Vehicle inertial navigation auxiliary positioning and deviation correction method in weak network environment

Abstract

The application belongs to the technical field of satellite navigation and positioning, and particularly relates to a vehicle inertial navigation auxiliary positioning and deviation rectification method in a weak network environment, which comprises the following steps: acquiring inertial measurement unit data to calculate a motion uncertainty index; combining a trend to adjust a detection frequency; embedding an inertial measurement unit pre-integration increment into a QUIC self-defined frame; optimally transmitting based on a performance model and the motion uncertainty index; solving and positioning by a server using a factor graph fusion; constructing a state transition constraint factor when a source connection identifier is changed; analyzing and propagating a pre-integration increment to predict a state using a previous state; constructing a residual term by combining a predicted state and an observed state; and outputting a position based on Mahalanobis distance and a penalty upper limit to adjust a covariance. The application combines motion perception and network prediction to optimally select a path, uses predicted information to construct a state transition constraint factor to smooth state mutation, eliminates positioning jumps, and guarantees the continuity of vehicle position output in a weak network environment.

Description

Technical Field

[0001] This invention relates to the field of satellite navigation and positioning technology. More specifically, this invention relates to a vehicle inertial navigation-assisted positioning and correction method in a weak network environment. Background Technology

[0002] To achieve reliable positioning in all weather conditions and scenarios, modern vehicle positioning systems generally adopt multi-source fusion navigation solutions. This solution improves positioning reliability by deeply integrating information from various sensors, including global navigation satellite systems, inertial navigation systems, high-precision maps, and vehicle wheel speed sensors. In this system, the inertial navigation system, with its high update rate and short-term high accuracy, can effectively compensate for the interruption or quality degradation of satellite navigation signals in obstructed environments such as urban canyons or tunnels. In the current cloud-based collaborative positioning architecture, the vehicle terminal needs to continuously upload massive amounts of information, such as raw inertial navigation data and satellite navigation observations, to the server in real time. Subsequently, the server's powerful computing platform executes complex fusion algorithms to obtain high-precision positioning results that are superior to those obtained by independent vehicle-mounted solutions.

[0003] However, in practical applications, vehicles inevitably experience severe fluctuations in network signals while moving at high speeds. These fluctuations typically originate from handovers between different cellular base stations or even between multiple network standards, causing the communication process to frequently fall into weak network environments characterized by high latency, high jitter, and high packet loss rates. This unstable data transmission state can cause delays, out-of-order delivery, or even interruptions in the sensor data streams uploaded to the server, thus posing serious safety risks to the decision-making and control of the autonomous driving system.

[0004] To address the transmission requirements in weak network environments, the QUIC protocol, built on UDP, theoretically offers good adaptability to vehicle data upload scenarios in such environments due to its low latency, multiplexing capabilities, and core connection migration mechanism. This connection migration mechanism allows the terminal to maintain its existing connection with the server using a connection identifier when its IP address or port changes. Compared to the traditional TCP protocol, this mechanism avoids the expensive connection reconstruction overhead, thus ensuring the logical continuity of the data stream. Despite these advantages, standard connection migration strategies are typically based on passive triggering of network layer changes, neglecting the criticality of application layer positioning data in different driving tasks. For example, when a vehicle performs maneuvers such as turning or acceleration / deceleration, the uncertainty of its motion state increases dramatically. In this case, the system's requirement for real-time continuity of inertial navigation data is far higher than in a straight-line constant-speed driving scenario. However, existing mechanisms cannot perceive the vehicle's real-time motion state, making it difficult to differentiate network path management strategies. Further complicating matters, even if the QUIC protocol achieves seamless migration at the transport layer, network path switching inevitably introduces instantaneous changes in data transmission characteristics. Because the factor graph fusion localization algorithm used by the server relies heavily on strict temporal relationships, even minor timestamp inconsistencies or state prediction errors between two data packets before and after migration can cause the optimization model to fail to converge or produce incorrect trajectory corrections. The existing QUIC mechanism does not yet provide effective cross-layer information to assist the application-layer algorithm in smoothing out such state abrupt changes caused by network migration. This directly limits the mechanism's practical effectiveness in ensuring the continuity and accuracy of localization services under extremely weak network conditions. Summary of the Invention

[0005] To address the technical problem that existing technologies do not provide any cross-layer information to assist application layer algorithms in smoothing out state changes caused by network migration, thus limiting the actual effectiveness of the positioning service under extremely weak network conditions, this invention provides a vehicle inertial navigation-assisted positioning correction method in weak network environments.

[0006] This invention provides a vehicle inertial navigation-assisted positioning and correction method in a weak network environment, comprising: acquiring raw data from the inertial measurement unit of the vehicle terminal and calculating the motion uncertainty index; adjusting the detection frequency based on the motion uncertainty index and the predicted trend of network path quality; when constructing QUIC data packets, directly serializing the pre-integration increment of the inertial measurement unit as state prediction information and embedding it into a custom frame; maintaining independent connection identifiers for multiple available network paths, evaluating the comprehensive transmission utility score of each path based on the historical performance model of each path issued by the server and the current motion uncertainty index, and selecting the connection identifier corresponding to the path with the highest score to transmit data; the server adopts... The factor graph integrates received inertial measurement unit (IMU) data, global navigation satellite system (GNSS) observations, and map information for positioning calculation. When the server detects a change in the source connection identifier of a data packet, it constructs a state transition constraint factor in the factor graph. The state transition constraint factor uses the previous connection state as an initial condition, parses the state prediction information embedded in the first data packet of the new connection to obtain the IMU pre-integration increment, and performs state propagation to obtain the predicted state. The predicted state and the initial state calculated based on the initial observation data of the new connection constitute a residual term. The covariance of the state transition constraint factor is adjusted according to the Mahalanobis distance of the residual term and a preset penalty upper limit, thereby outputting the vehicle position.

[0007] By adopting the above technical solution, this invention couples the vehicle motion state at the application layer with the transmission strategy at the network layer. During network switching, it rapidly transmits state prediction information through custom frames and constructs a state transition constraint factor with a penalty upper limit in the cloud. This mechanism not only effectively smooths out state abrupt changes caused by network migration and maintains the high temporal strictness of factor graph optimization, but also avoids divergence in optimization solutions under extreme errors, ensuring the absolute continuity and high reliability of autonomous driving positioning under extremely weak network conditions, and eliminating safety hazards.

[0008] Preferably, the steps of acquiring the raw data of the inertial measurement unit of the vehicle terminal, calculating the motion uncertainty index, and adjusting the detection frequency include: calculating the statistical variance based on the three-axis angular velocity and three-axis acceleration of the inertial measurement unit within a preset time window, and combining the statistical variance after dimensionless processing to form the motion uncertainty index; establishing a prediction model based on the recent round-trip time sample sequence of the network path, estimating the network quality change trend within the future time window, and obtaining the predicted network quality degradation degree; determining the detection frequency based on the motion uncertainty index and the predicted network quality degradation degree, and limiting the detection frequency between a preset minimum and maximum detection frequency.

