Inertial navigation system error correction method and system based on IMU and visual data optimization

Abstract

The invention discloses an inertial navigation system error correction method and system based on IMU and visual data optimization, and belongs to the technical field of inertial navigation. The problems of vision constraint misalignment, positioning precision reduction and the like caused by sparse feature points and tracking failure due to the fact that integral drift of IMU original data of unmanned aerial vehicle inertial navigation is intensified and monocular vision data is influenced by texture blurring and insufficient illumination are solved. Comprising the following steps: collecting original IMU data and visual data; performing down-sampling and double quaternion interpolation reconstruction on the original IMU data; performing illumination quality optimization and texture quality optimization on the visual data to obtain an enhanced reflectogram; performing shallow feature extraction, depth feature extraction and high-resolution image reconstruction on the enhanced reflection image; and inputting the reconstructed image into an inertial navigation system, constructing a positioning target function, and searching a minimum value of the positioning target function, thereby obtaining an optimal feature extraction position and realizing error correction. The method is suitable for unmanned aerial vehicle geographic surveying and mapping, environment monitoring and other scenes.

Description

Technical Field

[0001] This invention belongs to the field of inertial navigation technology, specifically relating to an error correction technology for inertial navigation systems. Background Technology

[0002] In low-altitude missions such as surveying, logistics, and monitoring, multi-rotor UAVs rely on autonomous navigation, and the accuracy of their autonomous navigation is crucial to their efficient operation. Inertial navigation systems (INS) serve as the core positioning unit for UAVs, leveraging their strong autonomy to calculate the carrier's pose in real time through inertial measurement units (IMUs). However, high-precision INS are limited by cost, hindering their widespread application, while low-cost INS accumulate significant errors over extended periods.

[0003] To improve inertial navigation accuracy, existing technologies mainly employ three types of methods: hardware structure optimization methods such as rotation modulation, software compensation algorithms, and multi-sensor fusion. Among these, visual-inertial fusion technology has become a research hotspot due to its strong environmental adaptability.

[0004] Although there has been extensive research on visual-inertial fusion technology at the back-end fusion algorithm level, existing solutions often overlook the impact of input data quality on navigation error correction. On the one hand, unsuppressed random noise in the raw IMU data can cause error accumulation after pre-integration; on the other hand, directly using the raw visual data acquired by a monocular camera can easily lead to problems such as texture blurring and low illumination of visual information in complex environments, resulting in failure of feature point detection and tracking, thereby weakening the role of visual constraints in inertial navigation error correction.

[0005] High-precision positioning and attitude estimation for unmanned aerial vehicles (UAVs) in complex environments are crucial for their flight safety and mission execution. However, the IMU data relied upon by their inertial navigation systems is often affected by noise, limiting positioning accuracy. Furthermore, with increasing operating time or environmental complexity, the drift errors of accelerometers and gyroscopes accumulate, further impacting the accuracy of position and attitude estimation. While traditional Kalman filtering can mitigate noise interference to some extent, it still suffers from poor filtering performance and difficulty in meeting the requirements of high-precision positioning when facing high-frequency noise.

[0006] In the process of integrating visual information into UAV inertial navigation for localization, scene feature points are acquired through visual information, depth information is calculated, and visual constraints are constructed for inertial navigation localization. However, the effectiveness of this process is closely related to the ambient illumination. Sufficient ambient illumination is conducive to feature point extraction, while in dim scenes, the image quality acquired by the visual sensor will be greatly reduced, feature points will become sparse and tracking will be unstable, seriously affecting the reliability of inertial navigation localization.

[0007] In UAV inertial navigation systems, visual constraints can be constructed by fusing scene image information acquired from visual sensors with IMU data to compensate for inertial navigation drift and improve positioning accuracy and flight stability. However, when the acquired image texture is blurry, it leads to a sparse number of visual feature points, reduced matching accuracy, and even feature point tracking failure. Such low-quality visual constraints not only fail to effectively correct accumulated inertial navigation errors but may also introduce interference, exacerbate navigation calculation errors, and ultimately significantly affect the UAV's pose estimation accuracy and flight control performance. Summary of the Invention

[0008] This invention provides an error correction method and system for inertial navigation systems based on IMU and visual data optimization. The purpose is to solve the problems of visual-inertial fusion technology used in UAV inertial navigation in complex environments such as high dynamics, low illumination or weak texture, where the integration drift is aggravated due to the failure of the original IMU data to effectively suppress random noise, and the sparse feature points and tracking failure caused by the influence of texture blur and insufficient illumination on monocular visual data, which in turn leads to inaccurate visual constraints and reduced positioning accuracy.

