags according to claim 1 is characterized in that: The roadside subsystem also includes an edge computing unit for performing tag status monitoring and cross-device target splicing.

3. The intelligent connected vehicle positioning system with dynamically adjustable visual tags according to claim 1 is characterized in that: The multispectral imaging module includes a visible light and near infrared dual camera; The positioning fusion controller is used to run the UniVG-R1 model and the adaptive Kalman filter; The V2X communication module receives roadside tag maps and auxiliary positioning information.

4. The intelligent connected vehicle positioning system with dynamically adjustable visual tags according to claim 1 is characterized in that: The dynamically adjustable visual tag is a layered coding structure, including a base layer, a dynamic layer and an information layer; The base layer adopts a fixed black and white checkerboard pattern; The dynamic layer is an electroluminescent film; The information layer embeds an NFC chip to store the absolute coordinates of the tag and the topological relationship of the adjacent tags; The dynamically adjustable visual tag IDs are as follows: Tag ID = area code + type code + coordinate check code The area code contains 8 bits, the type code contains 4 bits, and the coordinate check code contains 16 bits.

5. The intelligent connected vehicle positioning system with dynamically adjustable visual tags according to claim 1 is characterized in that: The dynamically adjustable visual mark has an asymmetrical L-shaped guide symbol set in the upper right corner, a built-in light sensor and a temperature sensor, and regularly reports the working status through LoRa wireless communication.

6. A method for positioning intelligent connected vehicles using dynamically adjustable visual tags, characterized in that: The positioning method of the intelligent connected vehicle positioning system using the dynamically adjustable visual tag collaboration described in any one of claims 1 to 5 includes: Step S1: obtaining original data; Step S2: preprocessing the original data; Step S3: Segment the preprocessed data using the SegVG pixel-level segmentation network to generate a pixel-level label mask; Step S4: performing confidence evaluation on the pixel-level label mask, preprocessed data, and original data to obtain the label 6DoF pose, point cloud matching result, and corresponding confidence; Step S5: inputting the label 6DoF pose and point cloud matching results into the spatiotemporal alignment module to perform spatiotemporal alignment of multi-source data; Step S6: performing adaptive weighted extended Kalman filtering (EKF) on the corresponding confidence and time-space aligned data to obtain the final fusion positioning.

7. The intelligent connected vehicle positioning system with dynamically adjustable visual tags as claimed in claim 6 is characterized in that: The raw data includes images, point clouds and IMU; The preprocessing includes illumination adaptive preprocessing, that is, applying a nonlinear mapping to the input image, as follows: I_out=α·tanh(β·(I_in-γ))+δ Where I_out and I_in represent the pixel grayscale value after preprocessing and the pixel grayscale value of the input image, respectively; α, β, γ, and δ represent the output gain coefficient, contrast adjustment factor, grayscale offset, and brightness compensation value, respectively; The confidence evaluation includes UniVG-R1 inference, laser SLAM feature extraction and IMU pre-integration.

8. The intelligent connected vehicle positioning system with dynamically adjustable visual tags as claimed in claim 6 is characterized in that: The multi-source data spatiotemporal alignment includes timestamp synchronization and coordinate uniform conversion; The timestamp synchronization uses the PTPv2 protocol to align the timestamps of the camera, IMU, and roadside data; Coordinate unified transformation refers to defining the transformation relationship between the global coordinate system and the local coordinate system, as shown below: [X_global]=[R]·[X_local]+[T] Among them, X_global represents the global coordinate system, X_local represents the local coordinate system, [R] represents the rotation matrix, and [T] represents the translation vector.

9. The intelligent connected vehicle positioning system with dynamically adjustable visual tags as claimed in claim 6 is characterized in that: The step S6 comprises: Step S6.1: Construct the state vector as follows: x=[position,velocity,attitude]^T Among them, x represents the state quantity, position, velocity and attitude represent position, speed and attitude respectively, and T represents transposition; Step S6.2: Define the observation sources, which are the visual tag positioning result z_vision, the laser SLAM point cloud matching z_lidar, and the inertial navigation output z_imu; Step S6.3: Generate a weight matrix based on the confidence evaluation network of UniVG-R1. The formula is as follows: W = f (ambient illumination, tag occlusion rate, vehicle speed) Where W and f represent the weight matrix and confidence evaluation function respectively; Step S6.4: Update the measurement equation: z=H·x+v v~N(0,W·R) Where z represents the observed quantity, H represents the transformation matrix, v represents the measurement noise, N represents the Gaussian noise distribution with mean 0 and covariance matrix WR, and R represents the observation noise covariance.

10. The intelligent connected vehicle positioning system with dynamically adjustable visual tags as claimed in claim 6, characterized in that: It also includes vehicle-road collaborative optimization steps; The vehicle-road collaborative optimization step includes cross-device target splicing and global positioning correction; The cross-device target splicing includes the following sub-steps: Step A: extracting target motion features in the overlapping area, wherein the motion features include acceleration vector and angular velocity change; Step B: Calculate the similarity. The formula is as follows: S=w1·cosθ+w2·exp(-|Δa|)+w3·IoU_traj Where S represents the similarity, w1, w2 and w3 represent different weight values, θ represents the heading angle difference between the two trajectories in the overlapping area, Δa represents the modulus of the acceleration vector difference between the two trajectories, and IoU_traj represents the intersection-over-union ratio of the trajectories in the overlapping area; Step C: Determine whether the current similarity S is greater than a preset threshold. If so, it is determined to be the same target, achieving seamless switching between devices.

Images