[0009] By adopting the above technical solution, the positioning instability and fusion chaos caused by frequent network switching when the vehicle is in violent movement are avoided, and the backup communication path can be discovered and established in time before the network deteriorates, thus solving the risk of data flow interruption caused by the traditional passive network switching mechanism.

[0010] Preferably, the step of directly serializing and embedding the pre-integration increment of the inertial measurement unit as state prediction information into a custom frame includes: defining a custom frame type in the QUIC protocol; obtaining the frame type from the time of the previous data packet transmission. Up to the current moment The inertial measurement unit pre-integrates the increment, and directly serializes the floating-point value corresponding to the pre-integration increment into a fixed-length binary bit stream data through network byte order; the binary bit stream data is filled into the custom frame as a payload; when constructing the QUIC data packet to be sent, this custom frame is inserted into the frame sequence of the data packet.

[0011] By adopting the above technical solution, the additional computing power overhead and deserialization latency caused by heavy data compression encoding are eliminated, and serialization and splicing are performed directly using standard network byte order. With only a very small and negligible increase in network load, high-precision and extremely low-latency transmission of prediction information is achieved, improving the real-time performance of vehicle-cloud collaborative computing in weak network environments.

[0012] Preferably, the evaluation of the comprehensive transmission utility score for each path includes: obtaining the historical performance model of each network path from the server, wherein the historical performance model includes average throughput, average round-trip time, and packet loss rate; converting the motion uncertainty index into a normalized risk weighting factor; calculating the original comprehensive transmission utility score based on the average throughput, average round-trip time, and packet loss rate; and for other available paths besides the current primary path, dividing the original comprehensive transmission utility score by the risk weighting factor to obtain the final comprehensive transmission utility score.

[0013] By adopting the above technical solution, risk penalties are actively applied when the vehicle is moving violently to reduce the attractiveness of alternative paths, suppressing frequent path switching operations that may lead to data interruption, avoiding link turbulence, and laying an extremely stable data transmission foundation for subsequent smooth positioning calculations on the server side.

[0014] Preferably, the step of performing state propagation to obtain the predicted state includes: on the server side, when it is detected that the source connection identifier of a QUIC data packet has changed to a new connection identifier, locking the state vector and covariance matrix of the previous data packet sent at the time associated with the old connection identifier; parsing the state prediction information from the custom frame of the first data packet of the new connection identifier to obtain... The inertial measurement unit pre-integration increment is calculated; based on the inertial navigation motion equation, the predicted state vector and covariance matrix at the current moment are calculated using the state vector at the previous moment and the analyzed inertial measurement unit pre-integration increment.

[0015] By adopting the above technical solution, accurate incremental compensation and prior auxiliary information are provided to the server during the data blind spots caused by network migration, which solves the problem of discontinuous motion state prediction caused by packet delay, out-of-order or interruption.

[0016] Preferably, the process of constructing the residual term includes: using global navigation satellite system observations associated with the first data packet of the newly connected network to calculate the initial state vector and covariance matrix at the current moment, wherein the initial state vector includes position and velocity components; and performing a difference operation between the propagated predicted state vector and the initial state vector calculated based on the global navigation satellite system to construct a state prediction deviation vector as the residual term.

[0017] Preferably, the covariance of the state transition constraint factor is adjusted based on the Mahalanobis distance of the residual term and a preset penalty upper limit, including: combining the state prediction deviation vector, the covariance matrix of the predicted state, and the covariance matrix of the initial state to calculate the squared Mahalanobis distance representing the consistency between prediction and observation; calculating a scaling factor based on the degree to which the squared Mahalanobis distance deviates from the confidence interval threshold, and truncating the scaling factor using the preset penalty upper limit when the deviation surges due to long-term network disconnection and the scaling factor tends to infinity; setting the magnitude of the covariance matrix of the state transition constraint factor as a function of the squared Mahalanobis distance adjusted by the scaling factor, so that the weight of the factor is reduced when the difference between prediction and observation increases, while maintaining effective information matrix constraints.

[0018] Preferably, the server uses factor graph fusion positioning calculation, which includes: constructing three types of factors in the factor graph, the first type being the inertial measurement unit pre-integration factor connecting the state vectors of two consecutive time points, the second type being the factor connecting the state vector of a single time point with the observation value of the global navigation satellite system, and the third type being the map matching factor that projects the calculated position onto the road centerline of the high-precision map; and calling a nonlinear least squares optimization algorithm to solve for the state vector.

[0019] Preferably, the step of calculating the statistical variance of the three-axis angular velocity and three-axis acceleration of the inertial measurement unit within a preset time window, and combining the statistical variance into a motion uncertainty index after dimensionless processing, includes: acquiring the original data of the inertial measurement unit within a preset sliding time window; calculating the angular velocity variance of the three-axis angular velocity and the acceleration variance of the three-axis acceleration respectively; dividing the angular velocity variance and the acceleration variance by a preset reference variance threshold for dimensionless processing, wherein the reference variance threshold is pre-calibrated by the vehicle under a straight-line uniform speed state; and weighting and summing the dimensionless angular velocity variance and acceleration variance based on a preset weighting coefficient to calculate the final motion uncertainty index.

[0020] Preferably, the step of establishing a prediction model based on recent round-trip time sample sequences of network paths to estimate the trend of network quality changes within a future time window and obtain the predicted degree of network quality degradation includes: using an autoregressive model to predict the round-trip time within a future time window, wherein the input of the autoregressive model is multiple consecutive round-trip time samples before the prediction time point, and the output is the predicted round-trip time value; calculating the difference between the predicted round-trip time value and the sample mean of the most recent round-trip time samples of a preset primary path; taking the maximum value between the difference and zero to obtain the predicted degree of network quality degradation.

[0021] The technical solution of the present invention has the following beneficial technical effects: This invention optimizes the detection frequency of backup network paths by combining vehicle motion state and network quality prediction, reducing network signaling overhead while ensuring connection migration reliability. By comprehensively evaluating the transmission utility of each network path, it achieves optimal selection of data transmission paths, improving the stability and reliability of data links. On the server side, when connection migration occurs, a state transition constraint factor is introduced into the factor graph model, utilizing state prediction information provided by the vehicle terminal to constrain and bridge the positioning solution state before and after the migration. Simultaneously, the system dynamically adjusts the confidence level of this constraint factor to suppress positioning result jumps and interruptions caused by network switching, solving problems such as decreased inertial navigation-aided positioning accuracy and trajectory discontinuity caused by network connection switching in weak network environments, ensuring excellent continuity and reliability of positioning output under complex network conditions. Attached Figure Description

[0022] Figure 1 This is a flowchart of a vehicle inertial navigation-assisted positioning and correction method in a weak network environment according to the present invention. Figure 2 This is a schematic diagram of constraint factor covariance adjustment in an embodiment of the present invention; Figure 3 This is a comparison chart of experimental results from embodiments of the present invention. Detailed Implementation

[0023] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0024] This invention discloses a vehicle inertial navigation-assisted positioning correction method in a weak network environment, referring to... Figure 1 This includes steps S1-S9: S1: Acquire raw data from the vehicle terminal's inertial measurement unit and calculate the motion uncertainty index.

