A vehicle position estimation method of a fusion filtering network and a computer readable medium

Abstract

The application discloses a vehicle position estimation method based on a fusion filtering network and a computer readable medium. The application divides a window into multiple groups of IMU training samples, and labels the real vehicle position labels of each group of IMU training samples; a vehicle position estimation network based on a fusion Kalman filtering model is constructed, and an optimized vehicle position estimation network is obtained through ADAM algorithm optimization training; real-time collected vehicle IMU data is combined with the optimized vehicle position estimation network to perform vehicle position prediction, and real-time vehicle predicted positions are obtained. The application has a good processing effect on vehicle-mounted inertia data containing high noise, and weakens the adverse effects of noise on the system; a noise estimation model is constructed to perform real-time estimation of perfectness constraint noise, reduces the dependence of a traditional filtering scheme on experience parameter setting, and effectively improves the migration ability and adaptability of a learning model.

Description

Technical Field

[0001] This invention belongs to the field of vehicle positioning and navigation technology, and particularly relates to a vehicle position estimation method using a fusion filter network and a computer-readable medium. Background Technology

[0002] Vehicles typically use GNSS / INS integrated navigation systems for positioning. However, due to the instability and obstruction of GNSS, and the significant cumulative error of INS, integrity constraints are often constructed using a vehicle motion model (assuming the vehicle's lateral and vertical velocities are approximately zero) and forward velocity to improve positioning performance. Traditional methods rely on wheel speed odometers to obtain forward velocity, requiring the installation of an odometer and specific equipment to read the vehicle's odometer information. Furthermore, the extrinsic parameters between the IMU and the vehicle's coordinate system also affect this constraint. This paper proposes a virtual odometer based on a deep residual shrinking network to replace the traditional wheel speed odometer, thus reducing the requirements for this approach.

[0003] Convolutional neural networks, recurrent neural networks, and graph neural networks are widely used in computer vision, pattern recognition, and localization and navigation. Most of these networks can be used to train a virtual odometer. However, due to factors such as road conditions and weather during vehicle operation, and the inherent characteristics of inertial sensors, the collected inertial data often contains a large amount of high- and low-frequency noise and interference. This poses a significant challenge to network training, leading to overfitting or getting trapped in local optima. This invention proposes a virtual odometer based on a deep residual contraction network to replace traditional wheel speed odometers. By setting soft and hard thresholds, an attention mechanism, and an identity path, this network automatically performs soft thresholding during training, adaptively learning to eliminate noise and redundant information, which has a significant advantage in processing data containing noisy samples.

[0004] Models that integrate inertial motion and integrity constraints generally fall into two categories: one is a specific model fusion method represented by Kalman filtering; the other utilizes neural network learning. The first method generally has a mature theoretical foundation, but it is limited by empirical parameter settings and struggles to adapt to dynamic data characteristics. Neural network learning generally lacks interpretability, faces training difficulties for complex problems, and exhibits poor transferability. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a vehicle position estimation method using a fusion filter network and a computer-readable medium.

[0006] The technical solution of this invention is a vehicle position estimation method using a fusion filter network, and the specific steps are as follows:

[0007] Step 1: Obtain vehicle raw IMU data and vehicle location data at multiple historical time points. Obtain vehicle IMU data at multiple historical time points by adding noise to the vehicle raw IMU data at multiple historical time points. Combine the vehicle IMU data at multiple historical time points with the vehicle location data and divide them into multiple groups of IMU training samples through a window. Label each group of IMU training samples with the real vehicle location label.

[0008] Step 2: Construct a vehicle position estimation network that integrates a Kalman filter model. Input each set of IMU training samples into the vehicle position estimation network for prediction to obtain the predicted vehicle position label for each set of IMU training samples. Combine the real vehicle position labels of each set of IMU training samples to construct a loss function model. Optimize the training using the ADAM algorithm to obtain the optimized vehicle position estimation network.

[0009] Step 3: Combine the real-time collected vehicle IMU data with the optimized vehicle position estimation network to predict the vehicle position and obtain the real-time predicted vehicle position.

[0010] Preferably, the multiple sets of IMU training samples mentioned in step 1 are specifically defined as follows:

[0011] data k =(imu (k-1)*(L+1)+1 ,imu (k-1)*(L+1)+2 ,...,imu (k-1)*(L+1)+L P (k-1)*(L+1)+L )

[0012] k∈[1,K]

[0013] Among them, data k Let L represent the k-th training sample group, K represent the number of training samples, and L represent the number of vehicle IMU data points in each IMU training sample group. (k-1)*(L+1)+i P represents the vehicle IMU data of the i-th vehicle in the k-th training sample, that is, the vehicle IMU data of the (k-1)*(L+1)+i-th historical time. (k-1)*(L+1)+L Let i represent the vehicle position at the (k-1)*(L+1)+Lth historical moment, where i∈[1,L].

