Vehicle-mounted high-precision positioning method and system based on combination of beidou and monocular fusion

Abstract

The application discloses a network distribution vehicle-mounted high-precision positioning method and system based on Beidou and monocular fusion, relates to the technical field of positioning and navigation, and comprises the following steps: obtaining the relative displacement vector of a vehicle in adjacent frames based on a time domain differential carrier phase model; virtually converting adjacent two monocular images into a dynamic binocular stereo vision model based on the relative displacement vector, and obtaining absolute depth information of image feature points; performing catenary fitting on power lines in the image, obtaining physical geometric features of a plane where the power lines are located, and establishing a physical gravity field geometric constraint; obtaining a constraint factor based on the physical gravity field geometric constraint, the time domain differential carrier phase model displacement constraint and the absolute depth information, constructing a tightly coupled factor graph model for optimizing the motion state of the vehicle, solving the tightly coupled factor graph model, and obtaining a fusion solution containing the position, speed and attitude of the vehicle. The application realizes high-precision and high-robustness positioning and attitude determination, and improves the network distribution inspection operation efficiency.

Description

Technical Field

[0001] This application relates to the field of positioning and navigation technology, specifically to a high-precision positioning method and system for distribution network vehicles based on BeiDou and monocular fusion. Background Technology

[0002] With the advancement of smart grid construction, the demand for automated inspection of distribution networks is increasing. Distribution network lines are typically distributed in urban streets, rural roads, and forest areas, creating extremely complex environments. Inspection vehicles require high-precision position, speed, and attitude information to accurately locate and map poles, transformers, and line fault points. Currently, inspection vehicles mainly rely on real-time dynamic differential technology. However, the distribution network environment has significant non-open sky characteristics. When passing through tree-lined avenues or densely built-up urban areas, BeiDou / GNSS signals are easily interfered with by non-line-of-sight propagation and multipath effects. This makes it difficult for real-time dynamic differential to fix integer ambiguity, causing positioning accuracy to drop instantly from centimeter-level to meter-level or even tens of meters. Furthermore, the error in the elevation direction is particularly severe, failing to meet the needs of refined inspection. To compensate for the inadequacy of satellite signals, monocular vision positioning has been extensively studied. Compared to expensive lidar or binocular cameras, monocular cameras are extremely low-cost and easy to integrate. However, monocular vision suffers from scale loss, and algorithms cannot directly determine the true physical distance from images. Although existing technologies attempt to recover the scale through inertial measurement units, low-cost MEMS inertial measurement units suffer from severe zero-bias drift, causing the calculated trajectory size to be severely distorted from the real world after the vehicle has traveled a long distance.

[0003] Existing technologies attempt to fuse BeiDou with visual positioning, but these fusion techniques also have various limitations. On the one hand, conventional techniques calculate BeiDou coordinates and visual pose separately and then weight them, performing fusion only at the result layer. If the BeiDou signal quality is poor, erroneous coordinates will directly deviate the visual trajectory, causing fusion failure. On the other hand, they ignore the extremely high relative accuracy of BeiDou carrier phase over short periods. When satellite signals cannot be used to calculate absolute position, carrier phase can still provide accurate velocity or displacement increment information, and this high-value data is often discarded in existing algorithms. At the same time, existing visual SLAM algorithms are generally general-purpose and not optimized for power distribution network scenarios. The most prominent feature in the power distribution network environment is the overhead power lines. Power lines are subject to gravity and exhibit a standard catenary shape, and the poles are strictly perpendicular to the ground. Existing technologies fail to incorporate these strong physical constraints into the positioning equations, ultimately resulting in insufficient continuity of positioning and accuracy of attitude in complex environments.

[0004] In conclusion, improving vehicle-mounted positioning accuracy in complex power distribution network environments, thereby increasing the efficiency of power distribution network inspection operations, is a problem that needs to be addressed. Summary of the Invention

[0005] The purpose of this application is to address the problem that existing technologies lack sufficient positioning continuity and attitude accuracy in complex distribution network environments, resulting in low efficiency of distribution network inspection operations. A high-precision vehicle-mounted positioning method and system for distribution networks based on BeiDou and monocular fusion is proposed. By utilizing the time-domain differential carrier phase model to provide absolute scale constraints and combining the physical geometric characteristics of the distribution network environment for fusion optimization, high-precision and robust positioning and attitude determination are achieved, thereby improving the efficiency of distribution network inspection operations.

[0006] To achieve the above objectives, the technical solutions adopted in the embodiments of this application are as follows: In a first aspect, embodiments of this application provide a high-precision positioning method for distribution network vehicles based on BeiDou and monocular fusion, including: The relative displacement vector of the vehicle in adjacent frames is obtained based on the time-domain differential carrier phase model. Based on the relative displacement vector, the monocular images of two adjacent frames are virtualized into a dynamic binocular stereo vision model to obtain the absolute depth information of the image feature points. The electric power lines in the image are fitted with catenaries to obtain the physical geometric features of the plane where the electric power lines are located in order to establish physical gravity field geometric constraints. Based on the physical gravity field geometric constraints, the displacement constraints of the time-domain differential carrier phase model, and the absolute depth information, constraint factors are obtained. A tightly coupled factor graph model is constructed to optimize the vehicle motion state. The tightly coupled factor graph model is solved to obtain a fusion solution that includes the vehicle position, velocity, and attitude.

[0007] In this scheme, the time-domain differential carrier phase model utilizes the millimeter-level relative displacement accuracy of the BeiDou carrier phase within a short time to virtualize two frames of images before and after vehicle movement into a fixed-baseline binocular image. This provides absolute scale information to the monocular vision system in real-time and dynamically without increasing hardware costs, fundamentally eliminating the scale drift problem in monocular vision odometry. Simultaneously, by utilizing the natural sag curve formed by power lines under gravity in the distribution network environment, the absolute gravity direction is calculated in reverse, providing an absolute physical attitude reference independent of inertial sensors, effectively suppressing attitude errors accumulated during long-term vehicle operation. By acquiring constraint factors and constructing an optimization objective function model through physical gravity field geometric constraints, time-domain differential carrier phase model displacement constraints, and absolute depth information, a deep complementarity of multi-source information is achieved in a tightly coupled manner. Among them, Beidou carrier phase provides a scale reference, vision provides rich environmental features, and power line physical geometry provides an attitude reference. The three mutually verify and complement each other, ultimately outputting high-precision and highly robust positioning results. This improves the system's anti-interference capability and positioning robustness in complex environments, and further significantly improves the efficiency of inspection operations.

[0008] Optionally, before obtaining the relative displacement vector of the vehicle in adjacent frames based on the time-domain differential carrier phase model, the process includes: acquiring a monocular image stream and BeiDou raw radio frequency data based on the vehicle terminal; synchronizing the monocular image stream and BeiDou raw radio frequency data by clock; extracting carrier phase observation values ​​based on the BeiDou raw radio frequency data; using carrier phase observation values ​​with an elevation angle greater than a preset elevation angle threshold and a signal-to-noise ratio greater than a preset signal-to-noise ratio threshold as target carrier phase observation values; and performing difference calculations on the target carrier phase observation values ​​of the same satellite in two consecutive epochs based on the time-domain continuity characteristics of the carrier phase to construct a simplified observation equation and use it as the time-domain differential carrier phase model.