[0009] In the first aspect, this invention proposes an error correction method for inertial navigation systems based on IMU and visual data optimization, comprising the following steps: S1: Acquire raw IMU data and visual data; S2: Downsampling and biquaternion interpolation reconstruction of the original IMU data; S3: Optimize the illumination quality and texture quality of the visual data to obtain an enhanced reflectance map; S4: Perform shallow feature extraction, deep feature extraction, and high-resolution image reconstruction on the enhanced reflectance map to obtain the reconstructed image; S5: Input the reconstructed image into the inertial navigation system, construct the localization objective function, find the minimum value of the localization objective function, thereby obtaining the optimal feature extraction position and realizing error correction.

[0010] Furthermore, a preferred embodiment is provided: S2 includes: The original IMU data is downsampled and converted into dual quaternions. The Euler method is used to update the state of the dual quaternions, and the angular increment and force increment are extracted to obtain optimized IMU data.

[0011] Furthermore, a preferred embodiment is provided: S3 includes: S31: Perform illumination component estimation and illumination map enhancement operations on the original image to obtain the optimal illumination map and illumination components; S32: Perform gamma correction on the illumination component to obtain a preliminary reflection map. Perform secondary noise reduction and enhancement on the preliminary reflection map, and then perform histogram stretching and contrast enhancement to obtain an enhanced reflection map.

[0012] Furthermore, a preferred embodiment is provided: the formula for estimating the illumination component is: , , , , in, The objective function to be minimized is used to solve for the illumination map. The location of the pixel. The x-coordinate is the pixel coordinate. The vertical coordinate is the pixel coordinate. To output the illumination map of the image, Let q be the location of the illumination map that needs to be solved at a neighboring pixel. The normalized values ​​of the input original image, This is a regularization parameter that controls the strength of the smoothing term. For pixels The neighborhood, For input image The weighting function controls the pixel and Smoothness intensity between, This is a constraint term for the difference in illumination between adjacent pixels.

[0013] Furthermore, a preferred solution is provided: a global smoothing filter is used to estimate the illumination components and enhance the illumination map of the original image; a reflectance enhancement module based on adaptive homomorphic filtering is used to perform secondary denoising enhancement on the preliminary reflectance map.

[0014] Furthermore, a preferred solution is provided: S4 uses the SwinIR algorithm to perform shallow feature extraction, deep feature extraction, and high-resolution image reconstruction on the enhanced reflectance map.

[0015] Furthermore, a preferred embodiment is provided: the localization objective function is determined based on the visual reprojection error and the IMU pre-integration error.

[0016] Secondly, this invention proposes an inertial navigation system error correction system based on IMU and visual data optimization. The system is implemented based on any one or more of the aforementioned schemes combined to provide an inertial navigation system error correction method based on IMU and visual data optimization. The system includes: Data acquisition module: used to acquire raw IMU data and visual data; Data processing module: used to downsample and reconstruct the raw IMU data using dual quaternion interpolation; Image optimization module: used to optimize the illumination quality and texture quality of visual data to obtain enhanced reflectance maps; Image reconstruction module: used to perform shallow feature extraction, deep feature extraction, and high-resolution image reconstruction on the enhanced reflectance map to obtain the reconstructed image; Error correction module: It is used to input the reconstructed image into the inertial navigation system, construct the positioning objective function, find the minimum value of the positioning objective function, thereby obtaining the optimal feature extraction position and realizing error correction.

[0017] Thirdly, the present invention also provides a computer device, the computer device including a memory and a processor, the memory storing a computer program, and when the processor runs the computer program stored in the memory, the processor executes an inertial navigation system error correction method based on IMU and visual data optimization according to any one or more of the above-described schemes.

[0018] Fourthly, the present invention also proposes a computer-readable storage medium for storing a computer program that executes an inertial navigation system error correction method based on IMU and visual data optimization as described in any one or more of the above-described schemes.

[0019] The IMU data downsampling and dual quaternion interpolation reconstruction method proposed in this invention effectively reduces the random error of IMU data by suppressing high-frequency noise and accurately reconstructing motion increments, thereby achieving inertial navigation error correction in long-term and high-dynamic scenarios.

