Underwater SLAM method and system combining region-level descriptor and LSH retrieval

Abstract

The invention discloses an underwater SLAM method and system combining a region-level descriptor and LSH retrieval, belongs to the technical field of autonomous navigation of underwater robots, and solves the problems that track plotting errors can be accumulated along with time due to particularity of an underwater environment, and a traditional SLAM method lacks an effective position sensing mechanism to suppress the drift, so that the navigation accuracy is poor. And meanwhile, an effective preprocessing mechanism for improving the image quality is lacked. The method comprises the following steps: acquiring and preprocessing a polar coordinate image; acquiring a prior pose of a current key frame, inputting the prior pose to a position sensing ICP, performing scanning matching on point clouds of adjacent key frames in sequence, and establishing a sequence constraint relation; extracting an occupied area from the historical key frame point cloud, constructing an area descriptor, carrying out neighbor retrieval, screening candidate key frames, verifying matching, and generating an effective loopback constraint; and constructing a factor graph, performing global graph optimization, accumulating and updating environment point clouds, and generating an underwater environment map. The method is suitable for the underwater robot in the scenes of ocean exploration, environment monitoring, search and rescue operation and the like.

Description

Technical Field

[0001] This invention belongs to the field of underwater robot autonomous navigation and positioning technology, specifically relating to an underwater SLAM method that combines region-level descriptors and LSH retrieval. Background Technology

[0002] With the development of marine resource development and marine scientific research, underwater robots are increasingly being used in marine exploration, environmental monitoring, search and rescue operations, and other fields. The underwater environment, characterized by its inability to be penetrated by GPS signals, low visibility, and complex and variable conditions, presents significant challenges to the autonomous navigation and positioning of underwater robots.

[0003] Current underwater SLAM relies heavily on acoustics in complex underwater environments. Existing underwater robot SLAM technologies primarily depend on acoustic sensors for environmental perception. However, these traditional methods face numerous technical bottlenecks in complex underwater environments, making it difficult to meet the demands for high-precision autonomous navigation. On one hand, traditional feature extraction methods lack sufficient accuracy in extracting target features from sonar images and are easily affected by underwater noise and multipath effects, resulting in poor feature point quality and directly impacting subsequent matching accuracy. On the other hand, existing loop closure detection methods are mainly based on geometric or statistical features, which are prone to mismatches in complex underwater environments and have high computational complexity, making them difficult to meet real-time requirements. Furthermore, due to the unique characteristics of the underwater environment, trajectory estimation errors accumulate over time. Traditional SLAM methods lack effective position awareness mechanisms to suppress this drift, and sonar images often suffer from low contrast, high noise, and interference from carrier reflections, lacking effective preprocessing mechanisms to improve image quality.

[0004] In summary, the existing technology has the following technical problems: 1. Traditional underwater SLAM methods are susceptible to noise and dynamic interference in sonar images, resulting in unstable feature extraction, sparse point cloud input with severe noise redundancy, and difficulty in supporting the accuracy requirements of subsequent matching and mapping.

[0005] 2. Conventional ICP algorithms lack effective constraints on pose priors and search range, which can easily lead to incorrect matching and cumulative drift, causing trajectory divergence and localization failure in long-term tasks.

[0006] 3. In the loop closure detection stage, most existing methods rely on a single geometric similarity metric, which makes it difficult to stably detect and verify historical loops in sparse sonar point clouds, resulting in difficulty in ensuring global consistency.

[0007] 4. In terms of map construction, single-resolution TSDF cannot balance long-distance coverage and close-range detail representation, which affects both the overall usability of navigation and reduces the visualization accuracy of local surfaces. Summary of the Invention

[0008] This invention addresses the need for high-precision autonomous navigation of underwater robots. To solve the problem that due to the special nature of the underwater environment, trajectory estimation errors accumulate over time, and traditional SLAM methods lack effective position awareness mechanisms to suppress this drift, as well as effective preprocessing mechanisms to improve image quality, this invention proposes an underwater SLAM method that combines region-level descriptors and LSH retrieval. Through improved feature extraction, position-aware scan matching, and loop closure detection based on occupied regions, the accuracy and robustness of the underwater SLAM system are significantly improved.