[0009] Optionally, obtaining the relative displacement vector of the vehicle in adjacent frames based on the time-domain differential carrier phase model includes: using the coordinate increment of the receiver phase center between two epochs as the solution target of the time-domain differential carrier phase model; solving the time-domain differential carrier phase model based on the least squares method to obtain the relative displacement result of the Beidou antenna phase center; and transforming the relative displacement result to the camera optical center coordinate system through lever arm effect compensation to generate a camera translation vector with absolute scale information to obtain the relative displacement vector of the vehicle in adjacent frames.

[0010] Optionally, the step of virtualizing two adjacent monocular images into a dynamic binocular stereo vision model based on relative displacement vectors to obtain the absolute depth information of image feature points includes: establishing a temporal binocular correspondence by using the current frame image as the virtual right eye image and the previous frame image as the virtual left eye image; converting the time-series image into a spatial binocular stereo vision model based on the temporal binocular correspondence to obtain a temporal binocular image pair; using the relative displacement vector of the vehicle in adjacent frames as the fixed baseline length between the virtual binocular cameras, compensating for the intrinsic parameters, distortion coefficients, and lever errors of the virtual binocular cameras according to the fixed baseline length to obtain a target binocular baseline with absolute physical scale; extracting candidate feature points matching between adjacent frames based on the temporal binocular image pair, constructing epipolar constraints based on the virtual baseline to filter the candidate feature points, and obtaining feature point pairs with a matching degree higher than a preset threshold; and performing triangulation calculation on each pair of feature points based on the target binocular baseline to obtain the real physical coordinates of the feature point pairs, thereby obtaining the absolute depth information of the image feature points.

[0011] Optionally, the step of fitting catenary curves to power lines in the image to obtain the physical geometric features of the plane containing the power lines in order to establish physical gravity field geometric constraints includes: performing instance segmentation of distribution network features on the monocular image stream acquired based on the vehicle terminal to obtain a distribution network feature map containing distribution network equipment features; extracting a set of power line arc segment pixels based on the distribution network feature map, using each power line pixel in the set as input to the standard catenary equation, and solving it using the least squares method to obtain catenary geometric parameters including the lowest point of the catenary, the tangent direction, and the vertical plane; constructing a catenary geometric equation based on the physical characteristics of the power line catenary, solving the catenary geometric equation according to the catenary geometric parameters to obtain the gravity direction vector in the camera coordinate system; and establishing absolute attitude constraint equations corresponding to the vehicle pitch angle and roll angle based on the gravity direction vector to obtain physical gravity field geometric constraints.

[0012] Optionally, the step of constructing a tightly coupled factor graph model to optimize vehicle motion state based on physical gravity field geometric constraints, temporal differential carrier phase model displacement constraints, and absolute depth information to obtain constraint factors includes: obtaining a reprojection error term based on the absolute depth information of image feature points; determining the optimization nodes of the factor graph based on the vehicle's position, attitude, and velocity at each moment, determining the auxiliary nodes of the factor graph based on map point coordinates, and using the optimization nodes and auxiliary nodes as state variables of the factor graph; obtaining displacement constraint factors to constrain vehicle position changes in adjacent frames based on temporal differential carrier phase displacement, and obtaining visual reprojection constraint factors to constrain the consistency between camera pose and map point coordinates based on the reprojection error term; constructing an optimization objective function corresponding to the factor graph with the minimum sum of displacement constraint factor residuals, visual reprojection constraint factor residuals, and physical gravity field geometric constraint factor residuals as the optimization objective, and using it as the tightly coupled factor graph model.

[0013] Optionally, solving the tightly coupled factor graph model to obtain a fusion solution including vehicle position, speed, and attitude includes: performing anomaly checks on the constraint factors of the tightly coupled factor graph model; correcting the tightly coupled factor graph model based on the anomaly check results; iteratively solving the corrected tightly coupled factor graph model based on a nonlinear factor graph optimization algorithm to obtain the optimal state variable that minimizes the sum of constraint factor residuals, thereby obtaining the vehicle's high-precision position, attitude, speed, and map point coordinates.

[0014] Optionally, the step of performing anomaly verification on the constraint factors of the tightly coupled factor graph model and correcting the tightly coupled factor graph model based on the anomaly verification results includes: calculating the position jump variable observed by BeiDou and obtaining the attitude / position change trend of the vehicle in the same period through visual power line geometry; when the dissimilarity between the position jump variable and the attitude / position change trend is greater than a statistical threshold, it is determined to be a first anomaly of non-physical motion; based on the first anomaly, the corresponding displacement constraint factor item is removed or its optimization weight is adjusted; based on the relative displacement of the vehicle in adjacent frames, the deviation of the current frame relative to the initial scale is calculated; if the deviation of consecutive frames relative to the initial scale is greater than the deviation threshold and the deviation is linearly positively correlated with the vehicle's travel distance, it is determined to be a second anomaly; based on the second anomaly, the global scale coefficient of the visual map is corrected, and the initial value of the map points is updated according to the corrected scale.

[0015] Optionally, the step of iteratively solving the modified tightly coupled factor graph model based on the nonlinear factor graph optimization algorithm to obtain the optimal state variable that minimizes the sum of constraint factor residuals, thereby obtaining the vehicle's high-precision position, attitude, speed, and map point coordinates, includes: calculating the residuals of displacement constraint factor, visual reprojection constraint factor, and physical gravity field geometric constraint factor based on the current iteration's state variables, and summing the current total objective function value; calculating the Jacobian matrix of all residuals with respect to the state variables, linearizing the nonlinear optimization problem into an incremental equation, and solving the incremental equation to obtain the state update amount; updating the state variables based on the current total objective function value and the state update amount, and calculating the updated total objective function value to determine whether the iteration termination condition is met. If it is met, the optimal state variable is output; and performing coordinate transformation, filtering, and outlier processing on the optimal state variable to obtain the high-precision positioning result, smoothed attitude angle, speed value, and distribution network feature point cloud in the geographic coordinate system.

[0016] Secondly, embodiments of this application provide a high-precision positioning system for distribution network vehicles based on BeiDou and monocular fusion, comprising: a carrier phase acquisition module, used to acquire the relative displacement vector of the vehicle in adjacent frames based on a time-domain differential carrier phase model; a depth point cloud acquisition module, used to virtualize two adjacent monocular images into a dynamic binocular stereo vision model based on the relative displacement vector, and acquire the absolute depth information of image feature points; a gravity geometric constraint module, used to fit catenary lines in the image to acquire the physical geometric features of the plane where the power lines are located to establish physical gravity field geometric constraints; and a global constraint optimization module, used to acquire constraint factors based on physical gravity field geometric constraints, time-domain differential carrier phase model displacement constraints, and absolute depth information, construct a tightly coupled factor graph model to optimize the vehicle's motion state, solve the tightly coupled factor graph model, and obtain a fusion solution including the vehicle's position, speed, and attitude.

