Positioning method and system for underwater environment

Abstract

The invention relates to the technical field of underwater robot navigation and environment perception, and discloses a positioning method and system for an underwater environment, and the positioning method comprises the steps: obtaining flow velocity profile data, temperature data, salinity data and depth data at a sampling point; obtaining a normalized flow field feature vector and a physical feature vector corresponding to each sampling point; calculating the cosine similarity of the normalized flow field feature vector and the physical feature similarity of the normalized three-dimensional physical feature vector; performing weighted fusion on the cosine similarity and the physical feature similarity to obtain fusion similarity; performing screening by combining the fusion similarity and a preset constraint rule to obtain a position revisit constraint factor; adding the position revisit constraint factors into the factor graph to construct a complete factor graph model; and solving the factor graph by using an iSAM2 algorithm to obtain a globally optimized position sequence of the underwater vehicle. According to the method, the accumulated drift error can be corrected, and the reliability of the track of the AUV in a video-sound limited scene is effectively guaranteed.

Description

Technical Field

[0001] This application relates to the field of underwater robot navigation and environmental perception technology, for example, to a positioning method and system for underwater environments. Background Technology

[0002] In the meso-oceanic environment, visual positioning methods are difficult to apply effectively due to the short propagation distance and severe obstruction of light. At the same time, the performance of acoustic positioning systems is also significantly limited due to multipath effects and sound wave attenuation. Since GPS signals cannot reach these areas, autonomous underwater vehicles (AUVs) cannot rely on satellite navigation for real-time positioning and can only rely on inertial navigation systems and environmental perception methods to estimate their own position.

[0003] However, existing dead reckoning methods primarily estimate position by continuously integrating internal sensor data. These methods are highly sensitive to sensor noise and environmental interference, and errors accumulate over time, reaching 3%-5% of the journey's distance in long-duration missions, leading to severe positioning drift. In the mid-ocean environment, positioning methods based on external acoustic base stations (including USBL, LBL, and SBL) are difficult to apply. First, sound waves at this depth are affected by layered structures such as thermohaline layers, causing changes in sound velocity and refraction, resulting in signal attenuation and multipath interference, reducing positioning accuracy from theoretical sub-meter levels to several meters or even tens of meters. Second, USBL and SBL rely on communication links from surface platforms, but long-distance sound wave transmission leads to signal energy attenuation, limiting their effective range. While LBL theoretically offers higher accuracy, the deployment and maintenance costs of seabed base stations are high, and the lack of fixed reference points in open water makes it difficult to meet the needs of large-scale and long-term missions. Traditional SLAM methods struggle to achieve effective localization and navigation in the mesosphere, primarily due to the relatively open environment and lack of stable, rich geographical or structural features, making it difficult to construct accurate and persistent environmental maps. Furthermore, the frequent dynamic changes in water bodies and the susceptibility of environmental features to alteration increase the difficulty of map maintenance and updates, reducing the localization accuracy and robustness of SLAM systems. Although AUVs are commonly equipped with CTD sensors and ADCPs that can acquire real-time underwater temperature, salinity, depth, and current field data, and physical oceanography studies indicate that the temperature, salinity, and current field structure of the mesosphere are controlled by large-scale circulation and mesoscale eddies, exhibiting relative stability and spatial variability over timescales of several days to several weeks, this environmental field information is not effectively utilized in current underwater localization and trajectory correction algorithms. There is a lack of methods to transform spatial distribution patterns of the current field into location features, and a lack of algorithmic frameworks to integrate multimodal environmental information for location revisit detection and global trajectory optimization, thus limiting the localization accuracy of underwater AUVs in visually and acoustically confined environments.

[0004] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general commentary, nor is it intended to identify key / important components or describe the scope of protection of these embodiments, but rather as a prelude to the detailed description that follows.

[0006] This disclosure provides a positioning method and system for underwater environments to solve the technical problems of low positioning accuracy and severe trajectory drift of AUVs in the mid-ocean environment in the prior art.

[0007] In some embodiments, a positioning method for underwater environments includes: S10, marine environment data acquisition and preprocessing: acquiring velocity profile data detected by an acoustic Doppler current profiler at sampling points, as well as temperature data, salinity data, and depth data detected by a CTD sensor, and preprocessing each data to remove abnormal data; S20, multimodal environment description construction: converting the velocity profile data into a flow field image, extracting features from the flow field image using a ResNet50 model, and normalizing the extracted features to obtain a normalized flow field feature vector corresponding to each sampling point; simultaneously, for each sampling point, using the temperature data, salinity data, and depth data... S30. Construct a three-dimensional physical feature vector and normalize it; S40. Position revisit detection based on multimodal feature fusion: calculate the cosine similarity of the normalized flow field feature vector and the physical feature similarity of the normalized three-dimensional physical feature vector; weight the cosine similarity and the physical feature similarity to obtain the fusion similarity; combine the fusion similarity and preset constraint rules to filter and obtain the position revisit constraint factor; S40. Optimization estimation based on factor graph: add the position revisit constraint factor to the factor graph to construct a complete factor graph model; use the iSAM2 algorithm to solve the complete factor graph to obtain the globally optimized position sequence of the underwater vehicle.

[0008] In some embodiments, a positioning system for underwater environments includes: a marine environment data acquisition and preprocessing module configured to acquire current profile data detected by an acoustic Doppler current profiler at sampling points, as well as temperature data, salinity data, and depth data detected by a CTD sensor, and preprocess each data to remove abnormal data; a multimodal environment description construction module configured to convert the current profile data into a flow field image, extract features from the flow field image using a ResNet50 model, and normalize the extracted features to obtain a normalized flow field feature vector corresponding to each sampling point; simultaneously, for each sampling point, the temperature data, salinity data, and depth data are used to... A three-dimensional physical feature vector is constructed and normalized. A position revisit detection module based on multimodal feature fusion is configured to calculate the cosine similarity of the normalized flow field feature vector and the physical feature similarity of the normalized three-dimensional physical feature vector. The cosine similarity and the physical feature similarity are weighted and fused to obtain a fused similarity. A position revisit constraint factor is obtained by combining the fused similarity with preset constraint rules. A factor graph-based optimization estimation module is configured to add the position revisit constraint factor to the factor graph to construct a complete factor graph model. The iSAM2 algorithm is used to solve the complete factor graph to obtain the globally optimized position sequence of the underwater vehicle.

