IMU (Inertial Measurement Unit) speed estimation method and system based on multi-scale time sequence feature coding and Mama

Abstract

The invention provides an IMU (Inertial Measurement Unit) speed estimation method and system based on multi-scale time sequence feature coding and Mama. The method belongs to the field of IMU speed estimation, and comprises the following steps: S1, acquiring original time sequence data which is generated by an IMU and contains three-axis acceleration and three-axis angular velocity, and intercepting a section of historical time sequence data with a fixed length by taking a current moment as an end point to construct a historical window; s2, inputting the original time sequence data in the historical window into a multi-scale one-dimensional convolution encoder, processing the input data by the encoder by using at least two convolution kernels with different sizes in parallel so as to extract time sequence characteristics under different receptive fields, and fusing output characteristics of each branch to obtain a potential characteristic sequence; and S3, inputting the potential feature sequence into a Mama model to carry out time sequence dependency relationship modeling, and extracting an output state of the Mama model at the end of the sequence as a solved three-dimensional speed output. According to the method, the IMU speed estimation precision can be improved.

Description

Technical Field

[0001] This application relates to the field of IMU velocity estimation, and more specifically, to an IMU velocity estimation method and system based on multi-scale temporal feature coding and Mamba. Background Technology

[0002] With the widespread adoption of microelectromechanical inertial measurement units (MEMS IMUs) in wearable devices, robots, autonomous platforms, and vehicle systems due to their small size, low power consumption, and low cost, motion estimation based on IMUs has become an important research direction in the field of positioning and navigation. IMUs can provide high-frequency motion information even without external base stations.

[0003] However, relying solely on raw measurements from a single IMU is susceptible to noise drift, time-varying zero bias, and differences in motion patterns in real-world scenarios, resulting in insufficient accuracy and long-term stability in velocity estimation. This limits its usability in applications such as high-precision positioning and behavior recognition. Therefore, to achieve high-precision IMU velocity estimation, further processing of the raw IMU data using effective IMU velocity estimation methods is necessary.

[0004] Currently, technical methods for IMU velocity estimation can be mainly divided into two categories: traditional methods based on physical models and filtering, and data-driven learning methods. For example, invention patent CN118311630B proposes a high-precision positioning method for unmanned loaders based on RTK and IMU, estimating IMU velocity through high-precision satellite observation and Kalman filtering. However, under conditions of satellite signal loss or obstruction, the velocity estimation accuracy of this method decreases significantly. Invention patent CN116558523A proposes a velocity estimation method and device based on convolutional neural networks, using CNN to regress the carrier velocity from single IMU data. However, this method has limited ability to extract IMU temporal features, resulting in limited velocity estimation accuracy.

[0005] Therefore, in order to achieve high-precision and robust IMU velocity estimation in complex environments such as satellite obstruction, there is an urgent need for an end-to-end method that balances long-term temporal dependency modeling capabilities with the efficiency of learning models, so as to improve estimation accuracy. Summary of the Invention

[0006] The purpose of this application is to provide an IMU velocity estimation method and system based on multi-scale temporal feature coding and Mamba, which can improve the accuracy of IMU velocity estimation.

[0007] This application is implemented as follows: Firstly, this application provides an IMU velocity estimation method based on multi-scale temporal feature encoding and Mamba, comprising the following steps: S1: Obtain the raw time series data containing triaxial acceleration and triaxial angular velocity generated by the IMU, and extract a fixed-length segment of historical time series data with the current time as the endpoint to construct a historical window; S2: Input the original temporal data within the historical window into a multi-scale one-dimensional convolutional encoder. The encoder processes the input data in parallel using at least two different sizes of convolutional kernels to extract temporal features under different receptive fields, and fuses the output features of each branch to obtain a latent feature sequence. S3: Input the latent feature sequence into the Mamba model to model the temporal dependency relationship, and extract the output state of the Mamba model at the end of the sequence as the desired three-dimensional velocity output.

[0008] Based on the first aspect, the specific steps for acquiring the raw time-series data containing triaxial acceleration and triaxial angular velocity generated by the IMU, and constructing a historical window by extracting a fixed-length segment of historical time-series data with the current time as the endpoint, include: exist t The expression for the IMU output vector at any given time is defined as: ; in, for t The output vector of the IMU at time 10:00. for t The output of the triaxial accelerometer at all times, for t Output of the three-axis angular velocity at any given time; Given the length of history, construct t The expression for the historical IMU measurement matrix at time point is: ; in, for t Historical IMU measurement matrix at any given time for t The output vector of the IMU at time 10:00. L For the length of history; An asymmetric input-output mapping strategy is adopted, using the historical sequence as input, and predicting only the three-dimensional velocity at the last time step. The expression for the asymmetric input-output mapping strategy is: ; in, for t Predicted three-dimensional velocity at time step for t Historical IMU measurement matrix at any given time This is a nonlinear mapping implemented by a deep learning network.

