Vehicle positioning method and related device

Abstract

The invention discloses a vehicle positioning method and a related device, and the method comprises the steps: carrying out the calculation based on IMU data, RTK data and the wheel speed pulse data of each wheel through a satellite inertial navigation fusion positioning frame, obtaining a reference deviation, carrying out the training of a long and short term artificial neural network LSTM model according to the IMU data and the reference deviation, and obtaining an IMU deviation estimation model, determining an estimated deviation of an inertial measurement unit according to the IMU data by using an inertial measurement unit IMU deviation estimation model, further determining an estimated position and an estimated attitude of the target vehicle according to pre-integration of the wheel speed pulse data of each wheel, and then determining the estimated position and the estimated attitude of the target vehicle on the basis of an error state Kalman filtering model. And determining a target position and a target attitude of the target vehicle according to the estimated deviation, the estimated position and the estimated attitude so as to reduce error accumulation of vehicle inertia recursion and improve the accuracy of vehicle positioning.

Description

Technical Field

[0001] This application relates to the field of high-precision vehicle mapping and positioning technology, and in particular to a vehicle positioning method and related apparatus. Background Technology

[0002] Vehicle localization refers to the process of determining a vehicle's position and orientation on a map. In applications of autonomous driving and advanced driver assistance systems (ADAS), high real-time positioning accuracy is required. Existing technologies place high demands on the vehicle localization capabilities of intelligent driving systems in complex environments.

[0003] Currently, related technologies typically involve acquiring inertial measurement unit (IMU) data and then using inertial sensor data for vehicle positioning. However, IMUs are affected by factors such as vehicle motion (e.g., vibration and shock) and temperature, leading to systemic biases and low accuracy of IMU data. This, in turn, causes a rapid accumulation of positioning errors, resulting in low vehicle positioning precision. Summary of the Invention

[0004] To address the aforementioned problems, this application provides a vehicle positioning method and related apparatus to solve the problem of low vehicle positioning accuracy caused by systematic errors in inertial measurement units.

[0005] The embodiments of this application disclose the following technical solutions:

[0006] In a first aspect, embodiments of this application provide a vehicle positioning method, the method comprising:

[0007] The inertial measurement unit (IMU) data and reference bias of the inertial measurement unit in the target vehicle are acquired. The reference bias is calculated based on the IMU data, real-time dynamic RTK data and wheel speed pulse data of each wheel of the target vehicle through a satellite inertial navigation fusion positioning framework.

[0008] Based on the inertial measurement unit (IMU) bias estimation model, the estimated bias of the inertial measurement unit is determined according to the IMU data and the reference bias. The IMU bias estimation model is obtained by training a long short-term artificial neural network (LSTM) model based on the training inertial measurement unit data and the reference bias.

[0009] The estimated position and attitude of the target vehicle are determined based on the wheel speed pulse data of each wheel.

[0010] Based on the error state Kalman filter model, the target position and target attitude of the target vehicle are determined according to the estimated bias, estimated position and estimated attitude.

[0011] Optionally, the IMU bias estimation model is trained in the following way:

[0012] Acquire training inertial measurement unit data and reference bias;

[0013] The training inertial measurement unit data and reference bias are normalized to obtain the normalized result;

[0014] The normalization result is input into the LSTM model to be trained to obtain the training estimation bias;

[0015] The loss function value is determined based on the training estimation bias and the reference bias;

[0016] The loss function value is used to train the LSTM model to be trained.

[0017] Optionally, acquiring training inertial measurement unit data and reference bias includes:

[0018] The data collected by the inertial measurement unit from the current time to the previous n time steps are used as the training data for the inertial measurement unit.

[0019] The reference deviation of the inertial measurement unit is determined based on the data from the inertial measurement unit, real-time dynamic RTK data, and wheel speed pulse data of each wheel of the target vehicle.

[0020] Optionally, the loss function value is determined based on the training estimation bias and the reference bias, including:

[0021] The accelerometer loss function and gyroscope loss function are determined based on the LSTM model to be trained;

[0022] The accelerometer loss function value is determined based on the training estimation bias, reference bias, and accelerometer loss function, and the gyroscope loss function value is determined based on the training estimation bias, reference bias, and gyroscope loss function.

[0023] The LSTM model to be trained is trained using the loss function value, including:

[0024] The LSTM model to be trained is trained using accelerometer loss function values ​​and gyroscope loss function values.

[0025] Optionally, the estimated position and estimated attitude of the target vehicle are determined based on the wheel speed pulse data of each wheel, including:

[0026] The wheel speed pulse count and linear velocity of each wheel are determined based on the wheel speed pulse data of each wheel.

[0027] The first estimated position and first estimated attitude of the target vehicle are determined based on the wheel speed pulse count and linear velocity of each wheel.

[0028] Optionally, based on the error state Kalman filter model, the target position and target attitude of the target vehicle are determined according to the estimation bias, estimated position, and estimated attitude, including:

[0029] Using the inertial measurement unit pre-integration model, the second estimated position and second estimated attitude of the target vehicle are determined based on IMU data;

[0030] Based on the second estimated position and the second estimated attitude, the predicted position and predicted attitude of the target vehicle are determined by the state transition equation in the error state Kalman filter model.

[0031] An observation model is constructed based on the estimation bias, the first estimated position, and the first estimated attitude.

[0032] The predicted position and attitude are updated using the observation model to obtain the target position and attitude of the target vehicle.

[0033] Optionally, the predicted position and predicted attitude are updated using the observation model to obtain the target position and target attitude of the target vehicle, including:

[0034] Determine the Kalman gain of the error Kalman model using the observation model;

[0035] The error state vector of the error Kalman model is updated based on the Kalman gain;

[0036] The predicted position and predicted attitude are updated using the updated error state vector to obtain the target position and target attitude of the target vehicle.

[0037] Secondly, embodiments of this application provide a vehicle positioning device, the device comprising:

[0038] The acquisition module is used to acquire the inertial measurement unit (IMU) data and reference deviation of the inertial measurement unit in the target vehicle; the reference deviation is calculated based on the IMU data, real-time dynamic RTK data and wheel speed pulse data of each wheel of the target vehicle through a satellite inertial navigation fusion positioning framework.

