A Multi-Sensor Elastic Fusion Positioning Method and Apparatus for Scene Awareness in Underwater Docking

Abstract

This invention discloses a scene-aware multi-sensor elastic fusion positioning method for underwater docking, comprising: a scene perception probability acquisition step: acquiring the scene perception probability of each visual channel for the cursor in the current scene according to the scene perception model, as the prior prediction probability of the positioning filter; a state interaction initialization step: obtaining the state vector and covariance matrix of each positioning filter through the prior prediction probability and fusion weights; a positioning filter activation number update and state transition step: dynamically adjusting the activation state of the positioning filter according to the scene perception probability of each visual channel for the cursor, and constructing the state transition probability matrix between the positioning filters; and an estimation fusion step: updating the state vector and covariance matrix according to the dynamic changes of the current scene and the real-time operating state of each active positioning filter. This invention uses a smooth, gradual soft switching mechanism to replace rigid hard switching, thereby eliminating positioning jitter.

Description

Technical Field

[0001] This invention relates to the field of AUV navigation and positioning technology, and in particular to a scene-aware multi-sensor elastic fusion positioning method and apparatus for underwater docking. Background Technology

[0002] Autonomous underwater vehicles (AUVs), as untethered, autonomously navigated underwater probes, have broad application prospects in marine scientific research, military, and commercial fields. However, the limited energy carried by AUVs prevents them from conducting long-duration, large-scale underwater patrols. To solve this problem, it is necessary to establish fixed docking stations underwater to enable AUVs to automatically guide themselves back to the dock for energy replenishment and information transmission.

[0003] Dock-going guidance technology is crucial for the successful docking of AUVs with underwater docking stations. Currently, the mainstream dock-going guidance methods include acoustic guidance, optical guidance, and electromagnetic guidance. Due to the significant limitations of single sensors, the use of multi-sensor information fusion for combined guidance has become an inevitable trend in AUV dock-going technology. Existing multi-sensor fusion solutions have addressed the AUV dock-going guidance problem to varying degrees, with the most representative solutions including:

[0004] Option 2: Chinese patent CN113034590A discloses a dynamic docking and positioning method for AUVs based on visual fusion. Its core is to first calculate the precise dimensions of the target marker in the underwater environment using a binocular vision system to eliminate systematic errors introduced by shore-based measurements. During the subsequent docking process, the system decomposes the binocular vision into two independent monocular vision channels, calculating two sets of independent positioning results using the previously calibrated dimensions. Finally, these two sets of results are weighted and fused to obtain the final positioning information. A key feature of this option is that the weighted fusion calculation method is determined based on the different distance stages of the AUV, employing a segmented weighting strategy based on a fixed distance threshold.

[0005] Option 2: Chinese patent CN116699985A discloses a multi-sensor combined docking guidance method for AUVs suitable for high-latitude polar regions. At long distances, USBL (USBL-based navigation) guides the AUV back to base. When the AUV enters close range, the system processes two data streams in parallel: one uses a trained BP neural network to fit the AUV's azimuth and distance using magnetic field data obtained from a magnetometer; the other fuses machine vision and USBL data through "complementary filtering based on visual confidence." This option sets a switching condition: if the visual system cannot recognize the cursor, the controller will only use the electromagnetic information data fitted by the neural network; if the visual system can recognize it, it will finally fuse the two data streams (neural network fitted value and acoustic-optical fusion value) through adaptive Kalman filtering.

[0006] While the aforementioned existing technologies improve the reliability of AUV docking to some extent through multi-sensor fusion, they still have significant shortcomings in the design of the fusion mechanism, especially in the face of complex and variable underwater environments, where their robustness and smoothness are insufficient. Existing technologies generally rely on a "hard switching" mechanism when fusing different sensors or algorithms, that is, changing the system's fusion logic based on a fixed, either-or threshold.

