Underwater vision-inertia-acoustic odometer positioning method fused with DVL

Abstract

The invention discloses an underwater vision-inertia-acoustic odometer positioning method fused with a DVL, and belongs to the technical field of robot navigation positioning. According to the method, firstly, an underwater image acquired by a camera and carrier motion data acquired by an IMU and a DVL are used as input information to carry out time alignment and space coordinate system conversion; performing feature point extraction on the image to construct a visual re-projection residual error, and dynamically adjusting a weight ratio of visual information based on the number of feature points and the re-projection residual error; calculating IMU data increment between adjacent frames based on pre-integration to construct an inertial residual error; calculating the displacement change between image frames according to the speed information measured by the DVL to construct a DVL residual error; establishing a nonlinear least square optimization function based on Bayesian maximum posteriori estimation; vision, inertia and acoustic speed measurement data are fused at the rear end through a sliding window mechanism, a multi-constraint joint optimization model is constructed, and the robustness and accuracy of a vision-inertia-acoustic odometer system in the underwater environment are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of robot navigation and positioning, and in particular relates to an underwater visual-inertial-acoustic odometer positioning method for an unmanned underwater vehicle. Background Art

[0002] Unmanned underwater vehicles have important applications in marine resource exploration and underwater facility operation and maintenance, but their navigation systems face severe challenges. Due to the rapid attenuation of electromagnetic waves underwater, the global satellite navigation system (GNSS) cannot work effectively. Although traditional baseline positioning methods such as long baseline (LBL) and ultra-short baseline (USBL) can provide accurate positioning, they require underwater calibration facilities to be arranged in advance, which limits the operating range and flexibility of the submersible. Although the inertial navigation system (INS) has the advantages of not relying on external signals and low cost, its positioning error will continue to accumulate over time. Especially in long-range missions, the error will significantly affect the accuracy, limiting its long-range application. In contrast, the vision-based state estimation method, that is, the visual odometry (VO), relies on real-time visual information for self-positioning and has good autonomy and adaptability. However, the complexity of the underwater environment, such as insufficient lighting, sparse textures, and noise interference, seriously affects the working effect of visual sensors. Underwater images usually have problems such as blur and low contrast, which makes feature extraction and matching difficult, which directly affects the positioning accuracy of the visual odometry. Although the problem can be solved to some extent by combining visual sensors with inertial measurement units (IMUs) to form visual inertial odometry (VIO), and the high-frequency, short-term precision motion information provided by IMUs can help suppress positioning errors to some extent, mature visual inertial systems on land still cannot achieve the expected results in complex underwater environments. Currently, few studies have combined visual odometry with commonly used acoustic sensors in underwater environments. Summary of the invention

[0003] Purpose of the invention: In order to overcome the deficiencies in the prior art and to address the impact of harsh underwater environments on visual inertial odometer positioning, the present invention provides a visual-inertial-acoustic odometer system integrating DVL to solve the problem of decreased positioning accuracy of visual inertial odometers in complex underwater environments. The method uses DVL to assist in system initialization, accelerates initial state convergence, restores key parameters such as scale factor and gravity direction, and couples DVL residuals to factor graph optimization in the back-end stage. The weights of the reprojection residuals are dynamically adjusted through an adaptive factor to reduce the impact of low-quality images. It meets the requirements for high-precision and strong robustness positioning of submersibles in complex underwater environments.