[0039] The estimation bias determination module is used to determine the estimated bias of the inertial measurement unit (IMU) based on the IMU bias estimation model, according to the IMU data and the reference bias. The IMU bias estimation model is obtained by training a long short-term artificial neural network (LSTM) model based on the training IMU data and the reference bias.

[0040] The estimation module is used to determine the estimated position and estimated attitude of the target vehicle based on the wheel speed pulse data of each wheel;

[0041] The vehicle localization module is used to determine the target position and target attitude of the target vehicle based on the error state Kalman filter model, according to the estimated deviation, estimated position and estimated attitude.

[0042] Thirdly, embodiments of this application provide a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the vehicle positioning method as described in the first aspect.

[0043] Fourthly, embodiments of this application provide a computer-readable storage medium storing instructions that, when executed on a terminal device, cause the terminal device to perform the vehicle positioning method as described in the first aspect.

[0044] Compared with the prior art, this application has the following beneficial effects:

[0045] The vehicle positioning method provided in this application embodiment determines the estimated deviation of the inertial measurement unit (IMU) based on the acquired IMU data and the acquired reference deviation using an inertial measurement unit (IMU) deviation estimation model. The reference deviation is obtained based on the IMU data, real-time dynamic RTK data, and wheel speed pulse data of each wheel of the target vehicle. The IMU deviation estimation model is obtained by training a long short-term artificial neural network (LSTM) model using training IMU data and the reference deviation. Further, the estimated position and estimated attitude of the target vehicle are determined based on the wheel speed pulse data of each wheel. Then, based on an error state Kalman filter model, the target position and target attitude of the target vehicle are determined according to the estimated deviation, estimated position, and estimated attitude.

[0046] This application takes into account the time correlation characteristics of deviations in IMU data. Therefore, it uses an IMU deviation estimation model trained on an LSTM model to accurately estimate the deviations generated by the inertial measurement unit. Furthermore, it uses wheel speed pulse data to constrain the determined estimated deviations, thereby reducing the accumulation of vehicle positioning errors and improving the accuracy of vehicle positioning. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of this application 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 this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 A schematic diagram of a vehicle positioning process provided in an embodiment of this application;

[0049] Figure 2A schematic diagram illustrating vehicle positioning as provided in an embodiment of this application;

[0050] Figure 3 A schematic diagram of an LSTM model provided in an embodiment of this application;

[0051] Figure 4 This is a schematic diagram of a vehicle positioning structure provided in an embodiment of this application. Detailed Implementation

[0052] As described earlier, research on vehicle positioning has revealed that the mapping, positioning, and planning control required for advanced intelligent driving applications place high demands on the positioning capabilities of intelligent vehicles in complex environments. Current vehicle positioning systems generally rely on satellite positioning or high-precision maps to ensure positioning accuracy, which greatly limits the application scenarios applicable to advanced intelligent driving functions. In particular, it presents significant challenges in urban environments with poor satellite signals, inability to provide timely coverage of high-precision maps, and failure of scene perception.

[0053] Currently, mass-produced vehicles are generally equipped with low-cost Mems inertial measurement units (IMUs). Vibrations and shocks during vehicle operation may lead to unstable measurement results, and temperature changes may cause temperature drift in inertial measurements. In addition, vehicles may produce nonlinear responses under dynamic conditions such as high-speed driving and frequent lane changes. In order to ensure the safety and convenience of intelligent driving functions, it is necessary to be able to calibrate the system error of the IMU itself online and isolate the influence of platform movement and external environment on the stability of IMU measurement. Otherwise, the system deviation in the IMU data will cause the positioning error to accumulate rapidly during integration.

[0054] To address the error accumulation problem of IMUs in practical applications, related technologies can mitigate these issues to some extent by fusing wheel speed sensor data and vehicle non-holonomic dynamic constraints. However, due to hardware cost considerations for mass-produced vehicles, wheel speed pulse resolution is generally low. The vehicle speed and angular velocity measurements calculated based on the Ackerman model are significantly affected by signal delay and wheel speed pulse noise, which cannot effectively constrain the accumulated error of IMU position estimation. This causes the vehicle's positioning error to accumulate rapidly in extremely complex environments, making it difficult to guarantee the reliability of positioning functions in urban environments, and consequently reducing vehicle positioning accuracy.

[0055] To address the aforementioned problems, embodiments of this application provide a vehicle positioning method and related apparatus. The method includes: determining the estimated deviation of the inertial measurement unit (IMU) based on acquired IMU data and an acquired reference deviation using an inertial measurement unit (IMU) deviation estimation model; wherein the reference deviation is obtained based on IMU data, real-time dynamic RTK data, and wheel speed pulse data of each wheel of the target vehicle; the IMU deviation estimation model is obtained by training a long short-term artificial neural network (LSTM) model based on training IMU data and the reference deviation; further determining the estimated position and estimated attitude of the target vehicle based on the wheel speed pulse data of each wheel; and then determining the target position and target attitude of the target vehicle based on an error state Kalman filter model, according to the estimated deviation, estimated position, and estimated attitude.

[0056] Therefore, taking into account the characteristics of the time correlation of deviations in IMU data, this application uses the IMU deviation estimation model obtained by training the LSTM model to accurately estimate the deviation generated by the inertial measurement unit, and further uses wheel speed pulse data to constrain the determined estimated deviation, thereby reducing the accumulation of vehicle positioning errors and improving the accuracy of vehicle positioning.

[0057] In other words, this application starts from the actual needs of urban intelligent driving applications for multi-sensor fusion positioning stability. Based on data-driven training, it considers the time correlation characteristics of IMU system bias and implements a deep inference of IMU bias based on long short-term memory artificial neural network (LSTM). It further integrates the vehicle odometer with low-resolution wheel speed pulse signal pre-integration calibration to achieve enhanced positioning and reduce deployment and maintenance costs.

[0058] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0059] Example 1

[0060] See Figure 1 This figure is a flowchart illustrating a vehicle positioning method provided in an embodiment of this application. Figure 1 As shown, the vehicle positioning method provided in this application embodiment may include:

[0061] S101: Acquire the inertial measurement unit (IMU) data and reference bias of the inertial measurement unit in the target vehicle.

[0062] The target vehicle refers to the vehicle whose location is to be determined.

