An inertial attitude and positioning method applied to a vehicle

Abstract

An inertial attitude positioning method applied to a car, characterized in that it comprises the following steps: step one, obtaining three-axis accelerometer data and three-axis gyroscope data and filtering out engine jitter and other irregular high-frequency noise through a low-pass or band-pass filter; step two, extracting high-dimensional features through a recurrent convolutional neural network for the filtered six-axis data to determine the carrier motion attitude at this time; step three, determining a motion equation through the carrier motion attitude, processing the motion equation through a Kalman filter to obtain the attitude, and updating the speed and navigation coordinates according to the attitude. The present application overcomes the defects of the prior art and accurately completes the inertial attitude positioning on the car.

Description

TECHNICAL FIELD

[0001] The application is an inertial attitude positioning method applied to a vehicle, and relates to the technical field of vehicle dynamic attitude estimation and position estimation, and particularly relates to a vehicle dynamic attitude estimation and position estimation method based on a multi-sensor inertial navigation system. BACKGROUND

[0002] With the development of MEMS technology, the availability of strapdown inertial navigation solutions using MEMS accelerometers and gyroscopes is increasingly strong, and at the same time has a huge cost advantage compared to other inertial navigation solutions. Today, high-level autonomous driving systems have become the focus of development in the current automotive industry. Most of these electronic systems require feedback based on vehicle attitude and position information. Therefore, the development of intelligent driving vehicle systems is based on accurate, stable, and real-time attitude and navigation information. Inertial navigation also has applications in intelligent driving, and the accuracy of the strapdown inertial navigation system in the current technology still has a lot of room for improvement. The existing strapdown inertial navigation solution is as shown in the prior art. Figure 1

[0003] Defects and deficiencies of the prior art:

[0004] 1. Currently, MEMS sensors cannot completely remove noise effects even after strict calibration due to process limitations.

[0005] 2. In addition to identifying the motion attitude of the carrier, identifying the motion type of the carrier is also very critical. The existing solution can only identify the attitude and cannot determine the motion type.

[0006] 3. The existing solution must have different calibration solutions and different attitude solving motion equations for different carriers, which consumes manpower and resources.

[0007] In summary, the existing technical solution has a low accuracy rate for attitude and position estimation, and cannot directly obtain the motion state of the carrier. SUMMARY

[0008] The purpose of the present application is to provide an inertial attitude positioning method applied to a vehicle to overcome the defects of the prior art.

[0009] An inertial attitude positioning method applied to a vehicle, characterized in that it comprises the following steps:

[0010] Step 1: Obtain three-axis accelerometer data and three-axis gyroscope data and filter out engine jitter and other irregular high-frequency noise through a low-pass or band-pass filter;

[0011] Step 2: Extract high-dimensional features from the filtered six-axis data through a recurrent convolutional neural network to determine the motion attitude of the carrier at this time;

[0012] ​Step three, determine the motion equation through the carrier motion posture, process the motion equation through the Kalman filter to obtain the posture, and update the speed and navigation coordinates according to the posture.

[0013] The filter is a finite impulse response filter, including but not limited to Butterworth filter and Chebyshev filter.

[0014] The finite impulse response filter is a Butterworth filter or a Chebyshev filter.

[0015] The filtered six-axis data is extracted through a recurrent convolutional neural network to determine the carrier motion posture at this time.

[0016] The recurrent convolutional neural network is to output the final posture as part of the input vector of the next time of the network.

[0017] The recurrent convolutional neural network uses no less than six convolution kernels, and the convolution operation result of each convolution kernel corresponds to a pooling layer and no less than two hidden layers.

[0018] The softmax function is characterized by calculating in the following way:

[0019]

[0020] Wherein, Vi is the output of the output unit before the classifier. i represents the category index, and the total number of categories is C. Si represents the ratio of the current element index to the sum of all element indexes, which converts the output value of multi-classification into a relative probability

[0021] The six convolution kernels each correspond to a motion feature, including but not limited to the relationship between lateral acceleration and angular velocity, and the direction and size change of gravity direction acceleration.