[0009] Firstly, the present invention aims to provide an underwater SLAM method that combines region-level descriptors with LSH retrieval, comprising the following steps: S1: Obtain the polar coordinate image and perform preprocessing to obtain a two-dimensional feature point cloud; S2: Obtain the prior pose of the current keyframe based on the track, input the prior pose into the position-aware ICP, scan and match the point clouds of adjacent keyframes in sequence within the restricted search space to obtain the relative pose, and establish the order constraint relationship. S3: Collect historical keyframes, use DBSCAN clustering to extract occupied regions from the point clouds of historical keyframes, construct region descriptors, and perform nearest neighbor retrieval based on LSH to filter candidate keyframes that are separated in time and verify the matching, and generate effective loop closure constraints. S4: Construct a factor graph based on the sequential constraints and effective closure constraints, and use GTSAM for global graph optimization. At the same time, use the optimized keyframe poses to accumulate and update the environmental point cloud to generate an underwater environment map.

[0010] Furthermore, a preferred solution is provided: the preprocessing includes: removing carrier reflection using background subtraction, removing speckle noise, enhancing contrast with CLAHE and reducing noise with Gaussian filtering, employing an improved CFAR detection method, combining OTSU segmentation and dynamic target removal to extract stable feature points, and converting them into Cartesian coordinate point clouds. After downsampling and isolated point removal, a two-dimensional feature point cloud is generated to provide input for SLAM.

[0011] Furthermore, a preferred embodiment is provided: S2 includes: S21: Obtain the initial pose based on the trajectory and pre-place the current frame point cloud into the coordinate system of the previous frame; S22: Limit the search radius of the matching based on the prior error, so that the correspondence is established only within a reasonable neighborhood; S23: Determine the pose covariance and inter-frame time, and stipulate that the nearest neighbor is only searched within the radius, and those outside the radius are not included in the registration; S24: Perform point-to-point ICP within a restricted neighborhood. In each round, KD-Tree is used to establish nearest neighbor correspondences, a distance threshold is set to eliminate far pairs, and the rigid body increment is estimated and the attitude is updated using the closed-form solution of weighted SVD. At the same time, the root mean square error is used as the convergence criterion, and upper limits are set for single-step translation and rotation to constrain the iteration stability. S25: After registration, the overlap rate and the final root mean square error are used as acceptance criteria. If the criteria are met, the relative pose is obtained and written as the order factor; if the acceptance fails, the solution is continued in the order of hierarchical backtracking.

[0012] Furthermore, a preferred solution is provided: In step S25, continuing the solution in a step-back order includes the following steps: First, switch to standard ICP to expand the available correspondences; if still insufficient, then use sampling-covariance ICP to enhance robustness to sparse or anisotropic point sets by first downsampling to equalize the distribution and then solving based on local covariance weighting; if still insufficient, use NDT to give a new coarse pose as the initial pose, and then return to restricted ICP for refinement.

[0013] Furthermore, a preferred embodiment is provided: S3 includes: S31: Cluster the two-dimensional feature point cloud of the current keyframe into several occupied regions according to density, and calculate a region-level descriptor for each region; S32: Build a nearest neighbor index for the region descriptor extracted from historical keyframes, and record the corresponding keyframe number in the index entry; S33: For each descriptor in the current frame, perform a nearest neighbor query to calculate the weighted Euclidean distance with the historical descriptors, and convert it into similarity. S34: Calculate similarity scores based on descriptors and accumulate the scores to the corresponding keyframes to form a weighted vote; S35: Perform position-aware ICP verification on candidate frames, using the pose calculated from the previous dead center as the initial value, establish neighborhood correspondence, iteratively update rigid body transformation, and control convergence with root mean square error, single-step translation and rotation upper limit. S36: Effectively determine loop relationships.

