An air-ground cooperative intelligent vehicle positioning method, system and device

Abstract

The application provides a ground-air cooperative intelligent vehicle positioning method, system and device, belongs to the intelligent transportation system and the vehicle-road cooperative perception technical field, aiming at the problems of large GPS positioning error, missing perception data and significant difference in motion characteristics of intelligent network connected vehicles and air vehicles in complex urban environment, first, the multi-source data is subjected to vehicle adaptive filtering, time sequence completion and missing mask processing; then, a wireless channel state feature extraction mechanism enhanced by global environment semantic features is introduced, and the multi-position wireless channel features are dynamically weighted and fused based on the real-time spatial geometric change relationship between the vehicle and the road side unit. The application can realize cooperative high-precision positioning of ground intelligent network connected vehicles and air vehicles in the same framework, and has strong robustness and engineering application value.

Description

Technical Field

[0001] This invention belongs to the technical fields of intelligent transportation systems, vehicle-road cooperative perception, and low-altitude economy, and in particular relates to a ground-air cooperative intelligent vehicle positioning method, system, and device. Background Technology

[0002] With the rapid development of intelligent connected vehicles and the low-altitude economy, intelligent connected vehicles and air transportation are showing a trend of collaborative operation in urban environments, which places higher demands on continuous and high-precision positioning. Existing intelligent vehicle positioning solutions mainly rely on satellite navigation systems such as GPS. However, in urban canyons, densely built-up areas, or complex traffic environments, GPS signals are easily affected by obstruction and multipath effects, resulting in a significant increase in positioning errors. Moreover, the error distribution exhibits non-Gaussian characteristics, which are difficult to suppress stably using traditional filtering methods.

[0003] On the other hand, intelligent connected vehicles are constrained by road topology and their motion patterns are relatively stable, while aerial vehicles have obvious three-dimensional high dynamic characteristics. Existing positioning methods are usually designed for single types of targets, making it difficult to take into account the positioning needs of different motion characteristics within a unified framework. In addition, existing multimodal fusion positioning methods mostly use simple feature stitching methods and lack physical modeling of wireless signal propagation characteristics and the spatial geometric relationship between multi-roadside units and vehicles. This leads to decreased accuracy and instability of positioning results under conditions of missing perception data, line-of-sight obstruction, or uneven wireless signal quality.

[0004] Therefore, there is an urgent need for a collaborative localization method that can integrate multimodal perception information, be compatible with the motion characteristics of ground and air intelligent vehicles, and have strong robustness and engineering feasibility. Summary of the Invention

[0005] In view of this, the present invention aims to provide a ground-air cooperative intelligent vehicle positioning method, system and device to at least solve one of the problems in the background art.

[0006] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0007] Firstly, this solution discloses a ground-air cooperative intelligent vehicle positioning method, including the following steps:

[0008] Heterogeneous data (including heterogeneity in source and mode, motion characteristics and noise type, and time-series missing patterns) collected from intelligent connected vehicles and air vehicles are preprocessed. Vehicle adaptive filtering is performed on the raw GPS data according to different motion characteristics. Time-series filtering and missing masking are performed on the channel state information obtained between the roadside unit and the target vehicle. At the same time, time-series environmental semantic features are completed on the multi-view RGB images at the roadside.

[0009] The wireless channel state features enhanced by extracting environmental semantic information are obtained by using a visual perception model to extract environmental semantic features from the multi-view RGB images, generating a global environmental representation by introducing a cross-view fusion mechanism with geometric priors, and injecting the environmental representation into the wireless signal feature extraction model through a progressive gating method to obtain high-dimensional wireless channel state features that fuse environmental semantic information.

[0010] Based on explicit physical modeling, the wireless signal characteristics of multiple roadside units are evaluated and dynamically weighted and fused. An explicit spatial location feature code is constructed based on the spatial geometric relationship between the vehicle and each roadside unit, which includes distance, direction, relative position relationship and road semantic confidence. The wireless channel state characteristics of multiple roadside units are then dynamically weighted and fused in time series based on the explicit spatial location feature code.

[0011] The fused wireless channel state features are spliced ​​and fused with the filtered and preprocessed GPS prior pose information. The residual position regression module is used to predict the pose correction of the intelligent vehicle. The absolute position accuracy, motion direction consistency, correction rationality and dynamic constraints are jointly optimized by a multi-objective joint loss function to output the final positioning result of the intelligent vehicle.

[0012] For ground-based intelligent connected vehicles, the positioning results are used to determine the motion state. When it is determined that the vehicle is stationary, its predicted positioning coordinates are locked.

[0013] Furthermore, the vehicle adaptive filtering includes: for ground-based intelligent connected vehicles constrained by road topology and whose motion is mainly two-dimensional, a Kalman filter algorithm is used to smooth the collected GPS data; for air vehicles with six degrees of freedom motion characteristics, an unscented Kalman filter algorithm is used to perform nonlinear state estimation on the collected GPS data, in order to adapt to the differences in motion models of different types of intelligent vehicles.