[0063] An Inertial Measurement Unit (IMU) is a device that integrates sensors such as accelerometers and gyroscopes to measure and monitor the linear acceleration and angular velocity of an object. Based primarily on Newton's laws of motion, the IMU measures the acceleration of the carrier in an inertial reference frame, integrates it over time, and transforms it to the navigation coordinate system, thereby obtaining information such as velocity, yaw angle, and position in the navigation coordinate system.

[0064] The inertial measurement unit contains three single-axis accelerometers and three single-axis gyroscopes. The accelerometers detect the acceleration signals of the object on three independent axes of the carrier coordinate system, while the gyroscopes detect the angular velocity signals of the carrier relative to the navigation coordinate system, thereby calculating the attitude of the object.

[0065] Wheel speed pulse data refers to the number of pulse signals generated each time the wheels of a target vehicle rotate during its operation.

[0066] Reference bias refers to the IMU bias used as a reference condition during the training phase. IMU bias is the difference between the measured value of the inertial measurement unit and the value when there are no interference terms.

[0067] It should be understood that, in the embodiments of this application, in order to better estimate the deviation of the inertial measurement unit, it is necessary to train the model parameters based on the reference deviation during the training process, so that the deviation of the inertial measurement unit estimated later is closer to the accurate value, thereby improving the accuracy of vehicle positioning.

[0068] The reference bias can be obtained using a satellite-inertial navigation fusion positioning solution framework based on real-time dynamic (RTK) data (training phase only), IMU data, and wheel speed pulse data of each wheel of the target vehicle. Real-time dynamic (RTK) data is a high-precision positioning data based on differential GPS. RTK technology, through real-time communication and data processing, can provide positioning accuracy at the centimeter or even sub-meter level.

[0069] It should be understood that, under good signal conditions, the reference deviation of the inertial measurement unit can be determined using IMU data, RTK data, and wheel speed pulse data of each wheel.

[0070] In one possible implementation, the reference bias can be obtained in the following way:

[0071] A1: Data synchronization and preprocessing.

[0072] To ensure the accuracy of the reference bias, the IMU data, RTK data, and wheel speed pulse data need to be synchronized in time before determining the reference bias. In some possible implementations, this can be achieved through data timestamp alignment and sampling rate matching. Furthermore, data preprocessing is performed, including but not limited to denoising, outlier removal, and outlier handling, without specific limitations here.

[0073] A2: Establish a kinematic model.

[0074] In this embodiment, RTK data can be used to obtain the global precise position, speed and attitude information of the target vehicle. Then, combined with the wheel speed pulse data of each wheel (i.e., wheel speed pulse), a kinematic model of the target vehicle is established. Furthermore, the speed and attitude of the target vehicle are estimated by the rotational speed of each wheel and the kinematic relationship of the target vehicle.

[0075] A3: Use kinematic models to estimate the bias.

[0076] As an example, the vehicle state can be estimated based on kinematic models, IMU data, RTK data, and wheel speed pulse data. The bias parameters of the IMU, including but not limited to zero bias error and scaling factor error, can be estimated using graph optimization algorithms, Kalman filtering, or other optimization algorithms.

[0077] A4: Real-time updates and corrections.

[0078] During vehicle movement, new IMU data, RTK data, and wheel speed pulse data are collected in real time. Then, the deviation parameters of the IMU are continuously estimated and updated using the above steps A1 to A3, and the final output result is used as the reference deviation.

[0079] In other words, in this example, since RTK data has high accuracy, the reference deviation of the inertial measurement unit can be estimated by combining RTK data, reducing the cumbersome process of traditional deviation estimation and improving the efficiency and accuracy of reference deviation estimation.

[0080] In another implementation, the reference deviation can be obtained in the following way:

[0081] B1: Data collection and synchronization.

[0082] First, real-time dynamic RTK data is acquired from a satellite positioning system (such as GPS or BeiDou). This RTK data can include high-precision position, velocity, and attitude information of the target vehicle. Simultaneously, IMU data output from the inertial measurement unit (IMU) is collected, which can include acceleration and angular velocity. Additionally, wheel speed pulse data for each wheel of the target vehicle is acquired, reflecting the vehicle's real-time motion state. Then, the IMU data, RTK data, and wheel speed pulse data are synchronized in time for subsequent comparison and analysis.

[0083] B2: Fusion of RTK data and IMU data.

[0084] Specifically, RTK data is used as a reference to calibrate the IMU data. Because RTK data provides high-precision position and velocity information, it can be used as the "true value" to evaluate IMU performance. By comparing the RTK data with the IMU data, the IMU's bias can be calculated, which may include zero-bias error and scaling factor error.

[0085] B3: Verification and calibration using wheel speed pulse data.

[0086] Wheel speed pulse data provides real-time motion information of the vehicle at each wheel, and can therefore be used to further verify and calibrate IMU data. Specifically, by analyzing wheel speed pulse data, the real-time speed and acceleration of the target vehicle can be estimated, and then compared with the IMU data to detect any IMU deviations.

[0087] In other words, in this embodiment, RTK data can be directly used as "true" data, and RTK data can be fused with IMU data to determine the reference deviation of the inertial measurement unit. Furthermore, wheel speed pulse data can be used for calibration to improve the accuracy and reliability of the reference deviation.

[0088] It should be understood that the reference deviation obtained by the satellite inertial navigation fusion positioning framework in step S101 of this application can be obtained under the condition of good satellite signal (or the training phase) so as to be used to train the long short-term artificial neural network LSTM model in the absence of high-precision map and satellite signal.

[0089] S102: Determine the estimated bias of the inertial measurement unit (IMU) based on the IMU data using a pre-trained IMU bias estimation model.

[0090] An inertial measurement unit (IMU) bias estimation model is a mathematical model used to estimate and correct errors generated by the IMU during measurement. Because IMUs are affected by various factors in practical applications, deviations exist between the measured values ​​and the true values ​​without interference. Therefore, the bias estimation model can be used to estimate the bias generated by the IMU and then use this bias to correct the IMU, thereby improving its measurement accuracy and reliability.

[0091] The IMU bias estimation model is obtained by training a Long Short-Term Memory (LSTM) artificial neural network model based on training inertial measurement unit data and reference bias. The LSTM model is a special type of recurrent neural network (RNN) structure designed to address the gradient vanishing and limited memory capacity issues encountered by traditional RNNs when processing long sequences of data. The LSTM model enhances the network's ability to capture long-term dependency information by introducing memory units and gating mechanisms.