[0025] In an optional embodiment, the vehicle terminal acquires raw measurement values ​​from the inertial measurement unit at a sampling frequency of not less than 100Hz via a controller area network bus or serial communication interface. These raw measurement values ​​include the triaxial acceleration output by the triaxial accelerometer and the triaxial angular velocity output by the triaxial gyroscope.

[0026] To calculate the motion uncertainty index, the vehicle terminal maintains a fixed-length sliding time window, which, for example, includes all inertial measurement unit (IMU) measurement data from the past 500 ms. Within this sliding time window, the variances of the three-axis specific force vector magnitude and the three-axis angular velocity vector magnitude are continuously calculated using the Welford online algorithm. Subsequently, the motion uncertainty index is obtained by normalizing the acceleration variance and the angular velocity variance separately, and then weighted and summing them.

[0027] In an alternative embodiment, the vehicle terminal retrieves raw data from the inertial measurement unit (IMU) within the most recent second in each calculation cycle, including triaxial angular velocities and triaxial accelerations acquired at a sampling rate of 100Hz. The statistical variance of the data for each axis within this sliding time window is calculated, i.e., the angular velocity variance of the triaxial angular velocities is calculated separately. , , and the acceleration variance of the three-axis acceleration , , Next, the statistical variance of each axis's data is divided by a preset baseline variance threshold for dimensionless processing, with the baseline variance threshold for angular velocity set to 0.01. The baseline variance threshold for acceleration is set to 0.1. The baseline variance threshold was obtained by calibrating the vehicle while it was traveling at a constant speed in a straight line. Finally, the dimensionless angular velocity variance was... Compared with dimensionless acceleration variance By performing a weighted summation, the final motion uncertainty index is calculated. The uncertainty index of this movement The value is usually between 0.5 and 10.

[0028] Specifically, the motion uncertainty index Satisfying the relation:

[0029] In the formula, Indicates the uncertainty index of motion; This represents the variance of the angular velocity after dimensionless processing; This represents the variance of acceleration after dimensionless processing; This represents the angular velocity weighting coefficient, which is 0.5 in this embodiment. This represents the acceleration weighting coefficient, which is 0.5 in this embodiment.

[0030] Understandably, the baseline variance thresholds for angular velocity and acceleration are obtained through an offline calibration process. Specifically, the test vehicle is controlled to maintain a constant straight-line speed (e.g., 60 km / h) on a smooth paved road surface for a preset time, e.g., five minutes. During this period, the raw output of the inertial measurement unit is acquired at a frequency of 100 Hz. The statistical average of the variances of the three-axis angular velocity and three-axis acceleration within this time period is calculated and multiplied by a preset safety margin coefficient (e.g., between 1.2 and 1.5). These values ​​are then used as the basic reference thresholds for determining whether the vehicle is in a stable motion state and are entered into the system configuration.

[0031] It is understandable that the specific values ​​of the aforementioned angular velocity and acceleration weighting coefficients are not absolutely unique constants, but rather are engineering settings based on the specific hardware performance of the vehicle-mounted inertial measurement unit. In actual industrial deployments, these weighting coefficients can be obtained through offline calibration and grid search. Specifically, based on historical datasets containing raw data from the inertial measurement unit collected by actual vehicles on typical road sections, a set of benchmark weighting parameters is obtained offline through optimization, with the objective function of minimizing motion state judgment error. Those skilled in the art can also make adaptive fine-tuning according to actual business logic. For example, when the factory high-frequency noise index of the inertial measurement unit sensor is high, the acceleration weighting coefficient can be appropriately lowered to reduce interference caused by high-frequency vibration. The above setting methods are all within the conventional technical capabilities of those skilled in the art.

[0032] Thus, this step, by quantifying the real-time motion state of the vehicle into an uncertainty index, provides an accurate evaluation basis for network management strategies, solving the problem of rigid network detection strategies caused by the inability to perceive motion state.

[0033] S2: Adjust the detection frequency based on the motion uncertainty index and the predicted trend of network path quality.

[0034] In an optional embodiment, the vehicle terminal maintains an independent network quality parameter time series for each network path, such as a 5G network and an LTE network. This time series includes measured round-trip time, packet loss rate, and throughput. The vehicle terminal then uses an autoregressive integral moving average model or a Kalman filter to predict the future trend of the network quality parameter time series. Alternate network path detection is achieved by sending path probe frames via the QUIC protocol, the transmission frequency of which is controlled by a timer. The probe period of this timer... Dynamic adjustments are made according to the following rules: when the motion uncertainty index... When the speed exceeds a preset high threshold, it indicates that the vehicle is moving violently. To avoid frequent switching that could cause positioning jumps, the detection cycle will be adjusted. Extend the detection period to reduce the detection frequency; when the predicted trend of any key network quality parameter is deteriorating, extend the detection period. Shorten it to a smaller value, such as 200ms, to speed up the discovery of available paths.

[0035] In a specific computational implementation, the vehicle terminal maintains the 20 most recent round-trip time samples of the primary path and uses a third-order autoregressive model as the prediction model to predict the round-trip time within the next second, thus obtaining the predicted round-trip time value. This third-order autoregressive model is a time series prediction network. Its core structure uses the round-trip time values ​​of the past three time points to linearly predict the round-trip time value of the current time point. The input to this third-order autoregressive model is the sequence of the nearest round-trip time samples along the network path, specifically the three consecutive round-trip time samples before the prediction time point. Round-trip time samples and round-trip time samples The output of this third-order autoregressive model is the predicted round-trip time value within a future time window. .

[0036] Specifically, the predicted round-trip time values ​​satisfy the following relationship:

[0037] In the formula, This indicates the predicted round-trip time value; Represents a constant term; This represents the first autoregressive coefficient; This represents the second autoregressive coefficient; This represents the third autoregressive coefficient; This represents the first round-trip time sample before the predicted time point; This represents the second round-trip time sample before the predicted time point; This represents the third round-trip time sample before the predicted time point.

[0038] It is understandable that the constant term and the first, second, and third autoregressive coefficients in the aforementioned third-order autoregressive model are obtained by the system through dynamic learning based on historical network state sequences. Specifically, the vehicle terminal or server maintains a fixed-length historical network round-trip time sample pool. For the time-series data in this sample pool, the system employs parameter estimation algorithms such as the least squares method or the Yurwalk equation to minimize the mean square error between the model's predicted output value and the historical actual round-trip time observations. This allows for real-time rolling fitting and solving of the constant term and the three autoregressive coefficients, ensuring that the prediction model closely follows the dynamic trends of current network fluctuations.

[0039] Obtain the predicted round-trip time value Then, calculate the predicted degradation rate of network quality. .