[0014] Furthermore, a preferred embodiment is provided: S4 includes: S41: Under the constraints of sequence and effective loop closure, perform global optimization on the keyframe map to achieve a consistent pose sequence; S42: Construct a map based on the optimized keyframe relationships; S43: Register the two-dimensional feature point cloud and complete the incremental update of two types of maps in a unified coordinate system. One type is occupied grid and the other is multi-resolution TSDF. S44: Using the optimized keyframe pose as a reference, the two-dimensional feature point cloud obtained in S1 is mapped onto two map representations, namely the occupied grid and TSDF, to generate an underwater environment map.

[0015] Furthermore, a preferred solution is provided: the occupied grid uses the observed ray as the basic geometric unit and is updated in a logarithmic probability form; the TSDF uses two sets of voxel grids to collaboratively fuse the point cloud.

[0016] Secondly, the purpose of this invention is to provide an underwater SLAM system that combines region-level descriptors and LSH retrieval. The system is implemented using an underwater SLAM method combining region-level descriptors and LSH retrieval as described in any one or more of the above-mentioned schemes. The system includes: Image processing module: Acquires polar coordinate images and performs preprocessing to obtain two-dimensional feature point clouds; Sequence constraint module: It is used to obtain the prior pose of the current key frame based on the track, input the prior pose into the position-aware ICP, scan and match the point clouds of adjacent key frames in sequence within the restricted search space to obtain the relative pose, and establish the sequence constraint relationship. The loop closure constraint module is used to collect historical keyframes, extract the occupied regions from the point cloud of historical keyframes using DBSCAN clustering, construct region descriptors, perform nearest neighbor retrieval based on LSH, filter candidate keyframes that are separated in time and verify the matching, and generate valid loop closure constraints. Map generation module: It is used to construct a factor graph based on sequential constraints and effective closure constraints, and uses GTSAM for global graph optimization. At the same time, the optimized keyframe poses are used to accumulate and update the environmental point cloud to generate robot trajectory and map.

[0017] Thirdly, the present invention aims to provide a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes an underwater SLAM method combining region-level descriptors and LSH retrieval according to any one or more of the above-described schemes.

[0018] Fourthly, the present invention aims to provide a computer-readable storage medium for storing a computer program that executes an underwater SLAM method combining region-level descriptors and LSH retrieval as described in any one or more of the above-described schemes.

[0019] Compared with the prior art, the advantages of the present invention are: The method described in this invention, by combining improved CFAR detection with OTSU adaptive thresholding in forward-looking sonar images, effectively suppresses noise and dynamic interference in complex underwater environments, extracting stable and dense feature point clouds. This method not only improves the usability of sonar observations but also generates high-quality two-dimensional point cloud input after point cloud transformation and downsampling, providing a solid foundation for subsequent SLAM matching and mapping. Within this framework, the system further utilizes DBSCAN clustering and occupancy region descriptors to quickly filter historical keyframes, achieving efficient retrieval and loop closure candidate selection for large-scale sonar data. This maintains long-term consistency between trajectory estimation and map construction in large-scale underwater missions with relatively low computational overhead.

[0020] This invention introduces position-aware ICP into sequential scan matching and loop closure verification. By dynamically controlling the search radius through adaptive thresholding, and combining this with neighborhood correspondence, convergence criteria, single-step translation, and upper limits for turning angles, it effectively suppresses incorrect matching and cumulative drift, thereby improving the reliability of relative pose estimation. Furthermore, the system proposes a dual-resolution TSDF fusion strategy in the map building stage. This involves using low-resolution voxels to ensure coverage in distant areas and high-resolution voxels to recover fine surfaces in nearby areas, combined with ray tracing updates to achieve a unified representation of the occupied grid and the TSDF map. This mechanism significantly improves map quality and robustness, enabling the system to stably output high-precision navigation trajectories and environmental maps even in environments with strong sonar noise and sparse observations.

