An unmanned vehicle high-precision positioning method and system based on Beidou and an LSTM network

Abstract

This invention discloses a high-precision positioning method and system for unmanned vehicles based on BeiDou and LSTM networks, belonging to the field of autonomous driving and intelligent navigation. The method includes: acquiring and preprocessing BeiDou PPP-RTK, inertial sensor, and binocular vision data to construct a 32-dimensional unified feature vector; inputting the feature vector into a three-layer bidirectional LSTM-DYC network to output position and attitude error compensation values ​​and short-term state trend prediction values, and adaptively adjusting the compensation weights of BeiDou, inertial navigation, and vision based on the BeiDou signal-to-noise ratio; in the extended Kalman filter, using the short-term state trend prediction values ​​to correct the predictions of the basic physical model to obtain a hybrid prediction state, and then updating the system by using the error-compensated BeiDou PPP-RTK positioning results and the initial visual positioning values ​​as observation values ​​to output a high-precision fused positioning result. The system also includes a cloud navigation subsystem and a command subsystem. It achieves sub-decimeter-level robust positioning, reliable communication across the entire domain, and intelligent collaborative scheduling of multiple vehicles for unmanned vehicles in complex environments.

Description

Technical Field

[0001] This invention relates to the fields of autonomous driving and intelligent navigation technology, and in particular to a high-precision positioning method and system for unmanned vehicles based on BeiDou and LSTM networks. Background Technology

[0002] With the rise and development of intelligent connected vehicle technology, the requirements for accuracy, reliability and intelligence of the entire "positioning-navigation-command" chain of driverless vehicles have been greatly improved. The BeiDou Navigation Satellite System (BDS) has achieved global networking. Its ground-based augmentation service supports centimeter-level positioning for BeiDou terminals, while its real-time precise point positioning (PPP) service enables decimeter-level static positioning in network-free environments. The next-generation high-precision positioning technology (PPP-RTK) combines the advantages of both ground-based augmentation and real-time precise point positioning technologies, achieving centimeter-level positioning even in network-free environments. However, in complex environments, relying solely on the BeiDou Navigation Satellite System cannot guarantee the positioning accuracy and stability of terminals. LSTM networks, with their time-series data modeling capabilities, are widely used in trajectory prediction and error compensation. By leveraging the fundamental positioning capabilities of the BeiDou Navigation Satellite System and combining them with the powerful time-series data modeling and prediction capabilities of LSTM networks, positioning errors of BeiDou signals in complex environments can be effectively compensated, improving the positioning accuracy of unmanned vehicles in complex scenarios. Simultaneously, by analyzing environmental and historical data through LSTM networks, the robustness of the navigation system is enhanced, enabling dynamic prediction of obstacle movement trajectories. Combined with the BeiDou grid code, which divides the geographic space into multiple regular grids, more accurate path planning can be achieved, avoiding the lag problems caused by traditional navigation relying on high-precision maps. To ensure the stability of the navigation system under various complex conditions, a multi-mode communication fusion approach is adopted in the communication link, combining terrestrial 5G communication networks with satellite short message communication and other communication modes. When the 5G signal is interrupted in remote areas or signal blind spots, it automatically switches to satellite communication mode to ensure the smooth transmission of positioning data and command instructions in real time, greatly improving the reliability of unmanned vehicle communication and reducing the risk of loss of control. Through digital twin technology, a digital model that is completely consistent with the real unmanned vehicle and its operating environment can be created in virtual space, reflecting the unmanned vehicle's status, position, speed and other information in real time. Edge computing technology pushes data processing and analysis capabilities down to edge nodes closer to the data source, reducing data transmission latency and improving the system's response speed. After the integration of the two, the unmanned vehicle command system can quickly analyze and make decisions at the edge nodes based on the real-time acquired unmanned vehicle status and environmental data, realizing precise command and dispatch of unmanned vehicles.

[0003] Existing technical solutions:

[0004] ① A high-precision unmanned vehicle localization method based on LSTM cooperative map matching (patent CN202211382263.4 invention disclosure): A high-precision unmanned vehicle localization method based on LSTM cooperative map matching is proposed. When the localization signal is interrupted, the localization position at the next moment is predicted by the LSTM localization model. Based on the predicted localization position, the localization position is corrected by the cooperative map matching algorithm to output a more accurate localization position. However, edge computing technology is not introduced, so the real-time performance of the operation cannot be guaranteed.

[0005] ② A 5G-based V2X vehicle-to-everything (V2X) secure communication system and method (patent CN112055330A): This system transmits unmanned vehicle data to the cloud via 5G to achieve multi-vehicle monitoring and route scheduling. However, it faces the risk of communication interruption in areas without 5G coverage (such as mountainous areas), cannot guarantee all-time communication, and does not incorporate digital twin technology, making it impossible to realize virtual scenario simulation and command and dispatch.

[0006] ③ A method and system for UAV 3D path planning based on Beidou grid code (Patent CN202411005946.7: proposes a method for realizing spatial grid subdivision and path planning based on GeoSOT, but does not combine LSTM error compensation).

[0007] Autonomous vehicles face several challenges when operating in complex environments:

[0008] 1. Decreased positioning accuracy in complex environments: In urban canyons, tunnels, mountainous areas and other scenarios, BeiDou signals are easily blocked and interfered with by multipath effects, which seriously affects positioning accuracy. The error of traditional positioning technology can reach 1 to 5 meters or even higher, which is difficult to meet the lane-level positioning requirements of unmanned vehicles.

[0009] 2. Low robustness of navigation systems: Existing navigation systems are prone to positioning drift or failure when BeiDou signal is lost (e.g., in tunnels), and cannot dynamically predict the movement trajectory of obstacles, resulting in delayed path planning and compromised system stability. Traditional path planning relies on the real-time performance and accuracy of high-precision map data, but existing map update mechanisms are lagging and cannot meet the needs of dynamic scenarios.

[0010] 3. Limited intelligence in command and decision-making: Traditional command systems can only achieve real-time monitoring and lack predictive analysis based on historical data and environmental characteristics, making it difficult to cope with the collaborative scheduling of unmanned vehicles in emergency scenarios. Summary of the Invention

[0011] The technical solution of this invention to solve the above-mentioned technical problems is to provide a high-precision positioning method for unmanned vehicles based on BeiDou and LSTM networks, comprising:

[0012] Step S1, Data Acquisition and Preprocessing:

[0013] The system acquires BeiDou PPP-RTK observation data, inertial sensor data, and binocular vision data from the vehicle-mounted terminal, and performs preprocessing on each data to obtain BeiDou feature data, inertial feature data, and visual feature data.

[0014] Step S2, Unified Feature Vector Construction:

[0015] The preprocessed BeiDou feature data, inertial feature data, and visual feature data are fused to construct a multi-dimensional unified feature vector;

[0016] Step S3, LSTM dynamic error correction:

[0017] The multidimensional unified feature vector is input into a pre-trained LSTM dynamic error correction network, namely the LSTM-DYC network. The LSTM-DYC network outputs error compensation values ​​for localization and attitude. The output of the LSTM-DYC network includes two parts: the first part is the error compensation values ​​for position and attitude, and the second part is the short-term state trend prediction value used for state prediction correction.

[0018] Step S4, Adaptive weight adjustment:

[0019] Based on the signal-to-noise ratio (SNR) of the BeiDou signal, the compensation weights from the three data sources (BeiDou, inertial navigation, and vision) in the error compensation value output by the LSTM-DYC network are dynamically adjusted.

[0020] Step S5, Extended Kalman Filter (EKF) fusion:

[0021] Construct the state vector of the extended Kalman filter and perform the following fusion steps:

[0022] State prediction: A preliminary state prediction is performed based on the physical model of inertial navigation to obtain the basic predicted state; at the same time, the basic predicted state is corrected by the short-term state trend prediction value output by the LSTM-DYC network to obtain the hybrid predicted state.

[0023] Observation update: The BeiDou PPP-RTK positioning result corrected by the error compensation value and the initial visual positioning value are used together as observation values ​​to update the hybrid prediction state and output the final high-precision fusion positioning result.

[0024] Furthermore, the multidimensional unified feature vector constructed in step S2 is 32-dimensional, wherein:

[0025] The BeiDou feature data is 8-dimensional, including: BeiDou PPP-RTK positioning residual, number of visible satellites, average signal-to-noise ratio, ionospheric delay correction residual, and tropospheric delay correction residual;

[0026] The inertial characteristic data is 12-dimensional, including: three-axis acceleration increment, three-axis angular velocity increment, attitude angle change, and three-axis zero-bias compensation residual;

[0027] The visual feature data is 12-dimensional, including: the matching deviation between static road feature points and high-precision maps, the confidence level of traffic sign recognition, the binocular parallax calculation error, and the deviation between the initial visual positioning value and the BeiDou positioning result.

[0028] Furthermore, the LSTM-DYC network in step S3 includes a 3-layer bidirectional LSTM network, with a 32-dimensional input layer, a first hidden layer containing 128 bidirectional neurons, and a second hidden layer containing 64 bidirectional neurons; the first part of the output layer is a 6-dimensional vector, including three-dimensional position error compensation value and three-dimensional attitude error compensation value; the second part of the output layer is a 9-dimensional vector, including the position increment, velocity increment, and attitude angle increment at the next moment, as the short-term state trend prediction value.

[0029] Furthermore, the adaptive weight adjustment in step S4 specifically involves:

[0030] When SNR ≥ 30dB, the compensation weights for BeiDou, inertial navigation, and vision are set to 60%, 25%, and 15%, respectively.

[0031] When 20dB ≤ SNR < 30dB, the compensation weights for BeiDou, inertial navigation, and vision are set to 30%, 30%, and 40%, respectively.

[0032] When SNR < 20dB, the compensation weights for BeiDou, inertial navigation, and vision are set to 10%, 30%, and 60%, respectively.