[0092] It should be understood that, due to the temporal correlation of biases in IMU data, and because the LSTM model, through the introduction of memory units and gating mechanisms, can effectively handle long-term dependency issues, the long-term trend of bias accumulation can be effectively captured and modeled by the LSTM model. That is, in this embodiment, by training an LSTM network as the IMU bias estimation model, the IMU bias can be estimated more accurately, and the data output by the inertial measurement unit can be corrected in real time, thereby improving the measurement accuracy and reliability of the inertial measurement unit.

[0093] It should be noted that, in this embodiment, since IMU data has temporal characteristics—that is, the data changes continuously over time—training an LSTM model using IMU data allows us to capture hidden temporal patterns and long-term dependencies within the IMU data, enabling more accurate estimation of IMU bias. Furthermore, during training, using a reference bias as a supervisory signal to guide the LSTM model's learning allows the output of the trained IMU bias estimation model to more closely approximate the true bias, thus improving the accuracy of bias estimation.

[0094] S103: Determine the estimated position and estimated attitude of the target vehicle based on the wheel speed pulse data of each wheel.

[0095] It should be noted that in related technologies, there is no effective way to constrain the error accumulation of the inertial measurement unit (IMU). That is, if the error accumulation of the IMU cannot be constrained when the estimation deviation is determined, the accuracy of vehicle positioning will still be reduced. Therefore, the estimated position and estimated attitude determined by wheel speed pulse data can be used to further constrain the accuracy. Considering that the resolution of vehicle wheel speed pulses is generally not high, the traditional wheel speed pulse method has the disadvantage of amplifying pulse noise when solving instantaneous velocity and angular velocity through differential time. By constraining the error accumulation of the IMU through pre-integration of the wheel speed pulse plane, the influence of the process noise of the wheel speed pulse on vehicle positioning is reduced. In addition, the stability and safety of vehicle positioning are also improved.

[0096] S104: Based on the error state Kalman filter model, the target position and target attitude of the target vehicle are determined according to the estimated deviation, estimated position and estimated attitude.

[0097] Kalman filtering is a recursive estimation method that can estimate the state of a dynamic system from a series of incomplete and noisy measurements. The error Kalman filter model is a model built upon the Kalman filtering method. Its focus is on iteratively correcting the vehicle localization error to obtain predictions of the target vehicle's position and attitude.

[0098] It should be understood that during vehicle positioning, sensor data (such as inertial measurement units and wheel speed sensors) often contains various types of noise and uncertainties, such as data delay, vehicle platform vibration, and inaccuracies in the vehicle motion model. Error Kalman filtering models can effectively handle these noises and uncertainties, continuously correcting errors through iteration to obtain more accurate vehicle positioning results.

[0099] The vehicle positioning method provided in this application can be applied to complex urban environments without satellite signals. Considering the time correlation characteristics of the deviation in IMU data, the deviation generated by the inertial measurement unit is accurately estimated by the IMU deviation estimation model obtained by training the LSTM model. Furthermore, the wheel speed pulse data is used to constrain the determined estimated deviation, so as to reduce the accumulation of vehicle inertial recursion error and improve the accuracy of vehicle positioning in complex urban scenes without satellite signals.

[0100] Based on the vehicle positioning method provided in the above embodiments, see [link to relevant documentation]. Figure 2 This figure is a schematic diagram of vehicle positioning provided in an embodiment of this application, combined with... Figure 2As shown below, the training process of the IMU bias estimation model (see Example 2), the process of calculating the second estimated position and the second estimated attitude using the wheel speed pulse pre-integration algorithm (Example 3), and the process of vehicle localization based on the error state Kalman filter model (Example 4) will be described in detail below, and will not be repeated here.

[0101] Example 2

[0102] In some possible implementations, the IMU bias estimation model of this application embodiment can be trained in the following ways:

[0103] B1: Acquire training inertial measurement unit data and reference bias.

[0104] Training inertial measurement unit (IMU) data refers to the IMU data used to train the LSTM model. Reference bias refers to the reference bias used to train the LSTM model. Training IMU data can be obtained from the training dataset, while the reference bias can be calculated jointly from the training IMU data and other data.

[0105] It should be noted that training inertial measurement unit (IMU) data and reference bias can be acquired under conditions of stable data and good signal strength to improve the accuracy of the reference bias. Similarly, training RTK data and training wheel velocity pulse data can also be acquired under conditions of stable data and good signal strength.

[0106] In addition, the scenario use case data corresponding to the training inertial measurement unit data, RTK data, and wheel speed pulse data can include various road scenarios commonly encountered in vehicle driving, such as straight driving, S-curves, right-angle bends, and vehicle U-turns, in order to cover as many vehicle driving scenarios as possible and improve the accuracy of bias estimation.

[0107] After acquiring the data, positioning software based on the open-source satellite positioning framework RTKLIB can be used to input training inertial measurement unit data, training RTK data, and training wheel speed pulse data, and output the deviation of the inertial measurement unit corresponding to the time series data as a reference deviation.

[0108] In one possible implementation, step B1 may include:

[0109] C1: Use the data collected by the inertial measurement unit from the current time to the previous n time steps as the training data for the inertial measurement unit.

[0110] Specifically, inertial measurement unit data can be acquired using a measurement window, which can be a measurement window Γ of size w. mjThe data within the window is used as inertial measurement unit (IMU) data. Here, m represents the start time of the measurement window, j represents the end time of the measurement window (or the current time), and w represents the size of the measurement window. It should be noted that the measurement window slides forward over time, moving forward one time step at a time, thus providing continuous IMU training data to the LSTM model.

[0111] It should be understood that, due to the time correlation characteristics of bias in IMU data, in the LSTM (Long Short-Term Memory) model, by acquiring the data collected by the inertial measurement unit from the current time to the previous n time steps as training data for the inertial measurement unit, the LSTM model can capture the characteristics of time-series data during training, especially the information on the evolution of bias over time, thereby enabling the LSTM model to better understand the evolution process of bias.

[0112] C2: Determine the reference bias of the inertial measurement unit based on the inertial measurement unit data during the training phase, real-time dynamic RTK data, and wheel speed pulse data of each wheel of the target vehicle.

