A SLAM method integrating wavelet transform features and curvature features

Abstract

This invention belongs to the field of localization and mapping technology, and discloses a SLAM method that integrates wavelet transform features and curvature features. Addressing the limitations of existing LIO-SAM algorithms in feature extraction, which rely solely on traditional curvature feature extraction, particularly in terms of scale uniformity and noise sensitivity, this invention innovatively introduces a wavelet transform feature extraction method. First, it accurately converts the 3D point cloud into a 2D matrix using polar coordinate depth projection, simultaneously removing invalid data with negative height values. Then, based on the dual-dimensional classification of the small matrix's density and change intensity, it uses db3 wavelet transform to decompose the point cloud signal into different wavelet features. Finally, these features are fused with traditional curvature features to output a fused point cloud feature matrix. This invention significantly improves the comprehensiveness and robustness of feature extraction while maintaining the integrity of the core LiDAR point cloud processing workflow, thereby enhancing mapping quality.

Description

Technical Field

[0001] This invention belongs to the field of positioning and mapping technology, and specifically relates to a SLAM method that integrates wavelet transform features and curvature features, which is particularly suitable for improving the accuracy and robustness of synchronous positioning and mapping in harbor environments. Background Technology

[0002] Simultaneous Localization and Mapping (SLAM) is the core of autonomous navigation and environmental perception for robots and autonomous driving platforms. It enables robots and autonomous driving platforms to achieve autonomous movement, obstacle avoidance, and path planning in complex harbor environments without relying on pre-built high-precision harbor maps. Current mainstream SLAM systems are mainly based on multi-source sensor fusion schemes such as LiDAR, visual sensors, and inertial measurement units (IMUs).

[0003] Lidar Inertial Odometry via Smoothing and Mapping (LIO-SAM) is a representative method in the current laser inertial odometry field, achieving high-precision real-time positioning and mapping through a tightly coupled factor graph optimization framework. However, in typical port operation environments, LIO-SAM faces the following challenges: 1. Its method of extracting edge and planar features using fixed thresholds struggles to effectively distinguish between static container structures and dynamic interference; 2. Curvature-based feature selection performs poorly in feature-poor areas such as long-distance, regularly arranged container yards; 3. It cannot adaptively adjust the feature extraction scale, resulting in poor feature extraction performance for long-distance point clouds. These shortcomings significantly limit the application effectiveness of the LIO-SAM method in complex dynamic environments, resource-constrained platforms, and long-term autonomous operation scenarios.

[0004] To address these shortcomings, some scholars have proposed using deep learning models to replace the traditional curvature thresholding method, extracting more discriminative features from point clouds to improve robustness in unstructured environments. Alternatively, they have proposed fusing semantic features with semantic segmentation networks to identify dynamic objects, filtering dynamic point clouds during feature extraction to reduce pose estimation interference. However, such neural network-dependent algorithms require GPU acceleration, suffer from poor real-time performance on embedded devices, high power consumption, and limited generalization ability.

[0005] Furthermore, some scholars have proposed feature extraction techniques based on multi-sensor deep fusion. Currently, this technique faces the dual constraints of rigid algorithmic frameworks and hardware performance bottlenecks. Existing fusion algorithms typically pre-determine fixed sensor input dimensions; adding a new sensor requires reconstructing the entire algorithm framework, leading to a non-linear increase in computational complexity. Simultaneously, system performance is limited by the worst-case sensor hardware specifications; for example, a low-cost IMU can directly drag down the overall performance of a high-precision LiDAR.

[0006] Wavelet Transform (WT) is a time-frequency analysis method that can simultaneously provide localized features of a signal in both the time and frequency domains. Compared to the traditional Fourier Transform, wavelet transform offers multi-resolution analysis capabilities, making it widely used in signal processing, image analysis, data compression, and pattern recognition. Summary of the Invention