[0007] In Scheme 1 (CN113034590A), the system relies entirely on fixed distance thresholds to switch weight calculation methods. When the AUV hovers near these threshold boundaries, the weight calculation logic frequently abruptly changes, easily leading to "jumps" and "oscillations" in the final output positioning results, affecting system stability. In Scheme 2 (CN116699985A), the system uses a binary logic switching, where changes to the system fusion strategy depend entirely on a single binary judgment: "whether the vision can recognize the cursor." When the cursor signal is at the edge of the recognition threshold due to water flow disturbances or occlusion, this binary switching also causes frequent starts and stops of the fusion framework, leading to instability and oscillations in the positioning results. These "hard switching" mechanisms lack smooth transitions and cannot achieve gradual "soft switching." Summary of the Invention

[0008] This invention provides a scene-aware multi-sensor elastic fusion positioning method and device for underwater docking. It uses a smooth and gradual "soft switching" mechanism to replace the rigid "hard switching" in the prior art, thereby eliminating positioning jitter and ensuring the smoothness of the AUV's docking trajectory and the stability of the control system.

[0009] The technical solution of the present invention is as follows:

[0010] A scene-aware multi-sensor elastic fusion localization method for underwater docking utilizes a Kalman filter framework for localization in both the far-end return phase and the end-of-course docking phase, while adding multi-channel visual localization in the end-of-course docking phase. The method includes the following steps:

[0011] Scene perception probability acquisition steps: Based on the scene perception model, obtain the scene perception probability of the cursor for each visual channel in the current scene, and use them as the prior prediction probability of the localization filter corresponding to each visual channel.

[0012] State interaction initialization steps: By integrating the prior prediction probabilities and fusion weights of each positioning filter, the state vector and covariance matrix of each positioning filter are obtained.

[0013] The steps for updating the number of activated positioning filters and the state transition are as follows: The activation state of the positioning filters is dynamically adjusted according to the scene perception probability of the cursor by each visual channel, and the state transition probability matrix between each positioning filter is constructed.

[0014] Estimation fusion steps: Update the state vector and covariance matrix based on the dynamic changes of the current scene and the real-time running state of each active localization filter.

[0015] This invention integrates the cursor scene perception probability output by the scene perception model and the real-time information likelihood of each positioning filter, and dynamically adjusts the activation state and fusion weight of the positioning filter, thereby replacing the traditional "hard switching" mechanism based on a fixed threshold and realizing a smooth "soft switching" of the positioning strategy.

[0016] Preferably, the multi-channel visual positioning has RGB three visual channels.

[0017] Preferably, the positioning filters include positioning filters for ultra-short baseline positioning during the long-range return phase and R-channel, G-channel, and B-channel positioning filters for the terminal docking phase. All positioning filters use dead reckoning from a Doppler log (DVL) and electronic compass as the core of the navigation calculation.

[0018] Further preferably, the state vector components of each positioning filter include the misalignment angle, position error, installation angle error with the compass, and Doppler speedometer scaling factor error.

[0019] Preferably, the timing update of each positioning filter follows an independent computation mechanism, including state prediction, covariance prediction, Kalman gain calculation, state vector update, and covariance update.

[0020] Preferably, in the scene perception probability acquisition step, the method for constructing the scene perception model includes: training a support vector machine (SVM) regression model using a training set; the training set contains a total of The image consists of four underwater light spot images, each containing three RGB channels and a grayscale channel. These four channels are independent of each other and correspond to binary classification labels. After training, a scene perception model for the corresponding channel is formed.

[0021] Regarding the first Frame image channels The multidimensional feature vector is ,in The corresponding cursor existence detection label is , The cursor is missing. Assuming the cursor exists; construct the objective function for the support vector machine regression model:

[0022] ;

[0023] ;

[0024] in, These are the Lagrange multiplier vectors that need optimization. It is a vector whose elements are all 1; The size is A positive semi-definite matrix whose elements ; kernel matrix , It is a nonlinear mapping function. This refers to the kernel width parameter; The regularization coefficient is used.

[0025] After solving this optimization problem, for the image channels ,according to Multidimensional feature vectors right tags Solution:

[0026] ;

[0027] in, This indicates the number of support vectors. It is a symbolic function;

[0028] The Platt scaling method is used to construct a sigmoid probability transformation function, which is used to transform the label... Convert to scene-aware probability :

[0029] ;

[0030] ;

[0031] Among them, parameters , Determined by fitting the training samples. This is a bias term.

[0032] Preferably, in the state interaction initialization step, the reconstructed state vector of a single localization filter and covariance matrix They are respectively:

[0033] ;

[0034] ;

[0035] Among them, the positioning filter Initial weight coefficients It depends on the prior prediction probability of the scene. With fusion weight estimation ; and represents the posterior state estimate and posterior covariance matrix at time k-1, respectively; n is the number of currently active localization filters.