[0021] This invention is applicable to underwater robots in scenarios such as marine exploration, environmental monitoring, and search and rescue operations. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating an underwater SLAM method combining region-level descriptors and LSH retrieval, as described in a specific embodiment of the present invention. Figure 2 This is a flowchart of S2 as described in a specific embodiment of the present invention; Figure 3 This is a flowchart of S3 as described in a specific embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the key inter-frame constraint relationships described in a specific embodiment of the present invention. Detailed Implementation

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

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

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

[0026] Implementation Method 1: An underwater SLAM method combining region-level descriptors and LSH retrieval includes the following steps: Step 1: Acquire polar coordinate images using forward-looking sonar and perform preprocessing, including: Background subtraction is used to remove carrier reflection, speckle noise removal, CLAHE contrast enhancement, and Gaussian filtering for noise reduction. Then, an improved CFAR detection combined with OTSU segmentation and dynamic target removal is used to extract stable feature points, which are then converted into Cartesian coordinate point clouds. After downsampling and outlier removal, a high-quality two-dimensional feature point cloud is generated to provide input for SLAM.

[0027] Specifically: First, raw image data in polar coordinates is acquired using forward-looking sonar. To remove background noise from carrier reflections and water waves, a background map updated over time is created and subtracted from the current frame, retaining only the true echo information. Simultaneously, small-scale smoothing and contrast enhancement are applied to reduce speckle interference and sharpen target boundaries. After obtaining a clean image, the system employs an improved CFAR adaptive detection method, automatically adjusting the detection threshold based on the surrounding noise level to ensure both weak target detection and false alarm suppression. This is combined with OTSU threshold segmentation to stabilize blocky targets while eliminating short-term dynamic interference such as bubbles. Finally, the detected pixels are directly analyzed using relational methods. The data is converted to Cartesian coordinates in meters, and then downsampled and isolated points are removed to form a two-dimensional feature point cloud with uniform density and less noise.

[0028] Specifically, the improved CFAR adaptive detection method includes the following steps: First, according to the given training units Protection unit False alarm rate of target Unified threshold factor calculation (CA by) GOCA / SOCA / OS use their respective false alarm functions to numerically solve for... To ensure that after fixing false alarms, symmetrical sliding windows are used on the polar coordinate amplitude graph. Each protection, both sides Estimating the background for each training sample ,according to Candidates are obtained through pixel-by-pixel decision-making. The improvements are reflected in three aspects: the threshold factor is... Unified calibration, statistics, and decision-making are all completed in the polar domain to avoid geometric distortion. Dynamic masks are applied to CFAR candidates by frame timestamp for subtraction and recording, forming traceable online quality inspection results.

[0029] First, standardize the threshold factors for various CFAR types, directly implementing fixed false alarms into the parameters: given the number of training units. Number of protection units and target false alarm rate For unit-average type in an exponential context:

[0030] Analytical solution For the strongest / minimum side average and ordinal statistical types, the roots are obtained by finding the values ​​of their respective false alarm functions. The testing phase will uniformly adopt the following methods:

[0031] As a pixel threshold, Background estimation for window training units.

[0032] Secondly, CFAR detection is performed using a symmetrical sliding window on the polar coordinate amplitude map. Centered on the pixel to be detected, two windows are placed to the left and right. Each protection unit is located on its outer left and right sides. Estimating the background using one training unit: The unit average is:

[0033] Lateral averages calculated separately And take according to the algorithm or Suppress unilateral interference; take the first sorted number for the sequence statistics. Level Sample As a stable background Compare amplitude pixel by pixel. With threshold Obtain CFAR candidate masks With candidate points, and load a dynamic mask with the same name according to the frame timestamp. Perform the deduction to obtain This is a result of online visualization and quality statistics, and this branch does not output the final point cloud.