[0004] Technical solution: To achieve the above purpose, the present invention adopts the following technical solution: The present invention provides an underwater visual-inertial-acoustic odometer positioning method integrating DVL, the method comprising the following steps: S1: The image information obtained by the monocular camera, the inertial information obtained by the IMU, and the three-dimensional velocity information measured by the DVL are used as input information. Feature extraction and matching are performed on each frame of the image to obtain feature points. At the same time, the inertial information between adjacent frames is pre-integrated by the IMU, and the DVL measurement value is associated with the image frame with the closest time relationship. S2: Based on the feature points obtained in step S1, the initial pose of the camera without scale constraints is restored by the monocular camera initialization method, and the DVL three-dimensional velocity information associated with the image frame time obtained in step S1 is introduced and compared with the initial pose without scale constraints restored by the monocular camera initialization method to obtain a rough estimate of the scale factor and the camera key frame velocity; S3: construct visual reprojection residuals, IMU inertial residuals and DVL velocity residuals respectively according to the feature points obtained in step S1, the IMU pre-integration results and the three-dimensional velocity information measured by the DVL, establish a MAP estimation problem considering the IMU inertial information and the DVL velocity information, maximize the posterior distribution and assume Gaussian noise, regard MAP estimation as an optimization problem, solve the optimization function based on the Levenberg-Marquardt nonlinear least squares optimization method, and restore the accurate system initial pose, velocity and key parameters based on the initial pose without scale constraints and the rough estimate of the camera key frame velocity obtained in step S2, wherein the key parameters include gravity direction and scale; S4: Based on the feature points obtained in step S1 and the visual reprojection residual constructed in step S3, an adaptive dynamic weight factor is designed to work together with the Huber robust kernel function to reduce the impact of image quality on the accuracy of pose estimation; S5: In the local optimization stage, a nonlinear target quadratic function is established based on the factor graph probability model and Bayesian maximum a posteriori inference. The DVL velocity residual is coupled to the local optimization factor graph, and the back-end multi-sensor joint cost function is constructed together with the reprojection residual and inertial residual. The optimal estimation of the attitude and position is obtained by minimizing the cost function.

[0005] Furthermore, the step S1 specifically includes the following steps: S1-1: Use a quadtree-based feature extraction algorithm to extract corner points and calculate feature descriptors. Treat the image as a node, divide the image into four equal parts, and store the feature point information of each part. At the same time, consider removing nodes without feature information. Repeat the segmentation and selection of the best quality feature points, and iterate this process until the number of feature points reaches no less than 100.

[0006] S1-2: The data acquisition frequency of IMU is much higher than that of camera. The change between adjacent image frames is calculated by pre-integration to align it with the image in time: in the inertial coordinate system Keyframe No. Frame and The pose, velocity and position of the frame are expressed as , , and , , , then the first Pose, velocity and position of the frame: , in, represents the exponential map, For the and Keyframes between frames, is the time interval between two key frames, denote the measured values ​​of angular velocity and acceleration respectively, is the bias of the gyroscope and accelerometer, including fixed bias and random walk. represents the noise assumed to conform to Gaussian distribution, Represents the gravity vector in the world coordinate system ,in , is the gravitational constant, Indicates Frame and The time interval between frames, , , Respectively represent The pose, speed, and position of the frame; S1-3: Pre-integration introduces incremental form Avoid reintegrating the measurements and only need to calculate the Frame and Changes in pose, velocity, and position between frames : , Among them, the superscript Represents the transpose of a matrix.

[0007] Furthermore, the goal of the system initialization in step S2 is to estimate the following state variables: : , in, is the scale factor, is the rotation matrix from the gravity coordinate system to the world coordinate system, are the biases of the accelerometer and gyroscope, is the scale-free velocity of each keyframe; When the key frames with IMU pre-integration and DVL measurement data reach 8 frames, the system enters the initialization process. The monocular camera initialization method obtains the initial pose without scale constraints by solving the homography matrix and the basic matrix. ,in The rotation matrix representing the posture, represents the position without scale constraints. The result of the monocular camera initialization method is fixed, assuming that the odometer initialization process has Key frames are involved, and the DVL velocity measurement information is converted to the inertial coordinate system according to the external calibration parameters, and the converted result is used as the initial reference value. At the same time, the velocity information obtained is compared with the initial pose without scale constraints restored by the monocular camera initialization method to obtain a rough estimate of the scale factor. : , in is the rotation matrix from the DVL coordinate system to the camera coordinate system, It is The velocity measurement of the frame DVL, is the translation estimate, Indicates Frame and The time interval between frames, Indicates the number of key frames involved in the initialization process.