[0036] Preferably, in the step of updating the number of activations of the positioning filter and the state transition, when the first... i Probability of cursor scene perception in each visual channel Greater than the preset threshold At that time, adjust the first i The localization filter corresponding to each visual channel is in an active state.

[0037] Preferably, in the steps of updating the number of activations of localization filters and state transition, the state transition probability matrix between each localization filter is... The construction rules are as follows:

[0038] ;

[0039] ;

[0040] ;

[0041] ;

[0042] ;

[0043] ;

[0044] in, for 3D matrix This represents the total number of positioning filters; express No. Row vectors are defined as localization filters. To the localization filter respectively The transition probability is then used to define the far-end return positioning filter as the first positioning filter, and the visual channel filters correspond to the others. One positioning filter; The probability of scene perception at time k is represented by the corresponding visual channel. This is the discount factor.

[0045] Preferably, the estimation fusion step includes:

[0046] Calculate the innovation likelihood of each localization filter separately. :

[0047] ;

[0048] Among them, the new information sequence , For the first i The observation values ​​of each localization filter at time k. Let k be the prior state estimate at time k. For measuring the transfer matrix; covariance of the innovation sequence. ,in Let k represent the observation noise covariance matrix at time k. Represent the prior covariance matrix;

[0049] ;

[0050] ;

[0051] in, For the first i The posterior probability of each localization filter; For the first i The prior prediction probability of each localization filter;

[0052] The posterior state estimates of each localization filter The weighted combination calculation yields the final fusion state estimate. :

[0053] ;

[0054] The covariance matrix of each positioning filter Calculate the temporal fusion covariance matrix :

[0055] .

[0056] This invention also provides a scene-aware multi-sensor elastic fusion positioning device for underwater docking, which uses a Kalman filter framework for positioning during both the remote return phase and the end-of-course docking phase, and employs multi-channel visual positioning during the end-of-course docking phase; it includes the following units:

[0057] Scene perception probability acquisition unit: Based on the scene perception model, it acquires the scene perception probability of the cursor for each visual channel in the current scene, and uses it as the prior prediction probability of the localization filter corresponding to each visual channel.

[0058] State interaction initialization unit: By integrating the prior prediction probabilities and fusion weights of each positioning filter, the state vector and covariance matrix of each positioning filter are obtained.

[0059] The localization filter activation quantity update and state transition unit dynamically adjusts the activation state of the localization filters based on the scene perception probability of the cursor by each visual channel, and constructs the state transition probability matrix between each localization filter.

[0060] Estimation fusion unit: Updates the state vector and covariance matrix based on the dynamic changes of the current scene and the real-time operating status of each active localization filter.

[0061] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0062] (1) Eliminate positioning jumps and oscillations to achieve smooth soft switching:

[0063] Existing technologies employ a fixed-distance hard handover mechanism. Near the handover point, remote positioning data is significantly suppressed or completely discarded, easily leading to positioning jumps and oscillations in the positioning results. This invention dynamically adjusts the weight coefficients of the remote walk-back positioning filter and the end-point multi-channel visual positioning filter through scene perception probability, achieving a smooth transition and soft handover in the positioning solution, thereby effectively eliminating the drawbacks of traditional hard handover.

[0064] (2) Significantly enhances fault tolerance to underwater interference and locator failure:

[0065] Existing methods relying solely on visual positioning lack effective data fusion mechanisms to achieve fault tolerance in the event of visual positioning failure or performance degradation. This invention introduces innovation and its covariance to characterize the working state of each positioning filter in real time, enabling the positioning system to dynamically adjust weights based on the real-time performance of the filters (rather than solely relying on external distance). This allows the system to maintain fault tolerance for positioning failures and system-level fault tolerance even in the face of complex and abnormal conditions such as sudden underwater optical interference, water attenuation, and lens obstruction. Attached Figure Description

[0066] Figure 1 A schematic diagram of the state update mechanism for the localization filter;

[0067] Figure 2 A schematic diagram of a multi-sensor elastic fusion positioning framework for scene perception. Detailed Implementation

[0068] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described below are intended to facilitate the understanding of the present invention and do not limit it in any way.