[0020] The visual information illumination optimization and super-resolution texture enhancement method proposed in this invention significantly improves image quality under low illumination and texture blurring conditions, enhances the stability of feature point detection and tracking, and strengthens the contribution of visual constraints to inertial navigation error correction.

[0021] The method proposed in this invention has extremely important application value in high-precision autonomous navigation tasks in many fields such as UAV geographic mapping, logistics distribution, and environmental monitoring, especially in complex scenarios. Attached Figure Description

[0022] Figure 1 This is an execution flowchart of the visual-inertial information-enhanced UAV inertial navigation error correction method according to a specific embodiment of the present invention; Figure 2 This is a flowchart illustrating the optimization process of inertial navigation IMU data downsampling and dual quaternion interpolation reconstruction according to a specific embodiment of the present invention; Figure 3 The flowchart illustrates the inertial navigation error correction method for low-light image enhancement based on Retinex theory, as described in a specific embodiment of the present invention. Figure 4 This is a flowchart of the inertial navigation error correction method based on SwinIR super-resolution texture enhancement as described in a specific embodiment of the present invention; Figure 5 This is a line graph showing the IMU's three-axis angular velocity before and after reconstruction, as described in a specific embodiment of the present invention. Figure 6 This is a line graph showing the IMU's three-axis acceleration before and after reconstruction, as described in a specific embodiment of the present invention. Figure 7 This is a comparison of the low-light (left) and enhanced (right) effects of the V1_03 sequence in a specific embodiment of the present invention. Figure 8 This is a comparison image of the original MH_01 sequence (left) and the image after 2x super-resolution (right) according to a specific embodiment of the present invention. Detailed Implementation

[0023] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application can also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0024] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0026] Implementation Method 1: An error correction method for inertial navigation systems based on IMU and visual data optimization is proposed. First, IMU data reconstruction optimization based on dual quaternion interpolation is performed to correct the error of UAV inertial navigation.

[0027] The basic steps of biquaternion interpolation are as follows: Figure 2 As shown, firstly, the sampling rate of the raw IMU data. The original data is downsampled at a ratio of 1:2 to suppress high-frequency noise. The calculation method is as follows: , (1) in To the downsampled angular velocity, To obtain the downsampled acceleration, the downsampled IMU data is converted into a dual quaternion, representing the motion from both rotational and translational perspectives. The representation of the dual quaternion is as follows: (2) (3) (4) (5) (6) in It is a double quaternion. For rotation quaternions, For translation quaternions, For dual operators, Angular velocity in rotational quaternion representation. To represent acceleration as a rotational quaternion, after converting the IMU data into a double quaternion, the Euler method is needed for double quaternion state updates. The update formula is as follows: (7) (8) in, Indicates discrete sequence number. This indicates the update time step.

[0028] Finally, discrete angular and force increments are extracted from the dual quaternion sequence to provide incremental information on motion changes for the inertial navigation algorithm. The actual angular velocity and acceleration are calculated and compared with cubic spline interpolation data to correct the interpolated IMU three-axis acceleration and angular velocity data. The angular increment is calculated through the relative change of adjacent rotating quaternions, and the force increment is calculated through the difference of translational quaternions. The formulas for calculating the angular and force increments are shown below: (9) (10) (11) in For angular increments, For the increase in force, It represents the imaginary part of the quaternion.

[0029] This implementation proposes a low-light image enhancement algorithm based on Retinex theory, enabling UAVs to establish more robust visual constraints for inertial navigation positioning in dim environments, thus facilitating further error correction for inertial navigation. The principle diagram of inertial navigation error correction based on Retinex theory low-light image enhancement is shown below. Figure 3 As shown.

[0030] Retinex theory posits that an image can be decomposed into a reflection component (the texture structure of the object itself) and an illumination component (ambient light intensity). The core idea of ​​the algorithm is to improve image quality by indirectly calculating and enhancing the reflection component. The calculation formulas for the original image, illumination component, and reflection component are shown below: (12) in The original image of the object. The reflection component of the target object that carries image detail information. The ambient light component is the illumination component. Compared to the original image, the reflection component can reduce the influence of shadow areas caused by the illumination component, thereby recovering more of the object's inherent information. However, theoretically, the reflection component cannot be directly obtained from the original image alone. Therefore, the illumination component is estimated from the original image, and the reflection component is extracted based on the theoretical framework shown in Equation (12) by utilizing the ratio between the illumination component and the original image.