[0007] The purpose of this invention is to propose a SLAM method that integrates wavelet transform features and curvature features. By introducing wavelet transform into the feature extraction process of LIO-SAM, the feature points extracted by traditional curvature are fully integrated with the wavelet feature points extracted by wavelet transform. This compensates for the shortcomings of traditional feature extraction, which relies solely on curvature calculation to extract features, and ultimately improves the quality of mapping.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A SLAM method that integrates wavelet transform features and curvature features includes the following steps: Step 1. Remove point cloud distortion from the acquired lidar point cloud data; Step 2. Extract point cloud features from each frame of LiDAR point cloud using traditional curvature and construct a traditional curvature feature matrix; Step 3. Before processing the point cloud features using wavelet transform, first transform the coordinates of each frame of the lidar point cloud, then divide the matrix after coordinate transformation into sub-matrices, and pre-classify each sub-matrix according to the point cloud density and rate of change. Step 4. Use wavelet transform to extract point cloud features from each frame of laser point cloud to obtain the wavelet feature matrix; First, convolve the high-pass and low-pass filters of the wavelet transform to obtain four feature judgment matrices. Then, convolve each feature judgment matrix with the submatrix of the corresponding category after pre-classification to obtain the corresponding feature matrix. Then, the obtained feature matrices are concatenated into wavelet feature matrices according to the order of the pre-divided sub-matrices; Step 5. Classify the traditional curvature feature matrix according to the obtained wavelet feature matrix point cloud category, and fuse the traditional curvature feature matrix and wavelet feature matrix to construct the fused point cloud feature matrix of each frame of laser point cloud; Step 6. Perform pose transformation on the fused point cloud feature matrix generated for each frame of laser point cloud. After all the fused point cloud feature matrices have been transformed, the laser factor map is obtained and tightly coupled with the previously acquired IMU data for optimization. Finally, the tightly coupled optimization result is fed into the input pose graph optimization to generate the global map.

[0009] Furthermore, based on the aforementioned SLAM method that integrates wavelet transform features and curvature features, this invention also proposes a computer device comprising a memory and a processor. The memory stores executable code, and when the processor executes the executable code, it implements the steps of the aforementioned SLAM method that integrates wavelet transform features and curvature features.

[0010] Furthermore, based on the aforementioned SLAM method that integrates wavelet transform features and curvature features, this invention also proposes a computer-readable storage medium storing a program that, when executed by a processor, implements the steps of the aforementioned SLAM method that integrates wavelet transform features and curvature features.

[0011] The present invention has the following advantages: As described above, this invention discloses a SLAM method that integrates wavelet transform features and curvature features. This method innovatively introduces wavelet transform into the feature extraction stage of LIO-SAM. Feature point clouds are extracted independently by curvature and wavelet transform, and then traditional curvature features are deeply fused with wavelet features. Wavelet features can capture high-frequency details and low-frequency trends of the point cloud and integrate them into the point cloud features, significantly reducing the probability of feature confusion, improving the accuracy of feature matching, reducing pose estimation errors, and improving the accuracy of positioning. It overcomes the limitations of insufficient information dimension of a single feature descriptor and improves the robustness of complex positioning scenarios. Secondly, this invention employs a pre-classification-guided targeted convolution mechanism. Based on the point cloud density and rate of change characteristics, the data is pre-classified into four categories, so that each sub-matrix only needs to be convolved with a specific feature judgment matrix to generate the corresponding feature matrix. This innovative strategy avoids the computational redundancy of four global convolutions required in traditional methods, improving computational efficiency, and is particularly suitable for real-time processing of large-scale point cloud data. Furthermore, this invention strengthens specific features, transforming the two-dimensional matrix... Divide into multiple sub-matrices ,right Perform targeted feature processing, classify beforehand, and emphasize This property effectively improves the signal-to-noise ratio and accuracy of feature extraction, enhances pose estimation stability, and thus improves the overall mapping quality. Targeted processing is applied to classified regions; the feature judgment matrix is ​​convolved only with specific sub-matrices, rather than with the entire region. Convolution followed by discarding useless information effectively suppresses interference from irrelevant features, resulting in cleaner and more reliable feature extraction results. Furthermore, this invention designs a regularized feature fusion strategy, employing differentiated fusion methods for different feature types. Minimum values ​​are taken for planar features to enhance smoothness, average values ​​are taken for edge features to maintain stability, and maximum values ​​are taken for corner features to highlight saliency. This refined fusion method significantly optimizes the quality of the final feature matrix. This invention achieves efficient point cloud processing from preprocessing to feature fusion without relying on complex computational frameworks or high computing power, reducing system hardware costs and power consumption, and avoiding the one-sided feature representation problem caused by the single strategy in traditional fusion methods. By introducing wavelet transform into the feature extraction process of LIO-SAM, feature points extracted using traditional curvature extraction are fully fused with wavelet feature points extracted using wavelet transform, overcoming the shortcomings of traditional feature extraction that relies solely on curvature calculation, ultimately improving the mapping quality. Attached Figure Description