[0014] The actual vehicle location of each IMU training sample in step 1 is specifically defined as follows:

[0015] label k =P (k-1)*(L+1)+L+1

[0016] Where, label k P represents the true vehicle location label of the k-th training sample. (k-1)*(L+1)+L+1 This represents the vehicle position at the (k-1)*(L+1)+L+1th historical moment;

[0017] Preferably, the vehicle position estimation network fused with the Kalman filter model in step 2 includes:

[0018] Velocity estimation depth residual network, noise estimation convolutional neural network, inertial navigation module, and fused Kalman filter model;

[0019] The velocity estimation depth residual network is connected to the fused Kalman filter model;

[0020] The noise estimation convolutional neural network is connected to the fused Kalman filter module;

[0021] The inertial navigation module is connected to the fused Kalman filter model;

[0022] Each set of IMU training samples is input into the velocity estimation depth residual network, the noise estimation convolutional neural network, and the inertial navigation module, respectively.

[0023] The velocity estimation depth residual network is composed of multiple depth residual convolutional modules cascaded in sequence;

[0024] Each depthwise residual convolutional module consists of a cascaded convolutional layer consisting of the first convolutional layer, the second convolutional layer, ..., the (K-1)th convolutional layer, a soft thresholding module, and the Kth convolutional layer.

[0025] The velocity estimation deep residual network calculates the depth residual for each group of IMU training samples by sequentially passing each deep residual convolution module, obtains the estimated velocity for each group of IMU training samples, and outputs it to the fusion Kalman filter model.

[0026] The (j-1)th depth residual convolution module uses the multi-channel data features extracted by the convolutional neural network as input to the j-th depth residual convolution module;

[0027] In the j-th depth residual convolution module, the multi-channel data features of the (j-1)-th depth residual convolution module are convolved sequentially through the 1st convolutional layer, the 2nd convolutional layer, ..., the (K-1)th convolutional layer to further extract the multi-channel high-dimensional data features of each group of IMU training samples, and output to the soft thresholding module.

[0028] The soft thresholding module sequentially processes the high-dimensional data features of each channel of each IMU training sample through absolute value processing, global averaging, and attention mechanism network processing to obtain each threshold vector of each IMU training sample. Each threshold vector of each IMU training sample is then processed by an activation function to obtain each activated threshold vector of each IMU training sample. Each activated threshold vector of each IMU training sample is then multiplied by the corresponding high-dimensional data features of the channel to obtain the denoised data features of each channel of each IMU training sample. The denoised data features of multiple channels of each IMU training sample are then output to the Kth convolutional layer.

[0029] The noise estimation convolutional neural network processes each group of IMU training samples to calculate the harmonic coefficient vector of each group. This harmonic coefficient vector is then multiplied by the integrity constraint noise empirical value vector to obtain the estimated noise vector for each group of IMU training samples. Finally, the estimated noise vector of each group of IMU training samples is output to the fused Kalman filter model.

[0030] The estimated noise vector for each group of IMU training samples is defined as follows:

[0031]

[0032] k∈[1,K]

[0033] Among them, (κ) 1,k ,κ 2,k ,κ 3,k ) represents the harmonic coefficient vector of the k-th IMU training sample, κ i,k ∈[-1,1], i∈[1,3], Let K represent the vector of empirical values ​​for the integrity constraint noise, and κ represent the number of training samples. 1,k κ represents the harmonic coefficient of the vehicle forward velocity noise in the harmonic coefficient vector of the k-th IMU training sample. 2,k κ represents the vehicle lateral velocity noise harmonic coefficient in the harmonic coefficient vector of the k-th IMU training sample. 3,k This represents the vehicle vertical velocity noise harmonic coefficient in the harmonic coefficient vector of the k-th IMU training sample. This represents an empirical value for the vehicle's forward speed noise. This represents an empirical value for vehicle lateral speed noise. This represents an empirical value for the vertical speed noise of the vehicle.

[0034] The inertial navigation module calculates the inertial measurement results of each group of IMU training samples using an inertial navigation algorithm, and outputs the inertial measurement results of each group of IMU training samples to the fused Kalman filter model.

[0035] The inertial measurement results for each group of IMU training samples are specifically defined as follows:

[0036] [φ k V k ,p k ]

[0037] k∈[1,K]

[0038] Where, φ k V represents the vehicle attitude as determined by the inertial measurement results of the k-th IMU training sample. k p represents the vehicle speed from the inertial measurement results of the k-th IMU training sample. k The vehicle position represents the inertial measurement result of the k-th IMU training sample, where K represents the number of training samples;

[0039] The fusion Kalman filter model calculates the predicted vehicle position for each IMU training sample by combining the estimated velocity, the estimated noise vector, the inertial measurement results, and the introduced incomplete observation constraints of each IMU training sample through fusion Kalman filtering.

