CN-121977543-A - IMU posture correction method, device, equipment and medium based on time sequence large model

CN121977543ACN 121977543 ACN121977543 ACN 121977543ACN-121977543-A

Abstract

The invention relates to the technical field of data processing, aims to solve the problem of low accuracy of inertial measurement unit attitude estimation in the prior art, and provides an IMU attitude correction method based on a time sequence large model. The method comprises the steps of preprocessing angular velocity data and acceleration data acquired by an IMU to obtain angular velocity time sequence data and acceleration time sequence data, inputting the angular velocity time sequence data and the acceleration time sequence data into a first time sequence big model to obtain a current motion mode, determining gesture correction parameters according to the current motion mode, carrying out integral calculation according to the angular velocity time sequence data to obtain initial gesture data, constructing an error prediction input sequence based on historical inertia measurement data, the motion mode and the gesture correction parameters, inputting the error prediction input sequence into a second time sequence big model to obtain a gesture error prediction result, and correcting and compensating the initial gesture data according to the gesture error prediction result to obtain corrected gesture data. The adaptability, stability and precision of the attitude estimation are improved by introducing the time sequence large model.

Inventors

ZHANG ZHI
LI WEI
XIONG ZHANG
CHEN HUI
ZHANG QINGQI
AI WEI

Assignees

宁波星巡智能科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260116

Claims (10)