[0014] Furthermore, the preprocessing of the channel state information collected by the roadside unit includes: sampling the wireless channel state information through a sliding time window, introducing mask identifiers for wireless channel data with time sequence missing data, and reconstructing the channel characteristics at the missing time points in conjunction with a time sequence completion network to obtain a continuous channel state feature sequence.

[0015] Temporal environmental semantic feature completion of multi-view RGB images of roadside units includes: using a temporal modeling network with self-attention mechanism, introducing learnable mask vectors to place missing frames, and combining position encoding to preserve the temporal contextual relationship of image features, so as to achieve temporal consistency reconstruction of environmental semantic features.

[0016] Furthermore, the cross-view fusion mechanism that introduces geometric priors includes: constructing a geometric feature vector based on the installation position and orientation parameters of the roadside camera, and using the geometric feature vector as a guide to fuse the environmental semantic features extracted from different viewpoints through a dynamic complementary cross-attention mechanism to generate a global environmental representation with spatial consistency.

[0017] The progressive gating method includes: setting a gating fusion structure in the wireless signal feature extraction model, assigning learnable weights to the injected environmental semantic features, and initializing the weights to zero at the beginning of model training. As the number of training iterations gradually increases, the model learns to use environmental semantic features to enhance wireless signal features, thereby gradually increasing the weights, thus avoiding interference to wireless signal features caused by forcibly fusing multimodal information in the early stage of training.

[0018] Furthermore, the explicit spatial location feature encoding includes at least the horizontal distance, height difference, relative azimuth angle, orientation alignment, lateral deviation, longitudinal deviation, approach trend, and co-route confidence between the vehicle and the roadside unit, which is used to characterize the spatial geometric relationship between the vehicle and the roadside unit during the wireless signal propagation process, and serves as the basis for quality assessment and weighted fusion of channel state features of multiple roadside units.

[0019] The time-series dynamic weighted fusion of the channel states of the multiple roadside units includes: calculating the weight coefficients of each roadside unit at different time steps based on the explicit spatial location feature encoding, and performing weighted fusion on the wireless channel state features corresponding to the time steps, so as to reduce the impact of roadside units with severe obstruction or poor signal quality on the vehicle positioning results.

[0020] Furthermore, the multi-objective joint loss function includes at least the absolute position error loss, motion direction consistency loss, GPS correction rationality constraint loss, and vehicle dynamics constraint loss, aiming to improve positioning accuracy while ensuring that the vehicle position prediction results meet the actual physical motion laws.

[0021] Furthermore, the motion state determination of ground-based intelligent connected vehicles includes: performing principal component analysis on the vehicle's positioning coordinates within a predetermined time window, determining whether the vehicle is in a stationary or moving state based on the variance distribution of the trajectory point cloud, and locking the vehicle's positioning coordinates when the motion mode is determined to be stationary.

[0022] Secondly, this solution discloses a ground-air cooperative intelligent vehicle positioning method system, including:

[0023] The data preprocessing module is used to perform vehicle adaptive filtering on GPS data of ground intelligent connected vehicles and air vehicles, and to perform temporal filtering, masking and missing completion on wireless channel status information and multi-view RGB images collected by roadside units.

[0024] The feature enhancement and extraction module is used to extract environmental semantic features from the multi-view RGB image and generate a global environmental representation based on a cross-view fusion mechanism with geometric prior. The environmental representation is then injected into the wireless signal feature extraction model through a progressive gating method to obtain high-dimensional channel state features that fuse environmental semantics.

[0025] The wireless signal quality assessment and feature fusion module is used to construct explicit spatial location feature codes based on the spatial geometric relationship between the vehicle and each roadside unit, and to perform time-series dynamic weighted fusion of the wireless channel state features of multiple roadside units based on the coding module.

[0026] The location regression prediction module is used to concatenate and fuse the weighted fused wireless channel state features with the filtered GPS prior pose information, and predict the positioning results of the intelligent vehicle based on the residual location regression network constrained by the GPS prior pose and the multi-objective joint loss function.

[0027] The motion state discrimination and post-processing module is used to discern the motion state of the positioning results of the ground intelligent connected vehicle and lock the positioning coordinates when the vehicle is determined to be stationary.

[0028] Furthermore, the data preprocessing module employs a Kalman filter algorithm for ground-based intelligent connected vehicles and an unscented Kalman filter algorithm for air vehicles to adapt to the noise characteristics and positioning requirements under different degrees of freedom of motion.

[0029] The data preprocessing module includes a channel state information filtering unit based on a sliding time window, and reconstructs the missing channel state features through a masking mechanism and a timing completion network;

[0030] The feature enhancement and extraction module includes a temporal completion sub-module based on a self-attention mechanism, which is used to reconstruct the temporal consistency of environmental semantic features of multi-view RGB images at the road end.

[0031] The cross-view fusion mechanism constructs geometric feature vectors based on the installation position and orientation parameters of the roadside camera, and uses the geometric feature vectors to guide the complementary fusion of cross-attention mechanisms of environmental semantic features from different viewpoints.

[0032] The progressive gating method sets learnable fusion weights for global environmental semantic features, initializes them to zero at the start of model training to avoid their influence on wireless signal features, and gradually increases their fusion ratio by learning to utilize different modal features during training.