[0040] The process of calculating the predicted vehicle position for each IMU training sample by fusing Kalman filtering is as follows:

[0041] The Kalman filter model system observation vector is constructed using the estimated velocity, non-perfection constraint, and estimated noise vector of each IMU training sample, as detailed below:

[0042]

[0043] k∈[1,K]

[0044] in, The observation vector representing the velocity of the k-th IMU training sample in the vehicle coordinate system is the system observation vector. Let R represent the estimation velocity of the k-th IMU training sample, [0,0] represent the introduced incomplete observation constraint, K represent the number of training samples, and R represent the estimation velocity of the k-th IMU training sample. k Let represent the estimated noise vector of the k-th IMU training sample, specifically defined as follows:

[0045]

[0046] k∈[1,K]

[0047] Where, dig[·] represents a diagonal matrix, (κ 1,k ,κ 2,k ,κ3,k ) represents the harmonic coefficient vector of the k-th IMU training sample, κ i,k ∈[-1,1], Let K represent the vector of empirical values ​​for the integrity constraint noise, and κ represent the number of training samples. 1,k κ represents the harmonic coefficient of the vehicle forward velocity noise in the harmonic coefficient vector of the k-th IMU training sample. 2,k κ represents the vehicle lateral velocity noise harmonic coefficient in the harmonic coefficient vector of the k-th IMU training sample. 3,k This represents the vehicle vertical velocity noise harmonic coefficient in the harmonic coefficient vector of the k-th IMU training sample. This represents an empirical value for the vehicle's forward speed noise. This represents an empirical value for vehicle lateral speed noise. This represents the empirical value of the vehicle's vertical speed noise, and K represents the number of training samples;

[0048] The system error vector of the fused Kalman filter model for each group of IMU training samples is calculated as follows:

[0049] The vehicle velocity from the inertial measurement results of each IMU training sample is transformed into a coordinate system to obtain the expression for the vehicle velocity in the vehicle coordinate system, as follows:

[0050]

[0051] Among them, V k The vehicle speed represents the inertial measurement result of the k-th IMU training sample. This represents the system velocity of the k-th IMU training sample in the navigation coordinate system. This represents the angular velocity observation value of the k-th IMU training sample in the carrier coordinate system. This indicates the length of the lever arm between the IMU coordinate system and the vehicle coordinate system. This represents the rotation matrix between the vehicle coordinate system and the IMU coordinate system; Let represent the rotation matrix between the navigation coordinate system and the IMU coordinate system of the k-th IMU training sample;

[0052] The velocity observation equations for each group of IMU training samples are as follows:

[0053]

[0054] in, V represents the velocity observation residual of the k-th IMU training sample. k The vehicle speed represents the inertial measurement result of the k-th IMU training sample. This represents the observation vector of the velocity of the k-th IMU training sample in the vehicle coordinate system;

[0055] By combining the vehicle velocity expression in the vehicle coordinate system from the inertial measurement results of each IMU training sample and the velocity observation equation of each IMU training sample, the system error vector of the fusion Kalman filter model of each IMU training sample is calculated by the standard Kalman filter method.

[0056] The system error vector of the fused Kalman filter model for each group of IMU training samples is specifically defined as follows:

[0057]

[0058] k∈[1,K]

[0059] Where, (δφ) k Let (δv) represent the vehicle attitude error, which is the system error vector of the fused Kalman filter model for the k-th IMU training samples. k Let (δp) represent the vehicle speed error in the system error vector of the fused Kalman filter model for the k-th IMU training samples. k The vehicle position error (δb) represents the system error vector of the fused Kalman filter model for the k-th IMU training samples. w ) k The zero-bias error of the gyroscope represents the system error vector of the fused Kalman filter model for the k-th IMU training samples, (δb) a ) k The zero-bias error of the acceleration in the system error vector of the fused Kalman filter model for the k-th IMU training samples is represented. The system error vector of the fused Kalman filter model for the k-th IMU training sample represents the rotational error between the carrier coordinate system and the vehicle coordinate system. The lever arm error between the carrier coordinate system and the vehicle coordinate system represents the system error vector of the fused Kalman filter model for the k-th IMU training sample, where K represents the number of training samples.

[0060] The predicted vehicle position for each IMU training sample is calculated by combining the vehicle position from the inertial measurement results of each IMU training sample with the position error of the system error vector of the fused Kalman filter model of each IMU training sample, as follows:

[0061]

[0062] Where, p k δp represents the vehicle position, calculated from the inertial measurement results of the k-th IMU training sample. kThe vehicle position error represents the system error vector of the fused Kalman filter model for the k-th IMU training samples. This represents the predicted vehicle position of the k-th IMU training sample;

[0063] The loss function model described in step 2 is specifically defined as follows:

[0064]

[0065] Where, label k This represents the true vehicle location label of the k-th IMU training sample. Let K represent the predicted vehicle position of the k-th IMU training sample, where K represents the number of training samples.