[0113] It should be understood that the process of obtaining the reference deviation can be found in the explanations of steps B1 to B3 above, and will not be repeated here.

[0114] B2: Normalize the training inertial measurement unit data and reference bias to obtain the normalized result.

[0115] It should be understood that the numerical ranges of information such as acceleration and angular velocity contained in the training inertial measurement unit (IMU) data may vary considerably, and the range of reference bias may also differ from that of the training IMU data. Therefore, normalization can transform numerical values ​​from different ranges to a common scale, enabling the LSTM model to process this data more effectively.

[0116] B3: Input the normalization result into the LSTM model to be trained to obtain the training estimation bias.

[0117] See Figure 3 , Figure 3 This is a schematic diagram of an LSTM model provided in an embodiment of this application. (In conjunction with...) Figure 3 As shown, the LSTM model to be trained may include t LSTM layers, t normalization layers, and t fully connected layers, where t is a positive integer greater than or equal to 1, and the output of each LSTM layer is used as the input of the next LSTM layer.

[0118] The input to each LSTM is a measurement window Γ of size w. mj The training measurement unit data (which can be simply referred to as data, such as...) Figure 3(as shown) and the deviation estimate b of the external independent fusion filtering module at the previous time step. i Where i is a positive integer 0 or greater than 0, for example, the input of the f-th LSTM layer is the measurement window Γ of the f-th window with size w. mj The bias estimate b obtained from the training measurement unit data and the (f-1)th LSTM layer. f-1 When i equals 0, that is, the input of the first LSTM layer is the measurement window Γ of the first window size w. mj The training measurement unit data and training reference error b0 are included. Among these, considering the bias estimation caused by the LSTM model encountering external fusion with other sensors, the training measurement unit data and bias estimation b0 need to be included. i Normalization is performed, and then the normalized result is passed to the LSTM layer with retained states hi and ci for the next bias estimation. Here, hi and ci represent the hidden state and the cell state, respectively.

[0119] The hidden state *hi* refers to the state that the LSTM layer updates at each time step, based on the current input and the hidden state and unit states from the previous time step. This hidden state is essentially an encoded representation of the input sequence at the current time step, capturing all the important information up to that point. The hidden state is crucial for the model's subsequent predictions and decisions.

[0120] The cell state ci is another important state in the LSTM layer, responsible for storing information over a long period. Unlike the hidden state, the cell state is designed to store information more stably and is less susceptible to gradient vanishing or exploding. Through carefully designed gating mechanisms (such as input gates, forget gates, and output gates), the LSTM layer can control which information should be added to the cell state, which information should be forgotten, and which information should be output to the hidden state.

[0121] The combination of hi and ci together constitutes the internal state of the LSTM layer at time step x. The working of hi and ci can capture long-term dependencies in sequential data (such as training inertial measurement unit data) and enable the LSTM layer to handle tasks with complex temporal dependencies.

[0122] B4: Determine the loss function value based on the training estimation bias and the reference bias.

[0123] Training estimation bias refers to the bias in the output of the LSTM model being trained, while reference bias refers to the bias in the reference values ​​used to train the LSTM model. The loss function value is an indicator that measures the difference between the training estimation error and the training reference error of the LSTM model. During training, the loss function value is used as the objective function of optimization algorithms (such as gradient descent). By iteratively adjusting the model parameters, the loss function value is minimized, thereby optimizing the model's predictive ability.

[0124] As one possible implementation, the loss function value can be obtained in the following way:

[0125] D1: Determine the accelerometer loss function and gyroscope loss function based on the LSTM model to be trained.

[0126] It should be understood that since IMU data includes accelerometer data and gyroscope data, in order to better train the LSTM model, two loss functions can be determined based on the LSTM model to be trained.

[0127] As an example, considering the dynamic changes in the actual environment in which the vehicle operates, the vehicle dynamics are affected by the physical characteristics of vibration and environmental changes, exhibiting highly nonlinear characteristics. Therefore, a deep learning IMU bias estimation model (i.e., an LSTM model) can be established. θ The definition is as follows:

[0128]

[0129] Among them, Γ mj Represents time t m At time t j Training inertial measurement unit data between and These are different times t n and t i Deviation estimation of the inertial measurement unit between time points. and These are different times t n and t i Reference deviation between time points.

[0130] Therefore, the following loss function can be established for the accelerometer and gyroscope:

[0131]

[0132] Among them, b k This represents the deviation estimate of the inertial measurement unit. This represents the reference bias, which can be specifically determined based on the loss function. Train the parameters of the independent network models separately.

[0133] It should be noted that, for the task of learning the IMU bias estimation model, the loss function is defined as the deviation between the estimated bias of the learned IMU bias estimation model and the reference bias. The reference bias during the training phase can be considered as being calculated through the traditional satellite inertial navigation fusion positioning framework. Thus, the Huber loss function is defined to train the IMU bias estimation models for both the gyroscope and accelerometer separately. The Huber loss operator is used to improve the robustness of model training and to handle potentially outlier data.

[0134] D2: Determine the accelerometer loss function value based on the training estimation bias, reference bias, and accelerometer loss function, and determine the gyroscope loss function value based on the training estimation bias, reference bias, and gyroscope loss function.

[0135] That is, in the embodiments of this application, after determining the accelerometer loss function and the gyroscope loss function, the accelerometer loss function and the gyroscope loss function can be further determined according to the training estimation bias and the reference bias, which facilitates the subsequent training of the LSTM model.

[0136] B5: Use the loss function value to train the LSTM model to be trained.

[0137] It should be noted that the loss function value is an important basis for measuring model performance, optimizing model parameters, selecting the best model, and preventing overfitting. In the practice of machine learning and deep learning, the reasonable selection and design of the loss function is beneficial to improving the predictive ability of the model.

[0138] Wherein, when the loss function values ​​are the accelerometer loss function values ​​and gyroscope loss function values ​​determined by steps D1 and D2, then step B5 may include: training the LSTM model to be trained using the accelerometer loss function values ​​and gyroscope loss function values.

[0139] Example 3

[0140] Based on the vehicle positioning method provided in the above embodiments, in related technologies, due to the low resolution of wheel speed pulses, traditional wheel speed pulses suffer from amplified pulse noise when solving instantaneous velocity and angular velocity through differential time. To solve this technical problem, in some possible implementations, step S103 may specifically include:

[0141] E1: Determine the wheel speed pulse count and linear velocity of each wheel based on the wheel speed pulse data of each wheel.