[0033] Furthermore, the state prediction in step S5 further includes:

[0034] The mixed prediction state is calculated using the following formula. :

[0035] ; in, Based on the prediction of the state by the basic physical model Let λ(SNR) be the short-term state trend prediction value output by the LSTM-DYC network, and let λ(SNR) be the adaptive weighting factor calculated based on the signal-to-noise ratio (SNR), with the following functional form:

[0036] .

[0037] Furthermore, the extended Kalman filter in step S5 also employs an adaptive process noise covariance matrix. Its design is as follows:

[0038] ;

[0039] in, The basic process noise covariance matrix, The average signal-to-noise ratio is α, and β are preset parameters used to actively amplify process noise in low signal-to-noise ratio environments.

[0040] Furthermore, the method also includes a closed-loop feedback step: the final high-precision fusion positioning result is fed back to the data preprocessing stage in step S1 in real time, which is used to correct the multipath error model of Beidou PPP-RTK, update the zero-bias calibration coefficient of the inertial sensor, and adjust the feature point matching confidence threshold of visual positioning.

[0041] A high-precision positioning, navigation, and command system for unmanned vehicles based on BeiDou and LSTM networks, comprising:

[0042] The vehicle positioning subsystem includes the fusion positioning results output by the high-precision positioning method for unmanned vehicles based on BeiDou and LSTM networks as described above;

[0043] The cloud-based navigation subsystem generates a global reference path from the current location to the target point based on the fused positioning results and the BeiDou grid code.

[0044] The command subsystem is used to map the received real-time status of unmanned vehicles into the virtual scene through NTP time synchronization and BeiDou grid code spatial alignment; it uses the MARL algorithm to perform sandbox simulation of the movement of multiple vehicles in the virtual scene over the next 5 to 30 minutes, simulating and detecting potential path intersections and traffic congestion conflicts; for the detected conflicts, it automatically generates optimization strategies including local detours, speed adjustments, or priority yielding, and converts them into standardized command instructions; the command instructions are sent to the corresponding unmanned vehicles through the communication subsystem, and execution feedback is received to update the virtual scene, forming a closed loop of "simulation-optimization-issuance-execution-verification".

[0045] This invention provides a high-precision positioning method and system for unmanned vehicles based on BeiDou and LSTM networks. Through deep collaboration between the vehicle-mounted terminal, cloud platform, communication links, and command center, it effectively solves key technical problems such as positioning drift, navigation lag, communication interruption, and scheduling blind spots in complex environments. Compared with existing technologies, this invention has the following significant advantages:

[0046] 1. Significantly improved positioning accuracy and robustness:

[0047] Sub-decimeter-level high-precision positioning: By constructing a 32-dimensional unified feature vector (fusing BeiDou PPP-RTK, IMU, and binocular vision data), and using a 3-layer bidirectional LSTM-DYC network to output 6-dimensional positioning and attitude error compensation values, combined with extended Kalman filter (EKF) for nonlinear optimal fusion, the positioning accuracy can reach ≤0.6m (95% confidence) in complex scenarios such as urban canyons and tunnels. The positioning error within 3 seconds of tunnel signal loss is ≤0.15m, and the attitude angle accuracy is ≤0.1°.

[0048] Dynamic error adaptive compensation: An adaptive weight adjustment mechanism based on BeiDou signal-to-noise ratio (SNR) is introduced (60% BeiDou weight when SNR≥30dB; 40% visual weight when SNR20~30dB; 60% visual weight when SNR<20dB), which enables the system to maintain an error compensation accuracy of ≤0.04m even in scenarios with drastic fluctuations in signal quality, which is significantly better than the traditional fixed weight fusion method.

[0049] EKF Hybrid State Prediction: Innovatively, the short-term state trend prediction value output by the LSTM-DYC network is added to the EKF state prediction stage. The trust in the physical model and the data-driven model is dynamically adjusted by the adaptive weight factor λ (SNR). At the same time, an adaptive process noise covariance matrix Qadaptive is designed so that the filter can automatically amplify process noise in low signal-to-noise ratio environments and effectively suppress the positioning drift caused by IMU integral divergence.

[0050] 2. The stability and obstacle avoidance capabilities of the navigation system are significantly enhanced:

[0051] Predictive local path planning: Based on the LSTM-TP network, the system outputs the vehicle trajectory (30 trajectory points) and spatiotemporal cost field for the next 3 seconds. Combined with the improved A* algorithm, it introduces dynamic constraints for the autonomous vehicle (minimum turning radius ≥ 5m, maximum acceleration ≤ 2m / s²) and high-cost regions of dynamic obstacles (cost factor adjustable), generating a smooth local path (curvature ≤ 0.2m⁻¹) within 100m ahead. The obstacle avoidance response delay is ≤ 200ms, realizing a tightly coupled architecture of "prediction as planning", which significantly improves the autonomous obstacle avoidance capability in dynamic environments.

[0052] Cloud-vehicle multi-scale path fusion: The cloud generates a macro-scale global path (100 km level) based on Beidou grid code, and the vehicle-mounted terminal combines the local path (100 m level) to perform multi-scale matching and deviation correction. The path tracking accuracy is ≤0.5m and the control frequency is ≥10Hz, ensuring seamless connection between long-distance navigation and short-distance obstacle avoidance.

[0053] 3. Communication reliability achieves full coverage:

[0054] Intelligent switching of heterogeneous links: Constructing a converged communication architecture of "5G as primary and BeiDou short message backup," and designing an intelligent switching algorithm based on historical signal strength trend prediction (such as using the ARIMA model to predict the probability of future 5G availability). When the 5G link is predicted to be unavailable and continues to exceed the threshold, it switches to BeiDou short message in advance to ensure the stable transmission of positioning data and command instructions in areas without network coverage, such as mountainous areas, tunnels, and remote areas, and avoid control failures caused by sudden link interruptions.

[0055] Low latency and high reliability transmission: Hardware-level time synchronization (error ≤ 1ms) is achieved using the IEEE1588 PTP protocol. Combined with the historical latency compensation algorithm of the edge computing unit, the impact of visual data transmission lag (about 30ms) is eliminated, ensuring the timeliness of real-time fusion of multi-source data and command closed loop.

[0056] 4. Significantly improved intelligent collaborative scheduling and decision-making capabilities:

[0057] Digital twin virtual-real synchronization: The command center constructs a 3D virtual scene that maps to the physical world in a 1:1 ratio. Through NTP time synchronization (deviation ≤50ms) and BeiDou grid code spatial alignment (error ≤0.5m), it achieves millisecond-level real-time mapping of multiple unmanned vehicle states, with virtual-real synchronization accuracy ≤50ms, providing a high-fidelity digital foundation for global monitoring and scheduling.

[0058] Multi-agent Sandbox Simulation: Employing a Multi-agent Reinforcement Learning (MARL) algorithm, the simulation measures the movement trajectories of multiple vehicles over the next 5-30 minutes in a virtual environment. It automatically detects conflicts such as path intersections, congestion, and violations, and generates optimization strategies (detours, speed adjustments, yielding, etc.). The simulation results are converted into standardized commands and issued to the corresponding autonomous vehicles within 200ms, forming a closed loop of "simulation → optimization → command issuance → execution → verification," significantly improving the safety and scheduling efficiency of group collaboration.

[0059] Offline training + online inference architecture: The MARL model is trained offline on a high-performance server cluster in the command center. After training, the parameters are fixed and deployed to the real-time inference server. The vehicle terminal is only responsible for its own vehicle positioning and local planning, and does not participate in multi-vehicle collaborative computing. This ensures the real-time performance of the inference while avoiding the computational burden on the vehicle terminal.

[0060] 5. System closed-loop feedback enables continuous online optimization:

[0061] Location result closed-loop feedback: The final fused location result is fed back to the data preprocessing stage in real time.

[0062] The multipath error model of BeiDou PPP-RTK was revised, which improved the accuracy of ionospheric / tropospheric delay correction by about 15%.

[0063] After updating the IMU zero-bias calibration coefficients, the drift error decreased by approximately 40% after 300 seconds of continuous operation.

[0064] Adjusting the confidence threshold for visual localization feature point matching reduces the false match rate by approximately 20%.

[0065] Prediction and planning are linked in a closed loop: the predicted trajectory output by the LSTM-TP network is directly used to construct the cost field of local path planning, while the new observation data after planning is executed is used to update the training samples of the LSTM network, forming a continuous learning closed loop of "perception-localization-prediction-planning-execution-feedback", and the system performance continuously improves itself over time.

[0066] 6. Cloud-based navigation efficiency and scalability are significantly enhanced:

[0067] High-efficiency indexing using BeiDou geogrid codes: Using BeiDou geogrid codes (GeoSOT) as the spatial index key for road networks, the efficiency of road network data retrieval is improved by 3 to 5 times compared to traditional latitude and longitude retrieval. A two-level path search strategy (16 levels of coarse-grained direction planning + 18 to 32 levels of fine-grained lane-level path generation) balances computational efficiency and path accuracy.

[0068] An improved A* algorithm incorporating LSTM traffic prediction: Future traffic flow parameters predicted by LSTM are introduced as heuristic weights into the path cost function, enabling the global path to proactively avoid congestion. The generated global path is smoothed using B-spline and redundant nodes are removed (compression rate approximately 50%), and a 16-level grid code corresponding to each 1km is embedded, facilitating rapid matching of local maps by the vehicle-mounted device.

[0069] Dynamic route update mechanism: Route updates are automatically triggered by location deviation (>20m), mileage (every 1km), or emergencies (accidents within 5km ahead), ensuring that global guidance adapts to traffic changes and road control in real time.

[0070] 7. Overall technical specifications are superior to existing solutions:

[0071] Positioning performance: Positioning error in urban canyon scenes ≤0.6m (95% confidence level), error within 3 seconds of tunnel signal interruption ≤0.15m, attitude angle accuracy ≤0.1°.