[0034] Then, the intensity-based thresholding branch is run in parallel, also in the polar domain: the local adaptive threshold is adopted as follows:

[0035] Generate pixel-level candidate masks Global Otsu threshold By maximizing inter-class variance:

[0036] The lower bound of the entire field is obtained, which is used to impose conservative constraints on intensity candidates when necessary. The current implementation does not fuse with the CFAR mask, and the pixel candidates for generating the point cloud come from... And can be subject to Lower bound constraints. Finally, the pixels retained from the intensity segmentation are mapped to Cartesian coordinates according to the device calibration relationship:

[0037] An initial point set is obtained; voxel downsampling and out-of-radius point removal are performed on the point cloud to obtain a geometrically consistent and evenly distributed two-dimensional feature point cloud for subsequent registration and SLAM.

[0038] Step 2: Based on the trajectory calculation, obtain the pose prior of the current keyframe, input the prior pose into the position-aware ICP, and complete the sequential scanning and matching of the point clouds of adjacent keyframes within the restricted search space to obtain robust relative poses and establish sequential constraint relationships.

[0039] Specifically: In step two, a reliable relative pose is estimated between adjacent keyframes. The initial pose is first calculated from the flight path. The current frame's point cloud is pre-placed into the coordinate system of the previous frame; then, the search radius for matching is limited based on prior error, ensuring that the correspondence is established only within a reasonable neighborhood. The pose covariance and inter-frame time are combined to provide the following:

[0040] And it is stipulated that only within the radius The nearest neighbor is searched within the register, and those outside the register are not included in the registration process, thus suppressing erroneous correspondences at the source.

[0041] Perform point-to-point ICP within a restricted neighborhood, that is, establish nearest neighbor correspondences using a KD-Tree in each round and set a distance threshold. Eliminating distant pairs, the rigid body increment is estimated using the closed-form solution of weighted SVD. The pose is updated; the root mean square error (RMSE) is used as the convergence criterion, and upper limits are set for single-step translation and rotation to constrain iterative stability. After registration, the overlap rate is used as follows: , The final RMSE is used as the acceptance criterion; if the standard is met, the relative pose is obtained. And write it as an ordinal factor.

[0042] If the acceptance test fails, the solution continues in a tiered rollback sequence: first, switch to standard ICP to expand the available correspondences; if still insufficient, use sampling-covariance ICP to enhance robustness to sparse or anisotropic point sets by first downsampling to equalize the distribution and then solving based on local covariance weighting; if still unsatisfactory, use NDT to provide a new coarse pose as... Then, the process returns to the restricted ICP refinement. This approach first relaxes the search to obtain usable solutions, and then converges to a refined solution under prior constraints. This keeps the computation focused on the most likely local range, reducing erroneous correspondences and drift propagation, and maintaining stable relative pose estimation in the context of sonar noise.

[0043] Step 3: For the historical keyframe point cloud, DBSCAN clustering is used to extract the occupied region, construct the region descriptor, and perform nearest neighbor retrieval based on LSH. Candidate keyframes that are separated in time are screened and the matching is verified, thereby generating effective loop closure constraints.

[0044] Specifically: In step three, based on the adjacent keyframe order constraints obtained in step two, it is further determined whether there is a loop between the current keyframe and historical keyframes. First, the two-dimensional feature point cloud of the current keyframe is... Clustered by density into several occupied regions And compute a region-level descriptor for each region. For any Record its number as Then the positional mean Covariance Maximum radius areal density ;Main direction The largest eigenvector used Ratio of eigenvalues Quantization, direction angle atan2 Then, the radial distance to the center is squared to obtain the bins. Final descriptor writing: , It remains stable despite changes in sonar viewing angle and local occlusion, making it suitable for coarse detection of similar regions. A nearest neighbor index is built for the region descriptors extracted from historical keyframes, and the corresponding keyframe number is recorded in the index entries. For each of the current frames... Nearest neighbor query calculation and history descriptor Weighted Euclidean distance And convert it to similarity: , Subsequently, during loop closure detection, the system calculates similarity scores based on these descriptors and accumulates these scores onto the corresponding keyframes, forming a weighted vote. To avoid misclassifying temporally close neighboring frames as loop closures, temporal separation constraints are also implemented. Only candidate frames that meet the time interval condition will enter the sorting and verification process.