[0022] The motion equation is processed through the Kalman filter, specifically including the following steps:

[0023] Construct a Kalman filter system model:

[0024]

[0025] Where: x(k)∈R nx(k) represents the state vector of the system at time k; i(k) represents the motion state distinguished by the previous stage neural network at time k; z(k) is in R m x(k) represents the state vector of the system at time k; i(k) represents the motion state distinguished by the previous stage neural network at time k; z(k) is in R n and v(k) is in R m is the process noise vector with covariance Q(k) and R(k) respectively;

[0026] Filter equation is established: including time update, measurement update;

[0027] The time update includes the prior estimation of the system state quantity and the prior estimation of the prediction error covariance matrix;

[0028] The measurement update includes the Kalman gain of the system, the posterior estimation of the system state quantity and the posterior estimation of the prediction error covariance matrix.

[0029] The beneficial effects of the present application are:

[0030] The present application solves three problems of the prior art: 1) At present, MEMS sensors cannot completely remove the noise influence even after strict calibration due to the process limitation. 2) In addition to identifying the motion posture of the carrier, the identification of the motion type of the carrier is also very critical. The existing scheme can only identify the posture and cannot determine the motion type. 3) The existing scheme must have different calibration schemes and different posture solving motion equations for different carriers, which consumes manpower and material resources.

[0031] Therefore, in the scheme provided by the present application, the time sequence of the sensor is analyzed, the noise in the specific frequency band is screened out, the interference is reduced, and the signal-to-noise ratio is improved. The motion state information in the time sequence is identified by the convolution structure and the multi-layer perception machine, so as to automatically select the motion equation under different carriers. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 , it is the flow chart of the existing strapdown inertial navigation technology;

[0033] Figure 2 , it is the frequency domain response curve of the frequency domain filter designed by the present application;

[0034] Figure 3 , it is the overall flow chart of the present application;

[0035] Figure 4 , it is the convolution recurrent neural network structure diagram (1) of the present application, which describes the convolution structure;

[0036] Figure 5, as the invention convolutional recurrent neural network structure diagram (2), describe the loop structure;

[0037] Figure 6 , as the invention step schematic diagram. DETAILED DESCRIPTION

[0038] An inertial positioning method applied to a car, characterized in that it comprises the following steps:

[0039] Step one, obtain three-axis accelerometer data and three-axis gyroscope data and pass through a low-pass or band-pass filter to filter out engine jitter and other irregular high-frequency noise;

[0040] Step two, extract high-dimensional features from the filtered six-axis data through a recurrent convolutional neural network to determine the carrier motion attitude at this time;

[0041] Step three, determine the motion equation through the carrier motion attitude, process the motion equation through a Kalman filter to obtain the attitude, and update the speed and navigation coordinates according to the attitude.

[0042] The filter is a finite impulse response filter, including but not limited to Butterworth filter and Chebyshev filter.

[0043] The finite impulse response filter is a Butterworth filter or a Chebyshev filter.

[0044] The filtered six-axis data is extracted through a recurrent convolutional neural network to determine the carrier motion attitude at this time, which uses a recurrent convolutional neural network to extract features from time sequences, obtains the probability of each motion state, and selects the maximum probability as the estimated value of the motion state at this time.

[0045] The recurrent convolutional neural network is a part of the input vector of the next time of the network with the final attitude output.

[0046] The recurrent convolutional neural network uses no less than 6 convolution kernels, and the convolution operation result of each convolution kernel corresponds to a pooling layer and no less than two hidden layers, and finally each convolution kernel output is output through a softmax function to estimate the prior probability of each motion state, and the motion state with the maximum probability is obtained by accumulation.