[0066] The present invention also provides a computer-readable medium storing a computer program executed by an electronic device, which, when run on the electronic device, enables the steps of the vehicle position estimation method of the fusion filter network.

[0067] The vehicle position estimation network based on the fusion Kalman filter model constructed in this invention has a better processing effect on high-noise data such as vehicle inertial data, reducing the adverse effects of noise on the system. A noise estimation model is constructed to perform real-time estimation of integrity-constrained noise, realizing a model / data jointly driven elastic fusion method, which reduces the dependence of traditional filtering schemes on empirical parameter settings and effectively improves the model's transferability and adaptability.

[0068] This invention proposes a virtual odometry based on a deep residual shrinking network to replace traditional wheel speed odometry, thus reducing the requirements for its use. By setting soft and hard thresholds, an attention mechanism, and an identity path, this virtual odometry network automatically performs soft thresholding during training, adaptively learning to eliminate noise and redundant information, which has significant advantages in handling speed estimations containing noisy samples.

[0069] The method of this invention, within a mature filtering framework, utilizes deep learning to dynamically adjust the proposed virtual odometer, forward velocity noise, and nonholonomic constraint noise, thus solving the problem that filtering methods are limited by empirical parameter settings and have difficulty adapting to dynamic data characteristics. It combines standard IMU equations with side information of wheeled vehicle dynamics and trains them within the filtering framework, mitigating the problems of traditional learning methods such as lack of interpretability, difficulty in training complex problems, and poor transferability. Attached Figure Description

[0070] Figure 1 : Flowchart of the method according to an embodiment of the present invention;

[0071] Figure 2 : Neural network structure diagram of an embodiment of the present invention. Detailed Implementation

[0072] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] In specific implementation, the method proposed in the technical solution of this invention can be automatically executed by those skilled in the art using computer software technology. System devices for implementing the method, such as computer-readable storage media storing the corresponding computer program of the technical solution of this invention and computer equipment including the computer program running the corresponding computer program, should also be within the protection scope of this invention.

[0074] The following is combined Figures 1-2 The technical solution of the method in the embodiments of the present invention is a vehicle position estimation method using a fusion filter network, as detailed below:

[0075] like Figure 1 The diagram shown is a flowchart of a method according to an embodiment of the present invention.

[0076] Step 1: Obtain vehicle raw IMU data and vehicle location data at multiple historical time points. Obtain vehicle IMU data at multiple historical time points by adding noise to the vehicle raw IMU data at multiple historical time points. Combine the vehicle IMU data at multiple historical time points with the vehicle location data and divide them into multiple groups of IMU training samples through a window. Label each group of IMU training samples with the real vehicle location label.

[0077] The multiple sets of IMU training samples mentioned in step 1 are specifically defined as follows:

[0078] data k =(imu (k-1)*(L+1)+1 ,imu (k-1)*(L+1)+2 ,...,imu (k-1)*(L+1)+L P (k-1)*(L+1)+L )

[0079] k∈[1,K]

[0080] Among them, data k This represents the k-th training sample group, where K = 100 represents the number of training samples, and L = 9 represents the number of vehicle IMU data points in each IMU training sample group. (k-1)*(L+1)+iP represents the vehicle IMU data of the i-th vehicle in the k-th training sample, that is, the vehicle IMU data of the (k-1)*(L+1)+i-th historical time. (k-1)*(L+1)+L Let i represent the vehicle position at the (k-1)*(L+1)+Lth historical moment, where i∈[1,L].

[0081] The actual vehicle location of each IMU training sample in step 1 is specifically defined as follows:

[0082] label k =P (k-1)*(L+1)+L+1

[0083] Where, label k P represents the true vehicle location label of the k-th training sample. (k-1)*(L+1)+L+1 This represents the vehicle position at the (k-1)*(L+1)+L+1th historical moment;

[0084] Step 2: Construct a vehicle position estimation network that integrates a Kalman filter model. Input each set of IMU training samples into the vehicle position estimation network for prediction to obtain the predicted vehicle position label for each set of IMU training samples. Combine the real vehicle position labels of each set of IMU training samples to construct a loss function model. Optimize the training using the ADAM algorithm to obtain the optimized vehicle position estimation network.

[0085] like Figure 2 As shown, the vehicle position estimation network fused with the Kalman filter model in step 2 includes:

[0086] Velocity estimation depth residual network, noise estimation convolutional neural network, inertial navigation module, and fused Kalman filter model;

[0087] The velocity estimation depth residual network is connected to the fused Kalman filter model;

[0088] The noise estimation convolutional neural network is connected to the fused Kalman filter module;

[0089] The inertial navigation module is connected to the fused Kalman filter model;

[0090] Each set of IMU training samples is input into the velocity estimation depth residual network, the noise estimation convolutional neural network, and the inertial navigation module, respectively.

