A high-precision positioning data processing method

Abstract

This invention discloses a high-precision positioning data processing method, belonging to the field of BeiDou satellite navigation and intelligent connected vehicle positioning technology. Addressing the problems of drastic drops in positioning accuracy due to non-line-of-sight interference with satellite signals in complex scenarios, and divergent fusion errors caused by the lack of hardware synchronization benchmarks for low-cost multi-sensor systems, this invention utilizes a spatiotemporal graph neural network to adaptively identify non-line-of-sight signals from BeiDou satellite observation data. It combines visual odometry geometric constraints to reconstruct the reflection path of contaminated pseudoranges, recovering the true geometric pseudorange from the data source. Furthermore, it employs spline curves to continuously parameterize the vehicle's pose trajectory, projecting asynchronous observation data from the global navigation satellite system, inertial measurement unit, and visual sensors onto the same continuous time benchmark. This constructs a continuous time factor graph and jointly optimizes the trajectory state of external participants, achieving high-precision, highly reliable continuous positioning without the need for hardware synchronization.

Description

Technical Field

[0001] This invention relates to the field of high-precision positioning data processing technology, specifically to a high-precision positioning data processing method that combines adaptive recognition of satellite navigation non-line-of-sight signals with optimization of continuous time factor maps from multi-source sensors. It can be applied to scenarios with high requirements for high-precision, high-resolution, and high-reliability continuous positioning, such as BeiDou satellite navigation, intelligent connected vehicles, low-altitude aircraft, and smart cities. Background Technology

[0002] With the rapid development of autonomous driving, low-altitude aircraft, and smart cities, high-precision continuous positioning has become an indispensable core technology in related fields. The BeiDou Navigation Satellite System (BDS) obtains the receiver's three-dimensional position by measuring observations such as pseudorange and carrier phase through receivers and combining them with satellite ephemeris calculations. To improve positioning accuracy and continuity, the industry generally adopts a multi-source fusion technology approach that integrates the Global Navigation Satellite System (GNSS), inertial measurement units (INS), and visual sensors. The GNSS provides an absolute position reference, the INS provides high-frequency attitude and acceleration information to compensate for the insufficient update rate of the GNSS, and the visual sensor provides relative pose constraints through visual odometry or visual simultaneous positioning and mapping (SSM) technology to suppress cumulative drift in inertial navigation. The fusion framework typically uses extended Kalman filtering or graph optimization to estimate the state of the multi-source observations. The aforementioned fusion positioning technology has been widely applied in downstream scenarios such as advanced autonomous driving, low-altitude aircraft navigation, and intelligent transportation, which has placed continuous demands on improving the accuracy, continuity, and reliability of positioning systems.

[0003] Currently, existing technical solutions that are similar to this invention mainly fall into two categories. The first category is satellite positioning methods based on anti-multipath filtering. After the receiver acquires the original pseudorange and carrier-to-noise ratio (CNR) observations, this type of method weights the observations according to the CNR or elevation angle, assigning lower weights or even directly eliminating suspected multipath or non-line-of-sight (NLS) signals. The remaining observations are then substituted into least squares or Kalman filtering for position calculation. Because this type of method adopts a weighting or elimination strategy for NLS signals, the number of visible satellites available for positioning calculation will be greatly reduced in strong obstruction scenarios such as urban canyons and multi-level interchanges. The accuracy attenuation factor corresponding to the satellite geometry will increase significantly, ultimately causing the positioning accuracy to plummet from the centimeter level to the tens of meters level, making it difficult to meet the continuity requirements of high-precision positioning. Moreover, this type of method usually only performs static weighting judgment on observations of a single epoch, failing to fully utilize the correlation information of satellite observation data in the spatial and temporal dimensions, and has limited ability to identify and utilize NLS signals.