[0033] The explicit spatial location feature coding network constructed by the wireless signal quality assessment and feature fusion module includes at least the horizontal distance, height difference, relative azimuth angle, orientation alignment, lateral deviation, and approach trend between the vehicle and the roadside unit.

[0034] The position regression prediction module uses a multi-objective joint loss function that includes absolute position error constraints, motion direction consistency constraints, GPS correction magnitude constraints, and dynamic constraints to train the GPS prior-based residual position regression network.

[0035] Thirdly, this solution discloses a ground-air cooperative intelligent vehicle positioning method and apparatus, including a processor and a memory. The memory stores a computer program, and when the computer program is run on the processor, the processor executes the ground-air cooperative intelligent vehicle positioning method.

[0036] Compared with existing technologies, the ground-air cooperative intelligent vehicle positioning method, system, and device described in this invention have the following advantages:

[0037] (1) This invention introduces wireless channel state features with enhanced environmental semantics and combines GPS prior residual position regression to effectively suppress the interference of multipath effect and non-line-of-sight propagation on positioning results in complex urban road environments and improve overall positioning accuracy.

[0038] (2) This invention adapts intelligent connected vehicles with two-dimensional planar motion and highly maneuverable air vehicles in three-dimensional space within the same vehicle positioning architecture through vehicle adaptive filtering strategy and explicit spatial position feature encoding. By using a unified architecture, it avoids the system complexity caused by the need to build multiple independent models due to differences in motion modes, and adapts to ground-air coordinated transportation systems in low-altitude economic scenarios;

[0039] (3) The present invention dynamically weights and fuses wireless signal features based on the temporal spatial geometric relationship changes between the vehicle and multiple roadside units, so as to suppress the adverse effects caused by roadside units where wireless signals are blocked or the channel quality is severely degraded, thereby improving the robustness of the vehicle positioning system that fuses multiple location wireless signal features.

[0040] (4) This invention determines the motion state of the predicted position coordinates of ground intelligent connected vehicles and locks the positioning coordinates when the vehicle is stationary, effectively suppressing position drift under stationary conditions and improving the continuity and practicality of positioning results. Attached Figure Description

[0041] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0042] Figure 1 This is a schematic diagram of the overall process provided for embodiments of this disclosure;

[0043] Figure 2 A schematic diagram of a wireless signal feature extraction process based on multi-view fusion and global environmental feature enhancement provided in an embodiment of this disclosure;

[0044] Figure 3 This disclosure provides a schematic diagram of residual location regression based on multi-location CSI feature fusion and GPS prior.

[0045] Figure 4 Comparison and Difference Analysis of Explicit Spatial Coding Features between Air Vehicles (AT4) and Ground Vehicles (Car20);

[0046] Figure 5 The experimental results of vehicle positioning for air vehicles (AT4) and ground vehicles (Car20) are shown in the table. (a) shows the GPS correction results for AT4, and (b) shows the GPS correction results for Car20.

[0047] Figure 6 This is a schematic diagram of the motion state determination and stationary locking area of ​​the ground vehicle (Car4) based on principal component analysis in the method disclosed herein. Detailed Implementation

[0048] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0049] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0050] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0051] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0052] The following describes in detail, with reference to the accompanying drawings and embodiments, a ground-air cooperative intelligent vehicle positioning method based on multimodal perception information fusion for low-altitude economic scenarios according to the present invention:

[0053] like Figure 1 As shown, the method provided in this embodiment mainly includes the following five steps:

[0054] Step S1: Heterogeneous multimodal data preprocessing: The core of this step is to adopt vehicle-adaptive filtering strategies for targets with different motion characteristics (ground vehicles and air taxis), and to perform time-series restoration on the perception data collected by the roadside, in order to address the lack of sensor data in complex urban road scenarios and the multiple noises in low-altitude economic scenarios.

[0055] 1) Data Preprocessing Based on Heterogeneous Filtering Strategy: For ground-based intelligent connected vehicles, considering their constraints on road topology, their motion models are typically straight-line or curved. This embodiment employs existing Kalman filtering, utilizing observation residuals to correct prior states, effectively smoothing high-frequency random noise in GPS caused by multipath effects in urban canyons. For air vehicles, considering their highly dynamic nonlinear motion characteristics with six degrees of freedom (6DoF) (including pitch, roll, and vertical takeoff and landing), this embodiment employs existing unscented Kalman filtering to capture the complex maneuvering trajectory characteristics of air targets.

[0056] 2) Perception data completion network based on self-attention mechanism: For frame-level missing frames that occur when RGB camera captures images and when channel state information is transmitted between roadside equipment and vehicles, a learnable mask vector is used to fill the missing frames. Sine and cosine position coding is introduced to preserve the temporal context information of the sequence. The features of the missing parts are reconstructed using the contextually valid frames through the multi-head attention mechanism of the Transformer encoder.