[0031] To optimize the illumination of UAV image information, the original image needs to pass through two modules in sequence: an illumination map estimation module based on a global smoothing filter and a reflectance map enhancement module based on an adaptive homomorphic filter.

[0032] First, in the illumination map estimation module of the global smoothing filter, the original image is subjected to illumination component estimation and illumination map enhancement operations using the global smoothing filter. Following the idea of ​​the global smoothing filter, this implementation simplifies the estimation of illumination components into an optimization problem. The calculation formula for illumination component estimation is as follows: (13) (14) (15) (16) in The objective function to be minimized is used to solve for the illumination map. The location of the pixel. and The x and y coordinates are pixels. To output the illumination map of the image, For the illumination map that needs to be solved, in the neighborhood pixels The value, For the input image in pixels Normalized intensity at that point The normalized values ​​of the input original image, This is a regularization parameter that controls the strength of the smoothing term. For pixels The neighborhood, For input image The weighting function controls the pixel and Smoothness intensity between, This is a constraint term for the difference in illumination between adjacent pixels.

[0033] Secondly, by analyzing the objective function By taking the derivative and setting its value to 0, the optimal illumination pattern can be obtained. The calculation formula is as follows: (17) (18) in As a unit array, The global index of the current pixel. To make every pixel in the entire image A column vector formed by stacking; To and The associated pixel indices. After solving for the optimal illumination map, the illumination components were obtained. The reflection component, i.e. the preliminary enhancement image, can be obtained according to equation (12). However, this can lead to uneven brightness distribution in the image, so gamma correction and dynamic compression are introduced. The brightness gradient in the highlight region achieves illumination equalization. The formula for calculating the reflectance after gamma correction is shown below: (19) in This is the gamma value, which is generally between 0.5 and 0.7.

[0034] Furthermore, the reflectance image obtained by equation (19) still suffers from low-frequency brightness shift and high-frequency texture confusion. Therefore, a reflectance image enhancement module based on adaptive homomorphic filtering is needed to perform secondary denoising enhancement on the reflectance image to obtain the enhanced image after low-light conditions. The first step of this module is to apply adaptive homomorphic filtering to the reflectance image to achieve residual low-frequency noise suppression and high-frequency detail enhancement. Let the input reflectance component be... Its height is Width is Then its logarithmic and frequency domain Fourier transforms are expressed as follows: (20) (twenty one) An adaptive homomorphic filter is constructed to suppress low-frequency noise while enhancing high-frequency details in the image. The principle and formula are shown below: (twenty two) (twenty three) (twenty four) In the formula The distance from the center point in the frequency domain coordinates. The low-frequency attenuation coefficient is empirically taken to be 0.5 based on the Nyquist sampling theorem. This is the high-frequency enhancement factor, empirically taken as 1.5. For adaptive cutoff frequency, The gradient magnitude of the reflection component is calculated using the Sobel operator. This is the resolution scaling factor, empirically taken as 20. Let be the numerical stability constant, and take . After filtering the frequency domain logarithmic signal of the reflection component using this filter, an inverse transform is performed to obtain the filtered reflection component. The calculation results are shown below: (25) (26) Finally, the filtered reflection component can effectively recover the texture information of the underexposed areas in the original image, but the overall contrast of the image is low. To improve the image contrast, a histogram stretching contrast enhancement process is performed by truncating extremely dark / bright areas and linearly stretching only the effective grayscale range. On one hand, a probability density function is defined. This involves quantizing the grayscale distribution of the current image to provide a basis for subsequent truncation, and defining the cumulative distribution function. To determine the cumulative probability distribution of grayscale values, the grayscale value corresponding to the truncation threshold is found. The calculation formula is as follows: (27) (28) in grayscale value The number of pixels, This represents the total number of pixels. On the other hand, it determines the effective grayscale range. This cuts off noise and overexposed areas while stretching the effective grayscale range to the full extent. The calculation formula is as follows: (29) (30) (31) in , .

[0035] This embodiment also proposes an inertial navigation error correction method based on SwinIR super-resolution texture enhancement. The principle diagram of the inertial navigation error correction method based on SwinIR super-resolution texture enhancement is shown below. Figure 4 As shown.