[0072] Real-time performance: LSTM-DYC network inference latency ≤80ms, EKF fusion frequency 10Hz, local path planning output frequency 10Hz, obstacle avoidance response latency ≤200ms.

[0073] Reliability: Intelligent switching between 5G and BeiDou short message ensures full-domain communication availability; the synchronization deviation between digital twin and real systems is ≤50ms; and the response time for multi-vehicle simulation is ≤3s.

[0074] Scalability: Supports smooth expansion from single unmanned vehicles to large-scale unmanned vehicle platoons. The Beidou grid code multi-scale segmentation system can adapt to different application scenarios, from centimeter-level precise parking to hundred-kilometer-level cross-regional scheduling.

[0075] In summary, this invention achieves a comprehensive breakthrough in the positioning accuracy, navigation stability, communication reliability, and group collaboration efficiency of unmanned vehicles in complex environments through an innovative four-in-one architecture of "high-precision robust positioning on the vehicle end + efficient global navigation in the cloud + intelligent heterogeneous communication + virtual-real collaborative scheduling in the command center". It can be widely used in fields such as autonomous driving, intelligent transportation, emergency command, and unmanned logistics, and has significant technological advancement and industrial application value. Attached Figure Description

[0076] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0077] Figure 1 This is a schematic diagram of the structure of the high-precision positioning method for unmanned vehicles based on BeiDou and LSTM networks described in this invention. Detailed Implementation

[0078] This invention proposes a high-precision positioning method and system for unmanned vehicles based on BeiDou and LSTM networks, aiming to provide a high-precision positioning method and intelligent command system for unmanned vehicles in complex environments (such as urban canyons, tunnels, and areas without network coverage).

[0079] The high-precision positioning method for unmanned vehicles based on BeiDou and LSTM networks proposed in this invention will be described below in specific embodiments:

[0080] Example 1:

[0081] In the technical solution of this embodiment, such as Figure 1 As shown, a high-precision positioning method for unmanned vehicles based on BeiDou and LSTM networks includes the following steps:

[0082] Step S1, Data Acquisition and Preprocessing:

[0083] The system acquires BeiDou PPP-RTK observation data, inertial sensor data, and binocular vision data from the vehicle-mounted terminal, and performs preprocessing on each data to obtain BeiDou feature data, inertial feature data, and visual feature data.

[0084] Step S2, Unified Feature Vector Construction:

[0085] The preprocessed BeiDou feature data, inertial feature data, and visual feature data are fused to construct a multi-dimensional unified feature vector;

[0086] Step S3, LSTM dynamic error correction:

[0087] The multidimensional unified feature vector is input into a pre-trained LSTM dynamic error correction network, namely the LSTM-DYC network. The LSTM-DYC network outputs error compensation values ​​for localization and attitude. The output of the LSTM-DYC network includes two parts: the first part is the error compensation values ​​for position and attitude, and the second part is the short-term state trend prediction value used for state prediction correction.

[0088] Step S4, Adaptive weight adjustment:

[0089] Based on the signal-to-noise ratio (SNR) of the BeiDou signal, the compensation weights from the three data sources (BeiDou, inertial navigation, and vision) in the error compensation value output by the LSTM-DYC network are dynamically adjusted.

[0090] Step S5, Extended Kalman Filter (EKF) fusion:

[0091] Construct the state vector of the extended Kalman filter and perform the following fusion steps:

[0092] State prediction: A preliminary state prediction is performed based on the physical model of inertial navigation to obtain the basic predicted state; at the same time, the basic predicted state is corrected by the short-term state trend prediction value output by the LSTM-DYC network to obtain the hybrid predicted state.

[0093] Observation update: The BeiDou PPP-RTK positioning result corrected by the error compensation value and the initial visual positioning value are used together as observation values ​​to update the hybrid prediction state and output the final high-precision fusion positioning result.

[0094] Furthermore, the multidimensional unified feature vector constructed in step S2 is 32-dimensional, wherein:

[0095] The BeiDou feature data is 8-dimensional, including: BeiDou PPP-RTK positioning residual, number of visible satellites, average signal-to-noise ratio, ionospheric delay correction residual, and tropospheric delay correction residual;

[0096] The inertial characteristic data is 12-dimensional, including: three-axis acceleration increment, three-axis angular velocity increment, attitude angle change, and three-axis zero-bias compensation residual;

[0097] The visual feature data is 12-dimensional, including: the matching deviation between static road feature points and high-precision maps, the confidence level of traffic sign recognition, the binocular parallax calculation error, and the deviation between the initial visual positioning value and the BeiDou positioning result.

[0098] Furthermore, the LSTM-DYC network in step S3 includes a 3-layer bidirectional LSTM network, with a 32-dimensional input layer, a first hidden layer containing 128 bidirectional neurons, and a second hidden layer containing 64 bidirectional neurons; the first part of the output layer is a 6-dimensional vector, including three-dimensional position error compensation value and three-dimensional attitude error compensation value; the second part of the output layer is a 9-dimensional vector, including the position increment, velocity increment, and attitude angle increment at the next moment, as the short-term state trend prediction value.

[0099] Furthermore, the adaptive weight adjustment in step S4 specifically involves:

[0100] When SNR ≥ 30dB, the compensation weights for BeiDou, inertial navigation, and vision are set to 60%, 25%, and 15%, respectively.

[0101] When 20dB ≤ SNR < 30dB, the compensation weights for BeiDou, inertial navigation, and vision are set to 30%, 30%, and 40%, respectively.

[0102] When SNR < 20dB, the compensation weights for BeiDou, inertial navigation, and vision are set to 10%, 30%, and 60%, respectively.

[0103] Furthermore, the state prediction in step S5 further includes:

[0104] The mixed prediction state is calculated using the following formula. :

[0105] ; in, Based on the prediction of the state by the basic physical model Let λ(SNR) be the short-term state trend prediction value output by the LSTM-DYC network, and let λ(SNR) be the adaptive weighting factor calculated based on the signal-to-noise ratio (SNR), with the following functional form:

[0106] .

[0107] Furthermore, the extended Kalman filter in step S5 also employs an adaptive process noise covariance matrix. Its design is as follows:

[0108] ;

[0109] in, The basic process noise covariance matrix, The average signal-to-noise ratio is α, and β are preset parameters used to actively amplify process noise in low signal-to-noise ratio environments.

[0110] Furthermore, the method also includes a closed-loop feedback step: the final high-precision fusion positioning result is fed back to the data preprocessing stage in step S1 in real time, which is used to correct the multipath error model of Beidou PPP-RTK, update the zero-bias calibration coefficient of the inertial sensor, and adjust the feature point matching confidence threshold of visual positioning.

[0111] This embodiment takes "BeiDou PPP-RTK high-precision positioning as the core, LSTM optimization as the support, and local reference path generation as the goal" as the core. It uses BeiDou PPP-RTK to provide reference positioning, inertial navigation to compensate for signal interruption defects, and vision to provide environmental constraints. It relies on edge computing to achieve real-time fusion of multi-source data, and simultaneously completes positioning error correction, short-term position prediction and local reference path planning, forming a closed loop of "perception-positioning-prediction-planning" on the vehicle end.

[0112] Step 1, Data Acquisition and Preprocessing:

[0113] 1. Time Synchronization: Hardware-level time alignment is achieved using the IEEE1588PTP protocol. Based on BeiDou time, time synchronization pulses with 10ms intervals are output to the IMU and visual sensors to ensure that the timestamp error of all data is ≤1ms. For visual data (transmission delay of about 30ms), the "historical delay statistics + pre-compensation algorithm" of the edge computing unit is used to correct and compensate the timestamp at the moment of receiving data to eliminate the impact of transmission lag.

[0114] 2. Spatial Synchronization: Eye-in-Hand calibration (a core technology in robot vision used to establish a fixed transformation relationship between the camera coordinate system and the robot's end effector coordinate system; essentially, it eliminates pose deviations between the camera and the robotic arm's end effector through mathematical modeling and experimental calibration, ensuring the robot can accurately convert visual perception information into its own motion control commands, achieving high-precision coordination of "vision-guided actions") determines the transformation relationship between each sensor and the unmanned vehicle's coordinate system. For example, the offset between the BeiDou antenna phase center and the unmanned vehicle's centroid is predetermined, and the positioning results are compensated and transformed to the unmanned vehicle's coordinate system in real time. The relationship between the visual sensors and the unmanned vehicle's coordinate system is obtained through checkerboard calibration, acquiring intrinsic parameters (focal length, distortion coefficients) and extrinsic parameters (rotation matrix R, translation vector T), ensuring that visual feature points are accurately mapped to the world coordinate system.

[0115] 3. BeiDou PPP-RTK data preprocessing:

[0116] The system receives BeiDou PPP-RTK differential correction data and combines it with observation data from the vehicle-mounted BeiDou receiver (pseudorange, carrier phase, Doppler shift) to perform real-time precise point positioning calculations. Preprocessing mainly includes:

[0117] Outlier removal: Outliers in the observations are removed based on the chi-square test to prevent outliers from contaminating subsequent fusion.

[0118] Cycle slip detection and repair: The carrier phase cycle slip is jointly detected by GF (geometrically independent) combination and MW (Melbourne-Wübbena) combination, and repaired by LAMBDA method to ensure continuous positioning.

[0119] Multipath error reduction: The current observation is corrected by using the sidereal-day filtering method based on the multipath repetition characteristics of the previous period.

[0120] Output: Outputs the initial positioning values ​​(P_bds), positioning residuals (x / y / z direction residuals), satellite visibility number, average signal-to-noise ratio (SNR), ionospheric / tropospheric delay correction residuals, etc. in the ENU coordinate system, for use in subsequent feature vector construction.

[0121] IMU data preprocessing is used to suppress inertial measurement errors and mainly includes:

[0122] Zero bias removal: Subtract the constant zero bias of the accelerometer and gyroscope from the average of the samples taken during the stationary period.