[0057] Specifically, a learnable mask vector is used for the location of missing frames. The feature sequence is padded and aligned temporally using sine and cosine positional encoding. This process leverages the long-term dependencies of valid context frames to reconstruct continuous environmental features. X is the input feature sequence (containing valid frames and masked missing frames), PE is the positional encoding used to preserve the temporal context information of the sequence, and TransformerEncoder represents a self-attention-based feature encoder. The completed feature sequence is shown below. This can be represented by the self-attention mechanism as:

[0058] ;

[0059] Step S2: Channel state feature extraction with enhanced environmental semantics. This step aims to enhance the robustness of wireless channel features by utilizing visual semantics. It should be noted that the "wireless signal feature extraction model" (such as WiFo) and "visual perception large model (such as DINOv2)" used in this embodiment can build a backbone network based on existing pre-trained architectures, but the core improvement of this invention lies in the design of a unique "geometrically guided fusion" and "progressive gating fusion mechanism".

[0060] 1) Geometric Prior-Guided Visual Feature Extraction: Semantic features of roadside multi-view images are extracted using a large visual model. A geometric-aware visual-semantic fusion module is introduced: Fourier feature encoding technology is used to extract the spatial position parameters of the camera (installation orientation angle). Mapped to a high-dimensional geometric vector: The geometric vector is injected into the feature sequence, and through geometric prior-guided visual feature extraction and fusion, the model can automatically aggregate complementary information from different perspectives based on the geometric position to generate a global environmental representation of the roadside base station.

[0061] 2) Progressive Gated Fusion Mechanism: To prevent visual perception features from disrupting the physical consistency of wireless signal features during the early stages of training, this embodiment designs a gated fusion structure. It employs a two-layer linear mapping layer, initializing the second layer to zero. By setting a group learning rate, the global visual features are... A decoder for a smoothly injected wireless base station model, wherein, To integrate the wireless channel features after incorporating environmental semantics, The original wireless channel state information features are represented by Linear, which represents the linear mapping layer. The overall calculation formula is:

[0062] ;

[0063] in Learnable gate parameters (Gate). During training initialization, let... This allows the model to prioritize learning single-mode wireless signal features; as training converges, Automatic adjustment (experiments show convergence to 0.3) enables the "soft" implantation of visual perception semantic information.

[0064] The geometrically prior-guided cross-view fusion module maps camera spatial position parameters into high-dimensional geometric vectors using Fourier feature encoding technology, specifically including:

[0065] 1) Using Fourier feature coding technique The normalized installation orientation angle of the roadside camera Angular values ​​mapped to a high-dimensional feature space:

[0066] ;

[0067] in, The mapped high-dimensional geometric feature vectors Original visual features The Fourier feature encoding function maps angles to high-dimensional sine and cosine vectors. The projected weight matrix is ​​a learnable matrix. This is a bias term.

[0068] 2) Geometry-Aware Semantic Alignment Unit (GSU) Guided by Prior Geometricity: To achieve accurate alignment of unstructured visual semantics from different perspectives, this embodiment constructs a GSU module. Unlike traditional feature stitching, this module uses spatial geometric vectors as indices to dynamically aggregate complementary information from multiple perspectives. Specifically, this module performs the following calculation process: First, it defines any target perspective feature to be enhanced as... (As the query vector), the feature set of the remaining auxiliary perspectives is defined as follows: (As a source of key-value pairs). Next, the affinity matrix between the target viewpoint and the auxiliary viewpoint in the geometric-semantic joint space is calculated. This process measures the degree of matching of features from different viewpoints through scaled dot product operations, and the calculation formula is expressed as:

[0069] ;

[0070] in, This represents the feature projection transformation operation. The scaling factor for the feature dimension. This represents the semantic contribution weight of the j-th auxiliary viewpoint to the i-th target viewpoint.

[0071] 3) Weighted Feature Aggregation: Based on the affinity matrix, the semantic features from the auxiliary perspective are weighted and recombined to generate a globally consistent environment representation vector.

[0072] ;

[0073] in, This represents the globally consistent environment characterization of the i-th perspective after aggregation. The feature transformation function (Value projection) is used. Through the above geometrically aware semantic alignment calculation, the model can automatically suppress geometric distortion caused by viewpoint bias, ensuring that the generated environment representation has topological consistency in 3D space. Finally, through the constructed gated fusion structure, global visual features are integrated. A decoder for injecting wireless signal feature extraction models is used to achieve the fusion and interactive enhancement of multimodal sensing data.

[0074] Step S3: Wireless signal feature quality assessment and dynamic weighted fusion based on explicit physical modeling. This step is the core innovation of this invention. To address the problem of the lack of interpretability of actual physical information in "black box" models, this embodiment constructs a 12-dimensional explicit spatial feature encoding to quantify the spatial geometric relationship and road semantic information between vehicle trajectories and road side facilities (RSFs).

[0075] 1) Calculation of 12-dimensional physical features: Based on the preprocessed GPS prior coordinates and roadside unit pose coordinates, the following key features are calculated:

[0076] 2D horizontal distance and height difference: Decoupling horizontal and vertical distances enables the system to automatically distinguish between ground vehicles (small height difference) and aerial vehicles (large height difference).

[0077] Lateral Deviation: Describes the distance the vehicle deviates from the center line of the RSF (Right Signal Frame). If the deviation is too large, it indicates that the vehicle is in the signal sidelobe area, and the weight should be reduced.