[0036] First, assume the input image is Then, through shallow feature extraction, deep feature extraction, and high-resolution image reconstruction using SwinIR, the input image can be... Reconstructed into high-resolution images The calculation formula is as follows: (32) (33) (34) in It is a 3×3 convolutional layer. The extracted shallow features, This is a deep feature extraction module, containing multiple RSTB blocks. For the extracted deep features, This is a high-quality reconstruction module. This step performs super-resolution reconstruction on the original image data, thereby enhancing the texture details of the visual information. Secondly, when the reconstructed image is input into the visual-assisted inertial navigation system (VIS), the errors in the visual component mainly stem from inaccurate feature extraction and matching. Low-resolution and noisy images can lead to errors in feature point extraction and matching. Assuming the true location of the feature points is... The feature extraction error is The observed feature point location and inertial navigation positioning target function As shown in the following formula: (35) (36) in This is due to visual reprojection error. This refers to the IMU pre-integration error. It describes the observed feature point locations when a high-resolution reconstructed image is obtained and input into the vision-assisted inertial navigation system. It will be closer to the true location of the feature point. Extraction error It will be reduced to To find a smaller initial value for the objective function as shown in the following equation, thus accelerating the convergence of the objective function in equation (36) and making it easier to converge to the optimal value: (37) This embodiment proposes an inertial navigation error correction method based on joint optimization of IMU and visual data. By downsampling and reconstructing IMU data using dual quaternion interpolation, and optimizing the illumination and texture quality of visual data, noise and distortion are suppressed from the data source, so as to achieve efficient correction of inertial navigation system errors and significantly improve the positioning accuracy and robustness of UAVs in high dynamic, low illumination or weak texture environments.

[0037] Implementation Method Two: This embodiment is a further illustrative example of the inertial navigation error correction method based on joint optimization of IMU and visual data described in Embodiment 1.

[0038] This implementation method mainly addresses the problem of insufficient navigation accuracy caused by IMU data noise and decreased visual data quality during multimodal perception fusion inertial navigation of UAVs in complex environments such as low light and blurred textures.

[0039] 1) Data and Experimental Setup This specific embodiment is based on the EuRoc UAV dataset open-sourced by a certain college. It verifies the optimization effect of the multimodal perception fusion method on the inertial navigation positioning accuracy in complex environments such as low light and blurred textures under the Ubuntu 18.04 system environment. The experiment selected six sequences covering different flight difficulties (slow / fast movement, well-lit / dark environments), including MH_01, MH_03, MH_05, V1_01, V1_02, and V1_03, for testing. IMU data quality was optimized by reconstructing with dual quaternions, while the SwinIR super-resolution algorithm was used to enhance visual texture details and low-light images were optimized based on Retinex theory, forming a comprehensive scheme for inertial navigation error correction.

[0040] The test platform is equipped with an Intel Core i7-13700KF processor and an NVIDIA GeForce RTX4070 graphics card. The pure inertial navigation calculation is based on the MSF framework, and the visual inertial navigation fusion adopts the ROS-free VINS-Mono architecture to balance the requirements of real-time performance and accuracy.

[0041] Positioning performance is quantitatively assessed using the EVO tool to calculate the absolute trajectory error (ATE), ensuring the industrial rigor of the validation process. The formula for calculating ATE is shown below: (38) in To estimate the trajectory, For the actual trajectory, The Euclidean distance (ATE) is the primary metric for measuring the deviation between the estimated and actual trajectories, measured in meters. It quantifies positioning accuracy by calculating the Euclidean distance between the estimated and actual poses at each time point, and is suitable for evaluating the global positioning accuracy of the entire trajectory. A smaller ATE indicates a smaller trajectory error and higher positioning accuracy.

[0042] 2) Inertial navigation error correction results of IMU data downsampling and dual quaternion interpolation This implementation optimizes IMU data quality through dual quaternion interpolation reconstruction, significantly suppressing the cumulative error of the inertial navigation system. Specifically, the original IMU data with a sampling rate of 100Hz is first downsampled at a 1:2 ratio to filter high-frequency noise. Then, the downsampled data is converted into a dual quaternion model, the motion state is updated using the Euler method, and angular and specific force increments are extracted to reconstruct the optimized IMU data. The optimized data is input into the MSF inertial navigation system without fused visual data, and the absolute trajectory error (ATE) is evaluated on six sequences in the EuRoc dataset: MH_01, MH_03, MH_05, V1_01, V1_02, and V1_03. The IMU acceleration and angular velocity data before and after reconstruction optimization are shown below. Figure 5 As shown in Table 1, the inertial navigation ATE data before and after reconstruction and optimization are as follows: Table 1. Comparison of MSF and average ATE of inertial navigation before and after reconstruction interpolation error correction.