[0008] Furthermore, the step S3 specifically includes the following steps: S3-1: After introducing the three-dimensional velocity information measured by DVL, the state variable Re-expressed as: , in, It represents the speed of each key frame measured by DVL. represents a rough estimate of the scale factor; Given the state and measurement values, a MAP estimation problem is established, where the posterior distribution is: , in is the likelihood distribution of the inertial acoustic measurement value under a given state, represents the prior of the state, Indicates The system status of the frame, represents the set of IMU measurements, represents the set of DVL measurements; considering the independence of IMU and DVL measurements, the likelihood problem is decomposed into: , in, Indicates Frame to IMU measurements of the frame, Indicates Frame to DVL measurement of the frame, represents the bias of the IMU, Indicates Frame rate, Indicates Frame rate; To obtain the MAP estimator, maximize its posterior distribution to find the optimal parameters. The formula is: , Assuming a Gaussian noise distribution, MAP estimation is equivalent to an optimization problem: , in, is the optimal solution of the cost function, is a collection of keyframes, and is the prior residual and its covariance matrix, and is the inertial residual and its covariance matrix, and is the DVL residual and its covariance matrix, Indicates Frame to The inertial residual of the frame, Indicates Frame to DVL residual of the frame; S3-2: According to the IMU pre-integration theory, the inertial residual is expressed as: , in, Indicates The rotation matrix of the frame, represents the scale factor, Indicates Frame to The inertial residual of the frame, , , Respectively represent Frame to The frame's attitude residual, velocity residual, and position residual, Indicates Frame to The transpose of the frame's pose changes, , Respectively represent Frame and The scale-free speed of the frame, Indicates Frame to The frame time interval, Indicates Frame to The speed of the frame changes, , Respectively represent Frame and The scale-free position of the frame, Indicates Frame to Frame position changes; S3-3: The DVL residual is derived as follows: First, calculate the camera from the DVL coordinate system Frame to Frame displacement: , The camera in the DVL coordinate system is Frame to The displacement of the frame, Represents the rotation matrix from the camera coordinate system to the DVL coordinate system, Indicates the camera coordinate system Frame to The rotation matrix of the frame, Represents the translation vector from the DVL coordinate system to the camera coordinate system, Indicates the camera coordinate system Frame to The frame's scale-free translation vector; Next, calculate the Frame to DVL residual of the frame : , in Indicates the DVL coordinate system Frame rate; Indicates Frame to The time interval between frames; S3-4: Use the Levenberg-Marquardt method to solve the initialization cost function and obtain accurate initial parameters.

[0009] Furthermore, the step S4 specifically includes the following steps: S4-1: The complex underwater environment leads to a sparse number of feature points and an increase in mismatches. When the image quality deteriorates, the number of feature points decreases and the reprojection error of each feature point increases. Based on this factor, an evaluation factor is designed to reflect the changing trend of underwater image quality, i.e., camera reliability. : , in and They are The set and total number of map points observed in the frame. Indicates taking the latest past of the current frame The average number of map points in a frame, Usually set to 5, Indicates When the underwater image quality deteriorates, the reprojection error of the current image frame increases relative to the past frames and the number of successfully tracked feature points decreases. Showing an increasing trend.

[0010] S4-2: Evaluation factors for camera reliability trend After that, the adjustment weight of the visual sensor is further calculated: , This formula allows the system to dynamically adjust the weights in the optimization according to the degree of sensor degradation. The Huber kernel function suppresses outliers at a single feature point. The weight of all visual information is adjusted during the optimization process.

[0011] Furthermore, in the local optimization stage described in step S5, the state variables to be estimated are: , in Indicates the first The IMU state of each key frame, Respectively represent The attitude, position and velocity of the keyframe IMU, Respectively represent The accelerometer and gyroscope biases for each keyframe, is the position of the 3D map point; The cost function of the visual-inertial acoustic optimization problem can be summarized as: , in, represents the optimal solution of the cost function, is the adaptive dynamic weight factor, is the visual reprojection residual, is the inertia residual, is the DVL residual. Expanding it into a joint cost function yields: , in, is a collection of keyframes, Represents a collection of map points, is the visual reprojection residual, represents the Huber kernel function, and They represent the IMU inertial residual and DVL residual respectively. The residuals of the three sensors are tightly coupled through Maximizing the benefits of each sensor increases the positioning accuracy and robustness of the system.