[0123] Temperature compensation: Based on a pre-calibrated temperature-zero bias curve, real-time temperature drift correction is performed on the sensor output.

[0124] Gravity compensation: The gravitational acceleration component is subtracted from the accelerometer specific force to obtain the true motion acceleration in the carrier coordinate system.

[0125] Noise filtering: Low-pass filters (such as Butterworth filters, with the cutoff frequency set according to the vehicle's vibration characteristics) are used to suppress high-frequency vibration noise.

[0126] Output: Outputs the compensated and filtered three-axis acceleration increment (10ms integral value), three-axis angular velocity increment, attitude angle change, and zero-bias compensation residuals of the accelerometer / gyroscope, for use in subsequent feature vector construction.

[0127] Visual positioning data preprocessing:

[0128] Visual positioning data preprocessing is used to extract road environment constraint features from binocular camera images, mainly including:

[0129] Image distortion correction: Radial and tangential distortion correction is performed on the original image using pre-calibrated camera intrinsic parameters (focal length, distortion coefficients).

[0130] Stereo matching and disparity calculation: Epipolar correction is performed on the left and right eye images, and the pixel-level disparity is calculated using the semi-global block matching (SGBM) algorithm to generate a depth map.

[0131] Feature point extraction and matching: ORB or SuperPoint feature points are used to match the pre-stored high-precision map on the vehicle (or the visual feature map constructed offline) to obtain the initial value of visual 3D positioning.

[0132] Semantic feature extraction: Real-time detection of semantic features such as lane lines, traffic signs, and curbs using lightweight deep learning models (such as YOLOv5+ lane line detection branch), and calculation of the matching deviation between the semantic features and the corresponding semantic features in the map.

[0133] Output: The output includes the initial value of visual 3D positioning (P_vis), the matching deviation between static road feature points and high-precision maps (x / y / z directions), the confidence level of traffic sign recognition, the binocular parallax calculation error, and the deviation between the initial value of visual positioning and the initial value of Beidou positioning, which are used for subsequent feature vector construction.

[0134] Step 2, Unified Feature Vector Construction:

[0135] The vehicle-mounted edge computing unit transforms the preprocessed BeiDou PPP-RTK data, IMU data, and visual positioning data from "2. Multi-source data preprocessing" into a 32-dimensional unified feature vector. The dimensions of the feature vector are defined as follows:

[0136] BeiDou features (8 dimensions): BeiDou PPP-RTK positioning residual (x / y / z direction), number of visible satellites, average SNR, ionospheric delay correction residual, tropospheric delay correction residual;

[0137] IMU features (12 dimensions): IMU three-axis acceleration increment (10ms integral value), IMU three-axis angular velocity increment, IMU attitude angle (roll / pitch / heading) change, IMU three-axis zero bias compensation residual;

[0138] Visual features (12 dimensions): matching deviation between static road feature points and high-precision maps (x / y / z directions), confidence level of traffic sign recognition, binocular parallax calculation error, and deviation between the initial value of visual 3D positioning and the result of BeiDou 3D positioning.

[0139] Step 3, LSTM dynamic error correction:

[0140] LSTM-DYC (Long Short-Term Memory Dynamic Yaw Correction Network) model inference: The LSTM-DYC network outputs error compensation values ​​for position (x / y / z) and attitude (roll / pitch / head), therefore, "dynamic error correction" is the core. The model is deployed on the vehicle's edge computing unit to output positioning and attitude error compensation values.

[0141] Model Structure: A 3-layer bidirectional LSTM-DYC network is used. The input layer is a 32-dimensional feature vector. The first hidden layer has 128 bidirectional neurons, and the second hidden layer has 64 bidirectional neurons. The output layer consists of two parts: the first part is a 6-dimensional vector (including BeiDou positioning error compensation values ​​x, y, z, in meters; IMU attitude error compensation values ​​roll / pitch / heading, in degrees), using ReLU (Rectified Linear Unit, one of the most commonly used activation functions in deep learning. Input: any real number x, output: if x>0, output (x); otherwise, output 0); the second part is a 9-dimensional short-term state trend prediction value. These are used for EKF state prediction correction. The 9-dimensional short-term state trend prediction values ​​are used in subsequent EKF mixed state predictions. It includes: location increment Speed ​​increment Attitude angle increment These increments represent the predicted state changes from the current time to the next time (i.e., 10ms later), used to correct the prediction results of the EKF fundamental physical model. This 9-dimensional state trend value is directly output by a dedicated prediction head set at the end of the LSTM-DYC network, without needing to be obtained through error compensation value differentiation or indirect calculation, and directly participates in the EKF mixed state prediction as a data-driven state trend term.

[0142] Model Training: The training set contains over 150,000 frames of multi-scene data (urban canyons, tunnels, mountainous areas, rainy / foggy days), with the localization ground truth from binocular cameras used as the baseline for error labeling. The Adam optimizer was used for training (Adaptive Moment Estimation is a widely used optimization algorithm that combines the advantages of momentum and adaptive learning rates, suitable for most deep learning tasks. The default learning rate is 0.001). After training, the localization error compensation accuracy on the validation set is ≤0.04m, and the pose error compensation accuracy is ≤0.02°.

[0143] Real-time inference and adaptive weight adjustment: The on-board edge computing unit inputs a set of 32-dimensional feature vectors to the LSTM-DYC network model every 10ms (or a predetermined first cycle), with an inference latency ≤80ms; the compensation weights are dynamically adjusted according to the BeiDou SNR: when SNR≥30dB, the BeiDou signal quality is good, and BeiDou multipath error compensation is the main method, with the compensation weights for BeiDou, IMU, and vision set to 60%, 25%, and 15%, respectively; when 20dB≤SNR<30dB, the BeiDou signal is subject to some interference, and the IMU drift compensation weight is increased, with the compensation weights for BeiDou, IMU, and vision set to 30%, 30%, and 40%, respectively; when SNR<20dB, the BeiDou signal is severely attenuated, and compensation is enhanced based on visual feature deviation, with the compensation weights for BeiDou, IMU, and vision set to 10%, 30%, and 60%, respectively.

[0144] Fusion localization using Extended Kalman Filter (EKF) and LSTM-DYC network:

[0145] This invention employs Extended Kalman Filter (EKF) to achieve deep fusion of multi-source sensor data and LSTM-DYC network results. Unlike traditional EKF, the core innovation of this invention lies in integrating the output of the LSTM-DYC network not only as an observation but also deeply into the state prediction stage of the EKF. Furthermore, it introduces an environment-aware adaptive noise covariance adjustment mechanism, thereby significantly improving the filter's robustness in scenarios with drastic signal fluctuations. The specific implementation steps are as follows:

[0146] a) Definition of State Vector: Define the state vector of the autonomous vehicle. ,in:

[0147] The three-dimensional position in the ENU coordinate system;

[0148] The three-dimensional velocity in the ENU coordinate system;

[0149] A unit quaternion representing attitude;

[0150] Zero bias of the IMU accelerometer;

[0151] This is the zero bias of the IMU gyroscope.

[0152] b) Innovative state prediction (computing Xpred and Ppred): Traditional EKF relies solely on the physical model of the IMU for state prediction, while this invention proposes a hybrid state transition model that combines data-driven and model-driven approaches.

[0153] Basic physics model predictions: First, based on IMU preprocessed data (specific force f after removing the influence of gravity)... b and angular velocity ω b Preliminary state prediction is performed using a discretized inertial navigation mechanical arrangement formula:

[0154] Speed ​​prediction: ;

[0155] Location prediction: ;

[0156] Attitude prediction: ;

[0157] Zero-biased prediction: ;in, It is composed of quaternions Constructing the Directional Cosine Matrix (DCM). The gravity vector in the ENU coordinate system. This is the IMU sampling period.

[0158] LSTM-DYC Trend Correction Term: Obtain the short-term trend prediction of the state at the next time step from the LSTM dynamic error correction model deployed on the edge computing unit. (This trend prediction can be obtained by differentiating the error compensation value of the LSTM output or by using a dedicated prediction head.)

[0159] Hybrid state prediction: Combining the above two methods to form the final predicted state vector. : ;in, It is an adaptive weighting factor based on the BeiDou signal-to-noise ratio (SNR), defined as: When the SNR is high (signal is good). Approaching 0, It mainly relies on the physical model; when the SNR is low (poor signal strength). Approaching 1, By placing greater trust in the trend predictions of LSTM, positioning drift caused by IMU integral divergence can be effectively suppressed.

[0160] Covariance matrix prediction: Predicting the covariance matrix Calculated using the following formula: ;in, It is the Jacobian matrix of the mixed state transition function with respect to the state vector X; It is the adaptive process noise covariance matrix, and its design is as follows: This design ensures that process noise is actively amplified in low SNR environments, causing the filter to reduce its trust in the IMU model and instead rely more on subsequent observation updates.

[0161] Observation equation construction: The BeiDou PPP-RTK positioning results compensated by the LSTM dynamic error correction model, along with the initial visual positioning values, are used as observation values. The observation equation is: in For the observation matrix, To observe noise.

[0162] Kalman gain calculation and state update: based on the prediction covariance matrix Observation matrix The Kalman gain K is calculated using the observation noise covariance matrix R (which can also be dynamically adjusted based on the real-time confidence levels of each data source), and the state vector X and covariance matrix P are updated. The final obtained X is the high-precision fusion positioning result of this invention.

[0163] LSTM-TP Network Localization Prediction and Local Path Generation, Short-Term Localization Prediction in LSTM-TP Networks:

[0164] A multi-source heterogeneous input LSTM-TP prediction model (Long Short-Term Memory Trajectory Prediction Network) with integrated attention mechanism is designed, enabling it not only to predict vehicle trajectories but also to proactively perceive dynamic environmental risks. Unlike the aforementioned LSTM-DYC network model used for dynamic error correction, the LSTM-TP network model in this step is specifically designed for future trajectory prediction, and its input feature design is as follows:

[0165] Multimodal input feature construction: The model's input is a sequence of historical data with a time step of 60 (i.e., 600ms). The input feature vector for each time step... It is 32-dimensional, and its specific structure is as follows:

[0166] Vehicle motion state (15 dimensions):

[0167] .