[0142] Wheel speed pulse count is the number of pulse signals generated per rotation of the tires during vehicle operation. As one possible implementation, wheel speed pulse count can be calculated as follows:

[0143]

[0144] Where, Δθ i Let n be the rotation angle of the wheel, N be the number of pulses per revolution of the wheel, and n be the number of pulses per revolution of the wheel. i Let be the pulse number of the i-th wheel, where i = 1 and 2 represent the left and right rear wheels, respectively.

[0145] Linear velocity refers to the velocity of any point on a wheel as it moves in a circle about a fixed axis. It is a physical quantity that describes the speed and direction of a particle's motion in the curvilinear motion of the wheel. In some possible implementations, the linear velocity of the wheel can be calculated as follows:

[0146] v i =rΔθ i / Δt;ω i =v i / r;

[0147] Among them, v i Let ω be the linear velocity of the wheel. i ω is the angular velocity, r is the wheel radius, and Δt is the time interval.

[0148] In one possible implementation, the cumulative linear velocity V of the wheel can also be calculated. i and cumulative angular velocity Ω i :

[0149]

[0150] Among them, V i Ω is the cumulative linear velocity of the wheel. i For the cumulative angular velocity, k represents the cumulative count.

[0151] E2: Determine the first estimated position and first estimated attitude of the target vehicle based on the wheel speed pulse count and linear velocity of each wheel.

[0152] The first estimated position refers to the position of the target vehicle at the current moment, and the first estimated attitude refers to the attitude of the target vehicle at the current moment. In one possible implementation, after determining the wheel speed pulse count and linear velocity of each wheel, the first estimated position and the first estimated attitude can be further determined by the following formula:

[0153] x i =x i-1 +V i cosθ i-1 Δt;

[0154] y i =y i-1 +V i sinθ i-1 Δt;

[0155]

[0156] Where, x i Let x be the x-coordinate of the target vehicle, and x0 be the initial value of the x-coordinate; y i Let y be the vehicle's y-coordinate, and y0 be the initial value of the ordinate; θ i V is the target vehicle's heading angle, θ0 is the initial value of the heading angle, and L is the wheelbase; r,i and V l,i These are the cumulative linear velocities of the rear right wheel and the rear left wheel, respectively.

[0157] In this embodiment, the relative pose and covariance of the time interval of the corresponding measurement window calculated by the wheel speed pulse pre-integration algorithm are less affected by pulse noise than the traditional method. Therefore, by using low-frequency wheel speed pulse pre-integration to constrain the estimation error of the IMU, the accuracy of vehicle positioning can be improved.

[0158] It should be noted that the wheel speed pulse pre-integration algorithm is a process used to process and transform wheel speed sensor data of a target vehicle, aiming to improve the accuracy of positioning and speed estimation. The wheel speed pulse pre-integration algorithm typically involves collecting wheel speed pulse data from wheel speed sensors and then processing this data to estimate the position and attitude of the target vehicle.

[0159] Example 4

[0160] Based on the vehicle positioning method provided in the above embodiments, as one possible implementation, step S104 may include:

[0161] F1: Using the inertial measurement unit pre-integration model, determine the second estimated position and second estimated attitude of the target vehicle based on IMU data.

[0162] The inertial measurement unit (IMU) pre-integration model is a method for processing inertial sensor data. The IMU pre-integration model aims to estimate the second estimated position and the second estimated attitude of a target vehicle by integrating continuous IMU measurements, without needing to know the initial state (i.e., the position and attitude at the previous moment).

[0163] It should be understood that, in the embodiments of this application, the determination of the second estimated position and second estimated attitude of the target vehicle at the current moment by using inertial measurement unit data utilizes the characteristics of the inertial measurement unit pre-integration model, takes into account the influence of noise and zero bias of the inertial sensor, and processes the noise and zero bias to improve the accuracy of the data.

[0164] F2: Based on the second estimated position and the second estimated attitude, the predicted position and predicted attitude of the target vehicle are determined through the state transition equation in the error state Kalman filter model.

[0165] The state transition equation describes the evolution of the target vehicle's state from the previous moment to the current moment. In other words, the predicted position and attitude of the target vehicle at the current moment can be determined by the state transition equation in the error state Kalman filter model.

[0166] F3: Construct an observation model based on the estimation bias, the first estimated position, and the first estimated attitude.

[0167] The observation model describes the relationship between the first estimated position, the first estimated attitude, and the estimation bias. The observation model can be used to update the error state Kalman filter model.

[0168] F4: Update the predicted position and attitude using the observation model to obtain the target position and attitude of the target vehicle.

[0169] It should be understood that, since the predicted position and attitude may be affected by the inertial measurement unit and have certain errors, it is necessary to use the observation model to update the predicted position and attitude in order to obtain the target position and attitude of the target vehicle.

[0170] As one possible implementation, step F4 may include:

[0171] G1: Determine the Kalman gain of the error Kalman model using the observation model.

[0172] Kalman gain is a core concept in Kalman filtering theory. Kalman gain refers to the change in uncertainty after each data fusion. If the uncertainty decreases after fusing new data, the gain is positive and helps improve prediction accuracy. Conversely, if the uncertainty increases, the gain is negative and has the opposite effect on the accuracy of system predictions. Here, "uncertainty" is quantified using the covariance of the estimated random variables in each iteration.

[0173] G2: Update the error state vector of the error Kalman model based on the Kalman gain.

[0174] Error-state Kalman filtering focuses primarily on the errors in the states, rather than the states themselves. Therefore, it is necessary to define an error-state vector, which typically includes position errors, attitude errors (such as quaternion errors or Euler angle errors), velocity errors, etc.

[0175] In one possible implementation, the error state vector of the ErEKF filtering model is defined as follows:

[0176]

[0177] Here, δx represents the error state vector in the error state Kalman filter model, which can consist of four components. This represents the estimated transformation matrix from the target vehicle's body to its navigation coordinates. and actual The misalignment angle between them, δr n Represents the estimated position vector and the actual position vector r n The difference between them, δα represents the estimated IMU to vehicle coordinate transformation matrix. and actual The misalignment angle between the accelerometer and gyroscope is δβ, which represents the bias deviation between the accelerometer and gyroscope, and δk represents the error of the estimated scaling factor.