[0017] The beneficial effects of this application are: 1. This application constructs a time-domain differential carrier phase model and utilizes the millimeter-level relative displacement accuracy of the BeiDou carrier phase in a short time to virtualize two frames of images before and after vehicle movement into a fixed baseline binocular image. Thus, without increasing additional hardware costs, it provides absolute scale information to monocular vision in real time and dynamically, fundamentally eliminating the scale drift problem of monocular vision odometry. 2. This application fully utilizes the physical and geometric characteristics of the natural drooping curve formed by the power line under the action of gravity in the distribution network environment, and reversely calculates the absolute gravity direction, providing an absolute physical attitude reference independent of the inertial sensor. This effectively suppresses the attitude error accumulated by the vehicle during long-term driving, avoids the problem of attitude instability of low-cost positioning systems under complex road conditions, and significantly improves the long-term stability and accuracy of positioning results. 3. This application achieves deep integration of BeiDou satellite signals, monocular visual observation and physical environment features by constructing a tightly coupled factor graph model that includes three constraint factors: displacement, reprojection and gravity geometry. This model utilizes the natural sparsity and flexibility of the factor graph structure, can effectively handle nonlinear observation models, and eliminates the cumulative error of a single sensor through a global optimization algorithm, thereby improving the accuracy and robustness of the positioning system and thus improving the efficiency of inspection operations. Attached Figure Description

[0018] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings.

[0019] Figure 1 A flowchart of a high-precision positioning method for distribution network vehicles based on BeiDou and monocular fusion provided in this application embodiment.

[0020] Figure 2 A schematic diagram of a distribution network vehicle-mounted high-precision positioning system module based on BeiDou and monocular fusion provided in an embodiment of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely one preferred embodiment of this application and are only used to explain this application. They do not limit the scope of protection of this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] Example 1: As Figure 1As shown, a high-precision positioning method for distribution network vehicles based on BeiDou and monocular fusion includes steps S1-S4, wherein: S1. Obtain the relative displacement vector of the vehicle in adjacent frames based on the time-domain differential carrier phase model.

[0023] In an optional embodiment, prior to step S1, the following steps are included: Based on the vehicle-mounted terminal, monocular image stream and Beidou raw radio frequency data are acquired. After clock synchronization between the monocular image stream and Beidou raw radio frequency data, carrier phase observation values ​​are extracted based on the Beidou raw radio frequency data. Carrier phase observations with elevation angles greater than a preset elevation angle threshold and signal-to-noise ratios greater than a preset signal-to-noise ratio threshold are used as target carrier phase observations. Based on the continuous characteristics of carrier phase in the time domain, the difference calculation is performed on the target carrier phase observation values ​​of the same satellite in two consecutive epochs to construct a simplified observation equation and use it as a time-domain differential carrier phase model.

[0024] In some embodiments, after extracting carrier phase observations, data from all satellites is not used directly; instead, a rigorous screening process is performed. Considering the complex environment of power distribution network inspections, low-elevation satellite signals are easily obstructed by trees and buildings, resulting in multipath effects, and the long paths through the atmosphere lead to significant ionospheric delay errors. Therefore, a preset elevation angle threshold, such as 15 degrees or 30 degrees, is set to exclude satellites with elevation angles below this threshold. Simultaneously, a preset signal-to-noise ratio (SNR) threshold, such as 40 dB-Hz, is set to exclude satellite observations with weak signal strength. This dual screening mechanism based on elevation angle and SNR effectively eliminates poor-quality observation data, ensuring that the carrier phase observations entering the solution stage have sufficient stability and reliability, thereby improving the robustness of the model.

[0025] Furthermore, in constructing the time-domain differential carrier phase model, the continuous nature of carrier phase observations in the time domain is utilized. For the same satellite, the integer ambiguity parameter is the same in the carrier phase observations of two consecutive epochs (i.e., the times corresponding to adjacent image frames). The ionospheric delay, tropospheric delay, and satellite clock bias change extremely little in a short period of time and can be regarded as constants. By performing time-domain difference on the observations of adjacent epochs of the same satellite and the same receiver, a new simplified observation equation is constructed. This difference operation directly eliminates systematic errors such as integer ambiguity, satellite clock bias, and ionospheric / tropospheric delay that are almost unchanged in a short period of time, thus constructing a minimally simplified observation model that only includes the changes in receiver position and receiver clock bias. This simplified observation equation is the core of the time-domain differential carrier phase model. By simplifying the observation equation, the complex integer ambiguity resolution process in traditional positioning methods is avoided, enabling the calculation of millimeter-level relative displacement vectors even in obstructed environments where integer ambiguity cannot be fixed, utilizing the high-precision relative change characteristics of the carrier phase.

[0026] In this embodiment, inferior observations are eliminated through dual screening using elevation angle and signal-to-noise ratio, and common errors and integer ambiguities are eliminated through inter-epoch difference, thus constructing a highly robust time-domain differential carrier phase model, laying a solid data foundation for obtaining high-precision relative displacement vectors in the future.

[0027] In an optional embodiment, step S1 includes: The coordinate increment of the receiver phase center between two epochs is used as the solution target of the time-domain differential carrier phase model. The time-domain differential carrier phase model is solved using the least squares method to obtain the relative displacement result of the phase center of the Beidou antenna; The relative displacement results are transformed to the camera optical center coordinate system through lever effect compensation to generate a camera translation vector with absolute scale information, so as to obtain the relative displacement vector of the vehicle in adjacent frames.

[0028] In this embodiment, the purpose of constructing a time-domain differential carrier phase model is to perform a difference operation on the carrier phase observations of the same satellite in two consecutive epochs, eliminating integer ambiguity parameters, receiver clock drift, and ionospheric delay errors. Specifically, a subset of BeiDou satellites with elevation angles higher than a preset threshold and stable signal-to-noise ratios is selected to provide high-quality, high-reliability input data for model construction. Abnormal satellite observations with low elevation angles and low signal-to-noise ratios are eliminated to avoid invalid data interfering with the subsequent construction and solution of the differential model, which is a prerequisite for ensuring the effectiveness of the model. The core unknowns in the model (receiver coordinate increments) are solved using the least squares method to obtain millimeter-level relative displacement results (i.e., the solution result of the BeiDou antenna phase center). The relative displacement results are then converted into a rigid virtual baseline in the camera coordinate system to obtain a camera translation vector (relative displacement vector) with absolute scale information. The relative displacement vector of the vehicle in adjacent frames is the rigid virtual baseline, providing input for the scale constraints of subsequent monocular vision. It should be understood that although the least squares method is used as an example here, in other embodiments, estimation methods such as Kalman filtering can also be used, as long as the position increment can be calculated from the observation equation. This process utilizes the millimeter-level observation accuracy of the BeiDou carrier phase, so that the calculated relative displacement result has extremely high relative accuracy.

[0029] Furthermore, regarding lever arm effect compensation, since the BeiDou receiving antenna is typically mounted on the vehicle roof while the monocular camera is mounted on the windshield, there is a physical distance between them, i.e., a "lever arm." When the vehicle undergoes attitude changes during driving (such as turning or bumping), the motion trajectories of the antenna phase center and the camera optical center are not the same. If the displacement of the antenna center is directly used as the displacement of the camera, it will introduce additional displacement errors, especially when the vehicle is turning. Therefore, lever arm effect compensation is necessary. Specifically, using the attitude information provided by the vehicle attitude sensor or the attitude information estimated by visual odometry, a rotation matrix from the vehicle coordinate system to the camera coordinate system is constructed. Through coordinate transformation formulas, the relative displacement result of the BeiDou antenna phase center is transformed to the camera optical center coordinate system. This process eliminates the motion differences caused by different installation positions and ensures the consistency of the displacement vector.