[0168] Contextual information (17 dimensions):

[0169] Visual obstacle features (9-dimensional): Relative positions of the three nearest dynamic obstacles (such as vehicles and pedestrians) in front. and relative velocity (i=1,2,3).

[0170] Local map constraints (5D): lane centerline curvature, distance between left and right lane lines, road width, and slope within 50m ahead.

[0171] Cloud-based command guidance (3D): This refers to the angle deviation, distance deviation, and suggested speed deviation of the macroscopic path point from the cloud navigation subsystem relative to the current vehicle heading. In scenarios such as the initial system startup (before the first global cloud path is received) or communication interruptions (both 5G and BeiDou short message services are unavailable), the "Cloud-based command guidance (3D)" feature of the LSTM-TP network will automatically be set to default values: angle deviation = 0° (maintain current heading), distance deviation = 0m (no specified target point), and suggested speed deviation = 0km / h (maintain current vehicle speed). At this time, the LSTM-TP network performs short-term trajectory prediction based entirely on the vehicle's motion state (15D) and environmental context information (visual obstacle features and local map constraints in 17D), relying entirely on local perception. It can still output the predicted trajectory for the next 3 seconds, ensuring the autonomous vehicle has basic autonomous obstacle avoidance and lane-keeping capabilities even without cloud guidance. Once communication is restored or the first cloud path is successfully issued, the cloud-based command guidance feature is immediately updated to the actual values, and the LSTM-TP network seamlessly switches to normal fusion prediction mode.

[0172] Network architecture with integrated attention mechanism:

[0173] Encoder: A two-layer unidirectional LSTM-TP network (32 neurons per layer) is used as the encoder to process the above 32-dimensional input sequence and extract the deep temporal features of the historical state.

[0174] Attention Layer: A Bahdanau attention layer is introduced after the encoder. This layer computes a weight vector. This is used to measure the importance of each time step in the input sequence. The weight calculation formula is: ; ;in, It is the hidden state of the encoder at time step t. It is the hidden state of the decoder at the previous moment. These are learnable parameters. This mechanism allows the model to automatically focus on the moments most critical to predicting future trajectories (e.g., before a curve or when approaching an obstacle).

[0175] Decoder: A single-layer unidirectional LSTM-TP network (32 neurons) is used as the decoder. Combined with the attention-weighted context vector, the predicted trajectory for the next 3 seconds (30 points) is generated step by step.

[0176] Output layer: The output layer is 9-dimensional and directly corresponds to the complete state at a future time step. .

[0177] Rolling Prediction and Cost Field Output:

[0178] Rolling Prediction: A rolling prediction strategy is used to generate a complete 30-point trajectory.

[0179] Cost Field Generation: In addition to outputting trajectory points, the model also outputs a predicted spatio-temporal cost field. This cost field is a three-dimensional grid (covering a 100m x 20m area in front of the vehicle, spanning 3 seconds), where the value of each grid cell represents the probability of a collision or high-risk event occurring at that spatio-temporal point. This cost field will be directly used as part of the heuristic function for improving the A* algorithm, realizing a tightly coupled architecture of "prediction as planning".

[0180] Local reference path generation:

[0181] Based on the localization prediction results of the LSTM-TP network and the visual obstacle recognition results (identified by visual matching), an improved A* algorithm is used to generate local paths. The A* algorithm (A-Star Algorithm) is a heuristic path search algorithm that efficiently finds the optimal path (usually the shortest path) from the starting point to the destination by combining the "actual cost from the starting point to the current node" and the "estimated cost from the current node to the destination." The core of the A* algorithm is to guide the search direction through an evaluation function, prioritizing the expansion of nodes "more likely to be close to the destination." The evaluation function is defined as: f(n) = g(n) + h(n), where: f(n) is the total estimated cost of node n (the total cost from the starting point through n to the destination); g(n) is the actual cost from the starting point to node n (the confirmed path length with no estimation error); h(n) is the estimated cost from node n to the destination (heuristic information, which needs to be designed to "not overestimate the actual cost" to ensure optimality).

[0182] Local reference path generation steps:

[0183] Path constraints: Based on visually extracted features such as lane lines and curbs, path constraints are set, such as lane width ≥ 3.5m and prohibited lane crossing areas;

[0184] Obstacle region modeling: The trajectories predicted by the LSTM-TP network for visually recognized dynamic obstacles, such as pedestrians and other autonomous vehicles, are transformed into obstacle regions. A high cost coefficient is set in the path cost function (default cost factor = 3.0, which can be dynamically adjusted according to the obstacle type and relative speed).

[0185] Path generation and optimization: The improved A* algorithm introduces autonomous vehicle dynamic constraints, such as minimum turning radius ≥ 5m and maximum acceleration ≤ 2m / s², and searches to generate a sequence of local path points within a range of 100m ahead (or a dynamic distance associated with the current vehicle speed) (generating 1 path point every 1m); the path is smoothed by B-spline curves so that the path curvature is ≤ 0.2m-1;

[0186] Path output: The edge computing unit outputs local path data (including path point coordinates, heading angle, suggested speed, etc.) once every 100ms, and sends it to the vehicle control system via CANFD (Controller Area Network with Flexible Data-Rate) bus.

[0187] Results output and closed-loop feedback:

[0188] Output results: Positioning results output: output frequency 10Hz, coordinate system is ENU (East-North-Sky) local coordinate system (coordinate transformation with the plane coordinate system xyz is required, the coordinate transformation formula is not described in this invention); positioning accuracy: urban canyon scene ≤0.6m (95% confidence), tunnel scene (within 3s of Beidou signal loss) ≤0.15m, attitude angle accuracy ≤0.1°;

[0189] Path result output: output frequency 10Hz, path point spacing 1m, obstacle avoidance response delay ≤200ms.

[0190] Closed-loop feedback: The fused positioning results are fed back to each preprocessing module in real time to achieve iterative error optimization.

[0191] Feedback is sent to the BeiDou PPP-RTK data preprocessing module: Multipath error model parameters are corrected based on historical results. Simulation calculations show an approximately 15% improvement in ionospheric and tropospheric delay correction accuracy.

[0192] Feedback is sent to the IMU data preprocessing module: the IMU zero-bias calibration coefficients are updated based on the attitude error compensation values. Simulation calculations show that after 300 seconds of continuous operation, the IMU drift error decreases by approximately 40%.

[0193] Feedback is sent to the visual positioning data preprocessing module: based on the deviation between the initial visual positioning values ​​and the fused positioning results, the feature point matching confidence threshold is adjusted. Simulation calculations show a reduction in the false matching rate of approximately 20%.

[0194] Example 2:

[0195] A high-precision positioning, navigation, and command system for unmanned vehicles based on BeiDou and LSTM networks, comprising:

[0196] The vehicle positioning subsystem includes the fusion positioning results output by the high-precision positioning method for unmanned vehicles based on BeiDou and LSTM networks as described above;

[0197] The cloud-based navigation subsystem generates a global reference path from the current location to the target point based on the fused positioning results and the BeiDou grid code.

[0198] The command subsystem is used to map the received real-time status of unmanned vehicles into the virtual scene through NTP time synchronization and BeiDou grid code spatial alignment; it uses the MARL algorithm to perform sandbox simulation of the movement of multiple vehicles in the virtual scene over the next 5 to 30 minutes, simulating and detecting potential path intersections and traffic congestion conflicts; for the detected conflicts, it automatically generates optimization strategies including local detours, speed adjustments, or priority yielding, and converts them into standardized command instructions; the command instructions are sent to the corresponding unmanned vehicles through the communication subsystem, and execution feedback is received to update the virtual scene, forming a closed loop of "simulation-optimization-issuance-execution-verification".

[0199] The vehicle-mounted positioning subsystem, cloud-based navigation subsystem, and command subsystem are based on converged communication transmission and intelligent link switching. Based on the converged communication architecture and intelligent link switching algorithm, a heterogeneous communication architecture of 5G network and BeiDou short message service is constructed to ensure reliable transmission of positioning data (including fused positioning results and status information), while supporting downlink feedback of cloud commands and global navigation information. The intelligent communication switching algorithm predicts signal strength based on historical signal data, dynamically selects the optimal link, and formulates a switching decision strategy based on the "availability probability" of the output 5G link and the "predicted transmission success rate" of the BeiDou short message service to ensure stable transmission of positioning results.

[0200] The cloud-based navigation subsystem combines real-time location and BeiDou grid code to generate a global reference path. The core of cloud-based navigation computation is based on the real-time positioning results and BeiDou grid code uploaded by the vehicle terminal, combined with a high-precision road network database and real-time traffic information to generate a global reference path from the current location to the target point. This path is typically a long-distance path that is updated on a scale of hundreds of kilometers or every minute. This process needs to consider path optimization (shortest time / distance), feasibility (compliance with road rules), and dynamic adaptability (responding to traffic changes); real-time location data analysis: the fused positioning results uploaded by the vehicle terminal are received by the cloud and converted to CGCS2000 coordinates, P=(lat,lon,h), with an accuracy typically at the centimeter to decimeter level; the positioning results include auxiliary information such as positioning timestamp (UTC millisecond level), positioning accuracy (95% confidence radius, unit: m), vehicle speed v (km / h), and heading angle (ψ) (unit: °, true north is 0°, increasing clockwise). The BeiDou grid code is a discretized, multi-scale regional location identification system developed based on the GeoSOT geospatial partitioning theory. It divides the Earth's surface into rectangular grids according to accuracy levels (from level 1 to 32, accuracy from 1000km to 1cm), with each grid corresponding to a unique code composed of numbers and letters, used for efficient indexing of geographic information.