[0009] Based on the first aspect, the original temporal data within the historical window is input into a multi-scale one-dimensional convolutional encoder. The encoder processes the input data in parallel using at least two different sizes of convolutional kernels to extract temporal features under different receptive fields, and the output features of each branch are fused to obtain the latent feature sequence. The specific steps include: Motion features at different time scales are extracted from the raw time-series data within the historical window. Several one-dimensional convolutional kernels are applied in parallel to obtain features at multiple scales, the expression of which is: ; in, for t Time of the first i Convolutional outputs at various scales For a one-dimensional convolution operator, for t Historical IMU measurement matrix at any given time n The scale number; The features obtained from parallel scaling are concatenated into a holistic feature set, and a unified latent expression is obtained through linear projection: ; in, This is the latent feature matrix after encoder projection. For the overall feature matrix, For the projection matrix, This is the bias vector.

[0010] Based on the first aspect, the specific steps of inputting the latent feature sequence into the Mamba model for temporal dependency modeling and extracting the output state of the Mamba model at the end of the sequence as the desired three-dimensional velocity output include: The latent feature sequences from the latent feature matrix are input into the Mamba model, and the time evolution of the hidden states is recursively modeled using the Mamba architecture, expressed as: ; in, This represents the hidden state vector. Let be the hidden state vector from the previous time step. For learnable parameter matrix, For observation vectors; After modeling, the output state at the end of the sequence of the Mamba model is taken as the desired 3D velocity output, and its expression is: ; in, for t Predicted three-dimensional velocity at time step This refers to the observation at the last time step within the current time window.

[0011] Based on the first aspect, the following steps are also included: S4: Perform end-to-end training on the Mamba model and output the IMU speed after training is complete.

[0012] Based on the first aspect, the specific steps for end-to-end training of the Mamba model and outputting the IMU speed after training include: During the offline training phase, mean squared error is used as the training loss for end-to-end optimization. The loss function expression is as follows: ; in, For loss function, The number of training samples, For the first i Predicted 3D velocity for each training sample. The corresponding true value velocity; After training is completed, perform forward computation according to steps S1–S3 to output the IMU speed in real time.

[0013] Based on the first aspect, the length of the history window is 0.1 seconds to 5 seconds.

[0014] Secondly, this application provides an IMU velocity estimation system based on multi-scale temporal feature coding and Mamba, comprising the following steps: The data input definition module is configured to acquire raw time-series data containing triaxial acceleration and triaxial angular velocity generated by the IMU, and extract a fixed-length segment of historical time-series data with the current time as the endpoint to construct a historical window; The temporal feature extraction module is configured to input the raw temporal data within the historical window into a multi-scale one-dimensional convolutional encoder. The encoder processes the input data in parallel using at least two different sizes of convolutional kernels to extract temporal features under different receptive fields, and fuses the output features of each branch to obtain a latent feature sequence. The modeling and velocity output module is configured to input the latent feature sequence into the Mamba model to model temporal dependencies, and extract the output state of the Mamba model at the end of the sequence as the desired three-dimensional velocity output.

[0015] Thirdly, this application provides an electronic device, comprising: Memory, used to store one or more programs; processor; The above method is implemented when one or more programs are executed by the processor.