[0012] Beneficial effects: Compared with the prior art, the present invention innovatively proposes an underwater visual inertial acoustic odometer positioning method integrated with DVL. This method is based on visual odometer technology, combines IMU and DVL for multi-sensor fusion, uses Bayesian maximum a posteriori probability estimation to model the system state estimation problem, optimizes the cost function under the factor graph framework, and designs an adaptive weight factor to evaluate the reliability of sensors in harsh environments, tightly coupling multi-source information. The underwater visual inertial acoustic odometer system will be able to achieve higher positioning accuracy and robustness in more complex underwater environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 is an overall block diagram of the present invention; Figure 2 This is a visualization comparison of the experimental results. Figure 2 In the figure, (a) is a trajectory comparison diagram of the method of the present invention and the ORB-SLAM3 method for underwater data sequence 1, (b) is a graph showing the RMSE variation over time of the method of the present invention and the ORB-SLAM3 method for underwater data sequence 1, (c) is a trajectory comparison diagram of the method of the present invention and the ORB-SLAM3 method for underwater data sequence 2, and (d) is a trajectory comparison diagram of the method of the present invention and the ORB-SLAM3 method for underwater data sequence 3; Figure 3 These are some actual underwater scene images taken during the experiment. Figure 3 In (a), part of the field of view is too far away and unclear, in (b), there are a lot of dynamic spots and shadows, in (c), there are fewer textures for feature extraction, and in (d), there is a cliff-like change in the terrain. These factors will have an adverse effect on the system positioning. DETAILED DESCRIPTION

[0014] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments.

[0015] The present invention provides an underwater visual inertial acoustic odometer positioning method integrating DVL. Figure 1 As shown, specifically including: S1: The image information obtained by the monocular camera, the inertial information obtained by the IMU, and the three-dimensional velocity information measured by the DVL are used as input information. Feature extraction and matching are performed on each frame of the image to obtain feature points. At the same time, the inertial information between adjacent frames is pre-integrated by the IMU, and the DVL measurement value is associated with the image frame with the closest time relationship. S2: Based on the feature points obtained in step S1, the initial pose of the camera without scale constraints is restored by the monocular camera initialization method, and the DVL three-dimensional velocity information associated with the image frame time obtained in step S1 is introduced and compared with the initial pose without scale constraints restored by the monocular camera initialization method to obtain a rough estimate of the scale factor and the camera key frame velocity; S3: construct visual reprojection residuals, IMU inertial residuals and DVL velocity residuals respectively according to the feature points obtained in step S1, the IMU pre-integration results and the three-dimensional velocity information measured by the DVL, establish a MAP estimation problem considering the IMU inertial information and the DVL velocity information, maximize the posterior distribution and assume Gaussian noise, regard MAP estimation as an optimization problem, solve the optimization function based on the Levenberg-Marquardt nonlinear least squares optimization method, and restore the accurate system initial pose, velocity and key parameters based on the initial pose without scale constraints and the rough estimate of the camera key frame velocity obtained in step S2, wherein the key parameters include gravity direction and scale; S4: Based on the feature points obtained in step S1 and the visual reprojection residual constructed in step S3, an adaptive dynamic weight factor is designed to work together with the Huber robust kernel function to reduce the impact of image quality on the accuracy of pose estimation; S5: In the local optimization stage, a nonlinear target quadratic function is established based on the factor graph probability model and Bayesian maximum a posteriori inference. The DVL velocity residual is coupled to the local optimization factor graph, and the back-end multi-sensor joint cost function is constructed together with the reprojection residual and inertial residual. The optimal estimation of the attitude and position is obtained by minimizing the cost function.

[0016] Preferably, in step S1, FAST corner point extraction, BRIEF descriptor calculation, pre-integration, time alignment and spatial coordinate system conversion are performed on the image data collected by the camera, the acceleration and angular velocity data obtained by the IMU, and the velocity data measured by the DVL respectively: S1-1: Use a quadtree-based feature extraction algorithm to extract corner points and calculate feature descriptors. Treat the image as a node, divide the image into four equal parts, and store the feature point information of each part. At the same time, consider removing nodes without feature information. Repeat the segmentation and select the best quality feature points, and iterate this process until there are no less than 100 feature points.

[0017] S1-2: The data acquisition frequency of IMU is much higher than that of camera. The change between adjacent image frames is calculated by pre-integration to align it with the image in time: in the inertial coordinate system Keyframe No. Frame and The pose, velocity and position of the frame are expressed as , , and , , , then the first Pose, velocity and position of the frame: , in, represents the exponential map, For the and Keyframes between frames, is the time interval between two key frames, denote the measured values ​​of angular velocity and acceleration respectively, is the bias of the gyroscope and accelerometer, including fixed bias and random walk. represents the noise assumed to conform to Gaussian distribution, Represents the gravity vector in the world coordinate system ,in , is the gravitational constant, Indicates Frame and The time interval between frames, , , Respectively represent The pose, velocity, and position of the frame.