[0030] In this embodiment, by transforming the high-precision relative measurement capability of BeiDou carrier phase into an absolute scale reference usable by the visual system, key input data is provided to solve the problem of missing scale in monocular vision, and deep integration of satellite signals and visual observation at the physical scale level is realized.

[0031] S2. Based on the relative displacement vector, virtualize two adjacent monocular images into a dynamic binocular stereo vision model to obtain the absolute depth information of image feature points.

[0032] In an optional embodiment, step S2 includes: The current frame image is used as the virtual right eye image, and the previous frame image is used as the virtual left eye image to establish a temporal binocular correspondence. Based on the temporal binocular correspondence, the conversion of time-series images into spatial binocular stereo vision models is completed, and temporal binocular image pairs are obtained. The relative displacement vector of the vehicle in adjacent frames is used as the fixed baseline length between virtual stereo cameras. The virtual stereo camera intrinsic parameters, distortion coefficients and lever error are compensated according to the fixed baseline length to obtain the target stereo baseline with absolute physical scale. Based on the temporal stereo image pair, candidate feature points for inter-frame matching are extracted, and epipolar constraints are constructed based on the virtual baseline to filter the candidate feature points and obtain feature point pairs with a matching degree higher than a preset threshold. Based on the target binocular baseline, triangulation calculation is performed on each pair of feature points to obtain the true physical coordinates of the feature point pairs, thereby obtaining the absolute depth information of the image feature points.

[0033] Specifically, traditional binocular stereo vision systems rely on two physically fixed cameras to calculate depth using a known fixed baseline length. However, the vehicle-mounted terminal in this embodiment only carries a single monocular camera, lacking a physical binocular baseline. To address this issue, this embodiment creatively proposes the concept of "dynamic binoculars" by utilizing vehicle motion. When the vehicle is moving, two frames captured by the camera at different times are spatially equivalent to the results of the same camera captured from two different locations. If the camera position at the previous moment (previous frame) is considered the virtual left eye, and the camera position at the current moment (current frame) is considered the virtual right eye, then these two frames constitute a virtual binocular stereo vision model. This method of transforming a temporal image sequence into a spatial stereo vision model enables the monocular camera to perceive three-dimensional space, laying the geometric foundation for subsequent depth calculation.

[0034] Furthermore, in binocular stereo vision, the baseline length is a key parameter determining the accuracy of depth measurement. By calculating the camera translation vector with absolute scale information, its length is directly used as the fixed baseline length between the virtual binocular cameras. Since this baseline length originates from high-precision measurements of the BeiDou carrier phase and has authentic metric units, it endows the virtual binocular model with absolute physical scale. Simultaneously, considering potential radial and tangential distortions in the camera lens, as well as installation errors between the camera and the BeiDou antenna, the system pre-calibrates and compensates in real-time for the camera's intrinsic parameter matrix, distortion coefficients, and boom errors, ensuring the accuracy of the virtual binocular model's geometric parameters. This is the core step in achieving absolute scale recovery in monocular vision.

[0035] Furthermore, a triangulation method is employed to recover the 3D coordinates of feature points. The basic principle is to utilize the similar triangle relationship in spatial geometry, based on the difference in pixel coordinates (parallax) between the feature points in the left and right frames, as well as the camera's focal length and baseline length, to calculate the depth distance from the feature points to the camera. The specific formula can be expressed as: Depth Z = (Baseline Length * Focal Length) / Parallax. Since the baseline length in the formula is a known quantity with real physical units, the calculated depth Z also directly has real physical units (such as meters), eliminating the need for scale normalization. By traversing all filtered feature point pairs, the real physical coordinates of a large number of feature points in the image can be obtained, thereby generating a sparse point cloud map with absolute scale. This process completely solves the inherent limitation of monocular vision in acquiring absolute depth information, enabling low-cost monocular cameras to achieve or even surpass the ranging performance of traditional binocular cameras.

[0036] In this embodiment, by converting time-series monocular images into a spatial virtual binocular model and using the absolute scale baseline provided by the BeiDou carrier phase for triangulation measurement, the monocular vision system is successfully endowed with the ability to perceive the scale of the real physical world. This not only avoids the use of expensive binocular cameras or LiDAR hardware, significantly reducing system costs, but also effectively avoids the scale drift problem that has long existed in monocular vision odometry, providing high-precision three-dimensional environmental perception data for power distribution network vehicle inspection.

[0037] S3. Perform catenary fitting on the electric field lines in the image to obtain the physical geometric features of the plane where the electric field lines are located in order to establish the physical gravity field geometric constraints.

[0038] In an optional embodiment, step S3 includes: Instance segmentation of distribution network features is performed on the monocular image stream acquired based on the vehicle terminal to obtain a distribution network feature map containing the features of distribution network equipment; Based on the power line feature map, extract the set of power line arc segments, take each power line pixel in the set as the input of the standard catenary equation, and solve it using the least squares method to obtain the catenary geometric parameters including the lowest point of the catenary, the tangent direction, and the vertical plane. Based on the physical properties of the catenary, the geometric equation of the catenary is constructed. The geometric equation of the catenary is solved according to the geometric parameters of the catenary to obtain the gravity direction vector in the camera coordinate system. Based on the gravity direction vector, establish absolute attitude constraint equations corresponding to the vehicle's pitch and roll angles to obtain the physical gravity field geometric constraints.

[0039] Specifically, the features of distribution network equipment include, but are not limited to, distribution network equipment such as poles, transformers, and power lines. The distribution network feature map refers to the image of the distribution network feature region containing the equipment features separated from the background. The image is processed by an instance segmentation network to reduce the interference of irrelevant background information, providing a clean data source for subsequent accurate geometric fitting.

[0040] Furthermore, since the electric field line appears as a long, thin curved segment in the image, its pixel-level geometric center line can be obtained using skeleton extraction or edge detection algorithms. According to physical principles, a flexible conductor, under the influence of gravity alone and with its two suspension points at or approximately the same height, will naturally droop in a manner strictly conforming to the catenary equation. The standard catenary equation is typically expressed as a hyperbolic cosine function, for example: y = a*cosh(x / a)+b, where a and b are undetermined coefficients. Substituting the extracted electric field line pixel coordinates into this equation constructs a set of error equations. Since the observed data (pixel coordinates) inevitably contains noise, the least squares method is used for parameter estimation to find the optimal parameter solution that minimizes the sum of the squared distances from all pixels to the fitted curve. By solving this solution, the precise geometric shape of the electric field line segment in the image coordinate system can be determined, including the position of the lowest point of the catenary, the tangent direction at the lowest point, and the parameters of the vertical plane in which the catenary lies. This process transforms discrete pixels into a continuous, physically meaningful geometric parameter model.

[0041] Specifically, the physical characteristics of electric field catenaries include: the plane in which the catenary lies must be a vertical plane containing the gravity vector, and the tangent direction at the lowest point of the catenary must be perpendicular to the direction of gravity, that is, it is in a horizontal state.