[0091] The velocity estimation depth residual network is composed of multiple depth residual convolutional modules cascaded in sequence;

[0092] Each depthwise residual convolutional module consists of a cascaded convolutional layer consisting of the first convolutional layer, the second convolutional layer, ..., the (K-1)th convolutional layer, a soft thresholding module, and the Kth convolutional layer.

[0093] The velocity estimation deep residual network calculates the depth residual for each group of IMU training samples by sequentially passing each deep residual convolution module, obtains the estimated velocity for each group of IMU training samples, and outputs it to the fusion Kalman filter model.

[0094] The (j-1)th depth residual convolution module uses the multi-channel data features extracted by the convolutional neural network as input to the j-th depth residual convolution module;

[0095] In the j-th depth residual convolution module, the multi-channel data features of the (j-1)-th depth residual convolution module are convolved sequentially through the 1st convolutional layer, the 2nd convolutional layer, ..., the (K-1)th convolutional layer to further extract the multi-channel high-dimensional data features of each group of IMU training samples, and output to the soft thresholding module.

[0096] The soft thresholding module sequentially processes the high-dimensional data features of each channel of each IMU training sample through absolute value processing, global averaging, and attention mechanism network processing to obtain each threshold vector of each IMU training sample. Each threshold vector of each IMU training sample is then processed by an activation function to obtain each activated threshold vector of each IMU training sample. Each activated threshold vector of each IMU training sample is then multiplied by the corresponding high-dimensional data features of the channel to obtain the denoised data features of each channel of each IMU training sample. The denoised data features of multiple channels of each IMU training sample are then output to the Kth convolutional layer.

[0097] The noise estimation convolutional neural network processes each group of IMU training samples to calculate the harmonic coefficient vector of each group. This harmonic coefficient vector is then multiplied by the integrity constraint noise empirical value vector to obtain the estimated noise vector for each group of IMU training samples. Finally, the estimated noise vector of each group of IMU training samples is output to the fused Kalman filter model.

[0098] The estimated noise vector for each group of IMU training samples is defined as follows:

[0099]

[0100] k∈[1,K]

[0101] Among them, (κ) 1,k ,κ 2,k ,κ 3,k ) represents the harmonic coefficient vector of the k-th IMU training sample, κ i,k ∈[-1,1], i∈[1,3], Let K represent the vector of empirical values ​​for the integrity constraint noise, and κ represent the number of training samples. 1,kκ represents the harmonic coefficient of the vehicle forward velocity noise in the harmonic coefficient vector of the k-th IMU training sample. 2,k κ represents the vehicle lateral velocity noise harmonic coefficient in the harmonic coefficient vector of the k-th IMU training sample. 3,k This represents the vehicle vertical velocity noise harmonic coefficient in the harmonic coefficient vector of the k-th IMU training sample. This represents an empirical value for the vehicle's forward speed noise. This represents an empirical value for vehicle lateral speed noise. This represents an empirical value for the vertical speed noise of the vehicle.

[0102] The inertial navigation module calculates the inertial measurement results of each group of IMU training samples using an inertial navigation algorithm, and outputs the inertial measurement results of each group of IMU training samples to the fused Kalman filter model.

[0103] The inertial measurement results for each group of IMU training samples are specifically defined as follows:

[0104] [φ k V k ,p k ]

[0105] k∈[1,K]

[0106] Where, φ k V represents the vehicle attitude as determined by the inertial measurement results of the k-th IMU training sample. k p represents the vehicle speed from the inertial measurement results of the k-th IMU training sample. k The vehicle position represents the inertial measurement result of the k-th IMU training sample, where K represents the number of training samples;

[0107] The fusion Kalman filter model calculates the predicted vehicle position for each IMU training sample by combining the estimated velocity, the estimated noise vector, the inertial measurement results, and the introduced incomplete observation constraints of each IMU training sample through fusion Kalman filtering.

[0108] The process of calculating the predicted vehicle position for each IMU training sample by fusing Kalman filtering is as follows:

[0109] The Kalman filter model system observation vector is constructed using the estimated velocity, non-perfection constraint, and estimated noise vector of each IMU training sample, as detailed below:

[0110]

[0111] k∈[1,K]

[0112] in, The observation vector representing the velocity of the k-th IMU training sample in the vehicle coordinate system is the system observation vector. Let R represent the estimation velocity of the k-th IMU training sample, [0,0] represent the introduced incomplete observation constraint, K represent the number of training samples, and R represent the estimation velocity of the k-th IMU training sample. k Let represent the estimated noise vector of the k-th IMU training sample, specifically defined as follows:

[0113]

[0114] k∈[1,K]