[0016] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0017] Compared with the prior art, this application has at least the following advantages or beneficial effects: This invention provides an IMU velocity estimation method and system based on multi-scale temporal feature encoding and Mamba. By extracting a fixed-length historical window with the current time as the endpoint, it can achieve real-time online estimation based on past and current data, suitable for applications requiring low-latency feedback, such as robot control and VR / AR interaction. By inputting the raw temporal data within the historical window into a multi-scale one-dimensional convolutional encoder, a latent feature sequence is obtained. Using convolutional kernels of different sizes in parallel allows for the simultaneous extraction of short-, medium-, and long-term motion features, providing the model with richer and more comprehensive input information. Different convolutional kernels have different receptive fields; the latent feature sequence obtained by fusing them has a much higher information density and expressive power than features extracted at a single scale. Performing multi-scale one-dimensional convolution followed by temporal modeling using the Mamba model, the combination of "multi-scale one-dimensional convolution + Mamba model" ensures a "comprehensive observation" of motion information by the model, laying a solid foundation for high-precision estimation. Subsequent temporal modeling using the Mamba model can more effectively learn and compensate for the inherent integral drift of the IMU. The model can learn drift patterns from the entire history window and perform implicit corrections on output. This improves the accuracy and robustness of velocity estimation. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of an embodiment of an IMU velocity estimation method based on multi-scale temporal feature encoding and Mamba according to this application; Figure 2 This is a flowchart of another embodiment of an IMU velocity estimation method based on multi-scale temporal feature encoding and Mamba according to this application; Figure 3 This is a comparison of the IMU velocity estimation method based on multi-scale temporal feature encoding and Mamba from this application with the results of IMU velocity estimation using a convolutional neural network method. Figure 4This is a schematic diagram of the structure of an IMU velocity estimation system based on multi-scale temporal feature coding and Mamba according to this application; Figure 5 This is a schematic diagram of the structure of an electronic device according to this application.

[0020] icon: 1. Data input definition module; 2. Temporal feature extraction module; 3. Modeling and speed output module; 4. Processor; 5. Memory; 6. Communication interface. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0022] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the various embodiments and features described below can be combined with each other.

[0023] Through long-term research and practice, the inventors have discovered that existing technologies for IMU (Inertial Measurement Unit) velocity estimation can be mainly divided into two categories: one is the traditional method based on physical models and filtering, and the other is the data-driven learning method. For example, invention patent CN118311630B proposes a high-precision positioning method for unmanned loaders based on RTK and IMU, which estimates IMU velocity through high-precision satellite observation and Kalman filtering. However, under conditions of satellite signal loss or obstruction, the velocity estimation accuracy of this method decreases significantly. Invention patent CN116558523A proposes a velocity estimation method and device based on convolutional neural networks, which uses CNN to regress the carrier velocity from single IMU data. However, this method has limited ability to extract IMU temporal features, resulting in limited velocity estimation accuracy.

[0024] Therefore, this application provides an IMU velocity estimation method based on multi-scale temporal feature encoding and Mamba, which can improve the accuracy of IMU velocity estimation.

[0025] Please refer to Figure 1 The proposed IMU velocity estimation method based on multi-scale temporal feature encoding and Mamba includes the following steps: S1: Obtain the raw time series data containing triaxial acceleration and triaxial angular velocity generated by the IMU, and extract a fixed-length segment of historical time series data with the current time as the endpoint to construct a historical window; Specifically, using the current time t as a baseline, a fixed time interval is traced back, and data from that fixed time interval is extracted as an analysis sample. When estimating the current velocity, the model relies only on past and present information, not future information, thus achieving real-time online estimation. This is suitable for applications requiring low-latency feedback, such as robot control and VR / AR interaction. The serialized data generated in this step is a prerequisite for processing in subsequent steps S2 and S3.

[0026] Preferably, the length of the history window is between 0.1 seconds and 5 seconds.

[0027] As one implementation method, step S1 can be completed in the following way: exist t The expression for the IMU output vector at any given time is defined as: ; in, for t The output vector of the IMU at time 10:00. for t The output of the triaxial accelerometer at all times, for t Output of the three-axis angular velocity at any given time; Given the length of history, construct t The expression for the historical IMU measurement matrix at time point is: ; in, for t Historical IMU measurement matrix at any given time for t The output vector of the IMU at time 10:00. L For the length of history; An asymmetric input-output mapping strategy is adopted, using the historical sequence as input, and predicting only the three-dimensional velocity at the last time step. The expression for the asymmetric input-output mapping strategy is: ; in, for t Predicted three-dimensional velocity at time step for t Historical IMU measurement matrix at any given time This is a nonlinear mapping implemented by a deep learning network.

[0028] S2: Input the original temporal data within the historical window into a multi-scale one-dimensional convolutional encoder. The encoder processes the input data in parallel using at least two different sizes of convolutional kernels to extract temporal features under different receptive fields, and fuses the output features of each branch to obtain a latent feature sequence. Specifically, the multi-scale one-dimensional convolutional encoder deploys multiple one-dimensional convolutional kernels of different widths (e.g., k=3, 5, 7). The size of the convolutional kernel directly determines its "receptive field." Small convolutional kernels (e.g., 3) focus on local, high-frequency details (e.g., instantaneous jitter, impact); large convolutional kernels (e.g., 7) focus on more global, low-frequency trends (e.g., slow changes in motion direction). The feature maps extracted from each branch are concatenated or weighted and fused along the channel dimension to form a "latent feature sequence" with higher information density and both local details and global trends. This setup ensures that the model can simultaneously perceive subtle changes in motion and macroscopic motion trends, providing a rich source of information for high-precision estimation. Traditional methods require manually designed filters to extract features in specific frequency bands. This approach allows the model to automatically learn the optimal feature representation, making it more intelligent and adaptive. By fusing multi-scale information, the model is more robust to noise and unimportant high-frequency fluctuations.