[0040] Specifically, the predicted degradation rate of network quality satisfies the following relationship:

[0041] In the formula, Indicates the predicted degree of network quality degradation; This represents the function that takes the maximum value. This indicates the predicted round-trip time value; This represents the sample mean of the most recent 20 round-trip times for the primary path.

[0042] Subsequently, the degradation rate was predicted based on network quality. and motion uncertainty index Calculate the detection frequency of the backup path. .

[0043] Specifically, the detection frequency satisfies the following relationship:

[0044] In the formula, Indicates the detection frequency; This indicates the basic detection frequency, which has a preset value of 1Hz. Indicates the predicted degree of network quality degradation; Indicates the uncertainty index of motion; This represents the network quality degradation feedback coefficient, which is 2 in this embodiment. The motion uncertainty feedback coefficient is represented by 1 in this embodiment; 1 is an operational constant.

[0045] It is understandable that the specific values ​​of the aforementioned network quality degradation feedback coefficient and motion uncertainty feedback coefficient are obtained through offline data training in real-vehicle network communication scenarios. In practical implementation, the system uses a hyperparameter optimization algorithm to determine the baseline values ​​based on the correlation data between detection frequency, network handover success rate, and positioning continuity in historical tests. Those skilled in the art can make targeted fine-tuning based on the hardware wake-up latency of the communication module installed in the vehicle and the preference for network sensitivity. For example, on terminals with limited communication resources, the aforementioned feedback coefficients can be appropriately lowered to reduce detection overhead.

[0046] After the calculation is completed, the vehicle terminal will detect the frequency. The calculation results are clamped within a preset frequency range of 0.2 Hz to 5 Hz to ensure timely detection and cost-effectiveness. The detection frequency... With motion uncertainty index The negative correlation is due to the high uncertainty of the inertial measurement unit itself when the vehicle is moving violently. Frequent network switching will exacerbate positioning jumps and fusion chaos. In this case, the detection frequency should be reduced to stabilize the current link; while the detection frequency With network quality prediction degradation The positive correlation is because when a network deterioration is predicted, it is necessary to proactively accelerate the detection of alternative paths to avoid connection interruption.

[0047] Thus, by combining motion index and network prediction to dynamically adjust the detection frequency, this step balances handover timeliness and link stability, and solves the risks of frequent disconnections and data delays caused by traditional passive handover mechanisms.

[0048] S3: When constructing QUIC data packets, the pre-integration increment of the inertial measurement unit is directly serialized as state prediction information and embedded into a custom frame.

[0049] In an optional embodiment, the state prediction information refers to the information since the last data packet transmission time. Up to the current moment The inertial measurement unit (IMU) pre-integration increment within the time interval. This IMU pre-integration increment refers to the relative attitude change, velocity increment, and position increment obtained by integrating the angular velocity and acceleration measured by the IMU between the two time points mentioned above, and contains a total of 9 floating-point values.

[0050] Considering the extreme sensitivity to latency and system resources in weak network communication environments, and to avoid the additional CPU computational overhead and serialization / deserialization latency caused by excessive compression encoding, this embodiment does not perform heavy compression processing on the floating-point number set. Instead, it directly uses standard network byte order (Big-Endian) to perform compact memory serialization and concatenation of the above floating-point values ​​to form a fixed-length binary bit stream of state prediction information. This approach minimizes the computational load and processing latency at both ends of the vehicle-cloud communication network while increasing the network load by a negligible amount.

[0051] After data serialization, the binary bitstream data is encapsulated into a custom QUIC protocol frame type. Specifically, an experimental type ID 0x3f01 is defined for the inertial measurement unit (IMU) prediction frame. This prediction frame structure includes the frame type, payload length, and a fixed-length IMU pre-integration increment. Then, the encapsulation is performed using the application programming interface (API) provided by the QUIC protocol stack, such as msquic or quiche. When constructing data packets at the vehicle terminal application layer, this custom frame object, along with regular STREAM frames, is submitted to the QUIC protocol stack. During packetization, the QUIC protocol stack treats this IMU prediction frame as an independent frame, placing it along with other frames into the payload area of ​​the QUIC data packet for overall encryption and transmission. Upon receiving the frame, the server prioritizes parsing the custom IMU prediction frame. Through simple network byte order deserialization, the IMU pre-integration increment can be quickly reconstructed for connection migration calculation.

[0052] Thus, this step directly serializes and encapsulates the inertial navigation increment into a custom QUIC frame through network byte order, reducing end-to-end transmission latency and the risk of packet loss and retransmission while avoiding excessive compression computational overhead.

[0053] S4: Maintain independent connection identifiers for multiple available network paths. Based on the historical performance model of each path issued by the server and combined with the current motion uncertainty index, evaluate the comprehensive transmission utility score of each path and select the connection identifier corresponding to the path with the highest score to transmit data.

[0054] In an optional embodiment, when establishing a multipath QUIC protocol multipath connection with the server, the vehicle terminal obtains multiple connection identifiers from the server via a new connection identifier frame, and associates different connection identifiers with various network paths, such as cellular networks and wireless LANs, to distinguish transmission channels. The server periodically sends a geographic location-based rasterized network performance map to the terminal. This network performance map model provides the expected bandwidth, latency, and packet loss rate for each network path in different regions. The vehicle terminal uses a ratio-based utility function to calculate the value of each path. Comprehensive transmission utility score The overall transmission utility score The score is proportional to throughput and inversely proportional to latency and packet loss rate, and a motion uncertainty risk penalty is only applied to non-current primary paths. Before each data transmission, the terminal recalculates the score of all network paths and sets the connection identifier field in the QUIC packet header to the connection identifier corresponding to the highest-scoring path, then transmits the data packet through the socket of that highest-scoring path.

[0055] Furthermore, the vehicle terminal periodically receives performance models of each available network path from the server. These performance models include the path... average throughput Average round trip time and packet loss rate The terminal will use the currently calculated motion uncertainty index. Converted to normalized risk weight factor This makes the normalized risk weight factor The range is between 1 and 2.5.

[0056] Specifically, the normalized risk weight factors satisfy the following relationship:

[0057] In the formula, Represents the normalized risk weighting factor; Indicates the uncertainty index of motion; The base offset value represents the risk weight; in this embodiment, the base offset value is 1. This represents the maximum penalty scaling ratio for the risk weight. In this embodiment, the maximum penalty scaling ratio is 1.5. The upper limit of the motion uncertainty index is 10 in this embodiment. The lower limit of the motion uncertainty index is represented by a reference value of 0.5 in this embodiment.

[0058] It is understandable that the aforementioned risk weight's base offset and maximum penalty scaling ratio are set based on engineering experience and the system's tolerance for frequent network switching. In practical applications, these parameters are determined by offline playback and optimization using extensive real-vehicle road test data in weak network environments to find the optimal balance between avoiding invalid link switching and ensuring real-time data transmission. Those skilled in the art can adjust these ratio parameters according to the specific autonomous driving safety redundancy requirements of different vehicle models.