[0115] Where, dig[·] represents a diagonal matrix, (κ 1,k ,κ 2,k ,κ 3,k ) represents the harmonic coefficient vector of the k-th IMU training sample, κ i,k ∈[-1,1], Let K represent the vector of empirical values ​​for the integrity constraint noise, and κ represent the number of training samples. 1,k κ represents the harmonic coefficient of the vehicle forward velocity noise in the harmonic coefficient vector of the k-th IMU training sample. 2,k κ represents the vehicle lateral velocity noise harmonic coefficient in the harmonic coefficient vector of the k-th IMU training sample. 3,k This represents the vehicle vertical velocity noise harmonic coefficient in the harmonic coefficient vector of the k-th IMU training sample. This represents an empirical value for the vehicle's forward speed noise. This represents an empirical value for vehicle lateral speed noise. This represents the empirical value of the vehicle's vertical speed noise, and K represents the number of training samples;

[0116] The system error vector of the fused Kalman filter model for each group of IMU training samples is calculated as follows:

[0117] The vehicle velocity from the inertial measurement results of each IMU training sample is transformed into a coordinate system to obtain the expression for the vehicle velocity in the vehicle coordinate system, as follows:

[0118]

[0119] Among them, V k The vehicle speed represents the inertial measurement result of the k-th IMU training sample. This represents the system velocity of the k-th IMU training sample in the navigation coordinate system. This represents the angular velocity observation value of the k-th IMU training sample in the carrier coordinate system. This indicates the length of the lever arm between the IMU coordinate system and the vehicle coordinate system. This represents the rotation matrix between the vehicle coordinate system and the IMU coordinate system; Let represent the rotation matrix between the navigation coordinate system and the IMU coordinate system of the k-th IMU training sample;

[0120] The velocity observation equations for each group of IMU training samples are as follows:

[0121]

[0122] in, V represents the velocity observation residual of the k-th IMU training sample. k The vehicle speed represents the inertial measurement result of the k-th IMU training sample. This represents the observation vector of the velocity of the k-th IMU training sample in the vehicle coordinate system;

[0123] By combining the vehicle velocity expression in the vehicle coordinate system from the inertial measurement results of each IMU training sample and the velocity observation equation of each IMU training sample, the system error vector of the fusion Kalman filter model of each IMU training sample is calculated by the standard Kalman filter method.

[0124] The system error vector of the fused Kalman filter model for each group of IMU training samples is specifically defined as follows:

[0125]

[0126] k∈[1,K]

[0127] Where, (δφ) k Let (δv) represent the vehicle attitude error, which is the system error vector of the fused Kalman filter model for the k-th IMU training samples. k Let (δp) represent the vehicle speed error in the system error vector of the fused Kalman filter model for the k-th IMU training samples. k The vehicle position error (δb) represents the system error vector of the fused Kalman filter model for the k-th IMU training samples. w ) k The zero-bias error of the gyroscope represents the system error vector of the fused Kalman filter model for the k-th IMU training samples, (δb) a ) k The zero-bias error of the acceleration in the system error vector of the fused Kalman filter model for the k-th IMU training samples is represented. The system error vector of the fused Kalman filter model for the k-th IMU training sample represents the rotational error between the carrier coordinate system and the vehicle coordinate system. The lever arm error between the carrier coordinate system and the vehicle coordinate system represents the system error vector of the fused Kalman filter model for the k-th IMU training sample, where K represents the number of training samples.

[0128] The predicted vehicle position for each IMU training sample is calculated by combining the vehicle position from the inertial measurement results of each IMU training sample with the position error of the system error vector of the fused Kalman filter model of each IMU training sample, as follows:

[0129]

[0130] Where, p k δp represents the vehicle position, calculated from the inertial measurement results of the k-th IMU training sample. k The vehicle position error represents the system error vector of the fused Kalman filter model for the k-th IMU training samples. This represents the predicted vehicle position of the k-th IMU training sample;

[0131] The loss function model described in step 2 is specifically defined as follows:

[0132]

[0133] Where, label k This represents the true vehicle location label of the k-th IMU training sample. Let K represent the predicted vehicle position of the k-th IMU training sample, where K represents the number of training samples.

[0134] Step 3: Combine the real-time collected vehicle IMU data with the optimized vehicle position estimation network to predict the vehicle position and obtain the real-time predicted vehicle position.

[0135] A specific embodiment of the present invention also provides a computer-readable medium.

[0136] The computer-readable medium is a server workstation;

[0137] The server workstation stores the computer program executed by the electronic device. When the computer program runs on the electronic device, it causes the electronic device to execute the steps of the vehicle position estimation method of the fusion filtering network according to the present invention.

[0138] It should be understood that any parts not described in detail in this specification belong to the prior art.