[0018] S1-3: Pre-integration introduces incremental form Avoid reintegrating the measurements and only need to calculate the Frame and Changes in pose, velocity, and position between frames : , Among them, the superscript Represents the transpose of a matrix.

[0019] Preferably, step S2 performs DVL-assisted visual-acoustic initialization to roughly estimate some initialization parameters.

[0020] The goal of system initialization is to estimate the following state variables : , in, is the scale factor, is the rotation matrix from the gravity coordinate system to the world coordinate system, are the biases of the accelerometer and gyroscope, is the scale-free velocity of each keyframe.

[0021] When the key frames with IMU pre-integration and DVL measurement data reach 8 frames, the system enters the initialization process. The monocular camera initialization method obtains the initial pose without scale constraints by solving the homography matrix and the basic matrix. ,in The rotation matrix representing the posture, represents the position without scale constraints. The result of the monocular camera initialization method is fixed, assuming that the odometer initialization process has The DVL velocity measurement information is converted to the inertial coordinate system according to the external calibration parameters and used as the initial reference value. The velocity information obtained is compared with the initial pose without scale constraint restored by the monocular camera initialization method to obtain a rough estimate of the scale factor. : , in is the rotation matrix from the DVL coordinate system to the camera coordinate system, It is The velocity measurement of the frame DVL, is the translation estimate, Indicates Frame and The time interval between frames, Indicates the number of key frames involved in the initialization process.

[0022] Preferably, in step S3, the inertia residual and the DVL residual are considered, and an objective function is established and solved in a factor graph framework to obtain the optimal value of the initial state parameter: S3-1: After introducing the three-dimensional velocity information measured by DVL, the state variable Re-expressed as: , in, It represents the speed of each key frame measured by DVL. represents a rough estimate of the scale factor; Given the state and measurement values, a MAP estimation problem is established, where the posterior distribution is: , in is the likelihood distribution of the inertial acoustic measurement value under a given state, represents the prior of the state, Indicates The system status of the frame, represents the set of IMU measurements, represents the set of DVL measurements; considering the independence of IMU and DVL measurements, the likelihood problem is decomposed into: , in, Indicates Frame to IMU measurements of the frame, Indicates Frame to DVL measurement of the frame, represents the bias of the IMU, Indicates Frame rate, Indicates Frame rate; To obtain the MAP estimator, maximize its posterior distribution to find the optimal parameters. The formula is: , Assuming a Gaussian noise distribution, MAP estimation is equivalent to an optimization problem: , in, is the optimal solution of the cost function, is a collection of keyframes, and is the prior residual and its covariance matrix, and is the inertial residual and its covariance matrix, and is the DVL residual and its covariance matrix, Indicates Frame to The inertial residual of the frame, Indicates Frame to DVL residual of the frame; S3-2: According to the IMU pre-integration theory, the inertial residual is expressed as: , in, Indicates The rotation matrix of the frame, represents the scale factor, Indicates Frame to The inertial residual of the frame, , , Respectively represent Frame to The frame's attitude residual, velocity residual, and position residual, Indicates Frame to The transpose of the frame's pose changes, , Respectively represent Frame and The scale-free speed of the frame, Indicates Frame to The frame time interval, Indicates Frame to The speed of the frame changes, , Respectively represent Frame and The scale-free position of the frame, Indicates Frame to Frame position changes; S3-3: The DVL residual reflects the accuracy of the velocity estimate and is derived as follows: First, calculate the camera from the DVL coordinate system Frame to Frame displacement: , The camera in the DVL coordinate system is Frame to The displacement of the frame, Represents the rotation matrix from the camera coordinate system to the DVL coordinate system, Indicates the camera coordinate system Frame to The rotation matrix of the frame, Represents the translation vector from the DVL coordinate system to the camera coordinate system, Indicates the camera coordinate system Frame to The frame's scale-free translation vector; Next, calculate the Frame to DVL residual of the frame : , in Indicates the DVL coordinate system Frame rate; Indicates Frame to The time interval between frames; S3-4: Use the Levenberg-Marquardt method to solve the initialization cost function, perform linear approximation on the estimated parameter vector, ignore the derivative terms above the second order, transform the problem into a linear least squares problem, and obtain accurate initial parameters.