[0069] This invention proposes a probabilistically driven elastic fusion positioning framework based on scene perception and the operating state of multiple filters. Its purpose is to address the positioning jumps and oscillations caused by fixed-distance or binary logic "hard switching" mechanisms in existing underwater positioning methods. By establishing the innovation vector of the dead reckoning Kalman filter to characterize the operating state of each positioning filter in real time, and by dynamically adjusting the weight coefficients of the remote return positioning filter and the multi-channel visual positioning filter (R / G / B) in conjunction with the probability of light beacon perception in the camera scene, a smooth transition in positioning solution and system-level fault tolerance are achieved.

[0070] 1. Positioning filter state update mechanism

[0071] like Figure 1 As shown, both the long-range return and final docking phases utilize a Kalman filter framework for positioning. The main difference lies in the configuration of the observation sensors: the long-range return phase employs an ultra-short baseline positioning system (USBL), while the final docking phase uses a multi-channel visual positioning module, corresponding to red (R), green (G), and blue (B) visual channels respectively. All positioning filters use dead reckoning via Doppler log (DVL) and electronic compass as the core of the navigation calculation, ensuring continuous positioning output under extreme conditions.

[0072] Each positioning filter (denoted by the superscript i) includes four types of error parameters as state vector components: misalignment angle, position error, compass mounting angle error, and DVL scaling factor error. The temporal update of the positioning filters follows an independent computational mechanism, specifically comprising five recursive stages: state prediction, covariance prediction, Kalman gain calculation, state update, and covariance update.

[0073] (1) State prediction

[0074] Based on the state estimate from the previous time step By performing time recursion using the system model, the prior state estimate for the current moment is obtained. :

[0075] ;

[0076] in, Let be the state transition matrix.

[0077] (2) Covariance prediction

[0078] Based on the posterior covariance matrix of the previous time step Derivation of the prior covariance at the current moment :

[0079] ;

[0080] in, Characterization process noise covariance matrix.

[0081] (3) Gain calculation

[0082] Kalman gain matrix Determine the weighting ratio between model predictions and sensor observations during the state correction process:

[0083] ;

[0084] in, Represents the observation noise covariance matrix. For measuring the transfer matrix.

[0085] (4) Status update

[0086] When acquiring observation values At that time, prior estimates and observation information are fused using Kalman gain to generate posterior state estimates. :

[0087] ;

[0088] (5) Covariance update

[0089] Synchronously update the posterior covariance matrix To reflect the uncertainty of the estimate after state correction:

[0090] .

[0091] Under normal sensor operating conditions, the innovation sequence should theoretically follow a zero-mean Gaussian distribution. By monitoring the innovation sequence, the real-time operating status of the observation sensor can be effectively evaluated. To support the dynamic evaluation of filter operating status by the flexible fusion framework, this invention focuses on monitoring the innovation sequence. and its covariance :

[0092] ;

[0093] .

[0094] 2. Probability-driven elastic fusion framework

[0095] Flexible fusion framework such as Figure 2 As shown, the framework mainly consists of four core modules: scene-aware probability acquisition, filter number update, estimation fusion, and state interaction initialization. This framework achieves soft switching between far / end-point positioning strategies through scene-aware probability and introduces innovation and its covariance to achieve fault tolerance in positioning failures.

[0096] (1) Scene perception probability acquisition

[0097] The core task of the scene-aware model is to detect the probability of a cursor's presence in a camera image, providing prior probabilities for multi-source navigation information fusion. Due to the real-time constraints of the underwater docking task, the scene-aware model architecture must be lightweight and have low latency. Furthermore, given that this problem is essentially a binary classification task, an SVM model is used to construct the classification decision boundary. The training set contains a total of... The image consists of four underwater light spot images, each containing RGB three channels and a grayscale channel. The four channels are independent of each other and correspond to binary classification labels. After training, this channel regression model is formed. For the first... Frame image Channels, multidimensional feature vectors are ,in The corresponding cursor existence detection label is ( The cursor is missing. (As the cursor exists). Based on soft-margin optimization theory, an SVM regression model is constructed, and its objective function can be expressed as:

[0098] ;

[0099] ;

[0100] in, For the weight vector, These are bias terms, which together constitute the hyperplane equation. , The variable to be held is st, and the expression after st is the constraint condition. This is a nonlinear mapping function. The kernel matrix of this model is constructed using radial basis function (RBF) kernels. ,in This is the kernel width parameter. This is the regularization coefficient, used to balance model complexity and training error; and These represent the p-th and q-th frames, respectively.