[0012] Figure 1 This is a flowchart of the SLAM method that integrates wavelet transform features and curvature features in an embodiment of the present invention. Detailed Implementation

[0013] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Example 1 In complex operating environments such as ports, traditional feature extraction methods based on fixed neighborhood curvature thresholds are difficult to effectively handle multi-scale environmental features. This embodiment addresses the problems of scale uniformity and noise sensitivity in the feature extraction stage of existing LIO-SAM algorithms by innovatively introducing a SLAM method that integrates wavelet transform features and curvature features. The method first accurately converts the 3D point cloud into a 2D matrix through polar coordinate depth projection, simultaneously removing invalid data with negative z values. Then, based on the "density + change intensity" dual-dimensional classification of the divided small matrices, the point cloud signal is decomposed into different feature sub-bands using db3 wavelet transform. Finally, it is fused with traditional curvature features to output a fused point cloud feature matrix. This significantly reduces interference from invalid data and feature matching ambiguity, providing a more reliable constraint basis for pose calculation, thereby improving the accuracy, stability, and robustness of the SLAM system in complex scenarios. While maintaining the integrity of the core process of lidar point cloud processing, the method of this invention constructs a multi-scale, multi-dimensional feature description system by innovatively embedding wavelet transform in the feature extraction stage and fusing it with traditional curvature features, which significantly improves the comprehensiveness and robustness of feature extraction.

[0014] like Figure 1 As shown, the SLAM method that integrates wavelet transform features and curvature features in this embodiment includes the following steps: Step 1. Remove point cloud distortion from the acquired lidar point cloud data.

[0015] To obtain the necessary data for the port, LiDAR, IMU, and GPS (optional) are used.

[0016] The acquired raw data is processed to remove point cloud distortion, obtain point cloud information, traverse each point of the LiDAR, extract the depth value from each point to the LiDAR, and generate a one-dimensional structured signal to ensure that the data is accurate and usable.

[0017] The point cloud information includes the start index, end index, and point cloud range of each scan line.

[0018] Step 2. Extract point cloud features from each frame of LiDAR point cloud using traditional curvature and construct a traditional curvature feature matrix.

[0019] Calculate the curvature features of the point cloud based on the results of step 1, and mark potential edge points and planar points, i.e., the geometric features of the object.

[0020] Edge points correspond to points with high curvature, representing objects with distinct edge features near harbors, such as containers, quay cranes, and ships; planar points correspond to smooth objects such as water surfaces and roads. For each point in the point cloud... Search within its neighborhood or nearest neighbor The covariance matrix is ​​calculated for each point, and the curvature is obtained through eigenvalue decomposition. The specific calculation process is as follows: Calculate the centroid The formula is as follows: .

[0021] in, express A nearby point, Represents point clouds, Represent the point cloud index and construct the covariance matrix. : .

[0022] Calculate the covariance matrix eigenvalues ,Sure ≥ ≥ ≥0; where, , , They are matrices Based on the eigenvalues ​​obtained by linear algebra methods, sort them in descending order.

[0023] Each point curvature The calculation is performed using the following formula: .

[0024] Calculate each point mean and standard deviation The low threshold is determined based on the mean and standard deviation. and high threshold : , .

[0025] Value less than For a planar point, greater than The first point is a corner point, and the rest are edge points.

[0026] Arranged in the order of calculation, from front to back each Each row forms a single line. The matrix, with row labels as Column label .get The values ​​are stored in a matrix among, The size is The matrix is ​​expressed as follows: .

[0027] Step 3. Before processing the point cloud features using wavelet transform, first transform the coordinates of each frame of the LiDAR point cloud, then divide the transformed matrix into sub-matrices, and pre-classify each sub-matrix according to the point cloud density and rate of change.