[0042] By utilizing this prior physical knowledge, the geometric equations of the catenary are constructed. Using a known camera imaging model, three-dimensional points are projected onto a two-dimensional image plane. Through the camera's intrinsic parameter matrix, the geometric parameters of the catenary fitted in the image can be projected back into three-dimensional space. By utilizing the orthogonality between the normal vector of the vertical plane containing the catenary and the gravity vector, and the perpendicularity between the tangent direction at the lowest point and the gravity vector, a system of linear equations is constructed. Solving this system allows for the calculation of the rotation of the camera coordinate system relative to the true gravity direction, thus obtaining the gravity direction vector in the camera coordinate system. This process is completely independent of the inertial measurement unit (IMU), relying solely on visually observed environmental physical characteristics to obtain an absolute gravity reference, exhibiting the excellent characteristic of not drifting over time.

[0043] Specifically, vehicle attitude is typically described by pitch and roll angles, which define the degree of tilt of the vehicle body relative to the horizontal plane. Since the direction of gravity is fixed (vertically downward) in the geographic coordinate system, the deviation angle between the gravity direction vector in the camera coordinate system and the standard gravity vector in the geographic coordinate system can be calculated. Based on this, an absolute attitude constraint equation is established, which can be expressed as minimizing the residual between the observed gravity vector and the estimated gravity vector at the desired attitude. This constraint equation is the physical gravity field geometric constraint. In subsequent optimization processes, this constraint can be used as a strong constraint factor, forcing the estimated pitch and roll angles of the vehicle to converge to near the true attitude derived from the electric field line geometry. This gravity constraint mechanism effectively eliminates the cumulative attitude error generated by the visual odometry when driving on bumpy roads, especially significantly suppressing elevation drift.

[0044] In this embodiment, power line features are accurately extracted using instance segmentation technology, and the gravity direction is solved by using a catenary physical model, thus constructing an absolute attitude constraint independent of inertial sensors. This fully utilizes the unique physical and geometric features of the power distribution network environment, providing the positioning system with an attitude reference that does not drift over time. This avoids the problem of attitude instability in low-cost positioning systems under complex road conditions, significantly improving the long-term stability and accuracy of positioning results.

[0045] S4. Based on the physical gravity field geometric constraints, the temporal differential carrier phase model displacement constraints, and the absolute depth information to obtain constraint factors, a tightly coupled factor graph model is constructed to optimize the vehicle motion state. The tightly coupled factor graph model is solved to obtain a fusion solution that includes the vehicle position, velocity, and attitude.

[0046] In an optional embodiment, in step S4, based on the physical gravity field geometric constraints, the temporal differential carrier phase model displacement constraints, and the constraint factors obtained from absolute depth information, a tightly coupled factor graph model for optimizing the vehicle's motion state is constructed, including: The reprojection error term is obtained based on the absolute depth information of image feature points; The optimal nodes of the factor graph are determined based on the vehicle's position, attitude, and speed at each moment, and the auxiliary nodes of the factor graph are determined based on the map point coordinates. The optimal nodes and auxiliary nodes are used as state variables of the factor graph. The displacement constraint factor used to constrain the vehicle's position change in adjacent frames is obtained based on the time-domain differential carrier phase displacement, and the visual reprojection constraint factor used to constrain the consistency between the camera pose and the map point coordinates is obtained based on the reprojection error term. With the goal of minimizing the sum of displacement constraint factor residuals, visual reprojection constraint factor residuals, and physical gravity field geometric constraint factor residuals, an optimization objective function corresponding to the factor graph is constructed and used as a tightly coupled factor graph model.

[0047] It should be noted that the above-described virtual stereo model calculates the true physical coordinates (i.e., map points) of the feature points. Using these 3D coordinates of the map points, combined with the estimated pose of the camera in the current frame, they are reprojected back onto the image plane through a camera projection model to obtain the projected pixel coordinates. The Euclidean distance between these projected coordinates and the actually observed feature point pixel coordinates is the reprojection error term. The reprojection error term directly reflects the degree of consistency between visual observation and vehicle motion state. It not only includes the positional error of the feature points but can also introduce uncertainty information in feature point extraction through the covariance matrix, thereby giving greater weight to high-precision observations in the subsequent solution of the tightly coupled factor graph model.

[0048] In some embodiments, the factor graph is a bidirectional graph model consisting of nodes and factors. In this embodiment, nodes represent state variables to be estimated. For the vehicle's motion state, the vehicle's position vector, attitude quaternion, and velocity vector at each moment (i.e., each keyframe moment) are defined as optimization nodes. These variables constitute the main body of the vehicle trajectory. Simultaneously, to achieve visual constraints, the three-dimensional coordinates of observed environmental feature points (i.e., map points) are defined as auxiliary nodes. By jointly modeling the vehicle state and environmental map points, the system can optimize the vehicle pose while simultaneously correcting the environmental map, achieving simultaneous localization and mapping.

[0049] In other embodiments, if the system integrates an inertial measurement unit (IMU), the state variables can be further extended to include accelerometer zero bias and gyroscope zero bias to eliminate systematic errors of the inertial devices.

[0050] Specifically, the "factors" in the factor graph represent the constraint relationships between observations and state variables. This embodiment constructs three key constraint factors: a displacement constraint factor, derived from the calculated BeiDou carrier phase relative displacement vector, which constructs a constraint equation for the position state variables of two adjacent frames, limiting the vehicle's position change over a short period and providing an absolute scale reference for the system; a visual reprojection constraint factor, derived from the reprojection error term, which connects the camera pose node and the map point node, constraining the vehicle's relative motion trajectory by minimizing the reprojection error; and a physical gravity field geometric constraint factor, derived from the calculated gravity direction vector, which constructs a constraint equation for the vehicle's attitude (pitch and roll angles), limiting the deviation of the vehicle's attitude relative to the gravity field. These three constraint factors construct a rigorous constraint network from the three dimensions of position, attitude, and observation consistency.

[0051] Specifically, the objective function of the factor graph optimization involves a weighted sum of the residuals of the three constraint factors, with the weight of each constraint factor determined by its observation noise covariance matrix. For example, the BeiDou carrier phase displacement observation has extremely high accuracy (millimeter-level), thus its corresponding weight is relatively large; while the visual reprojection error is greatly affected by the feature extraction accuracy, so its weight is relatively moderate. By solving this nonlinear optimization problem, the optimal estimates of vehicle position, speed, attitude, and map point coordinates can be obtained simultaneously. This method, which directly incorporates the original observation data into a unified optimization framework, avoids the information loss caused by hierarchical processing in traditional loosely coupled schemes and achieves deep fusion of multi-source information.

[0052] In this embodiment, a tightly coupled factor graph model incorporating three constraint factors—displacement, reprojection, and gravity geometry—is constructed to achieve deep fusion of BeiDou satellite signals, monocular visual observations, and physical environment features. This model leverages the inherent sparsity and flexibility of the factor graph structure to effectively handle nonlinear observation models and eliminates the cumulative error of a single sensor through a global optimization algorithm, significantly improving the accuracy and robustness of the positioning system.

[0053] In an optional embodiment, step S4 involves solving the tightly coupled factor graphical model to obtain a fused solution containing vehicle position, velocity, and attitude, including: Anomaly checks are performed on the constraint factors of the tightly coupled factor graphical model, and the tightly coupled factor graphical model is corrected based on the anomaly check results. The modified tightly coupled factor graph model is solved iteratively using a nonlinear factor graph optimization algorithm to obtain the optimal state variables that minimize the sum of constraint factor residuals, thereby obtaining the vehicle's high-precision position, attitude, speed, and map point coordinates.