[0023] Preferably, the step S4 establishes an adaptive factor reflecting the influence of underwater image quality on odometer positioning accuracy, allowing it to dynamically adjust the weight of visual information in the optimization process: S4-1: Due to the complex underwater environment, such as low illumination and contrast, sparse texture, etc., the number of feature points is small and mismatches increase. When the image quality deteriorates, the number of feature points decreases and the reprojection error of each feature point increases. Based on this factor, an evaluation factor is designed to reflect the trend of underwater image quality, i.e., camera reliability. : , in and They are The set and total number of map points observed in the frame. Indicates taking the latest past of the current frame The average number of map points in a frame, Usually set to 5, Indicates When the underwater image quality deteriorates, the reprojection error of the current image frame increases relative to the past frames and the number of successfully tracked feature points decreases. Showing an increasing trend.

[0024] S4-2: After evaluating the reliability of the visual sensor, the adjustment weight of the visual sensor is obtained: , This formula allows the system to dynamically adjust the weights in the optimization according to the degree of sensor degradation. The Huber kernel function suppresses outliers at a single feature point. The weight of all visual information is adjusted during the optimization process, through the dynamic weight factor The synergistic effect with the Huber kernel can suppress the reduction of positioning errors when image quality degradation occurs.

[0025] Preferably, the step S5 establishes a nonlinear target quadratic function based on the factor graph probability model and the Bayesian maximum a posteriori probability estimation, couples the information from the three sensors in the factor graph framework, and minimizes the joint cost function. Thus, the optimal estimation of the system posture state is completed. Step S5 specifically includes: In the local optimization phase, the state variables to be estimated for: , in Indicates the first The IMU state of each key frame, Respectively represent The attitude, position and velocity of the keyframe IMU, Respectively represent The accelerometer and gyroscope biases for each keyframe, is the position of the 3D map point; The cost function of the visual-inertial acoustic optimization problem can be summarized as: , in, represents the optimal solution of the cost function, is the adaptive dynamic weight factor, is the visual reprojection residual, is the inertia residual, is the DVL residual; Expand the optimal solution of the cost function into a joint cost function: , in, is a collection of keyframes, Represents a collection of map points, is the visual reprojection residual, represents the Huber kernel function, and Represents the IMU inertial residual and DVL residual. The residuals among the three sensors are tightly coupled. Maximizing the benefits of each sensor increases the positioning accuracy and robustness of the system.

[0026] In order to verify the underwater visual-inertial-acoustic odometer positioning method integrated with DVL proposed in this invention, an actual navigation test of an underwater robot was carried out. The test equipment includes: underwater robot, underwater camera, DVL, IMU, NVIDIA JetsonOrin Nano development board, GNSS-RTK, laptop, DC power supply, etc. Three trajectories of different lengths and paths were planned in the test. The positioning accuracy evaluation results of the compared traditional visual-inertial SLAM method, namely the ORB-SLAM3 method and the method of this invention are shown in Table 1: Table 1. Comparison results of positioning accuracy of three sequences between the ORB-SLAM3 method and the method of the present invention in actual underwater experiments:

[0027] The test result trajectory is visualized as follows Figure 2 As shown in Table 1 and Figure 2 It can be seen that the positioning accuracy estimation performance of the present invention is significantly improved compared with the ORB-SLAM3 method in complex underwater environments.

[0028] Figure 3 Some underwater images taken by the actual underwater experimental camera are shown. It can be seen that the low visibility in some areas will cause difficulties in feature extraction. This method can effectively suppress the adverse effects of low-quality underwater images by fusing the speed information obtained by the acoustic sensor DVL.