[0009] The positioning method and system for underwater environments provided in this disclosure can achieve the following technical effects: 1. Eliminating reliance on optical and acoustic signals to achieve positioning in environments with limited visibility and acoustics. This invention utilizes ADCP current profiles and CTD temperature, salinity, and depth data as sources of environmental perception, eliminating the need for optical vision or acoustic positioning base stations. It is suitable for complex underwater environments such as the mid-ocean sphere where light attenuation is severe and acoustic multipath effects are significant, thus expanding the operational range of AUVs. 2. Significantly corrects the cumulative drift error in dead reckoning based on the factor graph optimization framework. This invention employs the iSAM2 incremental factor graph optimization algorithm, which unifies the modeling of prior factors, odometry factors, and position revisit constraint factors, and effectively eliminates the cumulative positioning error generated by the long-term operation of the inertial navigation system through global optimal solution; 3. Location revisit detection effectively ensures the reliability of AUV trajectories in visually and aurally limited scenarios. In applications requiring AUVs to conduct large-scale round-trip patrols, such as marine environmental monitoring and underwater target search, the cumulative drift from dead reckoning can cause trajectories to deviate significantly from the predetermined routes, affecting the geolocation accuracy of mission data. This invention, through location revisit detection and joint factor map optimization, automatically identifies and corrects trajectory drift when the AUV passes through previously visited areas, ensuring the spatial positioning accuracy of collected data and enhancing the scientific value and usability of marine survey data.

[0010] The above general description and the description below are exemplary and illustrative only and are not intended to limit this application. Attached Figure Description

[0011] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations and drawings do not constitute a limitation on the embodiments. Elements having the same reference numerals in the drawings are shown as similar elements. The drawings are not to be scaled. And wherein: Figure 1 This is a schematic diagram of a positioning method for underwater environments provided in an embodiment of this disclosure; Figure 2 This is a flowchart of a positioning system for underwater environments provided in an embodiment of this disclosure; Figure 3 This is a schematic diagram comparing the trajectories of the present method and the conventional method provided in this embodiment of the disclosure; Figure 4 This is a schematic diagram comparing the point-by-point position error between the method provided in this embodiment and the conventional method. Figure 5 This is a schematic diagram of a positioning system for underwater environments provided in an embodiment of this disclosure.

[0012] Figure label: 50. A positioning system for underwater environments; 51. A marine environment data acquisition and preprocessing module; 52. A multimodal environment description construction module; 53. A location revisit detection module based on multimodal feature fusion; 54. An optimization estimation module based on factor graphs. Detailed Implementation

[0013] To provide a more detailed understanding of the features and technical content of the embodiments of this disclosure, the implementation of the embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for illustrative purposes only and are not intended to limit the embodiments of this disclosure. In the following technical description, for ease of explanation, several details are used to provide a full understanding of the disclosed embodiments. However, one or more embodiments may still be implemented without these details. In other cases, well-known structures and devices may be simplified in their depiction to simplify the drawings.

[0014] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this disclosure described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion.

[0015] Unless otherwise stated, the term "multiple" means two or more.

[0016] In this embodiment of the disclosure, the character " / " indicates that the objects before and after it are in an "or" relationship. For example, A / B means: A or B.

[0017] The term "and / or" describes an association between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or A and B.

[0018] The term "correspondence" can refer to an association or binding relationship. The correspondence between A and B means that there is an association or binding relationship between A and B.

[0019] Combination Figure 1 As shown, this disclosure provides a positioning method for underwater environments, including: S10, Marine environmental data acquisition and preprocessing: Acquire current profile data detected by acoustic Doppler current profiler at sampling points, as well as temperature data, salinity data and depth data detected by CTD sensor, and preprocess each data to remove abnormal data; S20, Multimodal Environment Description Construction: The flow velocity profile data is converted into a flow field image. The ResNet50 model is used to extract the features of the flow field image, and the extracted features are normalized to obtain the normalized flow field feature vector corresponding to each sampling point. At the same time, for each sampling point, a three-dimensional physical feature vector is constructed using temperature data, salinity data, and depth data and normalized. S30, Location revisit detection based on multimodal feature fusion: Calculate the cosine similarity of the normalized flow field feature vector and the physical feature similarity of the normalized three-dimensional physical feature vector; weight the cosine similarity and physical feature similarity to obtain the fusion similarity; combine the fusion similarity and preset constraint rules to filter and obtain the location revisit constraint factor. S40, Optimization estimation based on factor graph: Add position revisit constraint factors to the factor graph to construct a complete factor graph model; use the iSAM2 algorithm to solve the complete factor graph to obtain the globally optimized position sequence of the underwater vehicle.

[0020] First, execute S10, which involves acquiring and preprocessing marine environmental field data. This includes the following steps: S11, acquire current profile data detected by ADCP. Acoustic Doppler Current Profiler (ADCP) is a standard sensor for measuring ocean current profiles. Its working principle is to emit acoustic pulses into the water and calculate the current velocity at each layer based on the Doppler frequency shift of the echoes reflected from scatterers in different depth layers of the water.

[0021] S111, ADCP operating mode: ADCP employs a downward detection mode, emitting a sound beam downwards from the bottom of the AUV. Let the AUV at the current time... The position is ,in The horizontal position coordinates are For depth.

[0022] S112, Sampling parameter settings: Key parameters of ADCP include: 1. Detection depth range In the preferred embodiment of this method, the depth is set to 500 meters, but this can be adjusted according to mission requirements and ocean stratification. 2. Vertical sampling interval This method is set to 10 meters; 3. Time resolution: Related to the AUV's motion speed, this method uses trajectory point sampling, with time resolution calculated at intervals of [missing information]. Perform ADCP sampling once for each trajectory point, and optimize... .

[0023] S12, acquire temperature, salinity, and depth data from CTD detection. The CTD sensor acquires the physical properties of seawater by measuring its conductivity, temperature, and pressure (converted to depth).

[0024] At the current location of the AUV At that location, the CTD sensor measured: 1. Temperature (Unit: degrees Celsius, °C); 2. Salinity (Unit: Practical salinity unit, PSU); 3. Depth (Unit: meter, m).

[0025] Constructing a three-dimensional physical feature vector: .

[0026] S13, Perform quality control on the above-mentioned test data to ensure the reliability of the environmental field data, including: (1) Noise filtering For continuously sampled data, moving average filtering or Gaussian filtering can be used to remove high-frequency noise: , in, This is half the width of the filter window; This represents the estimated value of temperature (or physical quantities such as salinity and flow rate) after noise filtering at the t-th sampling time.