[0078] Approach trend: Based on the distance change between consecutive frames, determine whether the vehicle is approaching or moving away, simulating Doppler frequency shift characteristics.

[0079] Co-path confidence: The gain lobe diagram of the simulated directional antenna is obtained using an exponential decay model.

[0080] Features such as elevation angle, relative angle, longitudinal deviation, orientation alignment, and X / Y axis offset.

[0081] 2) Multi-location RSF wireless signal feature quality assessment and weighted fusion: The aforementioned 12-dimensional spatial location feature encoding is input into the quality assessment network, which outputs the temporal dynamic weights of RSFs at different locations. Dynamically weighted wireless signal feature fusion is then performed to generate the final high-dimensional semantic features of the wireless signal. This mechanism can automatically suppress signal interference from obstructed or non-line-of-sight (NLOS) base stations and fuses the vehicle's trajectory with the high-dimensional channel state representations of multiple roadside base stations based on physical information in the actual road scenario.

[0082] This method models the roadside base station (RSF) as a directional spatial reference node. The physical calculation definition of the 12-dimensional explicit spatial feature encoding is as follows:

[0083] Define vehicle GPS coordinates as The coordinates of the roadside base station RSF are , For vehicle orientation (Yaw). For RSF orientation (Yaw).

[0084] 1) 2D horizontal Euclidean distance between the vehicle and the RSF :

[0085] ;

[0086] 2) Vertical height difference between the vehicle and the RSF :

[0087] ;

[0088] 3) RSF pointing relative to the vehicle's perspective :

[0089] ;

[0090] 4) Orientation Alignment (Cosine Similarity, 1 indicates same direction, -1 indicates opposite direction) :

[0091] ;

[0092] 5) Longitudinal deviation (distance along the RSF line of sight) :

[0093] ;

[0094] 6) Lateral deviation (distance perpendicular to the RSF line of sight) :

[0095] ;

[0096] 7) Approaching trend (1 means approaching, -1 means moving away) :

[0097] ;

[0098] This feature is encoded using a sliding window. The current frame position. sgn is the starting frame position of the window, and sgn is the sign function (ternary: 1 for close, -1 for far, 0 otherwise).

[0099] 8) Confidence level of same path :

[0100] ;

[0101] in These are the weighting coefficients. Distance threshold The beamwidth parameter is used. The co-path confidence is calculated using a multi-factor exponential attenuation model to simulate the gain lobe diagram of the directional antenna. The calculation formula involves a nonlinear mapping between the lateral deviation of the vehicle relative to the roadside unit and the Euclidean distance.

[0102] 9) Absolute offset in the X-axis direction :

[0103] ;

[0104] 10) Absolute offset in the Y-axis direction :

[0105] ;

[0106] 11) Relative angle between the radio signal base station (RSF) and the vehicle :

[0107] ;

[0108] 12) Elevation angle (reflects 3D spatial relationships, especially for air targets) :

[0109] ;

[0110] Step S4: Residual intelligent vehicle pose regression prediction based on GPS prior. In this embodiment, the position regression adopts the "residual network learning" strategy. The position regression network predicts the correction between the preprocessed GPS observation and the real position, rather than the absolute coordinates. The residual learning mode can achieve good training results on small sample datasets and is suitable for rapid response of vehicle positioning in complex road scenarios.

[0111] Multi-objective joint loss function: To ensure that the prediction results during location regression conform to the physical laws of downstream vehicle localization tasks, the following loss function is constructed:

[0112] ;

[0113] in, This is the total loss value. The absolute position loss is a hybrid loss consisting of the mean absolute error and the mean square error, which balances noise resistance and convergence. For directional loss: the difference in angle is calculated using geodesic distance to solve the problem of periodic jumps in Euler angles (such as 179° and -179°). To correct the reasonableness loss of GPS, an asymmetric gradient is designed. If the predicted error is larger than the original GPS error, a heavy penalty is imposed; if the error decreases, no penalty is imposed. This avoids arbitrary modifications to the model under uncertainty and ensures a safe lower bound for vehicle positioning. Dynamic constraints: Thresholding is applied to the predicted velocity and acceleration to prevent non-physical "teleportation" and other abnormal predictions. For confidence regularization: Six-degrees-of-freedom (6DoF) pose confidence is introduced to avoid the model being overconfident. These are the balancing weight coefficients for each loss term.

[0114] The multi-objective joint loss function includes five sub-losses, which are specifically defined as follows:

[0115] Position vector: (Including coordinates in the X, Y, Z world coordinate system), Euler angle vectors: It includes roll (r), pitch (p), and yaw (y). This is the ReLU function, used to penalize portions exceeding a threshold. The location coordinates predicted by the model. These are the actual location coordinates. These are the preprocessed GPS observation coordinates. For true Euler angles, The predicted Euler angles.

[0116] 1) Hybrid Absolute Position Loss: A hybrid loss of Mean Absolute Error (MAE) and Mean Squared Error (MSE) is adopted to balance robustness to outliers (MAE) and smoothness of convergence (MSE).