Claims

1. An underwater visual-inertial-acoustic odometer positioning method integrating DVL, characterized in that: The method comprises the following steps: S1: The image information obtained by the monocular camera, the inertial information obtained by the IMU, and the three-dimensional velocity information measured by the DVL are used as input information. Feature extraction and matching are performed on each frame of the image to obtain feature points. At the same time, the inertial information between adjacent frames is pre-integrated by the IMU, and the DVL measurement value is associated with the image frame with the closest time relationship. S2: Based on the feature points obtained in step S1, the initial pose of the camera without scale constraints is restored by the monocular camera initialization method, and the DVL three-dimensional velocity information associated with the image frame time obtained in step S1 is introduced and compared with the initial pose without scale constraints restored by the monocular camera initialization method to obtain a rough estimate of the scale factor and the camera key frame velocity; S3: construct visual reprojection residuals, IMU inertial residuals and DVL velocity residuals respectively according to the feature points obtained in step S1, the IMU pre-integration results and the three-dimensional velocity information measured by the DVL, establish a MAP estimation problem considering the IMU inertial information and the DVL velocity information, maximize the posterior distribution and assume Gaussian noise, regard MAP estimation as an optimization problem, solve the optimization function based on the Levenberg-Marquardt nonlinear least squares optimization method, and restore the accurate system initial pose, velocity and key parameters based on the initial pose without scale constraints and the rough estimate of the camera key frame velocity obtained in step S2, wherein the key parameters include gravity direction and scale; S4: Based on the feature points obtained in step S1 and the visual reprojection residual constructed in step S3, an adaptive dynamic weight factor is designed to work together with the Huber robust kernel function to reduce the impact of image quality on the accuracy of pose estimation; S5: In the local optimization stage, a nonlinear target quadratic function is established based on the factor graph probability model and Bayesian maximum a posteriori inference. The DVL velocity residual is coupled to the local optimization factor graph, and the back-end multi-sensor joint cost function is constructed together with the reprojection residual and inertial residual. The optimal estimation of the attitude and position is obtained by minimizing the cost function.

2. The underwater visual-inertial-acoustic odometer positioning method integrating DVL according to claim 1 is characterized in that: Step S1 specifically adopts the following sub-steps: S1-1: A quadtree-based feature extraction algorithm is used to extract corner points and calculate feature descriptors. The image is regarded as a node, and the feature point information of each part is stored by dividing it into four equal parts. At the same time, nodes without feature information are removed. The segmentation and selection of the best quality feature points are repeated, and this process is iterated continuously until no less than 100 feature points are reached. S1-2: The data acquisition frequency of IMU is much higher than that of camera. The change between adjacent image frames is calculated by pre-integration to align it with the image in time: in the inertial coordinate system Keyframe No. Frame and The pose, velocity and position of the frame are expressed as , , and , , , then the first Pose, velocity and position of the frame: ， in, represents the exponential map, For the and Keyframes between frames, is the time interval between two key frames, denote the measured values ​​of angular velocity and acceleration respectively, is the bias of the gyroscope and accelerometer, including fixed bias and random walk. represents the noise assumed to conform to Gaussian distribution, Represents the gravity vector in the world coordinate system ,in , is the gravitational constant, Indicates Frame and The time interval between frames, , , Respectively represent The pose, speed, and position of the frame; S1-3: Pre-integration introduces incremental form Avoid reintegrating the measurements and only need to calculate the Frame and Changes in pose, velocity, and position between frames : ， Among them, the superscript Represents the transpose of a matrix.

3. The underwater visual-inertial-acoustic odometer positioning method integrating DVL according to claim 2 is characterized in that: The specific method of step S2 is as follows: The goal of system initialization is to estimate the following state variables : ， in, is the scale factor, is the rotation matrix from the gravity coordinate system to the world coordinate system, are the biases of the accelerometer and gyroscope, is the scale-free velocity of each keyframe; When the key frames with IMU pre-integration and DVL measurement data reach 8 frames, the system enters the initialization process. The monocular camera initialization method obtains the initial pose without scale constraints by solving the homography matrix and the basic matrix. ,in The rotation matrix representing the posture, represents the position without scale constraints; fix the result of the monocular camera initialization method, and assume that the odometer initialization process has The DVL velocity measurement information is converted to the inertial coordinate system according to the external calibration parameters, and the converted result is used as the initial reference value. At the same time, the velocity information is compared with the initial pose without scale constraint restored by the monocular camera initialization method to obtain a rough estimate of the scale factor. : ， in is the rotation matrix from the DVL coordinate system to the camera coordinate system, It is The velocity measurement of the frame DVL, is the translation estimate, Indicates Frame and The time interval between frames, Indicates the number of key frames involved in the initialization process.