[0004] The second category is multi-sensor fusion methods based on hardware synchronization triggering. In this type of method, the global navigation satellite system receiver, inertial measurement unit, and camera are timestamped by a unified second pulse hardware trigger signal. After hardware synchronization, the data from each sensor are input to a tightly coupled Kalman filter or discrete time factor graph at a fixed timestamp for joint state estimation. The system outputs filtered and fused position, velocity, and attitude information. However, low-cost mass-produced sensor arrays generally do not have a unified high-precision hardware synchronization trigger pulse. The crystal oscillators of the global navigation satellite system, inertial measurement unit, and camera are different and have dynamic clock drift that changes over time. The timestamp misalignment, coupled with the dynamic changes in spatial extrinsic parameters, introduces systematic time and space dual errors into the data association process of the conventional discrete-time Kalman filter framework. This causes the errors to accumulate and diverge rapidly over time. Especially in scenarios with long-term satellite signal obstruction, such as long tunnels, inertial navigation drift is difficult to suppress effectively. In addition, this type of method relies on additional hardware synchronization circuits and high-precision sensor combinations, which also increases the hardware cost of the system to some extent, making it unsuitable for large-scale deployment in mass-produced vehicles or low-cost platforms.

[0005] In summary, existing technologies are insufficient in terms of satellite signal anti-interference capabilities under complex obstruction scenarios, and also have significant shortcomings in terms of the time reference unification accuracy of low-cost multi-sensor fusion. There is an urgent need for a high-precision positioning data processing method that can eliminate non-line-of-sight and multipath errors from the data source, and achieve high-precision soft synchronization of multi-source heterogeneous data and online self-calibration of space extrinsic parameters without relying on hardware synchronization. Summary of the Invention

[0006] The technical problem this invention aims to solve is that in complex scenarios such as urban canyons, multi-level interchanges, and long tunnels, existing high-precision positioning solutions suffer from a sharp drop in positioning accuracy due to non-line-of-sight reflections and multipath interference of satellite signals. Furthermore, low-cost multi-sensor systems lack a unified hardware clock synchronization reference, resulting in dynamic clock drift and rapid accumulation and divergence of fusion positioning errors. Therefore, this invention provides a high-precision positioning data processing method that does not rely on high-precision hardware synchronization and expensive sensor combinations, can eliminate non-line-of-sight and multipath errors at the data source, and achieves high-precision soft synchronization of multi-source heterogeneous data and online self-calibration of spatial extrinsic parameters.

[0007] To address the aforementioned technical problems, this invention provides a high-precision positioning data processing method, comprising the following:

[0008] Acquire raw pseudorange observation data, carrier-to-noise ratio observation data, satellite elevation angle observation data and azimuth angle observation data collected by the BeiDou satellite navigation system, triaxial acceleration data and triaxial angular velocity data collected by the inertial measurement unit, and continuous image frame sequences collected by the visual sensor;

[0009] Graph nodes are constructed based on the observation data of each satellite within the same observation epoch. Connecting edges are established between the corresponding graph nodes based on at least one of the following conditions: the spatial geometric angle between satellites and the correlation of historical observations. A spatiotemporal graph is constructed by combining the graph data corresponding to multiple consecutive observation epochs.

[0010] A graph convolutional network is used to perform spatial neighborhood aggregation on the features of each graph node in the spatiotemporal graph to obtain node features that fuse spatial topological relationships. Then, a temporal coding network is used to encode the node features corresponding to multiple consecutive observation epochs in the time dimension to obtain node features that fuse temporal evolution information. Based on the node features, the classification probability corresponding to each satellite observation value is output, and the signal type of each satellite observation value is determined as either line-of-sight signal type or non-line-of-sight signal type.

[0011] For observations of non-line-of-sight signal types, feature point extraction and inter-frame matching are performed by combining continuous image frame sequences acquired by visual sensors to obtain the relative pose change of the carrier between adjacent observation epochs. Then, the local geometric contour of the occlusion boundary around the carrier is constructed by combining the initial pose of the carrier. The geometric relationship of the reflection path is calculated by combining the elevation angle and azimuth angle of the corresponding satellite to obtain the corrected pseudorange observations. Together with the original pseudorange observations of line-of-sight signal types, they form the corrected observation set.