[0047] The softmax function is characterized by being calculated in the following way:

[0048]

[0049] Wherein, Vi is the output of the pre-classifier output unit, i represents the class index, and the total number of classes is C, and Si represents the ratio of the index of the current element to the sum of the indexes of all elements, which converts the output value of the multi-classification into a relative probability

[0050] The six convolution kernels each correspond to a motion feature, and the motion feature includes but is not limited to the relationship between lateral acceleration and angular velocity, and the direction and size change of the acceleration in the direction of gravity.

[0051] The motion equation is processed through the Kalman filter, and the specific steps include the following steps:

[0052] A Kalman filter system model is constructed:

[0053]

[0054] Wherein: x(k) belongs to R n The state vector of the system at the k moment is represented; i(k) represents the motion state distinguished by the pre-stage neural network at the k moment; z(k) belongs to R m The observation vector of the system at the k moment is represented; f(·) is an n-dimensional vector function, h(·) is an m-dimensional vector function, f(·) and h(·) are nonlinear with respect to their arguments; w(k) belongs to R n And v(k) belongs to R m The process noise vector with covariance Q(k) and R(k) respectively;

[0055] The filter equation is established: including time update and measurement update;

[0056] The time update includes the prior estimation of the system state quantity and the prior estimation of the prediction error covariance matrix;

[0057] The measurement update includes the Kalman gain of the system, the posterior estimation of the system state quantity and the posterior estimation of the prediction error covariance matrix.

[0058] The present application provides a scheme for removing specific frequency noise, calculating the carrier attitude and motion state through a neural network and the six-axis output of an accelerometer and a gyroscope.

[0059] The scheme first removes specific noise such as engine vibration through a frequency domain filter, and the frequency domain response curve of the filter is as shown in Figure 2 .

[0060] The scheme applies a convolutional neural network to the data processing of the accelerometer and the gyroscope, improves the accuracy of attitude recognition through the learning ability of the neural network, and greatly reduces the cost of the scheme applied to different carriers.

[0061] The flow of the method is as shown in Figure 3.

[0062] To address the aforementioned problems, the technical solution provided by this invention is as follows: the time-series array obtained by the six-axis inertial measurement unit in t samplings is treated as a 6*t two-dimensional image. Using a sliding window approach, the 6*n data points from the first n time steps are input into a recurrent convolutional neural network. The network structure is shown below. Figure 4 , Figure 5 .

[0063] Neural networks use convolution operations, such as... Figure 4 As shown, features are extracted to a higher dimension, and different dimensional characteristics are obtained through a multilayer perceptron. Finally, the motion state with the highest probability is selected based on the probability of each motion state. At the same time, the final output motion state is also used as the input for the next time step, forming a cyclic structure.

[0064] The motion equations are determined based on the determined motion posture, and then the attitude matrix is ​​regressed by extended Kalman filtering. The velocity and position are then calculated from the attitude matrix.

[0065]

[0066] Where: x(k)∈R n Let i(k) represent the state vector of the system at time k; i(k) represent the motion state determined by the preceding neural network at time k; z(k)∈R m Let w(k) represent the observation vector of the system at time k; f(·) is an n-dimensional vector function, and h(·) is an m-dimensional vector function. Both f(·) and h(·) are nonlinear with respect to their independent variables; w(k)∈R n and v(k)∈R m It is a process noise vector with covariances Q(k) and R(k).

[0067] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0068] Example 1:

[0069] like Figure 3 As shown, the inertial navigation attitude determination and positioning method provided in this embodiment of the invention includes the following steps:

[0070] First, angular velocity and angular speed are obtained from the gyroscope and accelerometer, and the accelerometer is calibrated according to the specific force equation. Initial alignment is then performed based on the accelerometer readings.

[0071] The second step is to extract high-dimensional features from the time-series data using a convolutional recurrent neural network. High-dimensional features are extracted through the convolutional part of the network, and then the previous time-step pose is used as input to construct a finite state machine via the recurrent neural network.