[0029] As one possible implementation method, step 2 can be completed using the following approach: Motion features at different time scales are extracted from the raw time-series data within the historical window. Several one-dimensional convolutional kernels are applied in parallel to obtain features at multiple scales, the expression of which is: ; in, for t Time of the first i Convolutional outputs at various scales For a one-dimensional convolution operator, for t Historical IMU measurement matrix at any given time n The scale number; The features obtained from parallel scaling are concatenated into a holistic feature set, and a unified latent expression is obtained through linear projection: ; in, This is the latent feature matrix after encoder projection. For the overall feature matrix, For the projection matrix, This is the bias vector.

[0030] S3: Input the latent feature sequence into the Mamba model to model the temporal dependency relationship, and extract the output state of the Mamba model at the end of the sequence as the desired three-dimensional velocity output.

[0031] Specifically, Mamba receives the latent feature sequence output from step S2 and processes it step by step. At each step, it combines the current input with the previous hidden state, updates the state, and generates an output. This process allows information from earlier moments in the sequence to influence and propagate to later moments. Extracting the last output state of the sequence is the "decision point" of the entire scheme. After processing the entire sequence, the hidden state of the Mamba model at the last time step condenses all the contextual information of the entire historical window. Using this most information-rich vector as the basis for the final velocity estimation lays the foundation for obtaining higher-precision velocity in subsequent steps. Traditional IMU integration suffers from drift that grows quadratically over time due to noise and zero bias. However, Mamba, through its powerful sequence learning capabilities, can implicitly learn the drift pattern and "infer" a more accurate velocity from global contextual information, thereby suppressing or even correcting the drift. Furthermore, the original data, after transformations in steps S2 and S3, is directly mapped to a three-dimensional velocity vector [v_x, v_y, v_z]. The entire process is clean and efficient, requiring no complex post-processing.

[0032] As one possible implementation, step S3 can be completed using the following method: The latent feature sequences from the latent feature matrix are input into the Mamba model, and the time evolution of the hidden states is recursively modeled using the Mamba architecture, expressed as: ; in, This represents the hidden state vector. Let be the hidden state vector from the previous time step. For learnable parameter matrix, For observation vectors; After modeling, the output state at the end of the sequence of the Mamba model is taken as the desired 3D velocity output, and its expression is: ; in, for t Predicted three-dimensional velocity at time step This refers to the observation at the last time step within the current time window.

[0033] Please refer to Figure 2 As one implementation method of this embodiment, the IMU velocity estimation method based on multi-scale temporal feature encoding and Mamba further includes the following steps: S4: Perform end-to-end training on the Mamba model and output the IMU speed after training is complete.

[0034] The specific steps include: During the offline training phase, mean squared error is used as the training loss for end-to-end optimization. The loss function expression is as follows: ; in, For loss function, The number of training samples, For the first i Predicted 3D velocity for each training sample. The corresponding true value velocity; After training is completed, perform forward computation according to steps S1–S3 to output the IMU speed in real time.

[0035] In summary, this invention provides an IMU velocity estimation method and system based on multi-scale temporal feature encoding and Mamba. By extracting a fixed-length historical window with the current time as the endpoint, it can achieve real-time online estimation based on both past and current data, making it suitable for applications requiring low-latency feedback, such as robot control and VR / AR interaction. By inputting the raw temporal data within the historical window into a multi-scale one-dimensional convolutional encoder, a latent feature sequence is obtained. Using convolutional kernels of different sizes in parallel allows for the simultaneous extraction of short-, medium-, and long-term motion features, providing the model with richer and more comprehensive input information. Different convolutional kernels have different receptive fields; the latent feature sequence obtained by fusing them has a much higher information density and expressive power than features extracted at a single scale. Performing multi-scale one-dimensional convolution followed by temporal modeling using the Mamba model, the combination of "multi-scale one-dimensional convolution + Mamba model" ensures a comprehensive observation of motion information by the model, laying a solid foundation for high-precision estimation. Subsequent temporal modeling using the Mamba model can more effectively learn and compensate for the inherent integral drift of the IMU. The model can learn drift patterns from the entire history window and perform implicit corrections on output. This improves the accuracy and robustness of velocity estimation.