[0178] It should be noted that, in this embodiment of the application, the accuracy of vehicle positioning is improved by using the Kalman gain error state vector for updating, that is, by using wheel speed pulse data to constrain the error accumulation of the inertial measurement unit.

[0179] G3: Update the predicted position and predicted attitude using the updated error state vector to obtain the target position and target attitude of the target vehicle.

[0180] It should be understood that by introducing an error state vector to update the predicted position and predicted attitude, the target position and target attitude of the target vehicle can be estimated more accurately, thereby improving the accuracy of vehicle positioning.

[0181] Based on Embodiment 1, Embodiment 4 of this application uses an error state Kalman filter model and, based on estimated position and estimated attitude, corrects the vehicle state predicted by the IMU pre-integration model to construct an ErEKF filter model for inertial wheel speed vehicle-enhanced odometer, thereby reducing the accumulation of errors in vehicle inertial recursion and improving the accuracy of vehicle positioning.

[0182] Example 5

[0183] Based on the vehicle positioning method provided in the above embodiments, this application also provides a vehicle positioning device, see [link to relevant documentation]. Figure 4 This figure is a structural schematic diagram of the vehicle positioning device provided in an embodiment of this application, combined with... Figure 4 As shown, the vehicle positioning device 400 provided in this application embodiment may include:

[0184] The acquisition module 401 is used to acquire the inertial measurement unit (IMU) data and reference bias of the inertial measurement unit in the target vehicle; the reference bias is calculated based on the IMU data, real-time dynamic RTK data and wheel speed pulse data of each wheel of the target vehicle through the satellite inertial navigation fusion positioning framework.

[0185] The estimation bias determination module 402 is used to determine the estimated bias of the inertial measurement unit (IMU) based on the IMU bias estimation model, according to the IMU data and the reference bias. The IMU bias estimation model is obtained by training a long short-term artificial neural network (LSTM) model based on the training IMU data and the reference bias.

[0186] The estimation module 403 is used to determine the estimated position and estimated attitude of the target vehicle based on the wheel speed pulse data of each wheel;

[0187] The vehicle positioning module 404 is used to determine the target position and target attitude of the target vehicle based on the error state Kalman filter model, according to the estimated deviation, estimated position and estimated attitude.

[0188] As an example, the IMU bias estimation model is trained in the following way:

[0189] The acquisition unit is used to acquire training inertial measurement unit data and reference bias;

[0190] The normalization processing unit is used to normalize the training inertial measurement unit data and reference bias to obtain the normalization result;

[0191] The input unit is used to input the normalization result into the LSTM model to be trained to obtain the training estimation bias;

[0192] The loss function value determination unit is used to determine the loss function value based on the training estimation bias and the reference bias.

[0193] The training unit is used to train the LSTM model to be trained using the loss function value.

[0194] As an example, retrieving a unit includes:

[0195] The first acquisition subunit is used to use the data collected by the inertial measurement unit from the current time to the previous n time steps as training data for the inertial measurement unit.

[0196] The second acquisition subunit is used to determine the reference deviation of the inertial measurement unit based on the satellite positioning structure, according to the training inertial measurement unit data, the training real-time dynamic RTK data, and the training wheel speed pulse data of each wheel of the target vehicle.

[0197] As an example, the loss function value determination unit includes:

[0198] The loss function determination sub-unit is used to determine the accelerometer loss function and the gyroscope loss function based on the IMU bias estimation model.

[0199] The loss function value determination subunit is used to determine the accelerometer loss function value based on the training estimation bias, reference bias, and accelerometer loss function, and to determine the gyroscope loss function value based on the training estimation bias, reference bias, and gyroscope loss function.

[0200] The training unit is specifically used for:

[0201] The LSTM model to be trained is trained using accelerometer loss function values ​​and gyroscope loss function values.

[0202] As an example, estimation module 403 includes:

[0203] The calculation unit is used to determine the wheel speed pulse count and linear velocity of each wheel based on the wheel speed pulse data of each wheel;

[0204] The first estimation unit is used to determine the first estimated position and first estimated attitude of the target vehicle based on the wheel speed pulse count and linear velocity of each wheel.

[0205] As an example, a vehicle positioning module includes:

[0206] The second estimation unit is used to determine the second estimated position and the second estimated attitude of the target vehicle based on IMU data using the inertial measurement unit pre-integration model.

[0207] The prediction unit is used to determine the predicted position and predicted attitude of the target vehicle based on the second estimated position and the second estimated attitude, through the state transition equation in the error state Kalman filter model.

[0208] A construction unit is used to construct an observation model based on the estimation bias, the first estimated position, and the first estimated attitude.

[0209] The vehicle localization unit is used to update the predicted position and predicted attitude using the observation model to obtain the target position and target attitude of the target vehicle.

[0210] As an example, a vehicle positioning unit includes:

[0211] The Kalman gain determination subunit is used to determine the Kalman gain of the error Kalman model using the observation model;

[0212] The update sub-unit is used to update the error state vector of the error Kalman model based on the Kalman gain;

[0213] The vehicle localization subunit is used to update the predicted position and predicted attitude using the updated error state vector to obtain the target position and target attitude of the target vehicle.

[0214] The vehicle positioning device provided in this application has the same beneficial effects as the vehicle positioning method provided in the above embodiments, so it will not be described again.

[0215] This application also provides corresponding devices and computer storage media for implementing the solutions provided in this application.

[0216] The device includes a memory and a processor. The memory stores instructions or code, and the processor executes the instructions or code to enable the device to perform the vehicle positioning method described in any embodiment of this application.

[0217] The computer storage medium stores code, and when the code is executed, the device running the code implements the vehicle positioning method described in any embodiment of this application.

[0218] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for the device and equipment embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiments. The device and equipment embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components indicated as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0219] In the embodiments of this application, the terms "first" and "second" (if they exist) are used only as name identifiers and do not represent the order of first and second.