[0059] Next, for each path Vehicle terminals are based on average throughput Average round trip time and packet loss rate Calculate the overall transmission utility score .

[0060] Specifically, the overall transmission utility score satisfies the following relationship:

[0061] In the formula, This represents the overall transmission utility score; Indicates average throughput; Indicates the average round-trip time; Indicates packet loss rate; This represents the throughput weight; in this embodiment, the throughput weight is set to 1. This represents the round-trip time weight; in this embodiment, the round-trip time weight is set to 0.02. This represents the packet loss rate weight. In this embodiment, the packet loss rate weight is set to 10.

[0062] It is understandable that the specific values ​​of the throughput weight, round-trip time weight, and packet loss rate weight mentioned above are not fixed, but are obtained through an offline calibration process. Specifically, based on historical datasets of network quality parameters collected from real vehicles under various network standards, this set of benchmark weight parameters is obtained offline through optimization, with the comprehensive objective function of minimizing end-to-end communication latency and improving the positioning system's reception success rate. Meanwhile, those skilled in the art can adaptively fine-tune these parameters according to the real-time requirements of the current driving task. For example, when the autonomous driving system is in a state with extremely high data integrity requirements, the packet loss rate weight can be increased accordingly.

[0063] Finally, the final comprehensive transmission utility score is calculated based on the path usage status. If the path This is not the currently primary path; the overall transmission utility score of this path. It needs to be divided by the normalized risk weight factor To obtain the final comprehensive transmission utility score. If the path If the current primary path is used, then the final overall transmission utility score is... Directly equal to the original comprehensive transmission utility score This measure reduces the attractiveness of alternative routes when vehicles are in high-intensity motion, thus suppressing frequent path changes that could lead to data interruptions. The vehicle terminal then selects the final overall transmission utility score. Data is transmitted via the path with the highest score.

[0064] Specifically, the final overall transmission utility score satisfies the following relationship: If path This is not the currently used primary path. ; If path This is the current primary path. ; In the formula, This represents the final overall transmission utility score; This represents the overall transmission utility score; This represents the normalized risk weight factor.

[0065] Thus, this step, by integrating motion uncertainty risk penalty assessment to evaluate the comprehensive utility of each path, suppresses invalid link switching under violent motion, laying an extremely stable data transmission foundation for subsequent smooth solution.

[0066] S5: The server uses factor graph fusion to perform positioning calculations on the received inertial measurement unit data, global navigation satellite system observations, and map information.

[0067] In an optional embodiment, the server uses the GTSAM or CeresSolver library to construct a factor graph. The variable nodes in this factor graph represent the vehicle state at different times, including 3D position, 3D velocity, and 3D attitude. Specifically, three types of factors are constructed in the factor graph: the first type is the inertial measurement unit (IMU) pre-integration factor, which connects the state vectors of two consecutive times. Its residual is calculated by the IMU pre-integration theory. Pre-integration can aggregate the IMU measurement sequence within a time period as incremental constraints, eliminating the need for point-by-point integration.

[0068] Next, the second and third types of factors are constructed. The second type is the Global Navigation Satellite System (GNSS) factor, which is a univariate factor connecting the state vectors at a single moment, and its residual is the difference between the vehicle's position at that moment and the position observed by the GNSS. The third type is the HD-Map matching factor, which is also a univariate factor. It calculates the lateral error as the residual by projecting the calculated position of the state vector onto the road centerline of the high-precision map.

[0069] After constructing the factor graph, the server calls nonlinear least squares optimization algorithms such as Levenberg-Marquardt or Doglg to solve for the state vector that minimizes the weighted sum of squares of all factor residuals, thereby obtaining the optimal vehicle trajectory.

[0070] Thus, this step, by rigorously fusing multi-source sensor and map data through factor graphs, overcomes the vulnerability of a single sensor in complex occlusion or weak network environments, and solves the problems of inaccurate positioning and decreased reliability.

[0071] S6: When the server detects a change in the source connection identifier of a data packet, it constructs a state transition constraint factor in the factor graph.

[0072] In an optional embodiment, the server maintains a record of the currently active connection identifier. When the connection identifier in the header of the received QUIC data packet differs from the active connection identifier of the previous moment, a connection migration event is determined to have occurred. At this time, the server parses the embedded state prediction information from the custom inertial measurement unit prediction frame of the first data packet of the new connection, and directly recovers the inertial measurement unit pre-integration increment through network byte order deserialization.

[0073] Next, set the old connection timestamp The timestamp corresponding to the last data packet on the old connection, the old connection timestamp The corresponding optimized state is the optimized state. ; Set the new connection timestamp The timestamp corresponding to the first data packet on the new connection, the new connection timestamp The corresponding state vector to be optimized is the state to be optimized. The terminal transmits data in a custom frame. The time interval includes state prediction information encompassing relative attitude changes, velocity increments, and position increments. The server decodes the inertial measurement unit (IMU) pre-integration increment for that time interval from the IMU prediction frame of the first data packet of the newly established connection. Based on the IMU pre-integration increment, the optimized state is generated. Derivation of the new connection timestamp Predicted state .

[0074] Subsequently, the server calls the `add` function of the GTSAM library to add a new custom binary factor, namely the state transition constraint factor, to the factor graph. This state transition constraint factor logically connects the optimized states. and state to be optimized Two variable nodes to predict the state State to be optimized The difference is the residual, which realizes the state continuity constraint before and after the connection migration.

[0075] Specifically, the residual of the state transition constraint factor Calculate the expression that satisfies the following relation:

[0076] In the formula, The residual represents the state transition constraint factor; Indicates the predicted state; This indicates a state that needs optimization.

[0077] Thus, this step, by introducing the construction of before-and-after state constraint factors during connection migration, smoothly maps the physical layer network switching to the application layer algorithm, thus solving the risks of trajectory breakage and model divergence.

[0078] S7: The state transition constraint factor uses the previous connection state as the initial condition, parses the state prediction information embedded in the first data packet of the new connection to obtain the pre-integration increment of the inertial measurement unit, and performs state propagation to obtain the predicted state.

[0079] In an optional embodiment, when the server receives a data packet, it checks the source connection identifier. When the source connection identifier changes from the old connection identifier to the new connection identifier, it immediately locks the last moment associated with the old connection identifier, i.e., the moment the previous data packet was sent. state vector and covariance matrix .

[0080] Specifically, the state vector satisfies the following relation:

[0081] In the formula, This represents the state vector at the time the previous data packet was sent; This represents the position vector at the time the previous data packet was sent; This represents the velocity vector at the moment the previous data packet was sent; This represents the attitude vector at the moment the previous data packet was sent; This represents the matrix transpose operation.

[0082] Next, the server parses the custom inertial measurement unit prediction frame from the first data packet of the new connection identifier and then parses the payload of this custom frame. The server directly recovers the data from the time the previous data packet was sent by performing network byte order deserialization. Up to the current moment The inertial measurement unit pre-integration increment within the time interval includes attitude change, velocity increment, and position increment.