[0054] In an optional embodiment, the step of performing anomaly checks on the constraint factors of the tightly coupled factor graphical model and correcting the tightly coupled factor graphical model based on the anomaly check results includes: Calculate the position jump variable observed by Beidou, and obtain the attitude / position change trend of the vehicle in the same period through visual electric field geometry. When the dissimilarity between the position jump variable and the attitude / position change trend is greater than the statistical threshold, it is determined to be the first anomaly of non-physical motion. Based on the first anomaly, the corresponding displacement constraint factor item is removed or its optimization weight is adjusted. The deviation of the current frame relative to the initial scale is calculated based on the relative displacement of the vehicle in adjacent frames. If the deviation of consecutive frames relative to the initial scale is greater than the deviation threshold and the deviation is linearly positively correlated with the vehicle's travel distance, it is determined to be a second anomaly. The global scale coefficient of the visual map is corrected based on the second anomaly, and the initial values ​​of map points are updated according to the corrected scale.

[0055] Specifically, the first anomaly refers to abnormal observations caused by non-physical motion, and the second anomaly refers to visual scale drift. For the second anomaly, the global scale coefficient of the visual map is corrected using BeiDou absolute displacement, and a secondary scale check is performed using the identified standard span of the towers to eliminate accumulated residuals. The initial values ​​of all map points in the visual reprojection constraint factor are updated using the calibrated scale to correct the systematic bias of the visual reprojection residuals, or the weight of the calibrated visual factor is increased to enhance its constraint ability on optimization.

[0056] In some embodiments, the determination of the second anomaly includes: using two independent absolute scale benchmarks as the true values, continuously comparing them with the visually estimated scale, and determining whether it is a systematic drift based on the cumulative trend of the deviation, completely distinguishing it from a single burst of anomalies.

[0057] Specifically, the system simultaneously acquires a baseline value (the magnitude of the absolute translation vector of adjacent frames) and a visual estimate (the magnitude of the normalized translation vector of adjacent frames calculated by epipolar geometry in monocular vision). The baseline value is divided by the visual estimate to obtain the true scale coefficient of the current frame. The true scale coefficient of the current frame is subtracted from the initial scale calibrated during system initialization, and the absolute value is divided by the initial scale to obtain the deviation relative to the initial scale. If the deviation relative to the initial scale in a single frame jumps by more than 10%, it is considered an anomaly indicating visual tracking failure and is not included in the drift assessment. If the deviation relative to the initial scale continuously exceeds the 1% threshold for three or more consecutive frames, and the deviation shows a monotonically increasing / continuously deviating trend without regression, it is considered the start of scale drift. If the deviation continuously exceeds the limit for ten or more consecutive frames, and the deviation is linearly positively correlated with the vehicle's travel distance without sudden jumps, it is ultimately determined to be a systemic scale drift in monocular vision, triggering scale correction. It should be understood that the above-mentioned deviation jump range threshold and the number of consecutive frames need to be adaptively set according to actual needs to ensure the accuracy and reliability of the anomaly detection results.

[0058] In other embodiments, the anomaly type also includes constraint conflicts, i.e., inconsistencies in the consistency of multi-source independent observations. Constraint conflicts include, but are not limited to, position domain constraint conflicts and attitude domain constraint conflicts. Position domain constraint conflicts are determined using BeiDou displacement constraints and visual motion constraints. The relative displacement deviation between the two sources is calculated based on the absolute vehicle displacement vectors of adjacent frames obtained from BeiDou source data and the vehicle displacement vectors of adjacent frames obtained from the camera's visual source image data. When the relative displacement deviation between the two sources exceeds a threshold range (e.g., 50%) of the vehicle's maximum physical displacement, it exceeds the vehicle's physical motion law and is determined to be a BeiDou-visual position domain constraint conflict, triggering weight reallocation. Attitude domain constraint conflicts are determined based on the vehicle pitch and roll angles estimated by visual odometry or IMU, and the catenary gravity constraint attitude angles (i.e., physical gravity field geometric constraints) obtained through the absolute attitude constraint equations. First, determine whether the fit of the power line to the catenary is greater than the gravity reference reliability threshold (e.g., 0.99), whether there is no feature tracking failure in vision, and whether there is no severe impact anomaly in the IMU. If all of the above conditions are met, single-source anomalies are excluded. Then, the absolute deviation of the attitude angle is obtained based on the absolute value of the difference between the vehicle roll angle (or pitch angle) estimated by the visual odometry or IMU and the attitude angle constrained by the catenary gravity. If the rate of change of attitude angle in adjacent frames exceeds the threshold (e.g., 30° / s), it exceeds the limit that the attitude change on the bumpy road surface does not conform to the physical motion law of the vehicle. It is determined to be an attitude domain constraint conflict, triggering gravity reference arbitration and weight reallocation.

[0059] In other embodiments, the correction of the tightly coupled factor graph model also includes a two-way attitude feedback correction covering the first anomaly, the second anomaly, and constraint conflicts. That is, the correction method does not belong to a single outlier handling, systematic drift correction, or constraint conflict resolution, but is a composite attitude calibration covering the above three types of anomalies. The calibration logic is as follows: the vehicle attitude error is corrected using the catenary gravity absolute reference, and the electric power line catenary fitting accuracy is optimized using the calibrated attitude, thereby improving the reliability of the gravity reference and forming a two-way closed loop.

[0060] In some examples, attitude bidirectional feedback correction addresses different types of problems under varying error scenarios using different processing strategies. For instance, when a vehicle encounters a pothole or severe bumps, causing a sudden change in attitude in a single frame of the vision / IMU output, a catenary gravity absolute reference is used to identify and correct the sudden, non-physical attitude change, eliminating attitude outliers and preventing them from entering factor graph optimization. When, after long-distance driving, the cumulative zero bias of the vision / IMU gyroscope causes the pitch / roll angle to continuously and slowly deviate from the true value, a drift-free gravity absolute reference is used to continuously correct the accumulated zero bias of the attitude, suppressing the systematic drift of the attitude over time and providing unbiased initial attitude values ​​for the factor graph. When the attitude estimates of vision and IMU are continuously contradictory, with no single-source anomalies, resulting in a conflict in attitude domain constraints, a gravity physical reference is used as the sole arbitrator to assign weights to the conflicting vision / IMU attitude constraints, resolving consistency contradictions in multi-source observations.

[0061] In this embodiment, by modifying the initial attitude value and fixing the highest weight of the catenary gravity factor term, the attitude constraint of the optimization objective function is ensured to have an absolute physical reference, thus avoiding optimization divergence.

[0062] In an optional embodiment, the iterative solution of the modified tightly coupled factor graph model based on the nonlinear factor graph optimization algorithm to obtain the optimal state variable that minimizes the sum of constraint factor residuals, thereby obtaining the high-precision position, attitude, speed, and map point coordinates of the vehicle, includes: Based on the state variables of the current iteration, calculate the residuals of the displacement constraint factor, the visual reprojection constraint factor, and the physical gravity field geometric constraint factor, and summarize the current total objective function value. Calculate the Jacobian matrix of all residuals with respect to state variables, linearize the nonlinear optimization problem into an incremental equation, and solve the incremental equation to obtain the state update. The state variables are updated based on the current total objective function value and the state update amount, and the updated total objective function value is calculated to determine whether the iteration termination condition is met. If it is met, the optimal state variable is output. The optimal state variables are subjected to coordinate transformation, filtering, and outlier processing to obtain high-precision positioning results, smoothed attitude angles, velocity values, and power distribution network feature point clouds in the geographic coordinate system.