[0201] In route planning, the core function of BeiDou grid codes is to quickly locate spatial ranges and associate road network data. First, grid code conversion is performed. The real-time location P is converted into a BeiDou grid code, simultaneously generating multi-level grid codes for the surrounding area. Second, spatial indexing is implemented. The cloud-based road network database uses BeiDou grid codes as spatial index keys, enabling rapid querying of road network topology data (road, intersection, lane connection relationships), static attributes (road grade, speed limits, traffic restrictions), and dynamic attributes (real-time congestion coefficient, construction information) between the current location and target points. Compared to traditional latitude and longitude retrieval, BeiDou grid code-based retrieval efficiency is improved by 3-5 times. User input (e.g., "XX Building") or designation by the command subsystem (e.g., "XX Rescue Point") is converted into target coordinates Pt=(latt,lont,ht) via geocoding, generating a corresponding BeiDou grid code. User preferences (shortest time / shortest distance / avoiding highways), autonomous vehicle attributes (vehicle length / width / maximum speed, e.g., trucks needing to avoid height-restricted sections), and real-time traffic control information (e.g., temporary road closures) are used as cost function parameters for path planning. Based on the BeiDou grid codes corresponding to the starting and target points, the cloud extracts complete road network data within the coverage area of ​​the "starting point-target point" and integrates multi-source information to construct a weighted road network graph.

[0202] Determining the scope of basic road network topology extraction: Using the starting point P0 and the target point Pt as diagonals, extend the buffer area by 10% (e.g., extend a 200km path by 20km). Index all road elements within this area using Beidou grid codes, including nodes, intersections, road endpoints, edges, and road segments connecting two nodes.

[0203] Dynamic traffic information fusion: Real-time traffic data is linked through Beidou grid codes, with a sampling frequency of 3 minutes / time, and the "travel cost" at the roadside is updated based on the congestion coefficient and event impact;

[0204] Special area handling of restricted areas: Restricted areas (such as traffic-limited areas, controlled areas, etc.) are matched by Beidou grid code, and the corresponding nodes / edges are marked as "unreachable" on the road network map; for complex intersections such as interchanges and roundabouts, the internal lane connection relationship is refined to ensure route feasibility.

[0205] Global reference path generation algorithm:

[0206] Based on the preprocessed road network map, an improved A* algorithm (adapted to the hierarchical characteristics of BeiDou grid codes) is used to generate global paths.

[0207] Hierarchical path search: Utilizing the hierarchical structure of BeiDou grid codes with "precision from low to high" (e.g., level 1 → level 32), the search range is narrowed in two steps. First, a coarse search (low-precision grid) is performed within the level 16 grid (100m precision) to plan the general direction on a grid-by-grid basis (e.g., "from grid A → grid B → grid C"), quickly eliminating obviously unreasonable paths (e.g., detours). Then, a fine search (high-precision grid) is performed within the level 18-32 grid (10m-1cm precision) to search specific road segments based on the direction determined by the coarse search, generating paths that include lane-level details.

[0208] Cost function design (weighted evaluation index): Based on the evaluation function of the A* algorithm, the A* algorithm is improved by combining the distance of the Beidou grid code to ensure that the "no overestimation" characteristic is met and the path optimality is guaranteed. The traffic flow parameters predicted by LSTM are introduced as heuristic function weight coefficients.

[0209] Path generation and optimization: Initial path generation: An improved A* algorithm is used to search for the initial path from P0 to Pt, which is represented as a "node sequence" (e.g., n0→n1→n2→...→nk), where n0 corresponds to P0 and nk corresponds to Pt; Path smoothing is performed to eliminate "acute angle turns" in the initial path, and a smooth path that conforms to vehicle kinematic constraints is generated through curve fitting; Redundant node removal: Redundant nodes on straight segments are deleted to reduce the amount of path data (usually compressed by about 50%), which facilitates parsing by the vehicle end.

[0210] Output and distribution of global reference paths: The generated global reference paths need to be converted into a standardized format, containing sufficient details for the vehicle-mounted terminal to combine with local path planning to achieve path planning. Specific output content includes:

[0211] Path data format: Path point sequence: [(lat1, lon1, s1, v1, limit) ..., (lati, loni, si, vi, limit))), where (lati, loni) are the CGCS2000 coordinates of the path point; si is the cumulative distance from the starting point to the point, in meters; vi, limit is the suggested driving speed for this road segment, in km / h; Key event markers: embed event information in the path points, (latj, lonj, ..., event="turn right after 100m"); Beidou grid code index: add Beidou grid codes to the path segments (e.g., one 16-level grid code for each 1km path) to facilitate quick matching of local map data by the vehicle terminal.

[0212] Distribution Mechanism: Triggering Method: When the autonomous vehicle travels 1km or its position deviates by more than 20m, the cloud recalculates and distributes the path; Event Trigger: When a sudden congestion / accident is detected within 5km ahead, an alternative route is immediately generated and distributed.

[0213] The vehicle-mounted navigation system fuses the global reference path (macroscopic long-distance guidance at the 100km level) sent from the cloud with the local reference path (microscopic short-distance obstacle avoidance within 100 meters) generated in real time by the vehicle's sensors. This fusion is achieved through multi-scale path matching and deviation correction, ultimately transforming the data into control commands (steering, throttle, braking) that can be executed by the autonomous vehicle. This process must balance path tracking accuracy (≤0.5m) with real-time response speed (control frequency ≥10Hz).

[0214] Command Subsystem: Based on digital twin virtual mapping and sand table simulation; it constructs a virtual scene highly consistent with the physical world through digital twin technology, realizing real-time mapping of the position and status of unmanned vehicles, and relying on sand table simulation function to simulate multi-vehicle path planning and conflict resolution, providing visualized and predictable technical support for global scheduling decisions. This system must meet the core indicators of "virtual-real synchronization accuracy ≤ 50ms" and "multi-vehicle simulation response time ≤ 3s".

[0215] Construction of digital twin virtual scene: The virtual scene is the "digital foundation" for the command center's monitoring. It needs to integrate multi-source data to construct a three-dimensional environment that corresponds to the real world in a 1:1 ratio, including static geographic information, dynamic traffic elements, and digital twins of unmanned vehicles and real-time vehicle movement trajectories.

[0216] Real-time mapping of vehicle position and status: Through a "positioning data transmission - spatiotemporal alignment - rendering update" mechanism, millisecond-level synchronization between the physical unmanned vehicle and the virtual scene is achieved, ensuring that the "virtual unmanned vehicle" seen by the command center is completely consistent with the movement of the real unmanned vehicle. Multi-source data fusion transmission: The vehicle-mounted terminal's fused positioning results, vehicle status messages, and event trigger data are fused and processed to form an unmanned vehicle database. Spatiotemporal alignment and noise filtering: All data carries a UTC timestamp. The command center calibrates its local clock through an NTP (Network Time Protocol) server to ensure that the time deviation between the virtual scene and the physical world is ≤50ms, completing time alignment; the vehicle-mounted terminal's ENU coordinates (local coordinate system) are converted to the virtual scene's global coordinate system (CGCS2000), and the corresponding geographical area is quickly matched through Beidou grid codes, with a conversion error ≤0.5m, completing spatial alignment; noise filtering: a sliding window mean filter (window size 5 frames) is used on the positioning data to eliminate accidental jumps while retaining realistic dynamic behaviors such as rapid acceleration and sharp turns. Real-time rendering and updates of virtual scenes: The position and speed of the autonomous vehicle are updated at a frequency of 10Hz (once every 100ms) to ensure smooth movement of the virtual vehicle. LOD (Level of Detail) technology is used—a simplified model (outline only + ID) is displayed for distant vehicles, while a detailed model (including headlights and license plates) is displayed for close-up vehicles. At the same time, the maximum number of vehicles rendered per screen is limited (≤100 vehicles) to ensure a stable frame rate of over 30fps. Status labels and environmental information are updated at a frequency of 1Hz to balance rendering efficiency and information timeliness. Anomaly highlighting: When a vehicle malfunctions, the virtual model automatically flashes and generates an alarm pop-up, while the location of the anomaly is marked on the map (e.g., marked as a red flashing dot).

[0217] Sand table simulation function—multi-vehicle path simulation and conflict resolution: Sand table simulation serves as a "pre-launch tool" for the command center. Based on the current traffic conditions and the planned paths of unmanned vehicles, it simulates the movement trajectories of multiple vehicles over the next 5-30 minutes, proactively identifying potential conflicts (such as path intersections and congestion) and optimizing scheduling strategies. The Multi-Agent Reinforcement Learning (MARL) model of this invention adopts an "offline training + online inference" architecture. Model training is completed on the high-performance server cluster of the command center (using historical running data and simulation environment for iterative training). After training, the model parameters are solidified and deployed to the real-time inference server of the command center. The multi-vehicle movement simulation, conflict detection, and optimization strategy generation during the sand table simulation all run in real-time on the command center's inference server, without relying on the onboard edge computing unit. The onboard edge computing unit is only responsible for the localization, error compensation, and local path planning of its own vehicle and does not participate in the calculation of multi-vehicle collaborative reinforcement learning. This division ensures the real-time performance and reliability of the simulation calculations while avoiding unnecessary computational burden on the onboard units.

[0218] The algorithm employs Multi-Agent Reinforcement Learning (MARL, where multiple agents coexist in the same environment, each agent perceives the environment through local observations, performs actions, and obtains rewards, and learns and adjusts its own policy to ultimately achieve individual goals (such as maximizing its own benefits) or global goals (such as maximizing the team's total benefits)) to simulate the movement of multiple vehicles. The core is to achieve local collision avoidance and efficiency optimization while ensuring the "global path consistency" of each vehicle.