1. An IMU pose correction method based on a time sequence large model, which is characterized by comprising the following steps: Preprocessing angular velocity data and acceleration data acquired by an inertial measurement unit according to a preset sliding time window to obtain angular velocity time sequence data and acceleration time sequence data; Inputting the angular velocity time sequence data and the acceleration time sequence data into a pre-constructed first time sequence large model to obtain a current motion mode, and determining posture correction parameters according to the motion mode, wherein the posture correction parameters are used for indicating weight distribution and/or correction intensity adjustment of a posture correction process; performing integral calculation according to the angular velocity time sequence data to obtain initial attitude data; Constructing an error prediction input sequence according to the historical inertial measurement data of the inertial measurement unit, the motion mode and the posture correction parameters, and inputting the error prediction input sequence into a pre-constructed second time sequence large model to obtain a posture error prediction result; and correcting and compensating the initial posture data according to the posture error prediction result to obtain corrected posture data, wherein the first time sequence large model and the second time sequence large model are constructed based on a time sequence neural network structure.
2. The IMU posture correction method based on a time series large model according to claim 1, wherein the preprocessing of angular velocity data and acceleration data acquired by an inertial measurement unit according to a preset sliding time window to obtain angular velocity time series data and acceleration time series data includes: acquiring angular velocity data and acceleration data corresponding to each sampling moment and performing time alignment; Constructing a sliding time window according to the preset window length and the sliding step length, and writing the angular velocity data and the acceleration data in a plurality of continuous sampling periods into the window according to time sequence to obtain an initial angular velocity sequence and an initial acceleration sequence; performing outlier processing on the initial angular velocity sequence and the acceleration sequence; Filtering the angular velocity sequence and the acceleration sequence according to a preset filtering algorithm to obtain a filtered angular velocity sequence and an acceleration sequence, wherein the preset filtering algorithm comprises low-pass filtering and/or Kalman filtering; and carrying out standardization processing on the filtered angular velocity sequence and the acceleration sequence to obtain the angular velocity time sequence data and the acceleration time sequence data.
3. The IMU posture correction method based on a time series large model according to claim 1, wherein said inputting the angular velocity time series data and the acceleration time series data into a first time series large model constructed in advance, obtaining a current motion pattern, and determining posture correction parameters according to the motion pattern, comprises: Performing characteristic organization on the angular velocity time sequence data and the acceleration time sequence data to construct multichannel time sequence input data for representing motion changes in a preset sliding time window; Inputting the multichannel time sequence input data into the first time sequence big model, and outputting a motion mode classification result and/or a motion mode confidence level, wherein the motion mode comprises a static mode, a steady motion mode, a severe shaking mode or a specific motion mode; and selecting posture correction parameters corresponding to the motion mode from a preset parameter set according to the motion mode classification result and/or the motion mode confidence coefficient.
4. The IMU pose correction method based on time series large model according to claim 3, wherein said performing integral calculation according to said angular velocity time series data to obtain initial pose data comprises: Acquiring angular velocity data corresponding to the current sampling time and the previous sampling time according to the angular velocity time sequence data; calculating an average value of the angular velocity according to the angular velocity data of the current sampling time and the previous sampling time; And integrating the gestures according to the average angular velocity and the time interval between the adjacent sampling moments to obtain initial gesture data of the current sampling moment.
5. The IMU attitude correction method based on a time series large model according to claim 3, wherein said constructing an error prediction input sequence from the historical inertial measurement data of said inertial measurement unit, said motion pattern and said attitude correction parameters, and inputting said error prediction input sequence into a second time series large model constructed in advance, to obtain an attitude error prediction result, comprises: Acquiring a training sample, wherein the training sample comprises historical inertial measurement data, a motion mode label corresponding to the historical inertial measurement data, posture correction parameters and posture error labeling data; constructing a training input sequence according to the training sample, and inputting the training input sequence into a second time sequence large model to obtain an attitude error prediction output; calculating prediction loss according to the attitude error prediction output and the attitude error labeling data, determining a physical consistency constraint item based on rigid body kinematics constraint and/or dynamics constraint, and adding the physical consistency constraint item as a priori constraint or loss item into a training target; parameter updating is carried out on the second time sequence big model according to the prediction loss and the physical consistency constraint item, and a trained second time sequence big model is obtained; In an online prediction stage, acquiring historical inertial measurement data in a preset time window, and aligning with the current motion mode and the gesture correction parameters in time sequence; constructing an error prediction input sequence according to the historical inertial measurement data, the motion mode and the posture correction parameters; And inputting the error prediction input sequence into the trained second time sequence large model, and outputting an attitude error prediction result in a plurality of sampling periods in the future.
6. The IMU pose correction method based on time series large model according to claim 5, wherein said correcting and compensating said initial pose data according to said pose error prediction result, obtaining corrected pose data comprises: determining an attitude error compensation amount according to the attitude error prediction result; performing prediction compensation on the initial posture data based on the posture error compensation quantity to obtain compensated initial posture data; Estimating a gravity direction according to the acceleration data, and calculating a pitch angle and a roll angle corresponding to the gravity direction; Acquiring pitch angles and roll angles in the compensation initial attitude data; And carrying out complementary filtering fusion on the pitch angle and the roll angle and the pitch angle and the roll angle in the compensated initial posture data to obtain corrected posture data.
7. The IMU attitude correction method based on a time series large model according to claim 6, wherein the complementary filtering fusion is performed on the pitch angle and the roll angle and the pitch angle and the roll angle in the compensated initial attitude data, so as to obtain corrected attitude data, and the method comprises the following steps: Acquiring a filter fusion coefficient according to the attitude correction parameter, wherein the attitude correction parameter comprises an evaluation weight for the reliability of accelerometer data in a current motion mode, and determining the filter fusion coefficient through linear mapping or a preset lookup table according to the evaluation weight; Respectively weighting the pitch angle and the roll angle estimated by the accelerometer and the pitch angle and the roll angle in the compensated initial attitude data according to the filter fusion coefficient to obtain a fusion result; And updating the corrected pitch angle and roll angle according to the fusion result, and taking the updated pitch angle and roll angle as the corrected posture data.
8. An IMU pose correction device based on a time series large model, the device comprising: the data acquisition module is used for preprocessing the angular velocity data and the acceleration data acquired by the inertial measurement unit according to a preset sliding time window to obtain angular velocity time sequence data and acceleration time sequence data; The motion mode determining module is used for inputting the angular velocity time sequence data and the acceleration time sequence data into a pre-constructed first time sequence big model to obtain a current motion mode, and determining posture correction parameters according to the motion mode, wherein the posture correction parameters are used for indicating weight distribution and/or correction intensity adjustment of a posture correction process; the gesture acquisition module is used for carrying out integral calculation according to the angular velocity time sequence data to obtain initial gesture data; The error prediction module is used for constructing an error prediction input sequence according to the historical inertial measurement data of the inertial measurement unit, the motion mode and the posture correction parameters, and inputting the error prediction input sequence into a pre-constructed second time sequence large model to obtain a posture error prediction result; And the compensation and correction module is used for correcting and compensating the initial posture data according to the posture error prediction result to obtain corrected posture data, wherein the first time sequence large model and the second time sequence large model are constructed based on a time sequence neural network structure.
9. An electronic device comprising at least one processor, at least one memory, and computer program instructions stored in the memory, which when executed by the processor, implement the method of any of claims 1-7.
10. A storage medium having stored thereon computer program instructions, which when executed by a processor, implement the method of any of claims 1-7.