[0012] Spline basis functions are used to continuously parameterize the carrier's pose trajectory. Data from the BeiDou Navigation Satellite System, Inertial Measurement Unit (IMU) data, and Visual Sensor data are interpolated and mapped to a unified continuous time reference based on their respective local timestamps. Spline curve control points are used as trajectory variable nodes to be optimized. Satellite pseudorange factor, inertial pre-integration factor, and visual reprojection factor are constructed by combining the corrected observation set, IMU observation data, and continuous image frame sequences. Extrinsic parameter variable nodes and trajectory variable nodes are introduced together as state variables to be optimized. A joint cost function is constructed based on the observation residuals and observation noise covariance of each factor, and joint optimization is performed to obtain the optimized trajectory control points and extrinsic parameter calibration results. Then, the position, velocity, and attitude estimation results of the carrier are obtained by interpolation.

[0013] The present invention has the following beneficial effects:

[0014] 1. Existing satellite positioning methods based on anti-multipath filtering generally consider pseudorange observations contaminated by non-line-of-sight or multipath to be unusable erroneous data, and usually adopt processing strategies such as downweighting or direct rejection. This invention breaks through the above-mentioned technical inertia perception, and jointly models the spatial topological relationship and temporal evolution characteristics of BeiDou satellite observation data through a spatiotemporal graph neural network to achieve adaptive recognition of non-line-of-sight signals. Combined with the dynamic geometric constraints of visual odometry, the reflection path of contaminated pseudorange is reconstructed. Without rejecting satellite observations, multipath and non-line-of-sight errors are eliminated from the data source. This allows the system to maintain a large number of usable satellites and a good geometric configuration even in strong occlusion scenarios such as urban canyons. It significantly improves the accuracy and continuity of positioning results in complex scenarios, achieving the technical effect of high-precision and high-resolution positioning. Moreover, the entire reconstruction process does not rely on pre-built high-precision map data, reducing the system deployment and maintenance costs.

[0015] 2. Existing multi-sensor fusion methods based on hardware synchronization generally assume that the unification of time references for multiple sensor sources must rely on a unified high-precision hardware synchronization trigger pulse. This invention breaks through the above-mentioned conventional understanding by using spline curves to express the carrier's pose trajectory in a continuous time parameterized manner. It projects asynchronous and multi-frequency multi-source observation data onto a continuous time reference to achieve soft synchronization and performs online joint self-calibration of spatial extrinsic parameters. It achieves high-precision time alignment without relying on hardware synchronization trigger pulses, overcomes the dependence of existing hardware synchronization schemes on dedicated synchronization circuits, effectively suppresses inertial navigation drift divergence in long-term satellite signal blockage scenarios, reduces system hardware costs, improves the reliability and continuity of positioning results, and achieves the technical effect of high-reliability positioning.

[0016] 3. This invention utilizes the local support properties of spline basis functions to make the Hessian matrix corresponding to the joint cost function present a sparse structure, and combines a sliding time window edge-out mechanism to process historical states. Under the premise of ensuring optimization accuracy, it effectively controls the computational complexity of the system, meets the requirements of high-order intelligent driving and low-altitude aircraft for the real-time performance and mass production cost of the positioning system, and has good engineering feasibility and promotion and application value. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the overall architecture of the high-resolution positioning system of the present invention.

[0018] Figure 2 This is a flowchart of the spatiotemporal geometric graph construction and graph attention network recognition process of the present invention;

[0019] Figure 3 This is a flowchart of the dynamic credibility allocation and factor graph optimization process of this invention. Detailed Implementation

[0020] The following is in conjunction with the appendix Figure 1-3 The following specific embodiments provide a detailed description of the high-precision positioning data processing method provided by the present invention. The embodiments are only used to illustrate the technical solution of the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0021] I. System Overall Architecture

[0022] The high-precision positioning data processing method provided by this invention can be implemented by a high-precision positioning data processing device. The device consists of a data acquisition module, a non-line-of-sight signal recognition and geometric error dynamic reconstruction module, and a continuous time factor map optimization module. The modules are connected sequentially through a data bus to form a complete data processing link from the input of raw observation data to the output of the final positioning result.