[0027] (2) Spatial registration Ensure that ADCP and CTD data are aligned in time and space. Since the two types of sensors may be located at different positions on the AUV, spatial registration is required based on installation location deviations. , in, Here is the AUV attitude rotation matrix. The mounting offset of the CTD relative to the center of the AUV; Indicates the current position of the CTD; This indicates the current position of the AUV.

[0028] Then, S20 is executed to construct the multimodal environment descriptor. The flow velocity profile data is converted into a flow field image. Features of the flow field image are extracted using the ResNet50 model, and the extracted features are normalized to obtain a normalized flow field feature vector for each sampling point. Simultaneously, for each sampling point, a three-dimensional physical feature vector is constructed using temperature, salinity, and depth data, and then normalized. Specifically, the steps include: S21, converting velocity profile data into flow field images, i.e., image-based encoding of flow field data. Converting the numerical velocity profile data acquired by ADCP into image form allows for the extraction of spatial distribution pattern features of the flow field using mature computer vision techniques. Specifically, this includes: S211, Selection of velocity components. Taking into account the vertical velocity component... The magnitude is typically 1-2 orders of magnitude smaller than the horizontal velocity component (typical value: meters per second (m / s), and The measurement noise is relatively large and the signal-to-noise ratio is low. This method only uses the horizontal velocity component. and Perform image encoding. The advantages of doing so include: 1. Reduce noise interference and improve feature quality; 2. Reduce feature dimensionality and improve computational efficiency; 3. Horizontal flow velocity contains more environmental information and is more valuable for location identification.

[0029] S212, Determine the global color mapping range. To ensure the comparability of flow field images generated at different locations, a unified color mapping range must be used. The specific steps are as follows: Step a: Traverse all ADCP sampling points (assuming a total of...) (1 sampling point), collect all flow rate values: .

[0030] Step b: Calculate the global minimum and global maximum values ​​of each horizontal velocity component: , .

[0031] Step c: To avoid the influence of extreme outliers, quantile ranges are used instead of global maximum and global minimum values. , in, express The Quantiles are usually taken as .

[0032] S213, Color Mapping Function. It uses Jet chromatography (a rainbow spectrum from blue to red) to map flow rate values, normalized based on quantile ranges, to RGB colors. Jet chromatography has the following characteristics: 1. The colors are varied and highly distinguishable, making them suitable for expressing continuously changing physical quantities; 2. Strong color contrast facilitates CNN extraction of texture and edge features; 3. Widely used in the field of scientific visualization, with strong interpretability.

[0033] The specific mapping steps include: First, normalize the horizontal velocity component to... Interval: , .

[0034] Then, the color values ​​are mapped to RGB three-channel color values ​​using the Jet colorimetric function: , , in, These are the RGB color channel values.

[0035] Mathematical definition of Jet chromatographic functions: Jet chromatography in The RGB components within the interval are defined as follows: , , .

[0036] S214, Flow field image generation. The generated flow field image... It has the following design features: a. Image size and layout: 1. Total image size: Pixel, in a preferred embodiment of the present invention Pixels Pixel; 2. Number of channels: 3 (RGB); 3. Layout: Components and The components are displayed side by side, each occupying half the width of the image; 4. Depth direction: From top to bottom, corresponding to the shallowest to the deepest layers, each layer occupies a certain height. Pixel.

[0037] b. Pure design: To ensure that the CNN network learns only the spatial distribution pattern features of the flow field and is not influenced by artificial annotation information, the image must meet the following purity requirements: 1. Background color: Pure black (RGB value is...) ); 2. No coordinate axes, tick marks, or grid lines; 3. No text labels, titles, or legends; 4. There are no border lines between the color blocks; 5. No artificially added identification information.

[0038] S22, the ResNet50 model is used to extract features from the flow field image, and the extracted features are normalized. ResNet (Residual Network) solves the gradient vanishing problem in deep networks by introducing residual connections. ResNet50 contains 50 convolutional layers and has achieved excellent performance on the ImageNet large-scale image classification task. The features it learns have strong generality and transferability. Specifically, the steps are as follows: S221, Selection and Loading of the Pre-trained Model. This method uses a ResNet50 model pre-trained on the ImageNet dataset as the feature extractor to extract features from flow field images. The reasons for choosing ResNet50 include: 1. Moderate depth (50 layers), capable of extracting high-level semantic features, while maintaining acceptable computational complexity; 2. Pre-trained weights are publicly available, eliminating the need for training from scratch and saving computational resources; 3. The feature dimensions are reasonable (2048 dimensions), which contains rich information without being too redundant; 4. It performs well in various computer vision tasks and is the preferred backbone network for transfer learning.

[0039] Model structure modification: 1. Remove the last fully connected classification layer of ResNet50 (originally used for 1000-class classification in ImageNet); 2. Retained to the Global Average Pooling (GAP) layer; 3. The output feature dimension is 2048.

[0040] Feasibility analysis of transfer learning: Although the ImageNet dataset mainly contains natural images (animals, plants, objects, etc.), while flow field images belong to scientific visualizations, and there is a domain gap between the two, transfer learning is still effective for the following reasons: 1. The low-level features of CNNs (edges, textures, corners, etc.) are universal and applicable to various types of images; 2. The spatial distribution pattern of flow field images can be regarded as a special texture feature; 3. Pre-trained weights provide good initialization and can extract effective features even without fine-tuning.

[0041] S222, preprocessing the flow field image to adapt it to the input requirements of ResNet50, including: a. Size Adjustment: The standard input size of ResNet50 is... Pixels. The flow field image is converted from its original size using bilinear interpolation. Zoom to : .

[0042] b. Channel normalization: Standardization is performed using statistical parameters from the ImageNet dataset to match the data distribution of the pre-trained model. , in, Represents the color channel, with the normalization parameter as follows: .

[0043] These parameters are the mean and standard deviation of each channel in the ImageNet training set.

[0044] c. Tensor Transformation: Converts image data into tensor format, with dimensions arranged as follows: ,in , where is the number of channels; , where is the height and width.

[0045] S223, Input the preprocessed flow field image into the ResNet50 model, and extract features through forward propagation: , in, This is the original feature vector output by the global average pooling layer.