Description

IMU posture correction method, device, equipment and medium based on time sequence large model Technical Field The present invention relates to the field of data processing, and in particular, to a method, apparatus, device, and medium for correcting an IMU posture based on a time sequence large model. Background The inertial measurement unit (Inertial Measurement Unit, IMU for short) is an important sensor component for realizing motion perception and attitude estimation, and is widely applied to scenes such as unmanned aerial vehicles, intelligent nursing robots, vehicle navigation, augmented reality equipment, wearable systems and the like. IMUs are typically composed of gyroscopes and accelerometers for measuring angular velocity and linear acceleration, respectively. Attitude change information can be obtained by integrating angular velocity signals of the gyroscope, but because the sensor has error sources such as offset, temperature drift, noise and the like, the integration result can generate accumulated deviation along with time to form drift errors, thereby affecting the accuracy and stability of attitude calculation. Accelerometers can provide a gravity direction dependent attitude reference under low dynamic conditions, but their output is prone to drift from the gravity direction when dynamic motion, vibration or transient acceleration disturbances are present, resulting in unstable correction effects. Therefore, relying on gyroscopes or accelerometers alone is difficult to achieve long-term stable high-accuracy attitude estimation under complex conditions. In order to reduce error accumulation, the prior art generally adopts fusion algorithms such as complementary filtering, kalman filtering and the like, combines short-term dynamic response of a gyroscope with long-term stability constraint of an accelerometer, and corrects the attitude by setting filtering parameters. However, the fusion method of the fixed parameters or the fixed model is difficult to adapt to the actual scene of continuous change of the motion state, and error correction is easily introduced due to excessive dependence on the accelerometer under high dynamic state, or drift accumulation is easily caused due to insufficient correction under low dynamic state. Meanwhile, there is also a proposal for correcting inertial measurement data by using a neural network. For example, publication CN112985462a discloses a method and a device for correcting inertial measurement data based on a convolutional neural network model, which adopts a sliding window to segment IMU measurement data in a period of time, combines a position data sequence of a positioning module as a training sample input, obtains the convolutional neural network model through iterative training, and is used for outputting IMU estimation data at each moment to reduce measurement errors. The scheme is focused on learning and correcting inertial data, but the scheme still needs to face the problems that in the gesture resolving application, the error characteristic difference is large, the correction strategy is difficult to adaptively switch, the prediction and advance compensation mechanism is lack for short-time future drift trend, and the like. With the development of deep learning and time sequence modeling technology, the time sequence large model shows advantages in the aspects of mode identification and trend prediction of multichannel time sequences, and provides a new thought for realizing more adaptive posture correction under different motion states. Therefore, how to introduce a time sequence large model to analyze the IMU time sequence data and improve the adaptability and the accuracy of posture correction is a technical problem to be solved urgently. Disclosure of Invention In view of the above, the embodiments of the present invention provide a method, an apparatus, a device, and a storage medium for correcting an IMU posture based on a time-series large model, so as to solve the problem in the prior art that the accuracy of inertial measurement unit posture estimation is low. In a first aspect, an embodiment of the present invention provides a method for correcting an IMU pose based on a time-series large model, the method including: Preprocessing angular velocity data and acceleration data acquired by an inertial measurement unit according to a preset sliding time window to obtain angular velocity time sequence data and acceleration time sequence data; Inputting the angular velocity time sequence data and the acceleration time sequence data into a pre-constructed first time sequence large model to obtain a current motion mode, and determining posture correction parameters according to the motion mode, wherein the posture correction parameters are used for indicating weight distribution and/or correction intensity adjustment of a posture correction process; performing integral calculation according to the angular velocity time sequence data to obta