[0083] Subsequently, based on the motion equations of the inertial measurement unit (IMU), state propagation is completed based on the reconstructed IMU pre-integration increments. The vehicle's attitude, velocity, and position are directly updated by the IMU pre-integration increments, thus yielding the predicted state vector. The predicted state vector contains the predicted location component. and predicted velocity components Meanwhile, the server uses a first-order Taylor expansion to linearize the motion equations and calculates the state transition matrix. And propagate the covariance to obtain the covariance matrix of the predicted state. .

[0084] Specifically, the covariance matrix of the predicted state Satisfying the relation:

[0085] In the formula, The covariance matrix represents the predicted state; Represents the state transition matrix; Represents the covariance matrix of the previous time step; This represents the transpose of the state transition matrix; This represents the process noise covariance.

[0086] The process noise covariance mentioned above is directly set according to the factory calibration parameters of the inertial measurement unit (IMU) it is equipped with, or the random walk noise density of the gyroscope angular velocity and the random walk noise density of the accelerometer velocity of the IMU are measured offline by the Allan variance calibration method, and combined with the zero bias instability parameter, a diagonal matrix is ​​constructed to obtain the process noise covariance.

[0087] Thus, this step effectively compensates for the data blind spots during network outage migration by parsing the custom frame increment and combining it with the previous state for prior propagation, and solves the problem of discontinuous prediction caused by network latency or out-of-order delivery.

[0088] S8: The predicted state and the initial state calculated based on the initial observation data of the new connection are used to form the residual term.

[0089] In an optional embodiment, the initial observation data based on the new connection refers to the Global Navigation Satellite System (GNSS) observations carried in the new data packet. The server extracts the associated GNSS raw observation data from the first data packet of the new connection identifier. The server uses a separate point positioning algorithm, such as weighted least squares, to calculate the current time based solely on the aforementioned GNSS observations. Unsmoothed initial state estimation Initial state estimation Includes initial position and speed Components. Simultaneously, the server calculates the covariance matrix corresponding to this initial state estimate based on the satellite's geometric distribution and observation noise. .

[0090] Next, the server defines the residual term of the state transition constraint factor as the difference between the predicted state obtained based on the data propagation of the inertial measurement unit and the observed state calculated by the global navigation satellite system, and defines it as the state prediction deviation vector.

[0091] Specifically, the state prediction bias vector satisfies the following relationship:

[0092] In the formula, This represents the state prediction deviation vector; This represents the predicted state obtained from data based on inertial measurement unit (IMU) data; This represents the initial state estimate calculated based on observations from the global navigation satellite system.

[0093] Furthermore, during the differential operation, the predicted state obtained in the previous step based on the inertial measurement unit data is used... Positional components in and velocity components The initial position in the initial state calculated by the global navigation satellite system and speed Perform vector subtraction to construct a state prediction bias vector with dimensions 6 by 1. .

[0094] Specifically, the state prediction bias vector satisfies the following relationship:

[0095] In the formula, This represents the state prediction deviation vector; Represents the positional component in the predicted state; This represents the initial position calculated based on the global navigation satellite system. This represents the velocity component in the predicted state; This indicates the speed of calculation based on the global navigation satellite system; This represents the matrix transpose operation.

[0096] Thus, this step constructs an accurate state prediction bias vector by performing a difference operation between the predicted state and the observed initial state of the new connection, providing a rigorous mathematical benchmark for subsequent adaptive adjustment of constraint factor weights.

[0097] S9: The covariance of the state transition constraint factor is adjusted based on the Mahalanobis distance of the residual term and the preset penalty upper limit, thereby outputting the vehicle position.

[0098] In an optional embodiment, the squared value of the Mahalanobis distance, representing the consistency between the predicted and observed state distributions, is calculated by combining the state prediction bias vector, the covariance matrix of the predicted state, and the covariance matrix of the initial state calculated based on the global navigation satellite system. First, calculate the prior covariance matrix of the state prediction bias vector. The prior covariance matrix It is determined by both the observation noise of the global navigation satellite system and the uncertainty of the state propagation model.

[0099] Specifically, the prior covariance matrix satisfies the following relationship:

[0100] In the formula, Represent the prior covariance matrix; The 6x6 dimension sub-blocks in the covariance matrix representing the position and velocity of the predicted state; This represents the covariance matrix of the initial state calculated based on the global navigation satellite system.

[0101] After obtaining the prior covariance matrix Next, calculate the squared value of the Mahalanobis distance. .

[0102] Specifically, the squared value of the Mahalanobis distance satisfies the following relationship:

[0103] In the formula, Represents the squared value of the Mahalanobis distance; This represents the state prediction deviation vector; This represents the transpose of the state prediction bias vector; represents the inverse of the prior covariance matrix.

[0104] The square of the Mahalanobis distance It follows a chi-square distribution with 6 degrees of freedom. Based on this characteristic, a threshold based on the 95% confidence interval of the chi-square distribution is set. Combining Figure 2 As shown, based on the squared value of the Mahalanobis distance Relative to threshold The size is determined by using a weighting function such as the Huber function or the Tukeybiweight function to calculate a scaling factor. .

[0105] Specifically, the scaling factor satisfies the following relationship:

[0106] In the formula, Indicates the scaling factor; Represents the squared value of the Mahalanobis distance; This represents the threshold based on the 95% confidence interval of the chi-square distribution; This represents the function that takes the maximum value. This represents the adjustment coefficient, for example, this adjustment coefficient. The default value is 0.5; 1 and 0 are operational constants.

[0107] Understandably, the specific values ​​of the aforementioned adjustment coefficients are obtained through offline calibration based on the system's requirements for the isolation penalty strength of abrupt state errors. Specifically, based on an extreme test dataset incorporating artificial network packet loss and out-of-order delivery, the convergence and positioning smoothness of factor graph optimization under different penalty strengths are observed. An optimization algorithm is then used to determine the baseline parameters that best suppress position jumps. Those skilled in the art can make adaptive numerical adjustments based on the trajectory smoothness requirements of specific application scenarios.

[0108] scaling factor Its value is 1 when the estimates of the predicted and observed states are in good agreement. Next, the covariance matrix of the state transition constraint factor is constructed. The covariance matrix The amplitude setting is used A function of the adjusted squared Mahalanobis distance. In factor graph optimization, the information matrix of the state transition constraint factors. That is, the covariance matrix The inverse matrix. Due to the covariance matrix... Increase, information matrix The weight of the state transition constraint factor is reduced accordingly.