[0046] The main structural layers of ResNet50: 1. Initial convolutional layer: Convolution, stride 2, output 64 channels; 2. Max pooling layer: Pooling, step size 2; 3. Residual Module Group 1 (conv2_x): 3 residual blocks, 256 output channels; 4. Residual Module Group 2 (conv3_x): 4 residual blocks, 512 output channels; 5. Residual Module Group 3 (conv4_x): 6 residual blocks, 1024 output channels; 6. Residual Module Group 4 (conv5_x): 3 residual blocks, 2048 output channels; 7. Global average pooling: reduces spatial dimensions Pooling ; 8. Output: 2048-dimensional feature vector.

[0047] S224, Feature Normalization. To facilitate subsequent similarity calculations, the extracted features are L2 normalized to obtain the normalized flow field feature vector for each sampling point: , The L2 norm is defined as follows: , Normalized eigenvectors satisfy .

[0048] Advantages of normalization: 1. Eliminate the influence of feature vector magnitude, so that similarity measurement focuses only on direction; 2. Simplify cosine similarity calculation to inner product operation; 3. Improve numerical stability and avoid calculation errors caused by excessively large or small eigenvalues.

[0049] S225, Feature Storage and Management. For each ADCP sampling point on the trajectory. Store its corresponding normalized flow field feature vector. Build a feature database: , This database supports efficient similarity queries and location revisit detection.

[0050] S23, Extraction and normalization of physical features of temperature, salinity, and depth. Specifically, this includes: S231 utilizes temperature, salinity, and depth data to construct a three-dimensional physical feature vector. For each sampling point on the trajectory... Temperature measured by CTD sensor ,salinity and depth Composition of three-dimensional physical feature vectors: .

[0051] The distribution of temperature and salinity in the mesosphere is controlled by the following factors: 1. Large-scale circulation: such as the Gulf Stream and the Kuroshio Current, forming stable water mass boundaries; 2. Mesoscale vortices: tens to hundreds of kilometers in diameter, with lifespans of weeks to months, and possessing either a cold or warm core structure; 3. Thermocline and halocline: The temperature and salinity gradients change drastically within a specific depth range, forming a distinct stratified structure.

[0052] These factors lead to spatial heterogeneity in temperature, salinity, and depth characteristics, meaning that the combinations of temperature, salinity, and depth at different locations have certain uniqueness and can serve as an "environmental fingerprint" for location identification.

[0053] S232, Feature Statistical Analysis. To design a reasonable similarity measurement method, it is necessary to analyze the statistical characteristics of physical features across the entire trajectory dataset. Let's assume... For each sampling point, calculate the statistical indicators of each physical quantity: Mean: .

[0054] Standard deviation: , , .

[0055] Coefficient of variation: .

[0056] The coefficient of variation reflects the relative magnitude of change of each physical quantity and is used to assess its ability as a location distinguishing feature.

[0057] Then, S30 is executed for location revisit detection based on multimodal feature fusion: the cosine similarity of the normalized flow field feature vector and the physical feature similarity of the normalized 3D physical feature vector are calculated; the cosine similarity and physical feature similarity are weighted and fused to obtain the fused similarity; the fused similarity is combined with preset constraint rules for filtering to obtain the location revisit constraint factor. Specifically, the steps are as follows: S31, calculate the cosine similarity. For two normalized flow field feature vectors... and ( Their cosine similarity is defined as: , in, The angle between the two eigenvectors.

[0058] Geometric meaning: 1. The two vectors have the same direction. Their features are completely similar; 2. Two vectors are orthogonal. The features are unrelated; 3. The two vectors are in opposite directions. Their characteristics are completely opposite.

[0059] For non-negative features (such as CNN features after ReLU activation), the range of values ​​for cosine similarity is narrowed to... .

[0060] S32, Calculate the similarity matrix. To accelerate location revisit detection, the image similarity between all sampling point pairs can be pre-calculated to construct a similarity matrix: , This matrix has the following properties: 1. Symmetry: ; 2. Diagonal is 1: ; 3. Nonnegativity: .

[0061] Computational optimization: Matrix operations are used to accelerate similarity calculation. Let the feature matrix be... , its first Behavior ,but: .

[0062] This operation can be accelerated using efficient matrix multiplication libraries (such as BLAS), with a computational complexity of O(n log n). ,in .

[0063] S33, Calculate physical feature similarity, including: S331, calculate the Euclidean distance, which reflects the straight-line distance between two locations in a three-dimensional space of temperature, salinity, and depth. For two physical eigenvectors... and Calculate their Euclidean distance: .

[0064] S332, calculate the Gaussian kernel similarity function. The distance is converted into a similarity using a Gaussian kernel (also known as the radial basis function, RBF): .

[0065] Function characteristic analysis: 1. When When (the two points have exactly the same physical characteristics). ; 2. When hour, ; 3. Similarity decreases smoothly with increasing distance; the rate of decrease is determined by the bandwidth parameter. control.

[0066] S3321, Adaptive Bandwidth Parameters The determination of the Gaussian kernel bandwidth parameter is achieved through a data-driven approach. The basic idea is: The Gaussian kernel should be matched to the degree of data variation, ensuring its effective range covers a reasonable data distribution area. Specifically, this includes: Step a: Calculate the standard deviation of the physical quantities (temperature, depth, salinity) at each sampling point. , , .

[0067] Step b: Calculate the composite standard deviation: This value reflects the average degree of variation of the three physical quantities.

[0068] Step c: Set bandwidth parameters: ,in, As a scaling factor, in a preferred embodiment of the present invention, it is taken as... .

[0069] The theoretical basis for this setting is: based on the Gaussian distribution... "Principles", approximately 68% of the data falls on Within the range, approximately 95% of the data falls within... Within the range. Settings mean: 1. Distance is Point-to-point similarity within ; 2. Distance is Point-to-point similarity within ; 3. Distance exceeds Point-to-point similarity They can be considered dissimilar.

[0070] This setting ensures that the effective range of the Gaussian kernel matches the natural variation range of the data, neither too strict (resulting in almost no dissimilar point pairs) nor too lenient (resulting in a large number of mismatches).

[0071] S333, Calculate the weighted Euclidean distance. Considering the different dimensions and scales of change of the three physical quantities—temperature, salinity, and depth—the weighted Euclidean distance can be used: , The weighting coefficient can be set as the reciprocal of the standard deviation of each physical quantity: .

[0072] This is equivalent to standardizing each dimension before calculating the distance, so that each dimension contributes equally to the total distance.

[0073] S34, weighted fusion of cosine similarity and physical feature similarity yields a fused similarity. Specifically, this includes: S341, for sampling points and Its fusion similarity is defined as: , in: : Flow field image similarity (cosine similarity); Physical feature similarity (Gaussian kernel similarity); Image feature weights; Physical feature weights; Constraints: , .