[0043] The optimization effect of the method described in this embodiment is most significant in high-speed scenarios (such as MH_03), proving that the method effectively alleviates the noise accumulation problem in dynamic environments. Visualization results are shown below. Figure 5 The results further validate the noise suppression effect of IMU data reconstruction optimization. The smoothness and continuity of the reconstructed and interpolated data are significantly improved, achieving inertial navigation error correction.

[0044] 3) Inertial navigation error correction results optimized by visual information ① Results of inertial navigation error correction for visual information illumination optimization This embodiment significantly improves the reliability of visual constraints in dark environments through Retinex-driven visual illumination optimization. Specifically, low-illuminance enhancement is performed on six sequences from the EuRoc dataset: MH_01, MH_03, MH_05, V1_01, V1_02, and V1_03. First, the illumination and reflection components of the image are decomposed. The illumination map is optimized using gamma correction, and then the reflection component is enhanced using adaptive homomorphic filtering. The optimized visual data and the original IMU data are input into the VINS-Mono inertial navigation system, which fuses the visual information to generate a trajectory. The improvement in navigation accuracy is evaluated using the absolute trajectory error (ATE). The visual-inertial navigation ATE data before and after enhancement are shown in Table 2. Table 2. Comparison of average ATE of VINS-Mono inertial navigation system before and after visual information illuminance optimization and inertial navigation error correction.

[0045] The key breakthrough of the optimization method proposed in this invention lies in its application in dark environment scenes such as the MH_05 and V1_01 sequences, proving that the method effectively solves the problem of feature loss in low-light images and achieves inertial navigation error correction. Figure 6 As shown, the enhanced image significantly improves the discernibility of textures in dark areas.

[0046] ② Results of inertial navigation error correction for visual information texture enhancement This embodiment enhances visual texture details using the lightweight SwinIR super-resolution model, significantly improving the reliability of feature matching in weakly textured scenes. Specifically, six sequences from the EuRoc dataset undergo bicubic downsampling preprocessing and are input into the SwinIR network for 2x super-resolution reconstruction. The reconstructed high-resolution image and the original IMU data are then input into a VINS-Mono inertial navigation system that fuses visual information to generate a trajectory. The texture optimization effect is evaluated using the absolute trajectory error (ATE). A comparison of image effects before and after super-resolution is provided. Figure 7 As shown in Table 3, the experimental results are as follows: Table 3. Comparison of average ATE of VINS-Mono inertial navigation system before and after inertial navigation error correction for visual information texture enhancement.

[0047] Experimental results show that in static, weakly textured scenes such as the MH_01 sequence, super-resolution reconstruction reduces the system's ATE to a certain extent, indicating that the method proposed in this invention can reduce the system's trajectory error and achieve error correction of the inertial navigation system.

[0048] 4) Test results of UAV inertial navigation error correction method based on IMU and visual data optimization This system improves positioning accuracy in complex scenes by collaboratively optimizing IMU data quality, visual information illumination, and texture details. In practice, the triple-optimized multimodal data is input into the VINS-Mono inertial navigation system, which integrates visual data. The absolute trajectory error (ATE) is evaluated using six sequences from the EuRoc dataset. The experimental results are shown in Table 4. Table 4. Comparison of VINS-Mono inertial navigation average ATE before and after UAV inertial navigation error correction based on IMU and visual data optimization.

[0049] Experimental results show that the method proposed in this invention achieves a 39.3% accuracy improvement in simple scenarios, proving that the synergistic gain of IMU noise suppression and visual quality improvement can help correct inertial navigation system errors.