[0101] Through Lagrange dual transformation, the original optimization problem can be transformed into the following dual form:

[0102] ;

[0103] ;

[0104] in, It is a vector in which all elements are 1. It is the solution to the dual problem, also known as the Lagrange multiplier; The size is A positive semi-definite matrix whose elements , and Let and represent the feature vector and label of the q-th training sample, respectively. After solving this optimization problem, the predicted image... Channel Labels Solve based on its multidimensional feature vectors The regression decision function can be derived as follows:

[0105] ;

[0106] in, This indicates the number of support vectors. It is a symbolic function.

[0107] After obtaining the regression predictions, a probabilistic output model needs to be constructed for fusing navigation decisions. Based on the nonlinear mapping characteristics between the decision function's range and the posterior probability, this invention uses the Platt scaling method to construct a sigmoid probability transformation function to predict probabilities. Mathematical form:

[0108] ;

[0109] ;

[0110] Among them, parameters , Determined by fitting the training samples.

[0111] (2) State interaction initialization

[0112] At any moment Before each filter performs the single-filter state update described above, cross-filter parameter fusion reconstruction is required. This mechanism breaks through the traditional single-source inheritance mode and achieves initialization parameter optimization by integrating multi-filter state information. The state re-initialization process achieves hybrid state reconstruction by integrating the prior prediction probabilities of the multi-filter scenario and the fusion weight estimation. The corresponding reconstructed single-filter state vector and its covariance matrix can be expressed as:

[0113] ;

[0114] ;

[0115] in, and Represents the target filter The reinitialized posterior state and covariance; and Indicates the filters participating in the fusion The independent posterior states and covariance; These are the initial weight coefficients, representing the filter. State of the target filter The degree of contribution of the filter; Initialize weight coefficients Depends on prior prediction probability of the scene With fusion weight estimation Defined as:

[0116] ;

[0117] This represents the state transition probability from filter j to filter i.

[0118] (3) Filter quantity update and state transition

[0119] This module dynamically adjusts the number of activated filters based on scene perception probability. When the visual filter... When the sensor approaches the target cursor, the scene perception model outputs the cursor recognition probability for the corresponding channel. .like Exceeding the preset threshold If the visual filter is deemed to have reliable positioning capabilities, it is then incorporated into the flexible fusion framework. During a typical connection process, each visual filter dynamically activates and deactivates based on its scene perception state. The scene perception probability directly determines the state transition probability between filters. This parameter is used as a prior probability and weighted by fusing it with the corresponding filter's posterior probability. (Filter state transition probability matrix) The construction rules are as follows:

[0120] ;

[0121] in, for 3D matrix. express No. Row vectors are defined as filters. to the filter respectively The transition probability. At this point, the system contains... One visual channel filter, This represents the scene perception probability of the corresponding channel at time k. The far-end return filter (filter 1), acting as a filter during the cruise phase, has its weight decreases as the number of active channels increases. In the state where no visual filter is active, the following settings are made: .coefficient The value is set to 0.8. This design ensures that when the scene recognition probability of a certain channel is 1, the filter state initialization process avoids ignoring the contributions of other filters.

[0122] (4) Estimation of fusion

[0123] In the multi-filter fusion estimation process, it is necessary to comprehensively consider both dynamic changes in the scene and the operating state of individual filters. The scene perception probability is obtained through a matrix. This method addresses the fusion weight allocation mechanism. Individual filters may experience transient observation biases or model inaccuracies due to factors such as sensor anomalies. This approach dynamically adjusts the fusion weights by introducing innovation and its covariance. This establishes a two-factor decision mechanism for the fusion weights, where the weights of each filter are jointly determined by its real-time performance metrics and prior scene probability parameters.

[0124] Each filter independently calculates the likelihood of generating the new information. This parameter characterizes the operating state of a single filter, and its calculation is based on the derivation of the innovation vector and its covariance matrix:

[0125] ;

[0126] filter The posterior probability, i.e., the fusion weight. From the new likelihood With the transition probability matrix Joint derivation:

[0127] ;

[0128] in, Represents filter The scenario prior prediction probability is fused with historical weight distribution and inter-filter state transition probability. (matrix element); This indicates the specific filter currently calculating the fusion weights, and j represents the filter participating in the fusion. and Let represent the innovation likelihood of the fusion filter j and the fusion weight of the previous time step k-1, respectively.