[0074] S342, Weight parameter settings.

[0075] Preferably, an equal-weighting scheme is used. In the absence of prior knowledge, an equal-weighting scheme is employed: This is the most unbiased choice, giving equal importance to both characteristics.

[0076] Optionally, data-driven weight optimization can be employed. Weight parameters can be optimized based on experimental data. The optimization objective is defined as maximizing the F1 score for position revisit detection. , , The F1 score is defined as follows: , , in, This is a true positive result (retest at the correct location). This is a false positive (retesting at the wrong location). It is a false negative (a repeat visit after a missed test).

[0077] S35, Multi-Constraint Rule-Based Location Revisit Determination. Relying solely on similarity thresholds for revisit determination can easily lead to false positives and redundant detections. This method combines similarity fusion with preset constraint rules for screening, resulting in location revisit constraint factors. The preset constraint rules filter unreasonable revisit pair candidates from multiple dimensions such as time, space, and similarity, improving the reliability and practicality of the detection results.

[0078] The preset constraint rules include: basic location revisit determination conditions, advanced constraint rules, and location revisit performance indicators.

[0079] The basic location revisit criteria include: For the current sampling time and historical moments ( To determine whether they constitute revisit events, the following three basic conditions must be met simultaneously: Condition 1: Time interval constraint: , in, As the minimum time interval threshold, this method is set in a preferred embodiment. One sampling point.

[0080] Physical meaning: 1. AUVs will inevitably be in close proximity within a short period of time, and this proximity does not constitute a meaningful "revisit"; 2. The significance of location revisit detection lies in identifying revisit events with a large time span, thereby providing long-distance constraints; 3. Too short time intervals will result in a large number of trivial and worthless location revisit checks.

[0081] Condition 2: Similarity threshold constraint: , in, To achieve a similarity threshold, in a preferred embodiment of this method, a threshold is set based on the data distribution characteristics. .

[0082] Condition 3: Spatial distance constraint: , in, and For a moment and The horizontal position estimate (given by dead reckoning). As a spatial distance threshold, this method is set in a preferred embodiment. rice.

[0083] Physical meaning: 1. Although the environmental characteristics are similar, if the estimated distance between two locations is too far, they may be similar but different regions (pseudo-revisit). 2. Spatial distance constraints take advantage of the fact that dead reckoning remains reliable in the short term; 3. This constraint prevents mismatches over long distances due to environmental similarity.

[0084] Advanced constraint rules include: To further improve the quality of location revisit detection, this method designs the following high-level constraint rules: Rule A: Initial Protection Period: , in, The preferred embodiment of this method sets the initial protection period length as [the initial protection period length]. One sampling point.

[0085] Design rationale: 1. In the initial stage starting from the trajectory origin, historical data is scarce, making effective revisit matching impossible; 2. In the initial stage, dead reckoning errors are small and no revisit correction is required; 3. Avoid performing meaningless tests when data is insufficient.

[0086] Rule B: Revisit start interval constraint, let the detected revisit set be... For new candidates to revisit ,Require: , in, To minimize the interval between revisiting the starting point, the preferred embodiment of this method sets... One sampling point.

[0087] Design rationale: 1. If a revisit is detected in a certain area, multiple points in the vicinity of that area may also meet the revisit criteria; 2. However, these revisits are redundant, all pointing to the same area; 3. Limit the interval between historical matching points to avoid detecting too many dense revisits in the same area.

[0088] Rule C: Cooldown period after revisit, assuming the last time a revisit pair was detected was... Then the current sampling time Must meet: , in, The preferred embodiment of this method sets the cooling period length as follows: One sampling point.

[0089] Design rationale: 1. When an AUV enters a revisit area, multiple consecutive points may meet the revisit criteria; 2. Without restrictions, this will result in a large number of redundant revisit pairs; 3. The cooling-off period mechanism pauses detection after a revisit event is detected, allowing the AUV to leave the area before resuming detection.

[0090] Rule D: Single revisit principle for each point, for each current time... At most, only one revisit will be recorded, that is, the revisit with the first historical moment that meets the conditions. pair: .

[0091] Design rationale: 1. A current point may be revisited with multiple historical points (e.g., an AUV passes through the same area multiple times). 2. However, recording only one revisit is sufficient to provide adequate constraint information; 3. Recording the first (earliest historical point) allows for the largest possible time span, which is most beneficial for trajectory optimization; 4. Reducing the number of revisit pairs can decrease the complexity of the factor graph and improve optimization efficiency.

[0092] Location revisit performance metrics include: Detection Rate: , This indicates the percentage of sampling points where revisited pairs were detected out of the total sampling points. Among them, Represents a set The cardinality is the number of elements in the set, which is also the number of revisited sampling points detected.

[0093] Average Time Span: , This represents the average time interval between revisited pairs, reflecting the time scale of the revisit constraint.

[0094] Average Spatial Distance: , This represents the average estimated distance between pairs of revisited points (based on dead reckoning).

[0095] True Place Revisiting Accuracy: If a real trajectory is available as a reference, the true distance between revisited point pairs can be calculated: , in, For a moment The true location.

[0096] The detection rate reflects the coverage capability of location revisit detection; a higher value indicates more revisits are detected. A larger average time span indicates a longer time scale for location revisit constraints, which is more beneficial for suppressing long-term drift. A lower true revisit accuracy indicates that the detected revisit pairs are spatially closer to the true revisits. These indicators comprehensively reflect the quality of the location revisit detection module, ultimately manifested in the trajectory RMSE improvement rate after factor graph optimization in S40.

[0097] Finally, S40 is executed, based on factor graph-based optimization estimation: position revisit constraint factors are added to the factor graph to construct a complete factor graph model; the iSAM2 algorithm is used to solve the factor graph to obtain the globally optimized position sequence of the underwater vehicle. Specifically, the steps are as follows: S41, Constructing the Factor Graph Model. A factor graph is a probabilistic graphical model used to represent the factor decomposition structure of a multivariate joint probability distribution. In this method, the factor graph is used to represent the AUV trajectory optimization problem.

[0098] The components of a factor graph are: 1. Variable nodes (circles): Represent the state variables to be estimated, such as the pose of the AUV. ; 2. Factor nodes (boxes): Represent constraints or observations, such as prior information, odometer measurements, and location revisit constraints; 3. Edge: Connects variable nodes and factor nodes, representing the variables involved in the factor.