[0050] Implementation Method 3: An inertial navigation system error correction system based on IMU and visual data optimization, characterized in that the system is implemented based on an inertial navigation system error correction method based on IMU and visual data optimization as described in Embodiment 1, and the system includes: Data acquisition module: used to acquire raw IMU data and visual data; Data processing module: used to downsample and reconstruct the raw IMU data using dual quaternion interpolation; Image optimization module: used to optimize the illumination quality and texture quality of visual data to obtain enhanced reflectance maps; Image reconstruction module: used to perform shallow feature extraction, deep feature extraction, and high-resolution image reconstruction on the enhanced reflectance map to obtain the reconstructed image; Error correction module: It is used to input the reconstructed image into the inertial navigation system, construct the positioning objective function, find the minimum value of the positioning objective function, thereby obtaining the optimal feature extraction position and realizing error correction.

Claims

1. A method for error correction in inertial navigation systems based on IMU and visual data optimization, characterized in that, Includes the following steps: S1: Acquire raw IMU data and visual data; S2: Downsampling and biquaternion interpolation reconstruction of the original IMU data; S3: Optimize the illumination quality and texture quality of the visual data to obtain an enhanced reflectance map; S4: Perform shallow feature extraction, deep feature extraction, and high-resolution image reconstruction on the enhanced reflectance map to obtain the reconstructed image; S5: Input the reconstructed image into the inertial navigation system, construct the localization objective function, find the minimum value of the localization objective function, thereby obtaining the optimal feature extraction position and realizing error correction.

2. The inertial navigation system error correction method based on IMU and visual data optimization according to claim 1, characterized in that, S2 includes: The original IMU data is downsampled and converted into dual quaternions. The Euler method is used to update the state of the dual quaternions, and the angular increment and force increment are extracted to obtain optimized IMU data.

3. The error correction method for inertial navigation systems based on IMU and visual data optimization according to claim 1, characterized in that, S3 includes: S31: Perform illumination component estimation and illumination map enhancement operations on the original image to obtain the optimal illumination map and illumination components; S32: Perform gamma correction on the illumination component to obtain a preliminary reflection map. Perform secondary noise reduction and enhancement on the preliminary reflection map, and then perform histogram stretching and contrast enhancement to obtain an enhanced reflection map.

4. The inertial navigation system error correction method based on IMU and visual data optimization according to claim 3, characterized in that, The formula for estimating the illumination components is: ， ， ， ， in, The objective function to be minimized is used to solve for the illumination map. The location of the pixel. The x-coordinate is the pixel coordinate. The vertical coordinate is the pixel coordinate. To output the illumination map of the image, Let q be the location of the illumination map that needs to be solved at a neighboring pixel. The normalized values ​​of the input original image, This is a regularization parameter that controls the strength of the smoothing term. For pixels The neighborhood, For input image The weighting function controls the pixel and Smoothness intensity between, This is a constraint term for the difference in illumination between adjacent pixels.

5. The inertial navigation system error correction method based on IMU and visual data optimization according to claim 3, characterized in that, A global smoothing filter is used to estimate the illumination components and enhance the illumination map of the original image; a reflectance enhancement module based on adaptive homomorphic filtering is used to perform secondary denoising and enhancement on the preliminary reflectance map.

6. The inertial navigation system error correction method based on IMU and visual data optimization according to claim 1, characterized in that, The S4 algorithm is used to perform shallow feature extraction, deep feature extraction, and high-resolution image reconstruction on the enhanced reflectance map.

7. The inertial navigation system error correction method based on IMU and visual data optimization according to claim 1, characterized in that, The localization objective function is determined based on the visual reprojection error and the IMU pre-integration error.

8. An inertial navigation system error correction system based on IMU and visual data optimization, characterized in that, The system is implemented based on an inertial navigation system error correction method optimized by IMU and visual data as described in any one of claims 1-7, and the system includes: Data acquisition module: used to acquire raw IMU data and visual data; Data processing module: used to downsample and reconstruct the raw IMU data using dual quaternion interpolation; Image optimization module: used to optimize the illumination quality and texture quality of visual data to obtain enhanced reflectance maps; Image reconstruction module: used to perform shallow feature extraction, deep feature extraction, and high-resolution image reconstruction on the enhanced reflectance map to obtain the reconstructed image; Error correction module: It is used to input the reconstructed image into the inertial navigation system, construct the positioning objective function, find the minimum value of the positioning objective function, thereby obtaining the optimal feature extraction position and realizing error correction.

9. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes an inertial navigation system error correction method based on IMU and visual data optimization according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program that executes an inertial navigation system error correction method based on IMU and visual data optimization according to any one of claims 1-7.

Images