[0028] For each frame of laser point cloud data, a wavelet transform is performed.

[0029] Step 3.1. Convert the 3D point cloud into a 2D matrix using polar coordinate projection. LiDAR point clouds are three-dimensional data. Data preprocessing requires converting them from three-dimensional to two-dimensional, which is done using polar coordinate depth projection.

[0030] Polar coordinate projection is a common preprocessing method for converting 3D LiDAR point cloud data into a 2D matrix, and it is suitable for rotating LiDAR systems. This method converts the Cartesian coordinates of the point cloud... Determine the planar position of a point cloud in three-dimensional space. Determine the height position and convert it to polar coordinates. The data is quantized into a two-dimensional matrix, which facilitates subsequent wavelet feature extraction.

[0031] and As a matrix The units of rows and columns, Represents the polar radius. Represents the polar angle, simultaneously eliminating elements in the two-dimensional matrix. For invalid data with negative values, set the coordinate data of coordinates with negative z values ​​and no fill content to 0.

[0032] The parameter in the original three-dimensional coordinate system represents the height relative to the horizontal plane; from the three-dimensional coordinate system... Convert to a two-dimensional matrix, and convert the three-dimensional coordinates , Convert to and , Add it to the matrix as content, keeping it unchanged.

[0033] The coordinates of the 3D point cloud scanned by the lidar Convert to two-dimensional matrix coordinates And retain parameters ,in Representing 3D point clouds Coordinates in three directions.

[0034] .

[0035] in The result is calculated from r and θ, then = That is, by The transformed expression is calculated. , Representation matrix The element in the i-th row and j-th column represents the physical height of the laser point cloud.

[0036] definition Fill the two-dimensional matrix with content. After filling the two-dimensional matrix, find the content in the two-dimensional matrix. For coordinates with values ​​less than 0 and no fill, set all their data to 0 to obtain a two-dimensional matrix. The polar coordinate matrix is ​​as follows: .

[0037] Two-dimensional matrix The number of rows and columns is The parameter of the row is the polar radius. The parameter of the column is the polar angle. .

[0038] The size of the row is the maximum detection range of the lidar and The ratio of the column scale to the ratio of 360°. The ratio of . Representation matrix The Middle row and number Column elements, .

[0039] Step 3.2. Convert the two-dimensional matrix The matrix is ​​split into multiple sub-matrices using a row-first, column-second traversal method. Then, based on the point cloud density and rate of change characteristics in each sub-matrix, each sub-matrix is ​​pre-divided into four categories.

[0040] Step 3.2.1. Determine the point cloud density in each submatrix.

[0041] First, according to Size division of two-dimensional matrix traverse the two-dimensional matrix in row-major order. Multiple submatrices are obtained. The size of the submatrix is Two-dimensional matrix The size is , .

[0042] Step 3.2.2. Then, based on the point cloud density and rate of change, adjust each sub-matrix. Perform pre-classification.

[0043] I. First, determine the submatrix The point cloud density is determined as follows: If submatrix If the number of valid points exceeds 80% of all data, it is considered dense; otherwise, it is considered sparse.

[0044] II. Further determine each submatrix The rate of change characteristics of the midpoint cloud are determined as follows: In this embodiment, the submatrix is ​​judged. The magnitude of changes in the content within a matrix is ​​determined using a local-to-global approach. This involves considering changes within each submatrix. Selected from One (e.g., 100) valid points.

[0045] From each submatrix Random selection There are 10 valid points, and the points adjacent to each valid point are defined as surrounding number points.

[0046] Calculate the mean of the number points surrounding each valid point, and obtain the result through calculation. The mean of the surrounding digital points corresponds one-to-one with each valid point; the purpose of calculating the mean of the surrounding digital points is to replace the isolated value of a single point with the overall data characteristics of the local area, avoid the interference of random errors, and at the same time capture the true difference between the local and the global.

[0047] Will The average of the surrounding number points is further calculated and compared with... The ratio is calculated by taking the average of the valid points. Specifically, the ratio is calculated by taking the average of the valid points. The result of averaging the surrounding number points and dividing by The average of the valid points.