[0109] It is particularly important to note that if the network outage due to a weak network is prolonged, the predicted state upon network recovery may deviate significantly from the observed state of the global navigation satellite system due to the cumulative drift of the inertial measurement unit itself. This can lead to a substantial discrepancy in the calculated squared Mahalanobis distance. It surges exponentially. Without a penalty cap, this leads to... As it approaches infinity, the covariance matrix It also tends towards infinity, and its information matrix The value will rapidly approach 0. In back-end algorithms based on nonlinear least squares optimization, such as GTSAM or CeresSolver, a constraint factor that causes an information matrix to approach 0 will completely sever the state nodes of the mathematical model, thereby causing the Jacobian matrix to be reduced in rank, which will cause the optimization solver to collapse and diverge directly.

[0110] This invention introduces in the formula As a truncation mechanism, a dynamically adjustable penalty upper limit is set. This means that even when encountering extremely severe abrupt large error data, the system can still maintain a weak but absolutely non-zero continuous topological constraint between preceding and following connection states, while significantly reducing the weight of outlier data for isolation. This mechanism not only isolates gross errors and smooths trajectories but also mathematically eliminates the divergence problem in factor graph optimization, ensuring the stability of the system under complex operating conditions at the vehicle end. Finally, the covariance matrix after this adjustment mechanism is... After adding the state transition constraint factor to the factor graph, the optimization solution is executed again to output continuous and high-precision vehicle positions.

[0111] Understandably, in practical implementation, the preset penalty upper limit is determined by the minimum non-zero singular value allowed by the Jacobian matrix in the factor graph optimization solver. To ensure that the information matrix does not fall into ill-conditioned conditions, this embodiment uses offline extreme condition testing to define the value of the preset penalty upper limit as a fixed constant within the range of 100 to 500, for example, a value of 200. When the calculated scaling factor is greater than the preset penalty upper limit, the scaling factor is forced to be equal to the preset penalty upper limit, thereby significantly isolating severe jump errors while retaining a minimum positive definite information matrix constraint and maintaining the topological connectivity of the optimized graph structure.

[0112] Specifically, the covariance matrix satisfies the following relationship:

[0113] In the formula, The covariance matrix represents the state transition constraint factor; Indicates the scaling factor; The 6x6 sub-blocks in the covariance matrix representing the position and velocity of the predicted state.

[0114] Specifically, information matrix Calculate the expression that satisfies the following relation:

[0115] In the formula, Represents the information matrix; The covariance matrix representing the state transition constraint factor; This represents the inverse of the covariance matrix of the state transition constraint factors.

[0116] Thus, this step dynamically adjusts the constraint covariance based on Mahalanobis distance and penalty upper limit, while maintaining the non-zero topological constraint of the factor graph, preventing divergence and achieving continuous high-precision output.

[0117] To verify the actual effect of the vehicle inertial navigation-assisted positioning and correction method in the weak network environment, the present invention also provides a specific experimental verification embodiment.

[0118] Specifically, in the experimental setup, the test vehicle was equipped with an inertial measurement unit, a multi-frequency global navigation satellite system receiver, and a dual-mode communication module supporting 4G and 5G networks. The test vehicle traveled along a predetermined 3-kilometer route in a complex urban environment, and network switching events were manually triggered five times during the journey.

[0119] This experiment compared three schemes: the first was the basic scheme, which did not employ any state prediction or constraint mechanism; the second was the fixed-weight scheme, which used state prediction but the covariance of the state transition constraint factor was a fixed value; and the third was the scheme presented in this paper, which fully implemented the aforementioned edge-side detection frequency adjustment, custom frame direct serialization embedding, and cloud-side constraint weight adjustment mechanism based on Mahalanobis distance. The evaluation metrics for this experiment included root mean square error in localization, maximum position jump at handover time, and average network overhead.

[0120] like Figure 3 The figure shows a comparison of the experimental results of the three schemes. Regarding positioning accuracy, the root mean square error (RMSE) of the basic scheme is 2.53 meters, the fixed weight scheme improves the RMSE to 1.21 meters, while the RMSE of the scheme presented in this paper reaches 0.82 meters. In terms of the crucial handover smoothness indicator, the basic scheme experiences a maximum position jump of 5.8 meters during network handover, manifesting as a trajectory break; the fixed weight scheme suppresses this position jump to 2.1 meters, but it is still perceptible; the scheme presented in this paper, by adjusting the constraint strength, controls the maximum position jump within 0.3 meters, achieving a seamless transition. Regarding resource consumption, based on the 5kbps data transmission overhead of the basic scheme, the fixed weight scheme, due to its fixed detection and data reporting mechanism, increases its overhead to 7.5kbps; while the scheme presented in this paper, benefiting from the detection frequency adjustment based on the motion uncertainty index and network quality prediction, has an average network overhead of only 6.2kbps.

[0121] Compared to the basic scheme, this invention reduces the positioning error by 67.6% and the position jump by 94.8%, fully demonstrating the significant advantages of using embedded state prediction information for forward and backward connection state propagation. Compared to the fixed weight scheme, this invention further reduces the positioning error by 32.2% and the maximum jump by 85.7%, fully illustrating the crucial importance of the core innovation of the Mahalanobis distance adjustment factor graph constraint weight, which can effectively identify and isolate potential gross errors at the moment of switching, avoiding contamination of the overall trajectory optimization results.

[0122] Furthermore, the success of this invention stems from end-to-end collaborative planning. On the vehicle terminal side, the detection frequency adjustment mechanism accurately matches the cost of network detection with the vehicle's motion state and the changing trends of the network environment, saving over 17% of network resources while ensuring timely switching. Simultaneously, the simplified encapsulation of direct serialization avoids additional computational latency at both the end-to-end and cloud-side while maintaining absolute accuracy. On the server side, the constraint mechanism based on state prediction and measured residuals becomes crucial for achieving high-smoothness trajectory continuity. This endows the navigation system with stability and recovery capabilities when facing non-ideal network switching events, ensuring the accuracy and continuity of the output trajectory.

[0123] Figure 2 This diagram illustrates the covariance adjustment of the constraint factor. The vertical dashed line represents the confidence interval threshold set based on the chi-square distribution. As can be seen from the diagram, when the squared Mahalanobis distance is less than this confidence interval threshold (to the left of the vertical dashed line), the covariance amplitude remains at a low, constant baseline level, indicating a very high consistency between the predicted and observed states, and the constraint factor exerts a strong smoothing effect. Conversely, when the squared Mahalanobis distance exceeds this confidence interval threshold (to the right of the vertical dashed line), the covariance amplitude exhibits a strictly linear increasing trend with the deviation of the squared Mahalanobis distance, thus reducing its optimization weight in the factor plot. This diagram visually verifies the system's penalty mechanism for dynamically isolating errors in abrupt changes.

[0124] This invention also discloses a vehicle inertial navigation-assisted positioning and correction system in a weak network environment, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, a vehicle inertial navigation-assisted positioning and correction method in a weak network environment according to the present invention is implemented.