[0219] Conflict detection and optimization suggestions: Identify conflict types such as path intersections, congestion, and violations, and visualize the conflicts (mark the conflict location with a "red conflict box" in the virtual scene and highlight the involved autonomous vehicles).

[0220] Optimization strategy generation: Replan local detour routes for conflicting vehicles, and stagger conflict times by controlling vehicle acceleration / deceleration time. In scenarios where multiple vehicles are vying for road space, force low-priority vehicles to stop and give way.

[0221] Output of simulation results and command and control: The simulation results are displayed in the form of a timeline and trajectory animation, showing the multi-vehicle paths before and after optimization. Key indicators (such as total travel time, number of conflicts, and number of violations) are compared, and the simulation results are output. The optimization strategy is converted into standardized instructions (such as "Vehicle ID=10000, decelerate to 10km / h at position (30.145, 120.789) and maintain it for 5 seconds"). The instructions are sent to the corresponding unmanned vehicles through a converged communication link. After receiving the instructions, the vehicle-mounted terminal updates the local path planning, while the command center monitors the execution effect of the instructions in real time (such as "whether the unmanned vehicle decelerates according to the instructions").

[0222] The key point of this invention, and also the point to be protected:

[0223] This invention provides a high-precision positioning method and system for unmanned vehicles in complex environments (such as urban canyons, tunnels, and areas without network coverage). Through deep collaboration among four subsystems—the vehicle-mounted terminal, the cloud, the communication link, and the command center—it solves core problems in existing technologies such as positioning drift, navigation lag, communication interruption, and scheduling blind spots. The modules are not simply stacked; instead, they form a closed loop through designed data flow and control logic, producing technical effects superior to individual components. The core innovation of this invention lies not only in the technical details within each subsystem but also in the deep coupling and data loop between the four modules, constructing a dynamic closed-loop intelligent agent system: ① The LSTM on the vehicle-mounted terminal is not only used for local positioning compensation but its prediction results also support local obstacle avoidance and cloud-based traffic flow modeling; ② The global path generated by the cloud based on the BeiDou grid code efficiently guides the vehicle-mounted terminal's local map matching through grid indexing; ③ The link status of the communication system affects the vehicle-mounted terminal's positioning mode switching (e.g., enhancing IMU / visual weights when 5G is interrupted); ④ The command center's deduction results directly intervene in the vehicle-mounted terminal's path planning, while the vehicle-mounted terminal's execution feedback updates the digital twin.

[0224] The specific innovations are as follows: Vehicle-mounted positioning subsystem: Achieves sub-decimeter-level robust positioning and predictive local path planning.

[0225] 1. Multi-source sensor dynamic error correction method: This invention constructs a 32-dimensional unified feature vector, integrates preprocessed data from BeiDou PPP-RTK (8-dimensional), IMU (12-dimensional), and binocular vision (12-dimensional), and inputs it into a 3-layer bidirectional LSTM-DYC (dynamic error correction) network, outputting 6-dimensional positioning and attitude error compensation values ​​(x / y / z direction position error and roll / pitch / heading attitude error). The key lies in introducing an adaptive weight adjustment mechanism based on BeiDou signal-to-noise ratio (SNR): when SNR ≥ 30dB, BeiDou multipath error compensation is the main method (weight 60%); when 20dB ≤ SNR < 30dB, IMU drift compensation is enhanced (weight 30%); when SNR < 20dB, visual feature deviation is relied upon for enhanced compensation (weight 50%). This mechanism ensures that positioning error compensation accuracy of ≤ 0.04m can still be maintained in complex scenarios with drastic fluctuations in signal quality. Furthermore, the compensated values ​​output by the LSTM are combined with the original observations to construct the observation equations for the Extended Kalman Filter (EKF). Through a complete "state prediction—gain calculation—state update" process, nonlinear optimal fusion of multi-source heterogeneous data is achieved, overcoming the shortcomings of traditional weighted average fusion in handling temporal inconsistencies and error coupling. Further, in the state prediction stage of the EKF, the state trend output by the LSTM network is innovatively introduced as a correction term, and fused using an adaptive weighting factor λ (SNR) based on BeiDou SNR to form a hybrid state transition model. Simultaneously, an adaptive process noise covariance matrix Qadaptive is designed, whose value is dynamically adjusted according to the environmental SNR, thereby enhancing the robustness of the filter in harsh signal environments.

[0226] 2. Local Reference Path Generation and Closed-Loop Feedback Mechanism: Based on the fused positioning results described above, a two-layer unidirectional LSTM-TP (trajectory prediction) network is used to predict the vehicle's trajectory within the next 3 seconds (prediction error ≤ 0.5m). The predicted trajectory and the visually recognized dynamic obstacle information are used as inputs to the improved A* algorithm. This A* algorithm introduces autonomous vehicle dynamic constraints (minimum turning radius ≥ 5m, maximum acceleration ≤ 2m / s²) and maps the predicted dynamic obstacle trajectory to a high-cost region (cost factor = 3.0), generating a smooth local path within a 100m range ahead (curvature ≤ 0.2m⁻¹), with an obstacle avoidance response delay ≤ 200ms. More importantly, the system establishes a vehicle-side closed-loop feedback mechanism: the fused positioning results are fed back to the BeiDou, IMU, and visual preprocessing modules in real time, respectively used to correct the ionospheric / tropospheric delay model (improving accuracy by approximately 15%), update the IMU zero-bias coefficient (reducing drift error by approximately 40%), and adjust the feature matching threshold (reducing mismatch rate by approximately 20%), thereby achieving continuous online optimization of positioning and perception capabilities.

[0227] Cloud-based navigation subsystem: Enables efficient, dynamic, and scalable global path planning.

[0228] 1. A Hierarchical Path Search and Fusion Method Based on BeiDou Grid Code: The cloud utilizes BeiDou grid code (GeoSOT) to perform multi-scale subdivision of the Earth's surface and uses it as a spatial index key. After receiving CGCS2000 coordinates uploaded by the vehicle terminal, the system converts them into grid codes of corresponding precision, quickly retrieving road network topology, static attributes (speed limits, traffic restrictions), and dynamic traffic information (congestion coefficient) within the buffer area covering the "start point to target point," improving retrieval efficiency by 3-5 times compared to traditional latitude and longitude indexing. Furthermore, a two-level path search strategy is proposed: first, coarse-grained direction planning is performed on a level 16 grid (100m precision) to eliminate detours; then, fine-grained lane-level path generation is performed on level 18–32 grids (10m–1cm precision), balancing efficiency and accuracy.

[0229] 2. Improved A* Algorithm Integrating LSTM Traffic Forecasting: In the path cost function design, in addition to traditional distance and time, future traffic flow parameters predicted by the LSTM network are introduced as dynamic weight coefficients of the heuristic function h(n), giving path planning a forward-looking nature. For example, when it is predicted that congestion will occur on a certain road segment within the next 2 minutes, its toll cost is automatically increased, guiding the path to an alternative route. The generated global path is smoothed by B-spline and redundant nodes are removed (compression rate of approximately 50%), and then distributed in a standardized format (including coordinates, cumulative distance, suggested speed, event markers, and 16-level grid codes corresponding to each 1km), supporting rapid matching of local maps on the vehicle side. Path updates are triggered by positional deviation (>20m), mileage (per 1km), or sudden events (accidents within 5km ahead), ensuring the real-time nature and feasibility of global guidance.

[0230] Communication Subsystem: Achieving Seamless and Reliable Heterogeneous Communication Ability Across the Entire Domain: Constructing a converged communication architecture with 5G as the primary technology and BeiDou short message service as backup, and designing an intelligent switching algorithm based on signal trend prediction. This algorithm uses historical 5G RSRP (Reference Signal Received Power) sequences and a lightweight time series model (such as ARIMA) to predict signal strength changes within the next T seconds. When the predicted 5G link availability probability is lower than a threshold and the duration exceeds a set window, the system switches to the BeiDou short message link in advance, ensuring stable transmission of positioning data and command instructions in areas without network coverage (such as mountainous areas and tunnels), and avoiding control failures caused by sudden link interruptions.

[0231] Command Subsystem: Enables intelligent collaborative scheduling that combines virtual and real-world operations, allows for simulation, and provides verifiable results.

[0232] 1. Digital Twin-Driven Multi-Vehicle Collaborative Scheduling Closed Loop: The command center constructs a 3D digital twin scene that maps 1:1 to the physical world. Through NTP time synchronization (deviation ≤50ms) and BeiDou grid code spatial alignment (error ≤0.5m), millisecond-level virtual-real mapping of the unmanned vehicle status is achieved. Based on this, a multi-agent reinforcement learning (MARL) algorithm is used to simulate the movement trajectories of multiple vehicles in the virtual environment for the next 5–30 minutes, identifying risks such as path intersections and congestion conflicts in advance.

[0233] 2. Closed-Loop System for Sand Table Simulation – Command Issuance – Execution Monitoring: For conflicts identified during simulations, the system automatically generates optimization strategies (such as partial detours, speed adjustments, and priority yielding) and converts them into standardized commands (e.g., "ID=10000, decelerate to 10km / h at (30.145, 120.789) and maintain this speed for 5 seconds"), which are then issued to the corresponding vehicles via the converged communication link. After execution by the vehicle, its status update is fed back to the digital twin in real time, forming a complete closed loop of "simulation → optimization → command issuance → execution → verification". This mechanism can complete multi-vehicle path reconfiguration under sudden obstacles within 200ms, significantly improving the safety and scheduling efficiency of group collaboration.

[0234] The vehicle-mounted positioning subsystem collects observation data in real time (BeiDou raw data, inertial sensor data, binocular camera data, etc.) to achieve high-precision fusion positioning. It constructs a dynamic error correction model through an LSTM network deployed by the edge computing unit to achieve sub-decimeter-level positioning in complex scenarios. It also combines the unmanned vehicle's own position to achieve a local reference path and uploads the processed data to the cloud navigation subsystem via 5G or BeiDou short message.