[0045] Subsequently, position-aware ICP verification is performed on the candidate frames. The pose is then calculated from the previous dead center. Using the initial value, we follow the restricted ICP approach from step two, but adopt a more relaxed search radius in the loop-based scenario. Establish neighborhood correspondences within the model, and then iteratively update the rigid body transformation. The convergence is controlled by RMSE, single-step translation, and upper limits for rotation angle. Finally, the validity determination also follows the acceptance system from step two. Specifically, the loop relationship is considered valid only when the number of matching points exceeds a preset lower limit, the overlap ratio meets the requirements, the displacement amplitude does not exceed the allowable translation threshold, the rotation angle is within the specified range, and the overall error is lower than the convergence criterion. Loop constraints that meet these conditions are formally incorporated into the factor graph, participating in global optimization along with sequence constraints, thereby effectively offsetting accumulated drift and improving trajectory consistency.

[0046] Step four involves constructing a factor graph from sequential scan matching constraints and loop closure constraints, using GTSAM for global graph optimization, and then using the optimized keyframe poses to accumulate and update the environmental point cloud, generating a more consistent robot trajectory and map.

[0047] Specifically: the keyframe pose set is obtained through global optimization. Subsequently, the occupied grid is synchronously published and multi-resolution TSDF maps are maintained within the same coordinate system. Under the combined effect of the sequence constraints and closure constraints established in steps two and three, the system performs global optimization on the keyframe map to obtain a consistent pose sequence. Based on this, the system constructs a map using the optimized keyframe relationships as the core. The sequential constraints and loop closure constraints between keyframes are shown in the figure, which are used to ensure the global consistency of the trajectory and the closed-loop correction.

[0048] Subsequently, the 2D feature point clouds generated in step one are registered according to these optimized poses, and incremental updates of two types of maps are completed within a unified coordinate system: one is the occupancy grid, and the other is the multi-resolution TSDF. The occupancy grid uses the observation ray as the basic geometric unit and is updated in a log-probability manner. For each measurement, starting from the sensor pose and connecting to the echo hit position, the voxels along the ray path are treated as free space, and the neighborhood of the hit point is considered as occupied, and the updates are accumulated incrementally in a log-probability manner.

[0049] When a probabilistic form is required, by This is given. By piecing together discrete echoes into continuous constraints, both observational gaps are filled and the free and occupied regions are clearly distinguished. TSDF uses two sets of voxel grids to collaboratively fuse point clouds: a coarser grid is used at long distances to ensure coverage and efficiency, while a finer grid is used closer to the surface to depict details. Along the ray at the cutoff distance... Internal calculation of symbolic distance And perform weighted fusion, with voxel values ​​and weights according to:

[0050] Incremental updates are limited to the cutoff distance. Within, the signed distance value of voxels near the hit point. The existing characterization is progressively corrected by weighting the measurement confidence level and incident angle. During rendering, high-resolution meshes are prioritized to ensure local details, while low-resolution meshes are combined to maintain overall coverage, thereby obtaining clear surfaces in the foreground and maintaining a complete map in the background.

[0051] Based on step four, the system uses the keyframe pose sets obtained in step two and the optimization in step two. For reference, the observation point cloud obtained in step one is stably mapped to both occupied grid and TSDF map representations. This gradually accumulates to form a coherent, reliable, and navigation- and visualization-friendly underwater environment map while ensuring a globally consistent trajectory.