[0117] ;

[0118] 2) Loss of direction of motion: It is used to calculate geodesic distances and solve the problem of periodic angle jumps. It approximates the angle difference in linear space onto the unit circle through sin / cos transformation, thereby improving the perception capability of the position regression module.

[0119] ;

[0120] 3) GPS correction for reasonable loss: The maximum correction magnitude threshold for a single step is determined by the asymmetric gradient design. The first term penalizes "negative optimization" (i.e., the error increases after correction); the second term limits the single correction magnitude from exceeding the threshold.

[0121] ;

[0122] 4) Dynamic constraint loss: Based on kinematic constraints, it penalizes unreasonable acceleration, excessive linear velocity and angular velocity to prevent the vehicle from rotating violently and unphysically during prediction.

[0123] ;

[0124] in, These are the acceleration, linear velocity, and angular velocity of the predicted trajectory, respectively. Here, denoted as the physical limit thresholds for velocity and angular velocity, respectively; T is the total length of the sampling time window; and t is a specific time step of the trajectory within the sampling time window.

[0125] 5) Confidence regularization loss: guides the confidence level c to remain within a reasonable range.

[0126] ;

[0127] The first term guides the confidence level to approach 0.65, while the latter two terms penalize extreme cases where c > 0.95 or c < 0.05 (this loss function can be set according to the actual needs of different scenarios).

[0128] Step S5: Ground vehicle motion state determination and post-stationary locking processing. Vehicle trajectories are sampled using a sliding window (e.g., 40 frames). Principal Component Analysis (PCA) is performed on the predicted position coordinates of all vehicles. The first principal component (λ1) represents the main diffusion direction (maximum variance direction) of the trajectory point cloud, the second principal component (λ2) represents the secondary diffusion direction perpendicular to the first principal component, and the third principal component (λ3) represents the minimum diffusion direction perpendicular to the first two. A total variance threshold (e.g., 1m) is set. 2 By setting a diffusion threshold (e.g., 1m) and considering the continuity of actual vehicle motion, a minimum stationary segment trajectory duration (e.g., 20 frames) is determined to accurately determine the stationary motion state of ground vehicles. In a stationary state, the predicted coordinates are forcibly locked. This strategy applies only to ground vehicles; it is not applied to air vehicles.

[0129] The specific steps of the principal component analysis are as follows: calculate the three principal component eigenvalues ​​of the vehicle trajectory prediction coordinates within the sliding window, and determine whether the ground vehicle is stationary based on the total variance threshold, the point cloud diffusion threshold, and the vehicle stationary continuous time; if it is determined to be stationary, then perform stationary locking coordinate correction on the vehicle positioning coordinates (X, Y, Z).

[0130] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0131] Detailed description of the attached diagram:

[0132] like Figure 1 As shown, this invention discloses a ground-air cooperative intelligent vehicle positioning system architecture for low-altitude economic scenarios, whose execution logic is divided into five core stages:

[0133] 1) Preprocessing Layer: The system acquires raw GPS positioning coordinates, Channel State Information (CSI), and RGB images in real time. For ground vehicles (Cars), a linear Kalman filter (KF) is used to smooth high-frequency random noise; for air vehicles (Air Taxi, ATs) with highly dynamic nonlinear motion characteristics of six degrees of freedom (6DoF), an unscented Kalman filter (UKF) is used for state estimation. Simultaneously, a temporal completion network based on a self-attention mechanism is used to reconstruct missing data from the actual perceived data.

[0134] 2) Feature Extraction Layer: This layer extracts environmental semantics using a visual backbone network and combines it with spatial perspective encoding for multi-view fusion. A progressive gating fusion mechanism (Gate) is used to assign learnable weights to the environmental representations. Inject wireless signal feature extraction model to generate CSI features with enhanced environmental semantics.

[0135] 3) Spatial Computation Layer: Based on the filtered GPS pose and Road Side Facility (RSF) coordinates, a 12-dimensional explicit spatial physical feature encoding, including 2D horizontal distance, height difference, orientation alignment, and co-route confidence, is calculated. This encoding serves as a quality assessment index, and the channel state representations at multiple locations are dynamically weighted and fused through a CSI weighted fusion module.

[0136] 4) Regression prediction layer: A residual position regression network is used, with the spliced ​​fused features as input, to predict the pose correction between GPS observations and the actual trajectory.

[0137] 5) Post-processing layer: After outputting the initial corrected coordinates, the motion state discrimination based on principal component analysis is introduced for ground vehicles to identify the stationary state range and perform coordinate locking, and output the final high-precision positioning result.

[0138] like Figure 2 As shown in the figure, this diagram illustrates in detail the process of precise alignment and interactive fusion of heterogeneous modal data in geometric space:

[0139] 1) Geometrically Aware Alignment: Using Fourier feature encoding technology, the normalized mounting orientation angle of the camera is aligned. By mapping to a high-dimensional space, a cross-perspective attention mechanism based on geometric perception is constructed. By calculating the affinity matrix between the target perspective and the auxiliary perspective in the joint space, global environmental features from a broader perspective are obtained.