[0063] Specifically, at the beginning of each iteration, residuals for various constraint factors are calculated based on the currently estimated state variables (including vehicle position, attitude, velocity, and map point coordinates). For the displacement constraint factor, the residual is defined as the difference between the relative displacement vector observed by the BeiDou carrier phase and the relative displacement vector calculated based on the current state estimate, reflecting the consistency between satellite observation and the motion model. For the visual reprojection constraint factor, the residual is defined as the Euclidean distance between the pixel coordinates of the map point after projection onto the current camera pose and the actual observed pixel coordinates, reflecting the consistency between visual observation and the map. For the physical gravity field geometric constraint factor, the residual is defined as the angular deviation between the gravity direction vector calculated based on the current attitude estimate and the gravity direction vector obtained by fitting the electric field lines and catenary, reflecting the consistency between attitude estimation and physical geometric constraints. The magnitude of the objective function value directly reflects the quality of the current state estimate; a smaller value indicates a more accurate estimate.

[0064] In some embodiments, the iteration termination condition includes any one of the following three cases: first, the change in the objective function value is less than a preset first threshold (e.g., ...). The characterization optimization has reached an extreme point; secondly, the magnitude of the state update is less than the preset second threshold (e.g., The system achieves three conditions: 1) the state variables have stabilized; 2) the number of iterations reaches a preset upper limit (e.g., 50 times) to prevent infinite loops. When any of the above conditions are met, the iteration ends, and the system outputs the current state variables as the optimal state variables, namely the vehicle's high-precision position, attitude, speed, and map point coordinates.

[0065] Furthermore, the optimal state variables obtained from the optimization solution are usually defined in a local coordinate system (such as the camera coordinate system or the local tangent plane coordinate system). To meet the needs of practical applications, they need to be transformed to a geographic coordinate system (such as the WGS-84 coordinate system or the local engineering coordinate system). Specifically, using known coordinate transformation parameters, the vehicle position and map point coordinates are transformed to the geographic coordinate system. Simultaneously, to eliminate high-frequency noise, the output attitude angle and velocity values ​​are filtered, for example, using moving average filtering or low-pass filtering to obtain smooth attitude angle and velocity values. In addition, outlier processing is performed on the generated distribution network feature point cloud, removing isolated points with excessive reprojection errors or unreasonable depth values ​​to ensure that the final output distribution network feature point cloud has high accuracy and high reliability. Finally, the output contains a complete positioning result including the vehicle's high-precision position, smoothed attitude angles, velocity values, and the distribution network feature point cloud.

[0066] In this embodiment, an iterative solution process is employed, utilizing the Jacobian matrix to linearize the nonlinear problem. Furthermore, dynamic adjustment of the damping factor ensures the stability and speed of convergence, ultimately achieving the optimal estimation of the vehicle's state. Simultaneously, coordinate transformation and post-processing output high-precision positioning results that meet practical application requirements, ensuring the timeliness and efficiency of the inspection operation.

[0067] Based on the same inventive concept, this application also provides a high-precision positioning system for distribution network vehicles based on BeiDou and monocular fusion, corresponding to the method for high-precision positioning of distribution network vehicles based on BeiDou and monocular fusion. Figure 2 As shown, it includes: The carrier phase acquisition module is used to obtain the relative displacement vector of the vehicle in adjacent frames based on the time-domain differential carrier phase model. The depth point cloud acquisition module is used to virtualize two adjacent monocular images into a dynamic binocular stereo vision model based on the relative displacement vector, and to obtain the absolute depth information of image feature points. The gravity geometry constraint module is used to fit catenary lines in the image to obtain the physical geometric features of the plane where the electric field lines are located in order to establish physical gravity field geometric constraints. The global constraint optimization module is used to obtain constraint factors based on physical gravity field geometric constraints, temporal differential carrier phase model displacement constraints, and absolute depth information. It then constructs a tightly coupled factor graph model to optimize the vehicle's motion state, solves the tightly coupled factor graph model, and obtains a fusion solution that includes the vehicle's position, velocity, and attitude.

[0068] In this embodiment, through the coordinated operation of various modules, a high-precision scale reference is provided by the carrier phase acquisition module, low-cost 3D perception is achieved by the depth point cloud acquisition module, an attitude reference that does not drift over time is introduced by the gravity geometry constraint module, and finally, deep fusion of multi-source information is achieved through the global constraint optimization module. This architecture design not only ensures high positioning accuracy and robustness but also has extremely high hardware compatibility, requiring only a conventional vehicle-mounted monocular camera and a common BeiDou module, without relying on expensive dedicated hardware, thus reducing the deployment cost and maintenance difficulty of the power distribution network automated inspection system.

[0069] The above-described embodiments are preferred embodiments of this application and are not intended to limit the specific scope of this application. The scope of this application includes but is not limited to the specific embodiments described above. All equivalent changes made in accordance with the shape, structure, and method of this application are within the protection scope of this application.

Claims

1. A high-precision positioning method for distribution network vehicles based on BeiDou and monocular fusion, characterized in that: Includes the following steps: The relative displacement vector of the vehicle in adjacent frames is obtained based on the time-domain differential carrier phase model; Based on the relative displacement vector, two adjacent monocular images are virtualized into a dynamic binocular stereo vision model to obtain the absolute depth information of image feature points; Catenary fitting is performed on the electric field lines in the image to obtain the physical geometric features of the plane where the electric field lines are located in order to establish the geometric constraints of the physical gravity field. Based on the geometric constraints of the physical gravity field, the displacement constraints of the time-domain differential carrier phase model, and the constraint factors obtained from the absolute depth information, a tightly coupled factor graph model is constructed to optimize the vehicle motion state. The tightly coupled factor graph model is solved to obtain a fusion solution that includes the vehicle position, velocity, and attitude.

2. The high-precision positioning method for distribution network vehicles based on BeiDou and monocular fusion as described in claim 1, characterized in that: Before obtaining the relative displacement vector of the vehicle in adjacent frames based on the time-domain differential carrier phase model, the following steps are included: Based on the vehicle-mounted terminal, monocular image stream and Beidou raw radio frequency data are acquired. After clock synchronization between the monocular image stream and Beidou raw radio frequency data, carrier phase observation values ​​are extracted based on the Beidou raw radio frequency data. Carrier phase observations with elevation angles greater than a preset elevation angle threshold and signal-to-noise ratios greater than a preset signal-to-noise ratio threshold are used as target carrier phase observations. Based on the continuous characteristics of carrier phase in the time domain, the difference calculation is performed on the target carrier phase observation values ​​of the same satellite in two consecutive epochs to construct a simplified observation equation and use it as a time-domain differential carrier phase model.

3. The high-precision positioning method for distribution network vehicles based on BeiDou and monocular fusion as described in claim 2, characterized in that: The method of obtaining the relative displacement vector of the vehicle in adjacent frames based on the time-domain differential carrier phase model includes: The coordinate increment of the receiver phase center between two epochs is used as the solution target of the time-domain differential carrier phase model. The time-domain differential carrier phase model is solved using the least squares method to obtain the relative displacement result of the phase center of the Beidou antenna; The relative displacement results are transformed to the camera optical center coordinate system through lever effect compensation to generate a camera translation vector with absolute scale information, so as to obtain the relative displacement vector of the vehicle in adjacent frames.