[0072] Thirdly, construct the motion equation according to the specific force equation. The specific force equation is as follows:

[0073]

[0074] Fourthly, take the motion equation constructed according to the motion state and the specific force equation in the third step as the state equation of the Kalman filter. At this time, the form of the Kalman filter is as follows:

[0075]

[0076] Wherein, x(k)∈R n represents the state vector of the system at the k moment; i(k) represents the motion state distinguished by the previous neural network at the k moment; z(k)∈R m represents the observation vector of the system at the k moment; f(·) is an n-dimensional vector function, h(·) is an m-dimensional vector function, f(·) and h(·) are nonlinear with respect to their arguments; w(k)∈R n and v(k)∈R m are process noise vectors with covariance Q(k) and R(k) respectively.

[0077] Fifthly and sixthly, solve the attitude matrix by using the quaternion, and solve the position and velocity according to the attitude.

[0078] The differential equation form for solving the quaternion is as follows:

[0079]

[0080] Wherein, is the quaternion of the b coordinate system relative to the n coordinate system; is the derivative of ;

[0081]

[0082] Wherein, ω x represents the angular velocity around the pitch axis; ω y represents the angular velocity around the roll axis; ω z represents the angular velocity around the yaw axis. The matrix form of the quaternion motion equation is as follows:

[0083]

[0084] Wherein, T represents the sampling interval of the system; and the following is obtained:

[0085]

[0086] Solving the quaternary differential equations using the first-order Runge-Kutta method yields the discrete-time model as follows:

[0087]

[0088] Where, q k-1 Let q(0) represent the quaternion at time k-1, I be the identity matrix, set the initial quaternion q(0), and use this equation to recursively derive the value q of the quaternion at time k. k Then update the attitude matrix.

[0089] Example 2:

[0090] like Figure 2 As shown, noise at a specific frequency is first removed by a finite impulse response filter based on a specific frequency response curve, thereby avoiding noise interference caused by the three-axis sway of the vehicle body and engine vibration.

[0091] The solution applies convolutional neural networks to the data processing of an inertial measurement unit composed of accelerometers and gyroscopes. Through the learning ability of neural networks, the accuracy of attitude recognition is improved, the uncertainty of Kalman filtering caused by statistical errors is reduced, and the adaptation cost of the solution to different carriers is greatly reduced.

[0092] First, the time series array obtained by the six-axis inertial measurement unit in t samplings is considered as a two-dimensional time series of size 6*t. Then, using a sliding window approach, the 6*n data points from the first n time steps are input into a recurrent convolutional neural network.

[0093] Neural networks extract features to higher dimensions through convolution operations, acquire features of different dimensions through multilayer perceptrons, and finally select the most probable motion state based on the probability of each motion state. Simultaneously, the final output motion state also serves as the input for the next time step, forming a recurrent structure.

[0094] The motion equation is determined based on the determined motion attitude, and then the attitude matrix is ​​regressed by extended Kalman filter. The velocity and navigation coordinates are then calculated from the attitude matrix.

[0095] Example 3:

[0096] This solution is applied to the Volkswagen Tiguan model. Testing was conducted at a sampling frequency of 500Hz using both the three-axis accelerometer and the three-axis gyroscope.

[0097] The experiment was conducted under various complex motion scenarios. Steady-state motion included: uniform motion, stationary (engine started), and stationary (engine not started); transient motion included: acceleration, turning, and going uphill or downhill.

[0098] The data can be clustered by adopting a machine learning method, and a convolution kernel is determined by an artificial strategy.The state at any moment is regarded as a Markov process, convolution is carried out through a convolution neural network, and the posture at the current moment is determined.

[0099] Formulation of the artificial strategy: the weight of each axis possessed by different motion states; whether the motion state can be regarded as a Markov process.

[0100] In the test, a standard turntable and field test are combined, data objective description of statistical analysis results, comparison of two group means, analysis of multiple group variances, correlation analysis of variables, multiple regression prediction, and equation model construction of results are carried out.

[0101] The application discloses a method for vehicle body dynamic posture estimation and position estimation based on inertial sensors.