[0235] The cloud-based navigation subsystem receives data from multiple unmanned vehicles, combines it with the real-time location of the unmanned vehicles, Beidou grid codes, and road network data for fusion processing and analysis, obtains a global reference path, and feeds the results back to the vehicle-mounted positioning subsystem or command subsystem.

[0236] The communication subsystem has multiple communication methods, including 5G and BeiDou short message service. It adopts a converged communication architecture and intelligent communication switching algorithm, and automatically switches communication links based on signal strength prediction to realize information transmission between the vehicle-mounted positioning subsystem, the cloud navigation subsystem, and the command subsystem.

[0237] The command subsystem monitors the overall situation, constructs a virtual road scene that maps 1:1 to the physical world through digital twin technology, generates dynamic paths based on LSTM trajectory prediction and improved A* algorithm, compares them in real time with the planned paths on the vehicle end, and realizes multi-vehicle collaborative scheduling through sand table simulation of the command subsystem, thereby achieving high-precision, high-reliability positioning, navigation and intelligent command and control of unmanned vehicles.

[0238] This invention comprises six core steps: real-time fusion positioning (BeiDou PPP-RTK high-precision positioning + inertial navigation + vision) and LSTM optimization and local reference path generation at the vehicle end; positioning results transmitted to the cloud and command center via fusion communication transmission and intelligent switching links; cloud navigation calculation; global reference path transmitted to the vehicle end and command center via fusion communication transmission and intelligent switching links; and vehicle end combining global and local reference paths to achieve vehicle navigation and global monitoring and command control at the command center. To address the positioning drift or failure caused by BeiDou signal obstruction in complex environments, this invention uses BeiDou PPP-RTK high-precision positioning as a foundation, integrates inertial navigation (IMU) and binocular vision data, and constructs an adaptive error compensation model through an LSTM-DYC (Dynamic Error Correction) network deployed in edge computing units. Combined with an improved extended Kalman filter (EKF), it achieves deep fusion of multi-source heterogeneous data, ensuring decimeter-level robust positioning in scenarios such as urban canyons and tunnels. To address the issues of low robustness and insufficient dynamic obstacle prediction capabilities in navigation systems, this invention designs an LSTM-TP (Trajectory Prediction) network. Taking the vehicle's motion state, environmental context, and cloud-based macroscopic path as input, it outputs the vehicle trajectory and spatiotemporal cost field for the next 3 seconds. The prediction results are then used as heuristic information to improve A* local path planning, introducing autonomous vehicle dynamic constraints and high-cost regions of dynamic obstacles to achieve predictive autonomous obstacle avoidance and smooth local path generation, significantly improving the navigation system's performance. This invention addresses the stability of the system in dynamic and complex environments. To mitigate the issue of single communication links being prone to interruption in remote areas or signal blind spots, it constructs a converged communication architecture with "5G as the primary and BeiDou short message backup." It also designs an intelligent link switching algorithm based on historical signal strength trend prediction, automatically switching communication links according to the predicted 5G availability probability and BeiDou short message transmission success rate, ensuring real-time and reliable transmission of positioning data and command instructions across the entire domain. Furthermore, to address the limitations of intelligent command decision-making and the lack of multi-vehicle collaborative prediction capabilities, this invention constructs a digital twin scenario in the command center that maps 1:1 to the physical world. It utilizes a multi-agent reinforcement learning (MARL) algorithm to perform multi-vehicle movement simulations for the next 5-30 minutes, automatically detecting path intersections, congestion, and other conflicts, generating optimization strategies, and issuing standardized instructions to each unmanned vehicle through the converged communication link. This forms an intelligent collaborative scheduling closed loop of "simulation → optimization → issuance → execution → verification," achieving conflict resolution and efficiency optimization for group tasks.

[0239] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A high-precision positioning method for unmanned vehicles based on BeiDou and LSTM networks, characterized in that, include: Step S1, Data Acquisition and Preprocessing: The system acquires BeiDou PPP-RTK observation data, inertial sensor data, and binocular vision data from the vehicle-mounted terminal, and performs preprocessing on each data to obtain BeiDou feature data, inertial feature data, and visual feature data. Step S2, Unified Feature Vector Construction: The preprocessed BeiDou feature data, inertial feature data, and visual feature data are fused to construct a multi-dimensional unified feature vector; Step S3, LSTM dynamic error correction: The multidimensional unified feature vector is input into a pre-trained LSTM dynamic error correction network, namely the LSTM-DYC network. The LSTM-DYC network outputs error compensation values ​​for localization and attitude. The output of the LSTM-DYC network includes two parts: the first part is the error compensation values ​​for position and attitude, and the second part is the short-term state trend prediction value used for state prediction correction. Step S4, Adaptive weight adjustment: Based on the signal-to-noise ratio (SNR) of the BeiDou signal, the compensation weights from the three data sources (BeiDou, inertial navigation, and vision) in the error compensation value output by the LSTM-DYC network are dynamically adjusted. Step S5, Extended Kalman Filter (EKF) fusion: Construct the state vector of the extended Kalman filter and perform the following fusion steps: State prediction: A preliminary state prediction is performed based on the physical model of inertial navigation to obtain the basic predicted state; at the same time, the basic predicted state is corrected by the short-term state trend prediction value output by the LSTM-DYC network to obtain the hybrid predicted state. Observation update: The BeiDou PPP-RTK positioning result corrected by the error compensation value and the initial visual positioning value are used together as observation values ​​to update the hybrid prediction state and output the final high-precision fusion positioning result.

2. The high-precision positioning method for unmanned vehicles based on BeiDou and LSTM networks according to claim 1, characterized in that, The multidimensional unified feature vector constructed in step S2 is 32-dimensional, where: The BeiDou feature data is 8-dimensional, including: BeiDou PPP-RTK positioning residual, number of visible satellites, average signal-to-noise ratio, ionospheric delay correction residual, and tropospheric delay correction residual; The inertial characteristic data is 12-dimensional, including: three-axis acceleration increment, three-axis angular velocity increment, attitude angle change, and three-axis zero-bias compensation residual; The visual feature data is 12-dimensional, including: the matching deviation between static road feature points and high-precision maps, the confidence level of traffic sign recognition, the binocular parallax calculation error, and the deviation between the initial visual positioning value and the BeiDou positioning result.

3. The high-precision positioning method for unmanned vehicles based on BeiDou and LSTM networks according to claim 1, characterized in that, The LSTM-DYC network in step S3 includes a 3-layer bidirectional LSTM network. Its input layer is 32-dimensional, the first hidden layer contains 128 bidirectional neurons, and the second hidden layer contains 64 bidirectional neurons. The first part of the output layer is a 6-dimensional vector, including three-dimensional position error compensation value and three-dimensional attitude error compensation value. The second part of the output layer is a 9-dimensional vector, including the position increment, velocity increment, and attitude angle increment of the next time step, which serve as the short-term state trend prediction value.

4. The high-precision positioning method for unmanned vehicles based on BeiDou and LSTM networks according to claim 1, characterized in that, The adaptive weight adjustment in step S4 specifically involves: When SNR ≥ 30dB, the compensation weights for BeiDou, inertial navigation, and vision are set to 60%, 25%, and 15%, respectively. When 20dB ≤ SNR < 30dB, the compensation weights for BeiDou, inertial navigation, and vision are set to 30%, 30%, and 40%, respectively. When SNR < 20dB, the compensation weights for BeiDou, inertial navigation, and vision are set to 10%, 30%, and 60%, respectively.

5. The high-precision positioning method for unmanned vehicles based on BeiDou and LSTM networks according to claim 1, characterized in that, The state prediction in step S5 further includes: The mixed prediction state is calculated using the following formula. : ; in, Based on the prediction of the state by the basic physical model Let λ(SNR) be the short-term state trend prediction value output by the LSTM-DYC network, and let λ(SNR) be the adaptive weighting factor calculated based on the signal-to-noise ratio (SNR), with the following functional form: 。 6. The high-precision positioning method for unmanned vehicles based on BeiDou and LSTM networks according to claim 1, characterized in that, The extended Kalman filter in step S5 also employs an adaptive process noise covariance matrix. Its design is as follows: ； in, The basic process noise covariance matrix, The average signal-to-noise ratio is α, and β are preset parameters used to actively amplify process noise in low signal-to-noise ratio environments.

7. The high-precision positioning method for unmanned vehicles based on BeiDou and LSTM networks according to claim 1, characterized in that, The method further includes a closed-loop feedback step: the final high-precision fusion positioning result is fed back to the data preprocessing stage in step S1 in real time, which is used to correct the multipath error model of Beidou PPP-RTK, update the zero-bias calibration coefficient of the inertial sensor, and adjust the feature point matching confidence threshold of visual positioning.

8. A high-precision positioning, navigation, and command system for unmanned vehicles based on BeiDou and LSTM networks, characterized in that, include: The vehicle positioning subsystem includes a fusion positioning result output by the high-precision positioning method for unmanned vehicles based on BeiDou and LSTM networks as described in any one of claims 1 to 7. The cloud-based navigation subsystem generates a global reference path from the current location to the target point based on the fused positioning results and the BeiDou grid code. The command subsystem is used to map the received real-time status of unmanned vehicles into the virtual scene through NTP time synchronization and BeiDou grid code spatial alignment; it uses the MARL algorithm to perform sandbox simulation of the movement of multiple vehicles in the virtual scene over the next 5 to 30 minutes, simulating and detecting potential path intersections and traffic congestion conflicts; for the detected conflicts, it automatically generates optimization strategies including local detours, speed adjustments, or priority yielding, and converts them into standardized command instructions; the command instructions are sent to the corresponding unmanned vehicles through the communication subsystem, and execution feedback is received to update the virtual scene, forming a closed loop of "simulation-optimization-issuance-execution-verification".

Images