[0129] Final fusion state estimation Estimated values ​​from each filter Weighted combination implementation:

[0130] ;

[0131] Simultaneous derivation of the temporal fusion covariance matrix:

[0132] .

[0133] The main innovations of this invention are as follows:

[0134] (1) Probability-driven elastic fusion positioning framework

[0135] A probability-driven elastic fusion framework is constructed, comprising four core modules: scene perception probability acquisition, filter number update, estimation fusion, and state interaction initialization, to realize fusion weights. A two-factor decision-making mechanism. The weights are simultaneously determined by real-time performance metrics (news likelihood). ) and scenario prior probability This is a joint decision. Dynamically adjusting the number of activated filters will satisfy the cursor recognition probability threshold. The visual channel is incorporated into the fusion framework.

[0136] (2) Scene-aware far / end positioning strategy "soft handover" mechanism

[0137] A smooth, gradual "soft handover" for far / end positioning is achieved through scene-aware probabilistic detection, replacing the traditional fixed-distance "hard handover" mechanism, thereby eliminating positioning jumps and oscillations during handover. First, a scene-aware model (trained based on SVM and Platt scaling methods) is applied to detect the probability of the presence of light beacons in camera images in real time. The probability is then used to dynamically adjust the activation state and weight allocation of the remote return positioning filter and the terminal multi-channel visual positioning filter.

[0138] (3) Real-time evaluation and fault-tolerant mechanism of multi-filter operating state based on innovation vector

[0139] Establish and monitor the innovation vectors of each positioning filter (including the remote USBL / DVL and the end R / G / B visual channels) in real time. and its covariance matrix Derivation of innovation likelihood through innovation sequence This serves as a basis for characterizing the real-time operating status and performance indicators of a single-unit filter. The innovation likelihood is used as... Introduced into fusion weights In the calculation, dynamic weighting based on the real-time performance of the filter is implemented, thereby ensuring system-level fault tolerance when the performance of a single filter degrades or fails under complex working conditions such as water interference and optical attenuation.

[0140] The method of the present invention has the following advantages:

[0141] (1) Eliminate positioning jumps and oscillations to achieve smooth soft switching.

[0142] Existing technologies employ a fixed-distance hard handover mechanism. Near the handover point, remote positioning data is significantly suppressed or completely discarded, easily leading to positioning jumps and oscillations in the positioning results. This invention dynamically adjusts the weight coefficients of the remote walk-back positioning filter and the end-point multi-channel visual positioning filter through scene perception probability, achieving a smooth transition and soft handover in the positioning solution, thereby effectively eliminating the drawbacks of traditional hard handover.

[0143] (2) Significantly enhances fault tolerance to underwater interference and positioner failure.

[0144] Existing methods relying solely on visual positioning lack effective data fusion mechanisms to achieve fault tolerance in the event of visual positioning failure or performance degradation. This invention introduces innovation and its covariance to characterize the working state of each positioning filter in real time, enabling the positioning system to dynamically adjust weights based on the real-time performance of the filters (rather than solely relying on external distance). This allows the system to maintain fault tolerance for positioning failures and system-level fault tolerance even in the face of complex and abnormal conditions such as sudden underwater optical interference, water attenuation, and lens obstruction.

[0145] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A scene-aware multi-sensor elastic fusion positioning method for underwater docking, characterized in that, Both the remote return phase and the final docking phase utilize a Kalman filter framework for localization, with multi-channel visual localization added during the final docking phase; this includes the following steps: Scene perception probability acquisition steps: Based on the scene perception model, obtain the scene perception probability of the cursor for each visual channel in the current scene, and use them as the prior prediction probability of the localization filter corresponding to each visual channel. State interaction initialization steps: By combining the prior prediction probabilities of each positioning filter with the fusion weights, the state vector and covariance matrix of each positioning filter are obtained; The steps for updating the number of activated positioning filters and the state transition are as follows: The activation state of the positioning filters is dynamically adjusted according to the scene perception probability of the cursor by each visual channel, and the state transition probability matrix between each positioning filter is constructed. Estimation fusion steps: Update the state vector and covariance matrix based on the dynamic changes of the current scene and the real-time running state of each active localization filter.