[0052] The method proposed in this embodiment can stably extract environmental features and construct accurate trajectories and maps even under noise interference and sparse observation conditions. This method achieves high-precision positioning and mapping of underwater autonomous vehicles during long-term navigation through sonar image preprocessing and improved CFAR and OTSU feature extraction, position-aware ICP relative pose estimation, region descriptor loop closure detection, and dual-resolution TSDF fusion mapping. Figure 1 To the point of being responsive.

[0053] Implementation Method Two: An underwater SLAM system combining region-level descriptors and LSH retrieval is disclosed. The system employs an underwater SLAM method combining region-level descriptors and LSH retrieval as described in Embodiment 1. The system includes: Image processing module: Acquires polar coordinate images and performs preprocessing to obtain two-dimensional feature point clouds; Sequence constraint module: It is used to obtain the prior pose of the current key frame based on the track, input the prior pose into the position-aware ICP, and scan and match the point clouds of adjacent key frames in sequence within the restricted search space to obtain the relative pose and establish the sequence constraint relationship. The loop closure constraint module is used to collect historical keyframes, extract the occupied regions from the point cloud of historical keyframes using DBSCAN clustering, construct region descriptors, perform nearest neighbor retrieval based on LSH, filter candidate keyframes that are separated in time and verify the matching, and generate valid loop closure constraints. Map generation module: It is used to construct a factor graph based on sequential constraints and effective closure constraints, and uses GTSAM for global graph optimization. At the same time, the optimized keyframe poses are used to accumulate and update the environmental point cloud to generate robot trajectory and map.

Claims

1. An underwater SLAM method combining region-level descriptors and LSH retrieval, characterized in that, Includes the following steps: S1: Obtain the polar coordinate image and perform preprocessing to obtain a two-dimensional feature point cloud; S2: Obtain the prior pose of the current keyframe based on the track, input the prior pose into the position-aware ICP, scan and match the point clouds of adjacent keyframes in sequence within the restricted search space to obtain the relative pose, and establish the order constraint relationship. S3: Collect historical keyframes, use DBSCAN clustering to extract occupied regions from the point clouds of historical keyframes, construct region descriptors, and perform nearest neighbor retrieval based on LSH to filter candidate keyframes that are separated in time and verify the matching, and generate effective loop closure constraints. S4: Construct a factor graph based on the sequential constraints and effective closure constraints, and use GTSAM for global graph optimization. At the same time, use the optimized keyframe poses to accumulate and update the environmental point cloud to generate an underwater environment map.

2. The underwater SLAM method combining region-level descriptors and LSH retrieval according to claim 1, characterized in that, The preprocessing includes: removing carrier reflections using background subtraction, removing speckle noise, enhancing contrast with CLAHE and reducing noise with Gaussian filtering, using an improved CFAR detection method, combining OTSU segmentation and dynamic target removal to extract stable feature points, and converting them into Cartesian coordinate point clouds. After downsampling and isolated point removal, a two-dimensional feature point cloud is generated to provide input for SLAM.

3. The underwater SLAM method combining region-level descriptors and LSH retrieval according to claim 1, characterized in that, S2 includes: S21: Obtain the initial pose based on the trajectory and pre-place the current frame point cloud into the coordinate system of the previous frame; S22: Limit the search radius of the matching based on the prior error, so that the correspondence is established only within a reasonable neighborhood; S23: Determine the pose covariance and inter-frame time, and stipulate that the nearest neighbor is only searched within the radius, and those outside the radius are not included in the registration; S24: Perform point-to-point ICP within a restricted neighborhood. In each round, KD-Tree is used to establish nearest neighbor correspondences, a distance threshold is set to eliminate far pairs, and the rigid body increment is estimated and the attitude is updated using the closed-form solution of weighted SVD. At the same time, the root mean square error is used as the convergence criterion, and upper limits are set for single-step translation and rotation to constrain the iteration stability. S25: After registration, the overlap rate and the final root mean square error are used as acceptance criteria. If the criteria are met, the relative pose is obtained and written as the order factor; if the acceptance fails, the solution is continued in the order of hierarchical backtracking.