[0125] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0126] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A vehicle inertial navigation-assisted positioning and correction method in a weak network environment, characterized in that, include: The system acquires raw data from the vehicle terminal's inertial measurement unit (IMU) and calculates the motion uncertainty index. Based on the motion uncertainty index and the predicted trend of network path quality, the detection frequency is adjusted. When constructing QUIC data packets, the IMU pre-integration increment is directly serialized as state prediction information and embedded into a custom frame. Independent connection identifiers are maintained for multiple available network paths. Based on the historical performance model of each path issued by the server and combined with the current motion uncertainty index, the comprehensive transmission utility score of each path is evaluated, and the connection identifier corresponding to the path with the highest score is selected for data transmission. The server uses a factor graph to fuse the received IMU data, global navigation satellite system observations, and map information for positioning calculation. When the server detects a change in the source connection identifier of a data packet, a state transition constraint factor is constructed in the factor graph. The state transition constraint factor uses the previous connection state as an initial condition, parses the state prediction information embedded in the first data packet of the new connection to obtain the IMU pre-integration increment, and performs state propagation to obtain the predicted state. The process of obtaining the predicted state through state propagation includes: on the server side, when a change in the source connection identifier of a QUIC data packet to a new connection identifier is detected, locking the state vector and covariance matrix of the previous data packet sent at the time associated with the old connection identifier; parsing the state prediction information from the custom frame of the first data packet of the new connection identifier to obtain... The inertial measurement unit pre-integration increment is calculated; based on the inertial navigation motion equation, the predicted state vector and covariance matrix at the current moment are calculated using the state vector at the previous moment and the analyzed inertial measurement unit pre-integration increment. The predicted state and the initial state calculated based on the initial observation data of the new connection constitute the residual term; the covariance of the state transition constraint factor is adjusted according to the Mahalanobis distance of the residual term and the preset penalty upper limit, thereby outputting the vehicle position.

2. The vehicle inertial navigation-assisted positioning and correction method in a weak network environment according to claim 1, characterized in that, The steps of acquiring raw data from the inertial measurement unit (IMU) of the vehicle terminal, calculating the motion uncertainty index, and adjusting the detection frequency include: calculating the statistical variance based on the three-axis angular velocity and three-axis acceleration of the IMU within a preset time window, and combining the statistical variance after dimensionless processing to form the motion uncertainty index; establishing a prediction model based on the recent round-trip time sample sequence of the network path, estimating the network quality change trend within the future time window, and obtaining the predicted network quality degradation degree; determining the detection frequency based on the motion uncertainty index and the predicted network quality degradation degree, and limiting the detection frequency between a preset minimum and maximum detection frequency.

3. The vehicle inertial navigation-assisted positioning and correction method in a weak network environment according to claim 1, characterized in that, The step of directly serializing and embedding the pre-integrated increment of the inertial measurement unit as state prediction information into a custom frame includes: defining a custom frame type in the QUIC protocol; and obtaining the frame type from the time of the previous data packet transmission. Up to the current moment The inertial measurement unit pre-integrates the increment, and directly serializes the floating-point value corresponding to the pre-integration increment into a fixed-length binary bit stream data through network byte order; the binary bit stream data is filled into the custom frame as a payload; when constructing the QUIC data packet to be sent, this custom frame is inserted into the frame sequence of the data packet.

4. The vehicle inertial navigation-assisted positioning and correction method in a weak network environment according to claim 1, characterized in that, The evaluation of the comprehensive transmission utility score for each path includes: obtaining the historical performance model of each network path from the server, wherein the historical performance model includes average throughput, average round-trip time, and packet loss rate; converting the motion uncertainty index into a normalized risk weighting factor; calculating the original comprehensive transmission utility score based on the average throughput, average round-trip time, and packet loss rate; and dividing the original comprehensive transmission utility score by the risk weighting factor for other available paths besides the current primary path to obtain the final comprehensive transmission utility score.

5. The vehicle inertial navigation-assisted positioning and correction method in a weak network environment according to claim 1, characterized in that, The residual term is constructed by: using the global navigation satellite system observations associated with the first data packet of the new connection to calculate the initial state vector and covariance matrix at the current moment, wherein the initial state vector includes position and velocity components; and performing a difference operation between the propagated predicted state vector and the initial state vector calculated based on the global navigation satellite system to construct a state prediction deviation vector as the residual term.

6. The vehicle inertial navigation-assisted positioning and correction method in a weak network environment according to claim 1, characterized in that, The covariance of the state transition constraint factor is adjusted based on the Mahalanobis distance of the residual term and a preset penalty upper limit, including: combining the state prediction deviation vector, the covariance matrix of the predicted state, and the covariance matrix of the initial state to calculate the squared Mahalanobis distance representing the consistency between prediction and observation; calculating a scaling factor based on the degree to which the squared Mahalanobis distance deviates from the confidence interval threshold, and truncating the scaling factor using the preset penalty upper limit when the deviation surges due to long-term network disconnection and the scaling factor tends to infinity; setting the magnitude of the covariance matrix of the state transition constraint factor as a function of the squared Mahalanobis distance adjusted by the scaling factor, so that the weight of the factor is reduced when the difference between prediction and observation increases, while maintaining effective information matrix constraints.

7. The vehicle inertial navigation-assisted positioning and correction method in a weak network environment according to claim 1, characterized in that, The server uses factor graph fusion positioning calculation, which includes: constructing three types of factors in the factor graph. The first type is the inertial measurement unit pre-integration factor that connects the state vectors of two consecutive time moments. The second type is the factor that connects the state vector of a single time moment with the observation value of the global navigation satellite system. The third type is the map matching factor that projects the calculated position onto the road centerline of the high-precision map. The nonlinear least squares optimization algorithm is called to solve for the state vector.

8. The vehicle inertial navigation-assisted positioning and correction method in a weak network environment according to claim 2, characterized in that, The method of calculating the statistical variance of the three-axis angular velocity and three-axis acceleration of the inertial measurement unit within a preset time window, and combining the statistical variances into a motion uncertainty index after dimensionless processing, includes: acquiring the original data of the inertial measurement unit within a preset sliding time window, calculating the angular velocity variance of the three-axis angular velocity and the acceleration variance of the three-axis acceleration respectively; dividing the angular velocity variance and the acceleration variance by a preset reference variance threshold for dimensionless processing, the reference variance threshold being pre-calibrated by the vehicle under a straight uniform speed state; and weighting and summing the dimensionless angular velocity variance and acceleration variance based on preset weighting coefficients to calculate the final motion uncertainty index.

9. The vehicle inertial navigation-assisted positioning and correction method in a weak network environment according to claim 2, characterized in that, The method of establishing a prediction model based on recent round-trip time sample sequences of network paths to estimate the trend of network quality changes within a future time window and obtain the predicted degree of network quality degradation includes: using an autoregressive model to predict the round-trip time within a future time window, wherein the input of the autoregressive model is multiple consecutive round-trip time samples before the prediction time point, and the output is the predicted round-trip time value; calculating the difference between the predicted round-trip time value and the sample mean of the most recent round-trip time samples of the preset primary path; taking the maximum value between the difference and zero to obtain the predicted degree of network quality degradation.

Images