[0139] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A vehicle position estimation method using a fusion filter network, characterized in that, Includes the following steps: Step 1: Obtain vehicle raw IMU data and vehicle location data at multiple historical time points. Obtain vehicle IMU data at multiple historical time points by adding noise to the vehicle raw IMU data at multiple historical time points. Combine the vehicle IMU data at multiple historical time points with the vehicle location data and divide them into multiple groups of IMU training samples through a window. Label each group of IMU training samples with the real vehicle location label. Step 2: Construct a vehicle position estimation network that integrates a Kalman filter model. Input each set of IMU training samples into the vehicle position estimation network for prediction to obtain the predicted vehicle position label for each set of IMU training samples. Combine the real vehicle position labels of each set of IMU training samples to construct a loss function model. Optimize the training using the ADAM algorithm to obtain the optimized vehicle position estimation network. Step 3: Combine the real-time collected vehicle IMU data with the optimized vehicle position estimation network to predict the vehicle position and obtain the real-time predicted vehicle position. The multiple sets of IMU training samples mentioned in step 1 are specifically defined as follows: k∈[1,K] in, Indicates the first k Group training samples, K Indicates the number of training samples. L This indicates the number of vehicle IMU data points in each IMU training sample group. Indicates the first k The first group of training samples i The data from the vehicle's IMU, representing the data from the [number]th vehicle IMU. Vehicle IMU data at a historical moment, Indicates the first Vehicle location at a historical moment i∈[1,L] ; The actual vehicle location of each IMU training sample in step 1 is specifically defined as follows: in, Indicates the first k Real vehicle location labels for the training samples Indicates the first +1 historical vehicle position; The vehicle position estimation network fused with the Kalman filter model in step 2 includes: Velocity estimation depth residual network, noise estimation convolutional neural network, inertial navigation module, and fused Kalman filter model; The velocity estimation depth residual network is connected to the fused Kalman filter model; The noise estimation convolutional neural network is connected to the fused Kalman filter module; The inertial navigation module is connected to the fused Kalman filter model; Each set of IMU training samples is input into the velocity estimation depth residual network, the noise estimation convolutional neural network, and the inertial navigation module, respectively. The noise estimation convolutional neural network processes each group of IMU training samples to calculate the harmonic coefficient vector of each group of IMU training samples, multiplies the harmonic coefficient vector of each group of IMU training samples by the integrity constraint noise empirical value vector, and obtains the estimated noise vector of each group of IMU training samples; the estimated noise vector of each group of IMU training samples is then output to the fused Kalman filter model. The estimated noise vector for each group of IMU training samples is defined as follows: k∈[1,K] in, This represents the harmonic coefficient vector of the k-th IMU training sample. , ， This represents the vector of empirical values ​​for the integrity constraint noise. K Indicates the number of training samples. This represents the harmonic coefficients of the vehicle's forward velocity noise in the harmonic coefficient vector of the k-th IMU training sample. This represents the vehicle lateral velocity noise harmonic coefficient in the harmonic coefficient vector of the k-th IMU training sample. This represents the vehicle vertical velocity noise harmonic coefficient in the harmonic coefficient vector of the k-th IMU training sample. This represents an empirical value for the vehicle's forward speed noise. This represents an empirical value for vehicle lateral speed noise. This represents an empirical value for the vertical speed noise of the vehicle. The inertial navigation module calculates the inertial measurement results of each group of IMU training samples using an inertial navigation algorithm, and outputs the inertial measurement results of each group of IMU training samples to the fused Kalman filter model. The inertial measurement results for each group of IMU training samples are specifically defined as follows: k∈[1,K] in, The vehicle attitude represents the inertial measurement results of the k-th IMU training sample. The vehicle speed represents the inertial measurement result of the k-th IMU training sample. This represents the vehicle position as determined by the inertial measurement results of the k-th IMU training sample. K Indicates the number of training samples; The fusion Kalman filter model calculates the predicted vehicle position for each IMU training sample by combining the estimated velocity, estimated noise vector, and inertial measurement results of each IMU training sample with the introduced incomplete observation constraints through fusion Kalman filtering.

2. The vehicle position estimation method using a fusion filtering network according to claim 1, characterized in that: The velocity estimation depth residual network is composed of multiple depth residual convolutional modules cascaded in sequence; Each depthwise residual convolutional module consists of a cascaded convolutional layer consisting of the first convolutional layer, the second convolutional layer, ..., the (K-1)th convolutional layer, a soft thresholding module, and the Kth convolutional layer. The velocity estimation deep residual network calculates the depth residual for each group of IMU training samples by sequentially passing each deep residual convolution module, obtains the estimated velocity for each group of IMU training samples, and outputs it to the fusion Kalman filter model. The (j-1)th depth residual convolution module uses the multi-channel data features extracted by the convolutional neural network as input to the j-th depth residual convolution module; In the j-th depth residual convolution module, the multi-channel data features of the (j-1)-th depth residual convolution module are further extracted by convolution through the first convolutional layer, the second convolutional layer, ..., the (K-1)-th convolutional layer, and output to the soft thresholding module.