[0099] S42, Construct a probabilistic model for trajectory optimization. Assume the AUV trajectory is a state sequence. Each state Indicates the first Two-dimensional pose (position + heading) at each moment: , in: Horizontal position coordinates (meters); : Heading angle (radians) is defined as the clockwise angle relative to true north.

[0100] Optimization objective: Based on all observation data (Including prior, odometry, and revisit), solve for the maximum a posteriori (MAP) probability estimate of the state sequence: .

[0101] Based on Bayes' theorem and the conditional independence assumption, the above equation is equivalent to: .

[0102] Assuming the observation noise follows a Gaussian distribution, after taking the negative logarithm, MAP estimation is transformed into a nonlinear least squares problem: , in: : No. The residual function of each factor; The squared Mahalanobis distance weighted by covariance is defined as: .

[0103] S43. Construct a complete factor graph model to represent the trajectory optimization problem of an underwater vehicle. The factor graph contains three types of factors: Prior Factor, Odometry Factor / Between Factor, and Loop Closure Factor. The specific steps include: S431, Calculate the prior factors. The prior factors constrain the first state. The initial value represents prior knowledge of the starting position.

[0104] Residual function: , in, These are the observations of the initial pose.

[0105] Noise model: Assume that the errors of prior observations follow a zero-mean Gaussian distribution: .

[0106] The covariance matrix is ​​a diagonal matrix: .

[0107] In a preferred embodiment of this method, the following settings are provided. rice, radian.

[0108] S432, calculate the odometer factor. The odometer factor constrains the relative motion between adjacent states and is derived from dead reckoning.

[0109] Residual function: , in: : Relative transformation operations on the SE(2) group (detailed below); : Relative displacement measurements obtained from dead reckoning.

[0110] SE(2) relative transformation operation: Given two poses and Their relative transformation is defined as: , in: , , .

[0111] The first two equations will increment the position in the global coordinate system. Transform to local coordinate system (with) (where the direction is ), thus obtaining the forward displacement. and lateral displacement .

[0112] Angle difference Normalization is required to make it fall within the range of Within the range: .

[0113] The speed measured using a DVL velocimeter can be obtained directly from the sensor data integration. Angular velocity measured by IMU By integrating over time, we obtain: , in, This represents the sampling time interval.

[0114] Noise model: Odometer measurement errors mainly originate from sensor noise (DVL velocity error, IMU drift) and environmental interference (DVL misjudgment caused by ocean currents and changes in water density). Assume the measurement error follows a zero-mean Gaussian distribution: , The covariance matrix is: .

[0115] In a preferred embodiment of this method, the following settings are provided. to rice, to radian.

[0116] S433, the location revisit constraint factor is obtained from S30. The location revisit constraint factor requires that the two locations identified by the location revisit detection be as consistent as possible.

[0117] Residual function: For the detected revisit pairs ( Add a location revisit constraint: , in, For revisited measurements, the relative transformation between two poses is represented.

[0118] Preferably, the method for determining the revisited measurement value is zero constraint. It is assumed that the revisited points are at the same location, i.e., the relative transformation is zero. , The residual then simplifies to: , This constraint requires That is, the two poses should overlap as much as possible.

[0119] Applicable scenarios: 1. Location revisit detection is based on the similarity of environmental features; points with high similarity should theoretically be at the same location. 2. Simple to implement, requiring no additional calculation of relative pose; 3. A large noise variance needs to be set to tolerate actual location differences.

[0120] Preferably, the revisited measurement values ​​are determined using the true relative pose. If a true trajectory is available as a reference (for simulation verification only), the true relative transformation between the two positions can be calculated: , This constraint more accurately reflects the actual relationship between the two positions.

[0121] Applicable scenarios: 1. Simulation verification phase, used to evaluate the upper limit of the performance of the optimization algorithm; 2. In practical applications, it is usually impossible to obtain the true relative pose.

[0122] Noise model: The uncertainty of the revisit constraint stems from: 1. The non-uniqueness of environmental characteristics (different locations may have similar environmental characteristics); 2. Sensor measurement noise; 3. Temporal changes in the environmental field.

[0123] Assume that the location revisit measurement error follows a zero-mean Gaussian distribution: .

[0124] The covariance matrix is: .

[0125] Key parameter settings: Because the actual distance of the revisit point may be several meters to tens of meters (rather than strictly zero) under the zero-constraint approach, a large noise variance needs to be set to avoid optimization failure due to over-constraint. In the preferred embodiment of this method, the following settings are configured: .

[0126] S434, using the above three types of factor combinations, constructs a complete factor graph model. Assume... Each state node ( ), If there are 10 revisited pairs, then the factor graph contains: One prior factor: ; Odometry factors: ; Location revisit constraint factor: .

[0127] The overall optimization objective function is: .

[0128] S44. The iSAM2 algorithm is used to solve the complete factor graph to obtain the globally optimized position sequence of the underwater vehicle. Specifically, the following steps are included: S441, Set the key parameters of the iSAM2 algorithm. This method uses the GTSAM (Georgia Tech Smoothing and Mapping) library to implement iSAM2 optimization. The key parameter settings are as follows: a. Relinearization Threshold: When the update amount of a variable exceeds this threshold, relinearization is triggered. A smaller threshold ensures linearization accuracy but increases computational cost. b. Relinearize Skip: Setting it to 1 means checking for relinearization on every update. Setting it to a larger value (such as 5 or 10) can reduce the frequency of relinearization, speeding up computation but potentially reducing accuracy; c. The Factorization Method employs QR decomposition: QR decomposition has better numerical stability than Cholesky decomposition and is suitable for ill-conditioned matrices; d. Cache Linearized Point: Cache linearized factors to avoid redundant calculations and improve efficiency.

[0129] S442, using the configured iSAM2 algorithm to solve the trajectory optimization problem, obtain the current optimization estimation result.

[0130] S443, iterate through all time points from 0 to N, obtain all optimized poses of the underwater vehicle at all time points, and get the position sequence.

[0131] The code for extracting the optimal estimate of all state variables from the iSAM2 optimizer is as follows: result = isam.calculateEstimate() optimized_trajectory = [] for k in range(N+1): pose = result.atPose2(k) x_k = pose.x() y_k = pose.y() theta_k = pose.theta() optimized_trajectory.append([x_k, y_k, theta_k]).