2. The scene-aware multi-sensor elastic fusion positioning method for underwater docking according to claim 1, characterized in that, The multi-channel visual positioning system has three RGB visual channels.

3. The scene-aware multi-sensor elastic fusion positioning method for underwater docking according to claim 2, characterized in that, The positioning filters include positioning filters for ultra-short baseline positioning during the far-end return phase, and R-channel positioning filters, G-channel positioning filters, and B-channel positioning filters for the end-of-pipe docking phase.

4. The scene-aware multi-sensor elastic fusion positioning method for underwater docking according to claim 1, characterized in that, The timing updates of each localization filter follow an independent computational mechanism, including state prediction, covariance prediction, Kalman gain calculation, state vector update, and covariance update.

5. The scene-aware multi-sensor elastic fusion positioning method for underwater docking according to claim 1, characterized in that, Scene perception probability In the acquisition step, the method for constructing the scene-aware model includes: training a support vector machine regression model using a training set; the training set contains a total of The image consists of four underwater light spot images, each containing three RGB channels and a grayscale channel. These four channels are independent of each other and correspond to binary classification labels. After training, a scene perception model for the corresponding channel is formed.

6. The scene-aware multi-sensor elastic fusion positioning method for underwater docking according to claim 1, characterized in that, In the state interaction initialization step, the reconstructed state vector of a single localization filter and covariance matrix They are respectively: ； ； Among them, the positioning filter Initial weight coefficients It depends on the prior prediction probability of the scene. With fusion weight estimation and filters j to filter i State transition probability ; and represents the posterior state estimate and posterior covariance matrix at time k-1, respectively; n is the number of currently active localization filters.

7. The scene-aware multi-sensor elastic fusion positioning method for underwater docking according to claim 1, characterized in that, In the steps of updating the number of activations of the localization filter and state transition, when the first... i Probability of cursor scene perception in each visual channel Greater than the preset threshold At that time, adjust the first i The localization filter corresponding to each visual channel is in an active state.

8. The scene-aware multi-sensor elastic fusion positioning method for underwater docking according to claim 1, characterized in that, In the steps of updating the number of activations of localization filters and state transition, the state transition probability matrix between each localization filter is... The construction rules are as follows: ； ； ； ； ； ； in, for 3D matrix This represents the total number of positioning filters; express No. Row vectors are defined as localization filters. To the localization filter respectively The transition probability is then used to define the far-end return positioning filter as the first positioning filter, and the visual channel filters correspond to the others. One positioning filter; The probability of scene perception at time k is represented by the corresponding visual channel. It is the discount factor.

9. The scene-aware multi-sensor elastic fusion positioning method for underwater docking according to claim 1, characterized in that, The estimation fusion steps include: Calculate the innovation likelihood of each localization filter separately. : ； Among them, the new information sequence , For the first i The observation values ​​of each localization filter at time k. Let k be the prior state estimate at time k. For measuring the transfer matrix; covariance of the innovation sequence. ,in Let k represent the observation noise covariance matrix at time k. Indicates the first The prior covariance matrix of each localization filter; ； ； in, For the first i The posterior probability of each localization filter; For the first i The prior prediction probability of each localization filter; Represents filter j to filter i The state transition probability; The posterior state estimates of each localization filter The weighted combination calculation yields the final fusion state estimate. : ； The covariance matrix of each positioning filter Calculate the temporal fusion covariance matrix : 。 10. A scene-aware multi-sensor elastic fusion positioning device for underwater docking, characterized in that, Both the remote return phase and the final docking phase utilize a Kalman filter framework for localization, while the final docking phase incorporates multi-channel visual localization; it includes the following units: Scene perception probability acquisition unit: Based on the scene perception model, it acquires the scene perception probability of the cursor for each visual channel in the current scene, and uses it as the prior prediction probability of the localization filter corresponding to each visual channel. State interaction initialization unit: By integrating the prior prediction probabilities and fusion weights of each positioning filter, the state vector and covariance matrix of each positioning filter are obtained. The localization filter activation quantity update and state transition unit dynamically adjusts the activation state of the localization filters based on the scene perception probability of the cursor by each visual channel, and constructs the state transition probability matrix between each localization filter. Estimation fusion unit: Updates the state vector and covariance matrix based on the dynamic changes of the current scene and the real-time operating status of each active localization filter.

Images