[0048] If the ratio of the two is greater than the preset rate of change threshold, it is determined that the change is large; otherwise, it is determined that the change is stable.

[0049] The common range of empirical thresholds based on sample calibration is 1.2-2.0. Values ​​below 1.2 may misjudge slight fluctuations as significant changes, while values ​​above 2.0 may miss truly existing moderate changes.

[0050] In this embodiment, the rate of change threshold is set to 1.5, which serves as an intermediate value to effectively distinguish between "stable" and "significant change".

[0051] Based on the above judgment principles I and II, each submatrix is... Based on their characteristics, they are divided into four categories: If the submatrix is ​​sparse and its change is stationary, then it is marked as ; If the submatrix is ​​dense and changes smoothly, it is marked as ; If the submatrix is ​​sparse and varies greatly, it is marked as ; If the submatrix is ​​dense and varies greatly, it is marked as ; Among them, subscript Represents from a two-dimensional matrix The first division in the middle Submatrix .

[0052] Based on the matrix characteristics corresponding to the point cloud characteristics, when a lidar ray hits a flat object at a distance, the returned points are far apart and appear sparse. These surfaces are very flat, and their depth values ​​or normal directions hardly change.

[0053] The numerical changes in the matrix are very small, that is For objects such as ground, sea surface, ship sides, and container surfaces; when an object is very close to the lidar, the laser beam hitting it results in a very dense point cloud. Although the close distance leads to dense points, the surface remains smooth and curved, so the numerical changes are relatively stable. Alternatively, due to the greater distance, the point cloud itself is sparse, representing edges or corners. The depth values ​​or coordinates of adjacent points may jump, causing significant changes in the values ​​within the matrix. , This corresponds to the edges of objects; objects are close together, point clouds are dense, and a large amount of detail can be captured. The objects themselves have complex and uneven shapes, and the depth and direction of adjacent points differ greatly, resulting in drastic changes in matrix values. For small objects such as obstacles.

[0054] The main purpose of pre-classification in this step is to know in advance which class this submatrix mainly belongs to, but it is not yet known whether each point cloud belongs to this class. After convolution, due to the amplification and suppression effect, those point clouds with poor performance are suppressed, highlighting the point clouds that can represent this matrix.

[0055] Step 4. Use wavelet transform to extract point cloud features from each frame of laser point cloud to obtain a wavelet feature matrix. Each frame of laser point cloud will yield a wavelet feature matrix and a curvature feature matrix, and then proceed to step 5 once.

[0056] Specifically, the high-pass and low-pass filters of the wavelet transform are convolved to obtain four feature judgment matrices. Then, each feature judgment matrix is ​​convolved with the submatrix of the corresponding category after pre-classification to obtain the corresponding feature matrix.

[0057] The feature matrix is ​​used to determine the specific category of each point cloud in its corresponding submatrix, which is one of the four categories: LL, LH, HL, and HH. These four categories correspond to surfaces, horizontal edges, vertical edges, and details, respectively.

[0058] Then, the obtained feature matrix is ​​concatenated into a wavelet feature matrix according to the order of the pre-divided sub-matrices.

[0059] In this embodiment, the db3 wavelet is selected for feature extraction. The db3 wavelet uses the third-order vanishing moment, which can fully extract features while reducing the amount of computation.

[0060] Regarding the results obtained in step 3 The submatrix is ​​used for db3 wavelet feature extraction; based on the pre-classified submatrix, matrix multiplication is performed with the db3 high-pass and low-pass filters. The db3 filter length is 6.

[0061] The high-pass filter is: .

[0062] Low-pass filter: .

[0063] The following four feature judgment matrices are obtained by convolving the high-pass filter and the low-pass filter: , , , ; in Used for detecting the target matrix, i.e., the submatrix. The corner point, Used for detecting target matrix Horizontal edge point, Used for detecting target matrix The vertical edge point, Used for detecting target matrix Plane point.

[0064] Fill the edges of the above four feature judgment matrices with 0. size.