4. The high-precision positioning method for distribution network vehicles based on BeiDou and monocular fusion as described in claim 1, characterized in that: The process of virtualizing two adjacent monocular images into a dynamic binocular stereo vision model based on relative displacement vectors and obtaining the absolute depth information of image feature points includes: The current frame image is used as the virtual right eye image, and the previous frame image is used as the virtual left eye image to establish a temporal binocular correspondence. Based on the temporal binocular correspondence, the conversion of time-series images into spatial binocular stereo vision models is completed, and temporal binocular image pairs are obtained. The relative displacement vector of the vehicle in adjacent frames is used as the fixed baseline length between virtual stereo cameras. The intrinsic parameters, distortion coefficients and lever error of the virtual stereo cameras are compensated according to the fixed baseline length to obtain the target stereo baseline with absolute physical scale. Based on the temporal stereo image pair, candidate feature points for inter-frame matching are extracted, and epipolar constraints are constructed based on the virtual baseline to filter the candidate feature points and obtain feature point pairs with a matching degree higher than a preset threshold. Based on the target binocular baseline, triangulation calculation is performed on each pair of feature points to obtain the true physical coordinates of the feature point pairs, thereby obtaining the absolute depth information of the image feature points.

5. The high-precision positioning method for distribution network vehicles based on BeiDou and monocular fusion as described in claim 1, characterized in that: The step of fitting catenary lines to the electric field lines in the image to obtain the physical geometric features of the plane containing the electric field lines in order to establish physical gravitational field geometric constraints includes: Instance segmentation of distribution network features is performed on the monocular image stream acquired based on the vehicle terminal to obtain a distribution network feature map containing the features of distribution network equipment; Based on the power line feature map, extract the set of power line arc segments, take each power line pixel in the set as the input of the standard catenary equation, and solve it using the least squares method to obtain the catenary geometric parameters including the lowest point of the catenary, the tangent direction, and the vertical plane. Based on the physical properties of the catenary, the geometric equation of the catenary is constructed. The geometric equation of the catenary is solved according to the geometric parameters of the catenary to obtain the gravity direction vector in the camera coordinate system. Based on the gravity direction vector, establish absolute attitude constraint equations corresponding to the vehicle's pitch and roll angles to obtain the physical gravity field geometric constraints.

6. The high-precision positioning method for distribution network vehicles based on BeiDou and monocular fusion as described in claim 5, characterized in that: The tightly coupled factor graph model for optimizing vehicle motion state is constructed based on physical gravity field geometric constraints, temporal differential carrier phase model displacement constraints, and absolute depth information acquisition constraint factors, including: The reprojection error term is obtained based on the absolute depth information of image feature points; The optimal nodes of the factor graph are determined based on the vehicle's position, attitude, and speed at each moment, and the auxiliary nodes of the factor graph are determined based on the map point coordinates. The optimal nodes and auxiliary nodes are used as state variables of the factor graph. The displacement constraint factor used to constrain the vehicle's position change in adjacent frames is obtained based on the time-domain differential carrier phase displacement, and the visual reprojection constraint factor used to constrain the consistency between the camera pose and the map point coordinates is obtained based on the reprojection error term. With the goal of minimizing the sum of displacement constraint factor residuals, visual reprojection constraint factor residuals, and physical gravity field geometric constraint factor residuals, an optimization objective function corresponding to the factor graph is constructed and used as a tightly coupled factor graph model.

7. The high-precision positioning method for distribution network vehicles based on BeiDou and monocular fusion as described in claim 6, characterized in that: The solution to the tightly coupled factor graph model, obtaining a fused solution including vehicle position, velocity, and attitude, includes: Anomaly checks are performed on the constraint factors of the tightly coupled factor graphical model, and the tightly coupled factor graphical model is corrected based on the anomaly check results. The modified tightly coupled factor graph model is solved iteratively using a nonlinear factor graph optimization algorithm to obtain the optimal state variables that minimize the sum of constraint factor residuals, thereby obtaining the vehicle's high-precision position, attitude, speed, and map point coordinates.

8. The high-precision positioning method for distribution network vehicles based on BeiDou and monocular fusion as described in claim 7, characterized in that: The step of performing anomaly checks on the constraint factors of the tightly coupled factor graphical model, and correcting the tightly coupled factor graphical model based on the anomaly check results, includes: Calculate the position jump variable observed by Beidou, and obtain the attitude / position change trend of the vehicle in the same period through visual electric field geometry. When the dissimilarity between the position jump variable and the attitude / position change trend is greater than the statistical threshold, it is determined to be the first anomaly of non-physical motion. Based on the first anomaly, the corresponding displacement constraint factor item is removed or its optimization weight is adjusted. The deviation of the current frame relative to the initial scale is calculated based on the relative displacement of the vehicle in adjacent frames. If the deviation of consecutive frames relative to the initial scale is greater than the deviation threshold and the deviation is linearly positively correlated with the vehicle's travel distance, it is determined to be a second anomaly. The global scale coefficient of the visual map is corrected based on the second anomaly, and the initial values ​​of map points are updated according to the corrected scale.

9. The high-precision positioning method for distribution network vehicles based on BeiDou and monocular fusion as described in claim 7, characterized in that: The modified tightly coupled factor graph model is iteratively solved using a nonlinear factor graph optimization algorithm to obtain the optimal state variables that minimize the sum of constraint factor residuals, thereby obtaining high-precision vehicle position, attitude, velocity, and map point coordinates, including: Based on the state variables of the current iteration, calculate the residuals of the displacement constraint factor, the visual reprojection constraint factor, and the physical gravity field geometric constraint factor, and summarize the current total objective function value. Calculate the Jacobian matrix of all residuals with respect to state variables, linearize the nonlinear optimization problem into an incremental equation, and solve the incremental equation to obtain the state update. The state variables are updated based on the current total objective function value and the state update amount, and the updated total objective function value is calculated to determine whether the iteration termination condition is met. If it is met, the optimal state variable is output. The optimal state variables are subjected to coordinate transformation, filtering, and outlier processing to obtain high-precision positioning results, smoothed attitude angles, velocity values, and power distribution network feature point clouds in the geographic coordinate system.

10. A high-precision positioning system for distribution network vehicles based on BeiDou and monocular fusion, characterized in that: The method for high-precision vehicle-mounted positioning of power distribution networks based on BeiDou and monocular fusion as described in any one of claims 1-9 includes: The carrier phase acquisition module is used to obtain the relative displacement vector of the vehicle in adjacent frames based on the time-domain differential carrier phase model. The depth point cloud acquisition module is used to virtualize two adjacent monocular images into a dynamic binocular stereo vision model based on the relative displacement vector, and to obtain the absolute depth information of image feature points. The gravity geometry constraint module is used to fit catenary lines in the image to obtain the physical geometric features of the plane where the electric field lines are located in order to establish physical gravity field geometric constraints. The global constraint optimization module is used to obtain constraint factors based on physical gravity field geometric constraints, temporal differential carrier phase model displacement constraints, and absolute depth information. It then constructs a tightly coupled factor graph model to optimize the vehicle's motion state, solves the tightly coupled factor graph model, and obtains a fusion solution that includes the vehicle's position, velocity, and attitude.

Images