4. The visual-inertial-acoustic odometer positioning method integrating DVL according to claim 3 is characterized in that: The specific method of step S3 is as follows: S3-1: After introducing the three-dimensional velocity information measured by DVL, the state variable Re-expressed as: ， in, It represents the speed of each key frame measured by DVL. represents a rough estimate of the scale factor; Given the state and measurement values, a MAP estimation problem is established, where the posterior distribution is: ， in is the likelihood distribution of the inertial acoustic measurement value under a given state, represents the prior of the state, Indicates The system status of the frame, represents the set of IMU measurements, represents the set of DVL measurements; considering the independence of IMU and DVL measurements, the likelihood problem is decomposed into: ， in, Indicates Frame to IMU measurements of the frame, Indicates Frame to DVL measurement of the frame, represents the bias of the IMU, Indicates Frame rate, Indicates Frame rate; To obtain the MAP estimator, maximize its posterior distribution to find the optimal parameters. The formula is: ， Assuming a Gaussian noise distribution, MAP estimation is equivalent to an optimization problem: ， in, is the optimal solution of the cost function, is a collection of keyframes, and is the prior residual and its covariance matrix, and is the inertial residual and its covariance matrix, and is the DVL residual and its covariance matrix, Indicates Frame to The inertial residual of the frame, Indicates Frame to DVL residual of the frame; S3-2: According to the IMU pre-integration theory, the inertial residual is expressed as: ， in, Indicates The rotation matrix of the frame, represents the scale factor, Indicates Frame to The inertial residual of the frame, , , Respectively represent Frame to The frame's attitude residual, velocity residual, and position residual, Indicates Frame to The transpose of the frame's pose changes, , Respectively represent Frame and The scale-free speed of the frame, Indicates Frame to The frame time interval, Indicates Frame to The speed of the frame changes, , Respectively represent Frame and The scale-free position of the frame, Indicates Frame to Frame position changes; S3-3: The DVL residual is derived as follows: First, calculate the camera from the DVL coordinate system Frame to Frame displacement: ， The camera in the DVL coordinate system is Frame to The displacement of the frame, Represents the rotation matrix from the camera coordinate system to the DVL coordinate system, Indicates the camera coordinate system Frame to The rotation matrix of the frame, Represents the translation vector from the DVL coordinate system to the camera coordinate system, Indicates the camera coordinate system Frame to The frame's scale-free translation vector; Next, calculate the Frame to DVL residual of the frame : ， in Indicates the DVL coordinate system Frame rate; Indicates Frame to The frame interval; S3-4: Use the Levenberg-Marquardt method to solve the initialization cost function and obtain accurate initial parameters.

5. The underwater visual-inertial-acoustic odometer positioning method integrating DVL according to claim 4 is characterized in that The specific method of step S4 is as follows: S4-1: Design an evaluation factor that reflects the changing trend of underwater image quality, i.e. camera reliability : ， in and They are The set and total number of map points observed in the frame, Indicates taking the latest past of the current frame The average number of map points in a frame, Indicates When the underwater image quality deteriorates, the reprojection error of the current image frame increases and the number of feature points decreases relative to the past frames. Showing an increasing trend; S4-2: Evaluation factors for camera reliability trend After that, the adjustment weight of the visual sensor is further calculated. : ， The Huber kernel function suppresses outliers of a single feature point. The weight of all visual information is adjusted during the optimization process.

6. The underwater visual-inertial-acoustic odometer positioning method integrating DVL according to claim 5 is characterized in that: The specific method of step S5 is as follows: In the local optimization phase, the state variables to be estimated for: ， in Indicates the first The IMU state of each key frame, Respectively represent The attitude, position and velocity of the keyframe IMU, Respectively represent The accelerometer and gyroscope biases for each keyframe, is the position of the 3D map point; The cost function of the visual-inertial acoustic optimization problem can be summarized as: ， in, represents the optimal solution of the cost function, is the adaptive dynamic weight factor, is the visual reprojection residual, is the inertia residual, is the DVL residual; Expand the optimal solution of the cost function into a joint cost function: ， in, is a collection of keyframes, Represents a collection of map points, is the visual reprojection residual, represents the Huber kernel function, and They represent IMU inertial residual and DVL residual respectively.

Images