[0140] 2) Progressive injection mechanism: The gate structure will merge weights in the early stages of training. Initialized to 0.01, the network prioritizes convergence to the physical consistency model of the wireless signal, and then the global environment representation is automatically and smoothly injected with each training iteration. Finally, CSI features enhanced with spatial topology information are generated at the decoder.

[0141] like Figure 3 As shown, the position regression prediction module that integrates prior information and kinematic constraints is demonstrated:

[0142] 1) Feature fusion and residual learning: The module concatenates high-dimensional fused CSI features with GPS prior attitude as input to the residual regression network.

[0143] 2) Multi-objective joint loss optimization: The model is subject to the joint loss function. This constraint mechanism ensures that the prediction results strictly conform to the laws of physical motion while improving accuracy.

[0144] like Figure 4 As shown, comparative analysis verified the ability of multidimensional spatial physical feature encoding to represent vehicles with different motion modes. Quantitative comparison showed that the positioning task of air vehicles (AT4) is highly dependent on "elevation angle" (contribution 32.1%) and "altitude difference" (24.8%), while ground vehicles are significantly affected by "lateral offset" and "same-road confidence".

[0145] like Figure 5 As shown, the technical gains of this invention are verified by comparing the GPS prior data before and after filtering with the positioning data predicted by the model:

[0146] 1) Figure 5 (a) is a schematic diagram of the GPS correction results of AT4. The original 3D position error of the air vehicle (AT4) after unscented Kalman filtering is 2.347m. After correction by this system, it is reduced to 0.421m, with a correction rate of up to 82.1%, which effectively compensates for the high dynamic error in the low-altitude three-dimensional space.

[0147] 2) Figure 5 (b) is a schematic diagram of the GPS correction results for Car20. The X-axis error of the ground vehicle (Car20) after correction is reduced to 0.139m, and the comprehensive 3D position error is only 0.152m. The residual regression model, combined with environmental semantic enhancement, can effectively suppress the positioning deviation caused by multipath effects and non-line-of-sight (NLOS) propagation in urban canyon environments.

[0148] like Figure 6 The diagram shows the experimental results of stationary locking designed for ground scenarios: First, the eigenvalues, total variance, and diffusion of the Principal Component Analysis (PCA) decomposition are determined to be below a set threshold. Second, the regions sampled by multiple consecutive sliding windows must all be below the threshold. Finally, the continuous trajectory region is determined to be stationary, and the model's predicted coordinates are locked. This eliminates the coordinate drift problem in stationary vehicle positioning from a physical perspective.

Claims

1. A method for positioning a ground-air cooperative intelligent vehicle, characterized in that, Includes the following steps: The heterogeneous data collected from intelligent connected vehicles and air vehicles are preprocessed. Based on different motion characteristics, the raw GPS data is subjected to vehicle adaptive filtering. The channel state information obtained between the roadside unit and the target vehicle is subjected to time-series filtering and missing masking. At the same time, the temporal environmental semantic features of the multi-view RGB images at the roadside are completed. The wireless channel state features are enhanced by extracting environmental semantic information. Environmental semantic features are extracted from the multi-view RGB images using a visual perception model. A global environmental representation is generated by introducing a cross-view fusion mechanism with geometric priors. The environmental representation is then injected into the wireless signal feature extraction model through a progressive gating method. Based on explicit physical modeling, the wireless signal characteristics of multiple roadside units are evaluated and dynamically weighted and fused. An explicit spatial location feature code is constructed based on the spatial geometric relationship between the vehicle and each roadside unit, which includes distance, direction, relative position relationship and road semantic confidence. The wireless channel state characteristics of multiple roadside units are then dynamically weighted and fused in time series based on the explicit spatial location feature code. The fused wireless channel state features are spliced ​​and fused with the filtered and preprocessed GPS prior pose information. The residual position regression module is used to predict the pose correction of the intelligent vehicle. The absolute position accuracy, motion direction consistency, correction rationality and dynamic constraints are jointly optimized by a multi-objective joint loss function to output the final positioning result of the intelligent vehicle. For ground-based intelligent connected vehicles, the positioning results are used to determine the motion state. When it is determined that the vehicle is stationary, its predicted positioning coordinates are locked. Specifically, for the progressive gating method, a two-layer linear mapping layer is used, with the second layer initialized to zero. Global visual features are then processed by setting a group learning rate. A decoder for a smoothly injected wireless base station model, wherein, To integrate the wireless channel features after incorporating environmental semantics, Features of the original wireless channel state information Linear Representing the linear mapping layer, the overall calculation formula is: ； in The learnable gate parameter (Gate) is set during training initialization. This allows the model to prioritize learning the characteristics of single-mode wireless signals; Among them, the cross-view fusion mechanism of geometric prior uses spatial geometric vectors as indices to dynamically aggregate complementary information from multiple perspectives. This module performs the following calculation process: First, it defines any target view feature to be enhanced as... Define the feature set of the remaining auxiliary perspectives as Next, the affinity matrix between the target viewpoint and the auxiliary viewpoint in the geometric-semantic joint space is calculated. This process measures the degree of matching of features from different viewpoints through scaling dot product operations. The calculation formula is expressed as follows: ； in, This represents the feature projection transformation operation. The scaling factor is the feature dimension. Indicates the first j The auxiliary perspective for the first i Semantic contribution weights for each target perspective; The vehicle adaptive filtering includes: for ground-based intelligent connected vehicles constrained by road topology and whose motion is mainly two-dimensional, a Kalman filter algorithm is used to smooth the collected GPS data; for air vehicles with six degrees of freedom motion characteristics, an unscented Kalman filter algorithm is used to perform nonlinear state estimation on the collected GPS data. The explicit spatial location feature encoding includes at least the horizontal distance, height difference, relative azimuth angle, orientation alignment, lateral deviation, longitudinal deviation, and approach trend between the vehicle and the roadside unit. It is used to characterize the spatial geometric relationship between the vehicle and the roadside unit during the wireless signal propagation process and serves as the basis for quality assessment and weighted fusion of channel state features of multiple roadside units.