[0023] The data acquisition module includes a BeiDou satellite navigation system receiver unit, an inertial measurement unit, and a visual sensor unit. The BeiDou satellite navigation system receiver unit is used to acquire raw pseudorange, carrier phase, carrier-to-noise ratio, and satellite ephemeris observation data. The inertial measurement unit is a microelectromechanical system (MEMS) inertial measurement unit, used to acquire three-axis acceleration data and three-axis angular velocity data. The visual sensor unit can be a monocular camera or a binocular camera, used to acquire continuous image frame sequences. The above data acquisition modules can be integrated into vehicle-mounted domain controllers, airborne computing units of low-altitude aircraft, or other embedded platforms with data acquisition and computing capabilities. There is no requirement to configure a unified high-precision hardware synchronization trigger circuit between the sensors, thereby reducing the hardware complexity and deployment cost of the system.

[0024] The non-line-of-sight signal recognition and geometric error dynamic reconstruction module receives the original BeiDou observation data and the relative pose constraints output by the visual odometry, and outputs the corrected pseudorange observations. The continuous time factor map optimization module receives the corrected pseudorange, the original measurement data of the inertial measurement unit, and the visual feature observation data. It unifies the time reference of each sensor by parameterizing the continuous time trajectory of the spline curve, constructs and solves the factor map, and outputs high-precision, high-frequency continuous position, velocity, and attitude estimation results.

[0025] II. Adaptive Recognition and Dynamic Reconstruction Method for Non-Line-of-Sight Signals based on Geometric Errors

[0026] The core idea of ​​this method is to construct a dynamic spatiotemporal map from the BeiDou satellite observation data sequence, use a spatiotemporal map neural network to extract the spatial topological relationships and temporal evolution characteristics between satellite observations, achieve adaptive recognition of non-line-of-sight signals, and then combine the dynamic geometric constraints provided by visual odometry to reconstruct the geometric reflection path of pseudoranges identified as non-line-of-sight or multipath contamination, and restore their corresponding true geometric pseudoranges.

[0027] The receiver collects pseudorange, carrier-to-noise ratio, satellite elevation angle, and azimuth angle observation features of each visible satellite at each observation epoch. The observation feature vectors of each satellite within the same epoch are used as graph nodes. When the spatial geometric angle between two satellites is less than a preset angle threshold, or when the historical correlation between the observation values ​​of two satellites is greater than a preset correlation threshold, a connection edge is established between the corresponding graph nodes. At the same time, the graph sequence of multiple consecutive epochs is retained to construct a spatiotemporal graph. The preset angle threshold and preset correlation threshold can be preset and adjusted according to the specific application scenario of the receiver and the distribution characteristics of the satellite constellation.

[0028] After the spatiotemporal graph is constructed, for the graph corresponding to each epoch in the spatiotemporal graph, a graph convolutional network is used to aggregate the neighborhood information of the features of each graph node, resulting in a node feature representation that incorporates spatial topological relationships. The specific aggregation method is shown in the following formula:

[0029] ;

[0030] In the formula, Represents graph nodes In the Feature vectors in a layered network Represents graph nodes The set of adjacent nodes, Represents graph nodes With graph nodes The attention convergence coefficient between satellites is used to characterize the similarity of their spatial geometric relationship. Its value is calculated based on the spatial geometric angle between the satellites and the correlation with historical observations. Indicates the first Learnable weight matrices in layered networks This represents a non-linear activation function used to perform non-linear transformations on aggregated features. Graph convolutional networks can have multiple network layers, which aggregate neighborhood information of node features layer by layer, thereby expanding the spatial neighborhood range that a single node feature can perceive.