[0132] S444, Evaluation metric calculation. If a real trajectory is available as a reference, the root mean square error (RMSE) can be calculated for evaluation. , , in, Indicates: Root mean square error in the XY plane; Indicates: 3D root mean square error; Indicates: the optimized value of the X-axis position; Indicates: the true value of the X-axis position; Indicates: the optimized value of the Y-axis position; Indicates: the actual value of the Y-axis position; Indicates: the optimized value of the Z-axis position; Indicates: the true value of the Z-axis position.

[0133] RMSE reflects the overall deviation between the optimized trajectory and the true trajectory.

[0134] In summary, the positioning method provided in this embodiment constructs a multimodal environment description by fusing ADCP current profile data and CTD temperature, salinity, and depth data. It uses the ResNet50 model to extract visual features from the flow field image and combines them with the physical features of temperature, salinity, and depth for weighted fusion position revisit detection. A multi-constraint rule filtering mechanism ensures the reliability of the position revisit detection results. Furthermore, the iSAM2 incremental factor graph optimization framework is used to globally correct the cumulative drift error of dead reckoning. This enables AUVs to position themselves in underwater environments with limited visibility and sound without the need for external positioning infrastructure support, meeting the trajectory accuracy requirements of long-endurance unmanned operations such as deep-sea resource exploration, large-scale marine environmental monitoring, and underwater target search.

[0135] Experimental verification Experimental data and environment setup (1) Data source This method is validated using the ECCO (Estimating the Circulation and Climate of the Ocean) V4R4 ocean reanalysis dataset. ECCO is based on the ocean circulation model (MITgcm) and incorporates a large amount of satellite and field observation data.

[0136] ECCO dataset characteristics: 1. Spatial coverage: Global oceans; 2. Horizontal resolution: LLC90 grid, approximately 1° (approximately 111 km); 3. Vertical layering: 50 layers, with depths ranging from 5 meters to 5906 meters; 4. Time resolution: Daily average; 5. Includes variables such as three-dimensional flow velocity (U, V, W), temperature, salinity, and sea surface height.

[0137] (2) Data preprocessing and scaling The horizontal resolution of the raw ECCO data (approximately 111 kilometers) is much larger than the typical operating scale of AUVs (several kilometers to tens of kilometers). Direct use of this data would result in environmental characteristics changing too slowly and failing to reflect spatial heterogeneity.

[0138] The original data is horizontally scaled, with the original grid coordinates scaled at a ratio of 1:2000: .

[0139] After scaling, the grid spacing is approximately 55 meters, and the simulation area is approximately 5 kilometers × 5 kilometers, which matches the actual operating scale of AUVs.

[0140] (3) Combination Figure 3 and Figure 4 As shown, to verify the advantages of this method, it is compared with the traditional extended Kalman filter + RTS smoothing method.

[0141] from Figure 3 As can be seen, the red dashed line (determined dead reckoning trajectory) shows significant drift and expansion relative to the dark blue line (actual trajectory), with the trajectory expanding outwards overall. This indicates that the cumulative error of dead reckoning increases over time, resulting in severe distortion of the trajectory shape. While the light blue line (EKF+RTS trajectory) shows some improvement over dead reckoning, it still exhibits significant deviations in several sections, particularly on the left and lower sides of the trajectory. The green line (the trajectory of this method) almost perfectly overlaps with the actual trajectory, accurately following its movement, especially at multiple curves and intersections.

[0142] from Figure 4 As can be seen, the red dashed line (dead reckoning error) continuously increases with the advancement of the trajectory points, with the maximum error exceeding 400 meters, exhibiting a typical cumulative divergence trend. The blue line (EKF+RTS) performs reasonably well in the first half, but after point 1250, it experiences several large fluctuations, with the peak error approaching 300 meters, indicating insufficient stability during long-term operation. The green line (this method) maintains a consistently low error level, mostly within 50 meters, and the error rapidly decreases after the position revisit detection is triggered, indicating that the position revisit constraint effectively corrects the drift.

[0143] In summary, our proposed method significantly outperforms dead reckoning and the EKF+RTS method in both positioning accuracy and long-term stability. Dead reckoning suffers from unavoidable cumulative drift; while EKF+RTS can correct errors to some extent, it is limited by the local optima of the filter, making it prone to error divergence and instability during long-range operations. Our proposed method, through global optimization of the factor graph and the introduction of position revisit constraints, can consistently control the position error at a low level, validating the effectiveness and superiority of our method for AUV positioning in underwater environments with limited visibility and sound.

[0144] Combination Figure 5 This disclosure provides a positioning system 50 for underwater environments, including: a marine environment data acquisition and preprocessing module 51, configured to acquire current profile data detected by an acoustic Doppler current profiler at sampling points, as well as temperature data, salinity data, and depth data detected by a CTD sensor, and preprocess each data to remove abnormal data; and a multimodal environment description construction module 52, configured to convert the current profile data into a flow field image, extract features from the flow field image using a ResNet50 model, and normalize the extracted features to obtain a normalized flow field feature vector corresponding to each sampling point; simultaneously, for each sampling point, a three-dimensional physical feature vector is constructed using temperature data, salinity data, and depth data and normalized accordingly. The position revisit detection module 53, based on multimodal feature fusion, is configured to calculate the cosine similarity of the normalized flow field feature vector and the physical feature similarity of the normalized three-dimensional physical feature vector; weightedly fuse the cosine similarity and physical feature similarity to obtain a fused similarity; and filter the fused similarity and preset constraint rules to obtain position revisit constraint factors. The factor graph-based optimization estimation module 54 is configured to add the position revisit constraint factors to the factor graph to construct a complete factor graph model; and solve the factor graph using the iSAM2 algorithm to obtain the globally optimized position sequence of the underwater vehicle.

[0145] The specific implementation process of this system can be found in the description of the above method embodiments, and will not be repeated here.

[0146] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0147] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A positioning method for underwater environments, characterized in that, include: S10, Marine environmental data acquisition and preprocessing: Acquire current profile data detected by acoustic Doppler current profiler at sampling points, as well as temperature data, salinity data and depth data detected by CTD sensor, and preprocess each data to remove abnormal data; S20, Multimodal Environment Description Construction: The flow velocity profile data is converted into a flow field image, the features of the flow field image are extracted using the ResNet50 model, and the extracted features are normalized to obtain the normalized flow field feature vector corresponding to each sampling point; at the same time, for each sampling point, a three-dimensional physical feature vector is constructed using the temperature data, the salinity data, and the depth data and normalized. S30, Location revisit detection based on multimodal feature fusion: Calculate the cosine similarity of the normalized flow field feature vector and the physical feature similarity of the normalized three-dimensional physical feature vector; weight and fuse the cosine similarity and the physical feature similarity to obtain the fused similarity; combine the fused similarity with preset constraint rules to filter and obtain the location revisit constraint factor; S40, Optimization estimation based on factor graph: Add the location revisit constraint factor to the factor graph to construct a complete factor graph model; The iSAM2 algorithm is used to solve the complete factor graph to obtain the globally optimized position sequence of the underwater vehicle.