2. The method according to claim 1, characterized in that, Preprocessing the channel state information collected by the roadside unit includes: sampling the wireless channel state information through a sliding time window, introducing mask identifiers for wireless channel data with time sequence missing data, and reconstructing the channel characteristics of the missing time moments by combining a time sequence completion network. Temporal contextual semantic feature completion of multi-view RGB images of roadside units includes: using a temporal modeling network with a self-attention mechanism, introducing a learnable mask vector to fill in missing frames, and combining position encoding to preserve the temporal contextual relationship of image features; Specifically, a learnable mask vector is used for the location of missing frames. The process involves padding and using sine and cosine position coding to maintain temporal alignment. This process utilizes the long-term dependencies of context-valid frames to reconstruct continuous environmental features.

3. The method according to claim 1, characterized in that, The multi-objective joint loss function includes at least the absolute position error loss, motion direction consistency loss, GPS correction rationality constraint loss, and vehicle dynamics constraint loss.

4. The method according to claim 1, characterized in that, The motion state determination of ground-based intelligent connected vehicles includes: performing principal component analysis on the vehicle's positioning coordinates within a predetermined time window, determining whether the vehicle is in a stationary or moving state based on the variance distribution of the trajectory point cloud, and locking the vehicle's positioning coordinates when the motion mode is determined to be stationary.

5. A ground-air cooperative intelligent vehicle positioning system, based on the method described in any one of claims 1-4, characterized in that, include: The data preprocessing module is used to perform vehicle adaptive filtering on GPS data of ground intelligent connected vehicles and air vehicles, and to perform temporal filtering, masking and missing completion on wireless channel status information and multi-view RGB images collected by roadside units. The feature enhancement and extraction module is used to extract environmental semantic features from the multi-view RGB image and generate a global environmental representation based on a cross-view fusion mechanism with geometric prior. The environmental representation is then injected into the wireless signal feature extraction model through a progressive gating method to obtain high-dimensional channel state features that fuse environmental semantics. The wireless signal quality assessment and feature fusion module is used to construct explicit spatial location feature codes based on the spatial geometric relationship between the vehicle and each roadside unit, and to perform time-series dynamic weighted fusion of the wireless channel state features of multiple roadside units based on the coding module. The location regression prediction module is used to concatenate and fuse the weighted fused wireless channel state features with the filtered GPS prior pose information, and predict the positioning results of the intelligent vehicle based on the residual location regression network constrained by the GPS prior pose and the multi-objective joint loss function. The motion state discrimination and post-processing module is used to discern the motion state of the positioning results of the ground intelligent connected vehicle and lock the positioning coordinates when the vehicle is determined to be stationary.

6. The system according to claim 5, characterized in that: The data preprocessing module uses the Kalman filter algorithm for ground-based intelligent connected vehicles and the unscented Kalman filter algorithm for air vehicles to adapt to the noise characteristics and positioning requirements under different degrees of freedom of motion. The data preprocessing module includes a channel state information filtering unit based on a sliding time window, and reconstructs the missing channel state features through a masking mechanism and a timing completion network; The feature enhancement and extraction module includes a temporal completion sub-module based on a self-attention mechanism, which is used to reconstruct the temporal consistency of environmental semantic features of multi-view RGB images at the road end. The cross-view fusion mechanism constructs geometric feature vectors based on the installation position and orientation parameters of the roadside camera, and uses the geometric feature vectors to guide the complementary fusion of cross-attention mechanisms of environmental semantic features from different viewpoints. The progressive gating method sets learnable fusion weights for global environmental semantic features, initializes them to zero at the start of model training to avoid their influence on wireless signal features, and gradually increases their fusion ratio by learning to utilize different modal features during training. The explicit spatial location feature coding network constructed by the wireless signal quality assessment and feature fusion module includes at least the horizontal distance, height difference, relative azimuth, orientation alignment, lateral deviation, approach trend, and co-road confidence between the vehicle and the roadside unit. The position regression prediction module uses a multi-objective joint loss function that includes absolute position error constraints, motion direction consistency constraints, GPS correction magnitude constraints, and dynamic constraints to train the GPS prior-based residual position regression network.

7. A positioning device for a ground-air cooperative intelligent vehicle, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is run on the processor, it causes the processor to execute the ground-air cooperative intelligent vehicle positioning method according to any one of claims 1-4.

Images