[0036] To verify the correctness of this invention, a field test was conducted, and the performance parameters of the IMU in the experiment are listed in Table 1.

[0037] Table 1: Performance parameters of the IMU in the actual experiment

[0038] Three sets of field experiments were conducted using the vehicle-mounted positioning data acquired by the IMU corresponding to Table 1. IMU velocity estimation was performed using both the convolutional neural network velocity estimation method and the IMU velocity estimation method based on multi-scale temporal feature encoding and Mamba proposed in this application. The three-dimensional velocity estimation errors of the two methods are shown in Table 2.

[0039] Table 2: Comparison of 3D velocity estimation errors of different algorithms

[0040] In Table 2, Vx-RMSE is the root mean square error of the velocity component in the X-axis direction, Vy-RMSE is the root mean square error of the velocity component in the Y-axis direction, Vz-RMSE is the root mean square error of the velocity component in the Z-axis direction, and Mean is the mean of the overall root mean square error of the three-dimensional velocity vector estimation.

[0041] Furthermore, the applicant conducted IMU velocity estimation simulations using the convolutional neural network velocity estimation method and the IMU velocity estimation method based on multi-scale temporal feature encoding and Mamba proposed in this application. The simulation results are available for reference. Figure 3 .

[0042] From Table 2 and Figure 3 As can be seen, the algorithm proposed in this invention improves the accuracy of three-dimensional velocity estimation by an average of about 35% compared with convolutional neural networks, indicating that the method proposed in this invention has higher velocity prediction accuracy. Therefore, the effectiveness and correctness of the method provided in this invention are verified.

[0043] Please refer to Figure 4 This invention also provides an IMU velocity estimation system based on multi-scale temporal feature encoding and Mamba, comprising the following steps: Data input definition module 1 is configured to acquire raw time series data containing triaxial acceleration and triaxial angular velocity generated by the IMU, and extract a fixed-length segment of historical time series data with the current time as the endpoint to construct a historical window; The temporal feature extraction module 2 is configured to input the original temporal data within the historical window into a multi-scale one-dimensional convolutional encoder. The encoder processes the input data in parallel using at least two different sizes of convolutional kernels to extract temporal features under different receptive fields, and fuses the output features of each branch to obtain a latent feature sequence. The modeling and velocity output module 3 is configured to input the latent feature sequence into the Mamba model to model the temporal dependency relationship, and extract the output state of the Mamba model at the end of the sequence as the desired three-dimensional velocity output.

[0044] Specifically, for a detailed implementation of the IMU velocity estimation system based on multi-scale temporal feature coding and Mamba, please refer to the detailed implementation of the IMU velocity estimation method based on multi-scale temporal feature coding and Mamba, which will not be elaborated on here.

[0045] Please refer to Figure 5 The present invention also provides an electronic device, comprising: Memory 5 is used to store one or more programs; Processor 4; Processor 4 and memory 5 are connected via communication interface 6; When one or more programs are executed by processor 4, all or some of the above methods are implemented.

[0046] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor 4, implements all or part of the methods described above.

[0047] It will be apparent to those skilled in the art that this application is not limited to the details of the exemplary embodiments described above, and that this application can be implemented in other specific forms without departing from the spirit or essential characteristics of this application. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of this application is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within this application. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for IMU velocity estimation based on multi-scale temporal feature encoding and Mamba, characterized in that, Includes the following steps: S1: Obtain the raw time series data containing triaxial acceleration and triaxial angular velocity generated by the IMU, and extract a fixed-length segment of historical time series data with the current time as the endpoint to construct a historical window; S2: Input the original temporal data within the historical window into a multi-scale one-dimensional convolutional encoder. The encoder processes the input data in parallel using at least two different sizes of convolutional kernels to extract temporal features under different receptive fields, and fuses the output features of each branch to obtain a latent feature sequence. S3: Input the latent feature sequence into the Mamba model to model the temporal dependency relationship, and extract the output state of the Mamba model at the end of the sequence as the desired three-dimensional velocity output.