[0031] Furthermore, the spatial feature sequences of nodes corresponding to multiple consecutive epochs are input into the gated recurrent unit and encoded along the time dimension to obtain node feature representations that integrate historical temporal evolution information. This is used to capture the dynamic changes in satellite signal obstruction or reflection. The spatiotemporal graph neural network also includes a graph construction unit and a feature fusion layer. The graph construction unit is used to construct graph nodes from the observation feature vectors of different satellites within the same epoch and establish connecting edges. The feature fusion layer is used to concatenate and fuse the spatial features output by the spatial graph convolutional layer with the temporal features output by the gated recurrent unit to obtain spatiotemporal fused features.

[0032] The spatiotemporal fusion features are input into the fully connected classification layer, which outputs the classification probability corresponding to the current observation value of each satellite. This classification probability is then compared with a preset probability threshold. When the classification probability is greater than the preset probability threshold, the signal type of the corresponding satellite observation value is determined to be a non-line-of-sight signal type. When the classification probability is not greater than the preset probability threshold, the signal type of the corresponding satellite observation value is determined to be a line-of-sight signal type, and its original pseudorange observation value is directly retained. The aforementioned spatiotemporal graph neural network can be pre-trained offline using historical observation data with line-of-sight and non-line-of-sight labels. During training, the cross-entropy between the classification probability and the true label is used as the loss function to iteratively optimize the learnable parameters in the graph convolutional network, gated recurrent unit, and fully connected classification layer. After the network parameters are solidified, they are deployed on a vehicle-mounted or airborne computing platform for online real-time inference.

[0033] For satellite observations identified as non-line-of-sight signals, continuous image frames acquired by a visual sensor are used to estimate the relative pose change of the carrier between adjacent epochs through feature point extraction and inter-frame matching. Combined with the known initial pose of the carrier and kinematic constraints, the local geometric contour of the occlusion boundary of the surrounding buildings is constructed in real time, forming dynamic geometric constraints that do not rely on prior high-precision maps. Feature point extraction and inter-frame matching can be implemented using corner feature extraction algorithms with rotation invariance and scale invariance. The pose change between frames can be obtained by solving the essential matrix or homography matrix of the matched feature point pairs.

[0034] Combining the aforementioned local geometric contours with the corresponding satellite's azimuth and elevation angle information, the geometric relationship of the possible reflection path of the satellite signal is inverted, and the true geometric pseudorange estimate between the satellite and the receiver is reconstructed. The reconstructed relationship can be expressed as:

[0035] ;

[0036] In the formula, Indicates the reconstructed satellite Corrected pseudorange, This indicates the satellite data originally measured by the receiver. pseudorange observations, This represents the additional path error correction amount calculated from the reflection path geometry. Its value varies with the satellite elevation angle, azimuth angle, and local geometric constraints. Indicates satellite altitude angle, Indicates satellite azimuth angle, This represents the set of local geometric contour constraints constructed in real time as described above.

[0037] The original pseudorange observations corresponding to the line-of-sight signal type and the corrected pseudorange observations corresponding to the non-line-of-sight signal type are combined to form a complete set of corrected observations, which are then output to the continuous time factor map optimization module for subsequent fusion and solution.

[0038] III. Hardware-Free Synchronous Multi-Source Fusion Method Based on Continuous Time Factor Graph Optimization

[0039] This method uses spline curves to express the trajectory of the carrier motion in a continuous time parameterized manner. It projects BeiDou satellite navigation system data, inertial measurement unit data and visual observation data with different sampling frequencies and asynchronous timestamps onto the same continuous time reference, constructs a factor graph containing multiple observation factors and performs joint optimization solution, and simultaneously calibrates the spatial extrinsic parameters between various sensors online.