[0065] The attributes obtained after pre-classification are submatrix and the padded Convolution yields the feature matrix ; The attributes obtained after pre-classification are submatrix and the padded Convolution yields the feature matrix ; The attributes obtained after pre-classification are submatrix and the padded Convolution yields the feature matrix ; The attributes obtained after pre-classification are submatrix and the padded Convolution yields the feature matrix ; The characteristic matrix , , , These represent the object's face, horizontal edge, vertical edge, and details, respectively.

[0066] Since the submatrices have already been roughly classified in step 3, each submatrice In the feature judgment matrix When performing convolution, you only need to select features that match your own features. (Right now , , , That's fine.

[0067] By identifying differences in signal strength contained within the feature matrix itself, the system can achieve [the following]: The screening and verification process amplifies the features that meet the expectations and suppresses the influence of irrelevant noise. (Feature matrix) yes The data in was After amplification and suppression, the point cloud data of the planar region is... The matrix corresponds to high-energy coefficients, and the signals in these regions are amplified after convolution; while local noise or non-planar interference (such as small protrusions) is... Low-energy matrices are suppressed after convolution; similarly, , , After , , After suppressing the amplification of the data, it is possible to achieve the transformation from classification to precise classification.

[0068] This approach, compared to determining a two-dimensional matrix... The point cloud characteristics in the two-dimensional matrix The feature judgment matrix is ​​convolved four times directly, and then a two-dimensional matrix is ​​selected. The method of determining the type of midpoint cloud and discarding unnecessary convolution results significantly reduces computational redundancy and saves computation time. It also "enhances" specific features through convolution, making the target feature response more prominent and improving the signal-to-noise ratio of feature extraction. In harbor scenarios, the port area is complex while the sea surface is very simple, with noise interference from waves. Using "enhanced" convolution allows for clear distinction between the port and sea surface scenes, and the improved signal-to-noise ratio removes noise interference from waves, ensuring accurate harbor positioning and mapping.

[0069] Then, the calculated feature matrix results are processed according to the pre-segmented two-dimensional matrix. By concatenating the individual sub-matrices in the order they were obtained, the complete wavelet feature matrix is ​​obtained. , The expression for a matrix is: .

[0070] Step 5. Classify the traditional curvature feature matrix according to the point cloud category of the obtained wavelet feature matrix, and fuse the traditional curvature feature matrix and the wavelet feature matrix to construct the fused point cloud feature matrix of each frame of laser point cloud.

[0071] The wavelet feature matrix and the traditional curvature feature matrix are downsampled before feature fusion.

[0072] wavelet feature matrix A large number of adjacent elements often have strong correlations, leading to information duplication. Downsampling the matrix, sampling every other row and column, changes the original... Become The size becomes the original. By reducing the amount of data, processing efficiency is optimized, and excessive redundant information is avoided in the calculation, ultimately resulting in a matrix. The expression is as follows: = .

[0073] For example, there is one The matrix has 16 data points, and the formula is expressed as follows: .

[0074] Sample every other row and every other column, i.e., retain the data in rows 1 and 3 and columns 1 and 3. The sampled matrix is ​​as follows: .

[0075] The points used in traditional curvature calculations are derived from eigenvalues. To determine whether it is a plane point, an edge point, or a corner point.

[0076] Traditional curvature feature matrix With wavelet feature matrix The same downsampling method is used to obtain a curvature feature matrix of the same size as the feature extracted by wavelet transform. The expression is as follows: .

[0077] For curvature characteristic matrix The classification process is as follows: The wavelet feature matrix has been obtained through step 4. The midpoint cloud features belong to four specific categories: LL, LH, HL, and HH. These four categories correspond to surfaces, horizontal edges, vertical edges, and details, respectively.

[0078] Curvature feature matrix The point cloud categories are based on the wavelet feature matrix. The same subscript is mapped to the corresponding position for category classification and judgment. The point cloud values ​​in the data belong to which of the following categories: LL, LH, HL, and HH?

[0079] For curvature characteristic matrix After partitioning, different fusion methods are selected based on the category to which each submatrix belongs (i.e., LL, LH, HL, HH). The new feature matrix after fusion based on wavelet transform feature extraction and curvature feature extraction is denoted as... .