2. The positioning method for underwater environments according to claim 1, characterized in that, The preprocessing of each data point to remove outlier data includes: Use moving average filtering or Gaussian filtering to remove high-frequency noise from each data set. Based on the attitude rotation matrix of the underwater vehicle and the installation offset of the CTD sensor relative to the center of the underwater vehicle, spatial registration is performed on the installation position deviation of the CTD sensor so that the data detected by the acoustic Doppler current profiler and the CTD sensor are aligned in time and space.

3. The positioning method for underwater environments according to claim 1, characterized in that, Converting the velocity profile data into a flow field image includes: Traverse all sampling points and collect the horizontal velocity components of all the aforementioned velocity profile data. and : ;in, For sampling points; Calculate the global minimum value of each horizontal velocity component. and global maximum value ; The quantile range is used instead of the global minimum and the global maximum. The flow field image is generated by mapping the flow velocity values, which are represented by quantile ranges, to RGB colors using Jet chromatography.

4. The positioning method for underwater environments according to claim 1, characterized in that, The extraction of features from the flow field image using the ResNet50 model includes: The flow field image is preprocessed, including: size adjustment, channel normalization, and tensor transformation; The preprocessed flow field image is input into the ResNet50 model, and features are extracted through forward propagation. ， in, This is the original feature vector output by the global average pooling layer; This is the data obtained after normalization.

5. A positioning method for underwater environments according to claim 1, characterized in that, The calculation of the cosine similarity of the normalized flow field feature vector and the physical feature similarity of the normalized three-dimensional physical feature vector includes: The cosine similarity is: ， in, and The respective i , j The flow field feature vector is extracted and normalized from the flow field image of each sampling point; Construct a similarity matrix: Matrix operations are used to accelerate the calculation of the cosine similarity; Calculating the physical feature similarity includes: calculating Euclidean distance, Gaussian kernel similarity, and weighted Euclidean distance; The step of determining the bandwidth parameter of the Gaussian kernel similarity includes: Calculate the standard deviation of temperature, salinity, and depth data for each sampling point; Calculate the composite standard deviation based on the standard deviations described above; The bandwidth parameters ,in, Scaling factor The combined standard deviation is given.

6. A positioning method for underwater environments according to claim 5, characterized in that, The step of weightedly fusing the cosine similarity and the physical feature similarity to obtain a fused similarity includes: For sampling points i and j The fusion similarity is: ， in, The cosine similarity is given. Gaussian kernel similarity; Image feature weights; Weights for physical features; Constraints: , ; in, Alternatively, optimize based on experimental data. and .

7. A positioning method for underwater environments according to claim 6, characterized in that, The preset constraint rules include: basic location revisit determination conditions, advanced constraint rules, and location revisit performance indicators; wherein... The basic location revisit determination criteria include: Time interval constraints: ,in, The minimum time interval threshold, t At the current sampling time, i A historic moment; Similarity threshold constraint: ,in, To fuse similarity thresholds, The fusion similarity at the current sampling time; Spatial distance constraints: ,in, and For a moment and Horizontal position estimation, Spatial distance threshold; The advanced constraint rules include: Initial protection period: ,in, The initial protection period length; Location revisit event start interval constraint: Let the set of detected revisit pairs be . For new candidates to revisit ,Require: ,in, The minimum interval for revisiting the starting point; Cooldown period after revisit event: Let the last time a revisit was detected be... Then the current sampling time Must meet: ,in, This refers to the length of the cooling period; Single-point revisit principle: For each current sampling time... , with the first historical moment that meets the conditions pair, ; The location revisit performance metrics include: Detection rate: This indicates the percentage of revisited sampling points out of the total sampling points; where, Represents a set The cardinality; Average time span: , representing the average time interval between revisited pairs, reflects the time scale of the location revisit constraint; Average spatial distance: , representing the average estimated distance between pairs of revisited points; True revisit accuracy: ,in, For a moment The true location.

8. A positioning method for an underwater environment according to any one of claims 1 to 7, characterized in that, The construction of the complete factor graph model includes: Calculate prior factors: ,in, , which is the observed value of the initial pose; Calculate the odometer factor: ,in, For relative transformation operations on the SE(2) group; , which is the relative displacement measurement value obtained from dead reckoning; Using one of the aforementioned prior factors N odometry factors and M location revisit constraint factors Construct the factor graph model; The overall optimization objective function is: 。 9. A positioning method for underwater environments according to any one of claims 1 to 7, characterized in that, The iSAM2 algorithm is used to solve the complete factor graph to obtain the globally optimized position sequence of the underwater vehicle, including: Set the key parameters of the iSAM2 algorithm; The factor graph is solved using the configured iSAM2 algorithm to obtain the current optimized estimation results; Iterate through all time points from 0 to N to obtain all optimized poses of the global underwater vehicle at all time points, thus obtaining the global position sequence.

10. A positioning system for underwater environments, characterized in that, include: The marine environmental data acquisition and preprocessing module is configured to acquire current profile data detected by an acoustic Doppler current profiler at the sampling point, as well as temperature data, salinity data and depth data detected by a CTD sensor, and preprocess each data to remove abnormal data. The multimodal environment description construction module is configured to convert the flow velocity profile data into a flow field image, extract features from the flow field image using a ResNet50 model, and normalize the extracted features to obtain a normalized flow field feature vector corresponding to each sampling point; simultaneously, for each sampling point, a three-dimensional physical feature vector is constructed using the temperature data, the salinity data, and the depth data and normalized accordingly. The location revisit detection module based on multimodal feature fusion is configured to calculate the cosine similarity of the normalized flow field feature vector and the physical feature similarity of the normalized three-dimensional physical feature vector; to perform weighted fusion of the cosine similarity and the physical feature similarity to obtain a fusion similarity; and to combine the fusion similarity with preset constraint rules to obtain a location revisit constraint factor. The factor graph-based optimization estimation module is configured to add the location revisit constraint factor to the factor graph to construct a complete factor graph model. The iSAM2 algorithm is used to solve the complete factor graph to obtain the globally optimized position sequence of the underwater vehicle.

Images