3. The vehicle position estimation method using a fusion filtering network according to claim 2, characterized in that: The soft thresholding module sequentially processes the high-dimensional data features of each channel of each IMU training sample through absolute value conversion, global averaging, and attention mechanism network processing to obtain each threshold vector of each IMU training sample. It then processes each threshold vector of each IMU training sample through an activation function to obtain each activated threshold vector of each IMU training sample. Finally, it multiplies each activated threshold vector of each IMU training sample with the corresponding high-dimensional data features of the channel to obtain the denoised data features of each channel of each IMU training sample. The denoised data features of multiple channels of each IMU training sample are then output to the Kth convolutional layer.

4. The vehicle position estimation method using a fusion filtering network according to claim 1, characterized in that: The process of calculating the predicted vehicle position for each IMU training sample by fusing Kalman filtering is as follows: The Kalman filter model system observation vector is constructed using the estimated velocity, non-perfection constraint, and estimated noise vector of each IMU training sample, as detailed below: k∈[1,K] in, The observation vector representing the velocity of the k-th IMU training sample in the vehicle coordinate system is the system observation vector. This represents the estimated speed of the k-th IMU training sample. This represents the introduced non-integrity observation constraint. K Indicates the number of training samples. Let represent the estimated noise vector of the k-th IMU training sample, specifically defined as follows: k∈[1,K] in, Indicates a diagonal matrix. This represents the harmonic coefficient vector of the k-th IMU training sample. , This represents the vector of empirical values ​​for the integrity constraint noise. K Indicates the number of training samples. This represents the harmonic coefficients of the vehicle's forward velocity noise in the harmonic coefficient vector of the k-th IMU training sample. This represents the vehicle lateral velocity noise harmonic coefficient in the harmonic coefficient vector of the k-th IMU training sample. This represents the vehicle vertical velocity noise harmonic coefficient in the harmonic coefficient vector of the k-th IMU training sample. This represents an empirical value for the vehicle's forward speed noise. This represents an empirical value for vehicle lateral speed noise. This represents the empirical value of vehicle vertical speed noise. K This indicates the number of training samples.

5. The vehicle position estimation method using a fusion filtering network according to claim 4, characterized in that: The system error vector of the fused Kalman filter model for each group of IMU training samples is calculated as follows: The vehicle velocity from the inertial measurement results of each IMU training sample is transformed into a coordinate system to obtain the expression for the vehicle velocity in the vehicle coordinate system, as follows: in, The vehicle speed represents the inertial measurement result of the k-th IMU training sample. This represents the system velocity of the k-th IMU training sample in the navigation coordinate system. This represents the angular velocity observation value of the k-th IMU training sample in the carrier coordinate system. This indicates the length of the lever arm between the IMU coordinate system and the vehicle coordinate system. This represents the rotation matrix between the vehicle coordinate system and the IMU coordinate system; Let represent the rotation matrix between the navigation coordinate system and the IMU coordinate system of the k-th IMU training sample; The velocity observation equations for each group of IMU training samples are as follows: in, This represents the velocity observation residual of the k-th IMU training sample. The vehicle speed represents the inertial measurement result of the k-th IMU training sample. This represents the observation vector of the velocity of the k-th IMU training sample in the vehicle coordinate system; By combining the vehicle velocity expression in the vehicle coordinate system from the inertial measurement results of each IMU training sample and the velocity observation equation of each IMU training sample, the system error vector of the fusion Kalman filter model of each IMU training sample is calculated by the standard Kalman filter method. The system error vector of the fused Kalman filter model for each group of IMU training samples is specifically defined as follows: k∈[1,K] in, The vehicle attitude error represents the system error vector of the fused Kalman filter model for the k-th IMU training samples. The vehicle speed error represents the system error vector of the fused Kalman filter model for the k-th IMU training sample. The vehicle position error represents the system error vector of the fused Kalman filter model for the k-th IMU training samples. The zero-bias error of the gyroscope represents the system error vector of the fused Kalman filter model for the k-th IMU training samples. The zero-bias error of the acceleration in the system error vector of the fused Kalman filter model for the k-th IMU training samples is represented. The system error vector of the fused Kalman filter model for the k-th IMU training sample represents the rotational error between the carrier coordinate system and the vehicle coordinate system. The lever arm error between the carrier coordinate system and the vehicle coordinate system represents the system error vector of the fused Kalman filter model for the k-th IMU training sample. K This indicates the number of training samples.

6. The vehicle position estimation method using a fusion filtering network according to claim 5, characterized in that: The predicted vehicle position for each IMU training sample is calculated by combining the vehicle position from the inertial measurement results of each IMU training sample with the position error of the system error vector of the fused Kalman filter model of each IMU training sample, as follows: in, This represents the vehicle position as determined by the inertial measurement results of the k-th IMU training sample. The vehicle position error represents the system error vector of the fused Kalman filter model for the k-th IMU training samples. Indicates the first k Predicted vehicle positions from IMU training samples.

7. A computer-readable medium, characterized in that, It stores a computer program executed by an electronic device, which, when run on the electronic device, causes the electronic device to perform the steps of the method as described in any one of claims 1-6.

Images