[0040] Specifically, a set of control points distributed along the time axis is selected, and... The order spline basis functions provide a continuous parameterized representation of the carrier pose trajectory at any time. The pose can be represented as:

[0041] ;

[0042] In the formula, express The pose state of the carrier at all times. This indicates that the total number of control points has decreased by one. Indicates the first indivual The order spline basis functions take non-zero values ​​only within their local support intervals. Indicates the first The pose state of the control point to be optimized is such that since each spline basis function only takes a non-zero value in its corresponding local support interval, the adjustment of a certain control point only affects the trajectory shape in its local neighborhood and will not significantly disturb the global trajectory. This characteristic also provides a basis for subsequent sparsification solution.

[0043] The raw observation data collected by the BeiDou satellite navigation system, inertial measurement unit and visual sensor are mapped to a unified continuous time reference based on their respective local timestamps through the interpolation relationship of the spline basis functions mentioned above. This achieves the alignment of multi-source data with different sampling frequencies and different timestamp references on the same continuous time axis without relying on hardware synchronization trigger pulses, that is, soft synchronization.

[0044] Using spline curve control points as the state variable nodes to be optimized, satellite pseudorange factor, inertial pre-integration factor, and visual reprojection factor are constructed based on the observation data after time alignment. The satellite pseudorange factor connects the trajectory variable nodes at the corresponding time, representing the corrected pseudorange observation constraint. The inertial pre-integration factor connects two adjacent trajectory variable nodes, representing the inertial measurement constraint obtained by integrating the original measurement data of the inertial measurement unit. The visual reprojection factor connects the trajectory variable nodes at the corresponding time with the map feature point variable nodes, representing the visual observation constraint composed of the image feature point reprojection error. At the same time, extrinsic parameter variable nodes representing the spatial position and attitude relationship between each sensor are introduced and connected to the factor nodes corresponding to the satellite pseudorange factor, inertial pre-integration factor, and visual reprojection factor, respectively, and included in the set of variables to be optimized in the factor graph to achieve online joint self-calibration of spatial extrinsic parameters.

[0045] The observation residuals of each factor are weighted according to their corresponding observation noise covariance to construct an overall joint cost function. This cost function is then jointly optimized using a nonlinear least squares method for all state variables. The cost function is expressed as:

[0046] ;

[0047] In the formula, This represents the set of states to be optimized, consisting of spline curve control points and extrinsic parameters. This represents the state estimation result obtained from the optimization solution. Indicates the first Each factor is relative to the state The observation residuals Indicates the first The covariance matrix of the observation noise for each factor is used to weight the residuals. It represents the weighted square norm of the residuals calculated using the inverse of the covariance matrix as the metric matrix.

[0048] Considering the local support characteristics of spline basis functions, the Hessian matrix corresponding to the joint cost function exhibits a highly sparse structure. In this embodiment, a sparse matrix decomposition method is used to efficiently solve the optimization problem. At the same time, a sliding time window is set to marginalize historical control points outside the window, retaining their prior constraint information on the state within the window. This controls the system's computational complexity and maintains real-time performance over long periods, enabling the system to meet the computing power constraints of vehicle-mounted or airborne embedded platforms.

[0049] The spline curve control points obtained from the optimization solution at the current moment are substituted into the continuous trajectory expression, and the estimated results of the carrier's position, velocity and attitude at any desired moment are obtained by interpolation. These results are then used as the final high-precision positioning data output for use by downstream navigation, path planning and control systems.

[0050] IV. Technical Effect Description

[0051] Through the above processing, the method provided in this embodiment can eliminate multipath and non-line-of-sight errors from the data source in complex scenarios such as urban canyons, multi-level interchanges, and long tunnels without removing satellite observations. This allows the system to maintain a large number of available satellites and a good geometric configuration even in scenarios with strong obstruction. At the same time, it achieves high-precision soft synchronization of multi-source heterogeneous data and online self-calibration of spatial extrinsic parameters without the need for hardware synchronization trigger pulses. It effectively suppresses inertial navigation drift divergence in scenarios with long-term satellite signal obstruction, achieving the technical effect of high-precision, high-resolution, and high-reliability continuous positioning. It is suitable for application scenarios such as intelligent connected vehicles, low-altitude aircraft, and smart cities. Moreover, the system does not require additional high-precision hardware synchronization circuits, reducing system hardware costs.