2. The IMU velocity estimation method based on multi-scale temporal feature coding and Mamba as described in claim 1, characterized in that, The specific steps for acquiring raw time-series data containing triaxial acceleration and triaxial angular velocity generated by the IMU, and constructing a historical window by extracting a fixed-length segment of historical time-series data with the current time as the endpoint, include: exist t The expression for the IMU output vector at any given time is defined as: ； in, for t The output vector of the IMU at time 10:

00. for t The output of the triaxial accelerometer at all times, for t Output of the three-axis angular velocity at any given time; Given the length of history, construct t The expression for the historical IMU measurement matrix at time point is: ； in, for t Historical IMU measurement matrix at any given time for t The output vector of the IMU at time 10:

00. L For the length of history; An asymmetric input-output mapping strategy is adopted, using the historical sequence as input, and predicting only the three-dimensional velocity at the last time step. The expression for the asymmetric input-output mapping strategy is: ； in, for t Predicted three-dimensional velocity at time step for t Historical IMU measurement matrix at any given time This is a nonlinear mapping implemented by a deep learning network.

3. The IMU velocity estimation method based on multi-scale temporal feature coding and Mamba as described in claim 1, characterized in that, The original temporal data within the historical window is input into a multi-scale one-dimensional convolutional encoder. The encoder processes the input data in parallel using at least two different sizes of convolutional kernels to extract temporal features under different receptive fields, and fuses the output features of each branch to obtain a latent feature sequence. The specific steps include: Motion features at different time scales are extracted from the raw time-series data within the historical window. Several one-dimensional convolutional kernels are applied in parallel to obtain features at multiple scales, the expression of which is: ； in, for t Time of the first i Convolutional outputs at various scales For a one-dimensional convolution operator, for t Historical IMU measurement matrix at any given time n The scale number; The features obtained from parallel scaling are concatenated into a holistic feature set, and a unified latent expression is obtained through linear projection: ； in, This is the latent feature matrix after encoder projection. For the overall feature matrix, Let be the projection matrix. This is the bias vector.

4. The IMU velocity estimation method based on multi-scale temporal feature coding and Mamba as described in claim 3, characterized in that, The specific steps of inputting the latent feature sequence into the Mamba model for temporal dependency modeling and extracting the output state of the Mamba model at the end of the sequence as the desired three-dimensional velocity output include: The latent feature sequences from the latent feature matrix are input into the Mamba model, and the time evolution of the hidden states is recursively modeled using the Mamba architecture, expressed as: ； in, Let this be the hidden state vector at the current time step. Let be the hidden state vector from the previous time step. For learnable parameter matrix, For observation vectors; After modeling, the output state at the end of the sequence of the Mamba model is taken as the desired 3D velocity output, and its expression is: ； in, for t Predicted three-dimensional velocity at time step This refers to the observation at the last time step within the current time window.

5. The IMU velocity estimation method based on multi-scale temporal feature coding and Mamba as described in claim 1, characterized in that, It also includes the following steps: S4: Perform end-to-end training on the Mamba model and output the IMU speed after training is complete.

6. The IMU velocity estimation method based on multi-scale temporal feature coding and Mamba as described in claim 5, characterized in that, The specific steps for end-to-end training of the Mamba model and outputting the IMU speed after training include: During the offline training phase, mean squared error is used as the training loss for end-to-end optimization. The loss function expression is as follows: ； in, For loss function, The number of training samples, For the first i Predicted 3D velocity for each training sample. The corresponding true value velocity; After training is completed, perform forward computation according to steps S1–S3 to output the IMU speed in real time.

7. The IMU velocity estimation method based on multi-scale temporal feature coding and Mamba as described in claim 1, characterized in that, The length of the history window is 0.1 seconds to 5 seconds.

8. An IMU velocity estimation system based on multi-scale temporal feature coding and Mamba, characterized in that, Includes the following steps: The data input definition module is configured to acquire raw time-series data containing triaxial acceleration and triaxial angular velocity generated by the IMU, and extract a fixed-length segment of historical time-series data with the current time as the endpoint to construct a historical window; The temporal feature extraction module is configured to input the original temporal data within the historical window into a multi-scale one-dimensional convolutional encoder. The encoder processes the input data in parallel using at least two different sizes of convolutional kernels to extract temporal features under different receptive fields and fuses the output features of each branch to obtain a latent feature sequence. The modeling and velocity output module is configured to input the latent feature sequence into the Mamba model to model temporal dependencies, and extract the output state of the Mamba model at the end of the sequence as the desired three-dimensional velocity output.

9. An electronic device, characterized in that, include: Memory, used to store one or more programs; processor; When the one or more programs are executed by the processor, the method as described in any one of claims 1-7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method as described in any one of claims 1-7.

Images