[0220] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that all or part of the steps in the methods of the above embodiments can be implemented by means of software plus a general-purpose hardware platform. Based on this understanding, the technical solution of this application can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as a read-only memory (ROM) / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, a server, or a network communication device such as a router) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0221] In this application, the collection and processing of relevant data must strictly comply with the requirements of national laws, administrative regulations, and mandatory national standards. Informed consent or separate consent from the entity obtaining personal information (hereinafter referred to as the "subject") is required, and subsequent data use and processing can only proceed within the scope authorized by laws, regulations, and the subject. Relevant requirements include, but are not limited to:

[0222] (i) The individual consent of the subject must be obtained, and the subject may set the consent period independently;

[0223] (ii) Provided that the vehicle’s driving safety is ensured, the data collection status is indicated to the subject in an appropriate manner, so as to facilitate the subject to terminate the data collection.

[0224] (iii) If the subject requests the deletion of some or all of the data related to the subject, the processor shall delete the relevant data within ten working days;

[0225] (iv) Data processors may collect data related to the subject only if the data has the purpose of enhancing driving safety, has the purpose of directly serving the subject, and / or is sufficiently necessary; data related to the subject includes, but is not limited to, biometric data (such as fingerprints, voiceprints, faces, heart rhythms, etc.), data for enhancing driving safety, intelligent driving data, navigation data, etc.

[0226] (v) The necessity of data collection and its impact on the subject can be communicated through prominent means such as user manuals, in-vehicle display panels, voice, text and vehicle-related applications;

[0227] (vi) If this application involves the collection of the subject's voice, the voice will be converted into text in the vehicle and uploaded in text form only, without involving personal voiceprint information, and the subject's personal information will be encrypted and converted before uploading.

[0228] The above description is merely one specific embodiment of this application, but the scope of protection of this application 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 this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A vehicle positioning method, characterized in that: The method comprises: Acquire the inertial measurement unit (IMU) data and reference deviation of the inertial measurement unit in the target vehicle; the reference deviation is obtained by solving the IMU data, the real-time dynamic RTK data and the wheel speed pulse data of each wheel of the target vehicle through the satellite inertial navigation fusion positioning framework; Based on an inertial measurement unit (IMU) bias estimation model, an estimated bias of the inertial measurement unit is determined according to the IMU data; the IMU bias estimation model is obtained by training a long-term and short-term artificial neural network (LSTM) model according to the training inertial measurement unit data and the reference bias; Determining the estimated position and estimated posture of the target vehicle according to the wheel speed pulse data of each wheel; Based on the error state Kalman filter model, the target position and the target attitude of the target vehicle are determined according to the estimated deviation, the estimated position and the estimated attitude.

2. The method according to claim 1, characterized in that The IMU bias estimation model is trained in the following way: Acquiring the training inertial measurement unit data and the reference deviation; Normalizing the training inertial measurement unit data and the reference deviation to obtain a normalized result; Inputting the normalized result into the LSTM model to be trained to obtain a training estimation deviation; Determine a loss function value according to the training estimation deviation and the reference deviation; The LSTM model to be trained is trained using the loss function value.

3. The method according to claim 2, characterized in that The step of obtaining training inertial measurement unit data and reference deviation includes: Using the data collected by the inertial measurement unit from the current moment to the previous n moments as training inertial measurement unit data; Based on the satellite positioning data, a reference deviation of the inertial measurement unit is determined according to the training inertial measurement unit data, the training real-time kinematic RTK data and the training wheel speed pulse data of each wheel of the target vehicle.

4. The method according to claim 2, characterized in that: The determining of the loss function value according to the training estimated deviation and the reference deviation comprises: Determining an accelerometer loss function and a gyroscope loss function based on the LSTM model to be trained; Determining an accelerometer loss function value according to the training estimated deviation, the reference deviation, and the accelerometer loss function, and determining a gyroscope loss function value according to the training estimated deviation, the reference deviation, and the gyroscope loss function; The step of training the LSTM model to be trained by using the loss function value includes: The LSTM model to be trained is trained using the accelerometer loss function value and the gyroscope loss function value.

5. The method according to claim 1, characterized in that Determining the estimated position and estimated posture of the target vehicle according to the wheel speed pulse data of each wheel includes: Determine the wheel speed pulse count and linear speed of each wheel according to the wheel speed pulse data of each wheel; A first estimated position and a first estimated attitude of the target vehicle are determined based on the wheel speed pulse counts of the respective wheels and the linear velocity.

6. The method according to claim 5, characterized in that: Based on the error state Kalman filter model, determining the target position and the target attitude of the target vehicle according to the estimated deviation, the estimated position and the estimated attitude, including: Determine a second estimated position and a second estimated attitude of the target vehicle based on the IMU data using an inertial measurement unit pre-integration model; Determining the predicted position and predicted attitude of the target vehicle through a state transfer equation in the error state Kalman filter model according to the second estimated position and the second estimated attitude; constructing an observation model according to the estimated deviation, the first estimated position and the first estimated posture; The predicted position and the predicted attitude are updated using the observation model to obtain the target position and the target attitude of the target vehicle.

7. The method according to claim 6, characterized in that The method of updating the predicted position and the predicted attitude by using the observation model to obtain the target position and the target attitude of the target vehicle includes: Determining a Kalman gain of the error Kalman model using the observation model; Updating the error state vector of the error Kalman model according to the Kalman gain; The predicted position and the predicted attitude are updated using the updated error state vector to obtain the target position and the target attitude of the target vehicle.

8. A vehicle positioning device, characterized in that: The device comprises: An acquisition module is used to acquire IMU data and reference deviation of an inertial measurement unit in a target vehicle; the reference deviation is obtained by solving the IMU data, real-time dynamic RTK data and wheel speed pulse data of each wheel of the target vehicle through a satellite inertial navigation fusion positioning framework; An estimated deviation determination module is used to determine the estimated deviation of the inertial measurement unit according to the IMU data and the reference deviation based on an inertial measurement unit IMU deviation estimation model; the IMU deviation estimation model is obtained by training a long short-term artificial neural network LSTM model according to the training inertial measurement unit data and the reference deviation; An estimation module, used for determining an estimated position and an estimated posture of the target vehicle according to the wheel speed pulse data of each wheel; The vehicle positioning module is used to determine the target position and target attitude of the target vehicle according to the estimated deviation, the estimated position and the estimated attitude based on the error state Kalman filter model.

9. A computer device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the vehicle positioning method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, and when the instructions are executed on a terminal device, the terminal device executes the vehicle positioning method according to any one of claims 1 to 7.

Images