[0052] V. Alternative Examples

[0053] In another embodiment, the non-line-of-sight signal recognition stage may not use a spatiotemporal graph neural network architecture, but instead adopt a temporal attention network based on a self-attention mechanism. This network directly performs self-attention modeling on the observation feature sequences of multiple epochs and multiple satellites to capture the spatiotemporal correlation of satellite signals and output the non-line-of-sight signal determination probability. This can also achieve the adaptive recognition effect of non-line-of-sight signals without relying on prior high-precision maps, and can be used as an alternative implementation of the spatiotemporal graph neural network module.

[0054] In another embodiment, the multi-source data time reference unification step can also avoid using spline curves for continuous time parameterization. Instead, Gaussian process regression can be used to model the carrier motion trajectory in continuous time. The kernel function of the Gaussian process can be used to interpolate and estimate the uncertainty of asynchronous observation data. This can also achieve the effect of multi-source data soft synchronization without relying on hardware synchronization trigger pulses, and can be used as an alternative to the continuous time trajectory expression method.

[0055] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A high-precision positioning data processing method, characterized in that, include: Acquire raw pseudorange observation data, carrier-to-noise ratio observation data, satellite elevation angle observation data and azimuth angle observation data collected by the BeiDou satellite navigation system, triaxial acceleration data and triaxial angular velocity data collected by the inertial measurement unit, and continuous image frame sequences collected by the visual sensor; Graph nodes are constructed based on the observation data of each satellite within the same observation epoch. Connecting edges are established between the corresponding graph nodes based on at least one of the following conditions: the spatial geometric angle between satellites and the correlation of historical observations. A spatiotemporal graph is constructed by combining the graph data corresponding to multiple consecutive observation epochs. The features of each graph node in the spatiotemporal graph are aggregated using a graph convolutional network to obtain node features that integrate spatial topological relationships. The node features corresponding to multiple consecutive observation epochs are encoded in the time dimension using a temporal coding network to obtain node features that fuse temporal evolution information; Based on the node characteristics of the fused temporal evolution information, the classification probability corresponding to each satellite observation value is output, and the signal type of each satellite observation value is determined to be either a line-of-sight signal type or a non-line-of-sight signal type based on the classification probability. For observations whose signal type is determined to be a non-line-of-sight signal type, feature point extraction and inter-frame matching are performed based on the continuous image frame sequence to obtain the relative pose change of the carrier between adjacent observation epochs. The local geometric contour of the occlusion boundary around the carrier is constructed by combining the initial pose of the carrier. Based on the local geometric profile and the elevation and azimuth angles of the corresponding satellite, the geometric relationship of the reflection path of the corresponding satellite signal is calculated to obtain the corrected pseudorange observation value. The original pseudorange observations, whose signal type is determined to be a line-of-sight signal, and the corrected pseudorange observations constitute the corrected observation set. The pose trajectory of the carrier is expressed in continuous time parameterization using spline basis functions. Based on the local timestamps of the observation data of the BeiDou satellite navigation system, the observation data of the inertial measurement unit, and the observation data of the visual sensor, the observation data of the BeiDou satellite navigation system, the observation data of the inertial measurement unit, and the observation data of the visual sensor are interpolated and mapped to a unified continuous time reference. Using spline curve control points as trajectory variable nodes to be optimized, a satellite pseudorange factor is constructed based on the corrected observation set after time reference unification, an inertial pre-integration factor is constructed based on the observation data of the inertial measurement unit, a visual reprojection factor is constructed based on the continuous image frame sequence, and extrinsic variable nodes representing the spatial position and attitude relationship between sensors are introduced. The extrinsic variable nodes and the trajectory variable nodes are used together as state variables to be optimized. A joint cost function is constructed based on the observation residuals and corresponding observation noise covariance of the satellite pseudorange factor, the inertial pre-integration factor, and the visual reprojection factor, and the joint optimization solution is performed on the state variables to be optimized to obtain the optimized trajectory control points and external parameter calibration results. Interpolation is performed based on the optimized trajectory control points to obtain the position, velocity, and attitude estimation results of the carrier.