[0102] The application provides a scheme for removing specific frequency noise, estimating a motion state of a carrier by using outputs of an accelerometer and a gyroscope through a neural network, and finally calculating a carrier posture, velocity and position through Kalman filtering and a quaternion.

[0103] The scheme firstly removes specific noise such as engine vibration through a frequency domain filter, and retains a signal.The convolution neural network is applied to data processing of an inertial measurement unit composed of the accelerometer and the gyroscope, the learning ability of the neural network is used, the accuracy of posture recognition is improved, Kalman filtering is improved, and the cost of the scheme applied to different carriers is greatly reduced.

[0104] Firstly, a time series array obtained by a six-axis inertial measurement unit at t times of sampling is regarded as a 6*t size two-dimensional time series.Through a sliding window mode, 6*n data at the previous n times are input into a recurrent convolution neural network,

[0105] The neural network extracts features to a higher dimension through convolution operation, obtains different dimension features through a multilayer perception machine, finally selects a motion state with the largest possibility according to probabilities of each motion state, and simultaneously the finally output motion state is also input into the next moment to form a loop structure.

[0106] According to the determined motion posture, a motion equation is determined, an extended Kalman filter is used to regress the posture matrix at this moment, and the velocity and navigation coordinates are solved through the posture matrix.

[0107] The above-described embodiments are only preferred embodiments for fully illustrating the present application, and the protection scope of the present application is not limited thereto.Any equivalent replacement or transformation of the present application based on the present application is within the protection scope of the present application.The protection scope of the present application is subject to the claims.

Claims

1. An inertial positioning method for a vehicle, characterized in that, It comprises the following steps: Step one, obtain three-axis accelerometer data and three-axis gyroscope data and filter out engine jitter and other irregular high-frequency noise through a low-pass or band-pass filter; Step two, extract high-dimensional features from the filtered six-axis data through a recurrent convolutional neural network to determine the carrier motion posture at this time; Step three, determine the motion equation through the carrier motion posture, process the motion equation through a Kalman filter to obtain the posture, and update the speed and navigation coordinates according to the posture; The recurrent convolutional neural network is used to extract features from time sequences, obtain the probability of each motion state, and select the maximum probability as the estimated value of the motion state at this time; The recurrent convolutional neural network is a part of the input vector of the next time of the network with the final posture output as the input vector; The recurrent convolutional neural network uses no less than 6 convolution kernels, and the convolution operation result of each convolution kernel corresponds to a pooling layer and no less than two hidden layers, and finally each convolution kernel output is individually output through a softmax function to estimate the prior probability of each motion state, and the motion state with the maximum probability is obtained by accumulation; The softmax function is characterized by calculating as follows: Where, Vi is the output of the output unit before the classifier; i represents the class index, and the total number of classes is C; Si represents the ratio of the exponential of the current element to the sum of the exponents of all elements, which converts the output value of multi-classification into a relative probability; The six convolution kernels each correspond to a motion feature, and the motion feature includes the relationship between lateral acceleration and angular velocity, and the direction and size change of the acceleration in the direction of gravity; The motion equation is processed through a Kalman filter, which comprises the following steps: Construct a Kalman filter system model: wherein: represents the state vector of the system at time instant i (k) represents k the motion state identified by the previous stage neural network at time instant represents the observation vector of the system at time instant is a vector function of dimension a vector function of dimension a vector function of dimension and are both nonlinear in their arguments; and are process noise vectors with covariance and respectively. Establish a filter equation, including time update and measurement update; The time update includes the prior estimation of the system state quantity and the prior estimation of the prediction error covariance matrix; The measurement update includes the Kalman gain of the system, the posterior estimation of the system state quantity, and the posterior estimation of the prediction error covariance matrix.

2. The inertial positioning method according to claim 1, applied to a vehicle, characterized in that, The filter is a finite impulse response filter.

3. The inertial orientation positioning method applied to a vehicle according to claim 2, characterized in that, The finite impulse response filter is a Butterworth filter or a Chebyshev filter.

Images