4. The underwater SLAM method combining region-level descriptors and LSH retrieval according to claim 3, characterized in that, In step S25, continuing the solution in the step-back sequence includes the following steps: First, switch to standard ICP to expand the available correspondences; if still insufficient, then use sampling-covariance ICP to enhance robustness to sparse or anisotropic point sets by first downsampling to equalize the distribution and then solving based on local covariance weighting; if still insufficient, use NDT to give a new coarse pose as the initial pose, and then return to restricted ICP for refinement.

5. The underwater SLAM method combining region-level descriptors and LSH retrieval according to claim 1, characterized in that, S3 includes: S31: Cluster the two-dimensional feature point cloud of the current keyframe into several occupied regions according to density, and calculate a region-level descriptor for each region; S32: Build a nearest neighbor index for the region descriptor extracted from historical keyframes, and record the corresponding keyframe number in the index entry; S33: For each descriptor in the current frame, perform a nearest neighbor query to calculate the weighted Euclidean distance with the historical descriptors, and convert it into similarity. S34: Calculate similarity scores based on descriptors and accumulate the scores to the corresponding keyframes to form a weighted vote; S35: Perform position-aware ICP verification on candidate frames, using the pose calculated from the previous dead center as the initial value, establish neighborhood correspondence, iteratively update rigid body transformation, and control convergence with root mean square error, single-step translation and rotation upper limit. S36: Effectively determine loop relationships.

6. The underwater SLAM method combining region-level descriptors and LSH retrieval according to claim 1, characterized in that, S4 includes: S41: Under the constraints of sequence and effective loop closure, perform global optimization on the keyframe map to achieve a consistent pose sequence; S42: Construct a map based on the optimized keyframe relationships; S43: Register the two-dimensional feature point cloud and complete the incremental update of two types of maps in a unified coordinate system. One type is occupied grid and the other is multi-resolution TSDF. S44: Using the optimized keyframe pose as a reference, the two-dimensional feature point cloud obtained in S1 is mapped onto two map representations, namely the occupied grid and TSDF, to generate an underwater environment map.

7. The underwater SLAM method combining region-level descriptors and LSH retrieval according to claim 6, characterized in that, The occupied grid uses the observed ray as the basic geometric unit and is updated in a logarithmic probability form; the TSDF uses two sets of voxel grids to collaboratively fuse the point cloud.

8. An underwater SLAM system combining region-level descriptors and LSH retrieval, characterized in that, The system is implemented using an underwater SLAM method combining region-level descriptors and LSH retrieval as described in any one of claims 1-7. The system includes: Image processing module: Acquires polar coordinate images and performs preprocessing to obtain two-dimensional feature point clouds; Sequence constraint module: It is used to obtain the prior pose of the current key frame based on the track, input the prior pose into the position-aware ICP, scan and match the point clouds of adjacent key frames in sequence within the restricted search space to obtain the relative pose, and establish the sequence constraint relationship. The loop closure constraint module is used to collect historical keyframes, extract the occupied regions from the point cloud of historical keyframes using DBSCAN clustering, construct region descriptors, perform nearest neighbor retrieval based on LSH, filter candidate keyframes that are separated in time and verify the matching, and generate valid loop closure constraints. Map generation module: It is used to construct a factor graph based on sequential constraints and effective closure constraints, and uses GTSAM for global graph optimization. At the same time, the optimized keyframe poses are used to accumulate and update the environmental point cloud to generate robot trajectory and map.

9. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes an underwater SLAM method combining region-level descriptors and LSH retrieval according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program that executes an underwater SLAM method combining region-level descriptors and LSH retrieval as described in any one of claims 1-7.

Images