2. The method according to claim 1, characterized in that, The method of using a graph convolutional network to perform spatial neighborhood aggregation on the features of each graph node in the spatiotemporal graph includes: For graph node i in the spatiotemporal graph, the features of each neighboring node in the set of neighboring nodes of graph node i and the attention aggregation coefficient between graph node i and each neighboring node are weighted and summed, and the features of graph node i in the next network layer are obtained through a nonlinear activation function. The attention aggregation coefficient is used to characterize the similarity of the spatial geometric relationship between graph node i and its corresponding neighboring nodes. The temporal coding network is a gated recurrent unit used to encode the spatial feature sequences of nodes corresponding to multiple consecutive observation epochs along the time dimension.

3. The method according to claim 1, characterized in that, The construction of the local geometric contour of the occlusion boundary around the carrier based on the initial pose of the carrier includes: The local geometric contour is constructed in real time based on the initial pose of the carrier, kinematic constraints, and the relative pose change. The construction process of the local geometric contour does not rely on pre-constructed high-precision map data.

4. The method according to claim 1, characterized in that, The spline basis function is a d-order spline basis function. The trajectory variable nodes are distributed along the time axis. Each d-order spline basis function takes a non-zero value only in the corresponding local support interval. The carrier pose at any time is obtained by weighted summation of the spline basis function that takes a non-zero value in the local support interval and the corresponding trajectory variable node, where d is the order of the spline basis function.

5. The method according to claim 1, characterized in that, The extrinsic variable nodes are connected to the factor nodes corresponding to the satellite pseudorange factor, the inertial pre-integration factor, and the visual reprojection factor, respectively. During the joint optimization solution of the state variables to be optimized, the online calibration results of the spatial extrinsic parameters between sensors are obtained simultaneously.

6. The method according to claim 1, characterized in that, Also includes: The sparse structure of the Hessian matrix corresponding to the joint cost function is determined based on the local support characteristics of the spline basis function, and the sparse matrix decomposition method is used to solve the state variables to be optimized. A sliding time window is set, and historical trajectory variable nodes outside the sliding time window are marginalized. The prior constraint information obtained from the marginalization is used to solve the state variables within the sliding time window.

7. The method according to claim 1, characterized in that, The step of determining the signal type of each satellite observation value as either a line-of-sight signal type or a non-line-of-sight signal type based on the classification probability includes: The classification probability is compared with a preset probability threshold. When the classification probability is greater than the preset probability threshold, the signal type of the corresponding satellite observation is determined to be a non-line-of-sight signal type. When the classification probability is not greater than the preset probability threshold, the signal type of the corresponding satellite observation value is determined to be the line-of-sight signal type.

8. A high-precision positioning data processing device, characterized in that, It includes a data acquisition module, a non-line-of-sight signal recognition and geometric error reconstruction module, and a continuous time factor graph optimization module; The data acquisition module is used to acquire raw pseudorange observation data, carrier-to-noise ratio observation data, satellite elevation angle observation data and azimuth angle observation data collected by the BeiDou satellite navigation system, triaxial acceleration data and triaxial angular velocity data collected by the inertial measurement unit, and continuous image frame sequences collected by the visual sensor. The non-line-of-sight signal recognition and geometric error reconstruction module is used to perform the step of obtaining the corrected set of observations based on the original pseudorange observation data in the method of any one of claims 1 to 7; The continuous time factor graph optimization module is used to perform the step of obtaining the position, velocity and attitude estimation results of the carrier based on the modified observation set in the method of any one of claims 1 to 7.

9. The apparatus according to claim 8, characterized in that, It also includes a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images