[0080] The LL class embodies the planar features of the point cloud, and the selection matrix is ​​used. and Smaller numbers are added to the new feature matrix ; LH and HL classes are in an intermediate state in terms of numerical value, and are calculated using matrices. and The average value is added to the new feature matrix. ; HH class reflects corner features, selection matrix and Larger numbers are added to the new feature matrix ; Curvature feature matrix With wavelet characteristic matrix After all point cloud values ​​are blended, the blended value is obtained. .

[0081] Wavelet transform excels at capturing the multi-scale structure of point clouds by characterizing the point cloud matrix through feature judgment matrices, while traditional curvature transform focuses on feature curvature to describe the degree of geometric bending, accurately quantifying planar points, edge points, and corner points. The two are highly complementary in terms of information dimensions. Feature fusion of wavelet transform and traditional curvature transform can more comprehensively characterize the geometric attributes of point clouds, effectively avoiding information blind spots from single features, reducing scene misjudgment rates, and better supporting harbor positioning and mapping.

[0082] Step 6. Perform pose transformation on the fused point cloud feature matrix generated from each previous frame of laser point cloud. After all the fused point cloud feature matrices have been transformed, the laser factor map is obtained, and then it is tightly coupled with the previously acquired IMU data for optimization.

[0083] Finally, the tightly coupled optimization result is fed into the input pose graph optimization to generate the global map.

[0084] This invention breaks through the limitations of traditional point cloud feature extraction methods that rely solely on geometric features (such as curvature and normal vectors). It innovatively integrates multi-scale frequency domain features, simultaneously extracting the macroscopic spatial distribution and overall morphology of the point cloud (primarily through the generation of LL components) and the microscopic geometric abrupt changes and fine structures (through LH, HL, and HH components) via wavelet transform. This significantly improves the robustness of feature matching in dynamic environments. This invention uses db3 wavelet transform, embedding it into the feature extraction stage of LIO-SAM and fusing it with features extracted using traditional curvature extraction. This makes the point cloud features richer and more accurate, ultimately resulting in macroscopic structural features (such as the horizontal reference of the sea level) that SLAM mapping can capture, accurately distinguishing static details (such as continuous vertical signals at the edge of a ship and multi-directional abrupt changes at platform corners) from dynamic noise (such as discrete and irregular reflections of waves and random movement of drifting buoys). This leads to a 3D environmental map with stronger global consistency, less dynamic interference, and more accurate detail restoration. While maintaining the integrity of the core process of lidar point cloud processing, the method of this invention significantly improves the comprehensiveness and robustness of feature extraction, enhances mapping quality, helps subsequent processes generate more accurate and clearer 3D maps, better extracts effective environmental features, and enhances the system's adaptability to different environments.

[0085] Example 2 This embodiment 2 describes a computer device including a processor and one or more memories. Executable code is stored in the memories, and when the processor executes the executable code, it implements the steps of the SLAM method for fusing wavelet transform features and curvature features described in embodiment 1 above.

[0086] In this embodiment, the computer device can be any device or apparatus with data processing capabilities, and will not be described in detail here.

[0087] Example 3

[0088] This embodiment 3 describes a computer-readable storage medium storing a program that, when executed by a processor, implements the steps of the SLAM method for fusing wavelet transform features and curvature features as described in embodiment 1 above.

[0089] The computer-readable storage medium can be an internal storage unit of any device or apparatus with data processing capabilities, such as a hard disk or memory, or an external storage device of any device with data processing capabilities, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc.

[0090] Of course, the above description is only a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.

Claims

1. A SLAM method integrating wavelet transform features and curvature features, characterized in that, Includes the following steps: Step 1. Remove point cloud distortion from the acquired lidar point cloud data; Step 2. Extract point cloud features from each frame of LiDAR point cloud using traditional curvature and construct a traditional curvature feature matrix; Step 3. Before processing the point cloud features using wavelet transform, first transform the coordinates of each frame of the lidar point cloud, then divide the matrix after coordinate transformation into sub-matrices, and pre-classify each sub-matrix according to the point cloud density and rate of change. Step 4. Use wavelet transform to extract point cloud features from each frame of laser point cloud to obtain the wavelet feature matrix; First, convolve the high-pass and low-pass filters of the wavelet transform to obtain four feature judgment matrices. Then, convolve each feature judgment matrix with the submatrix of the corresponding category after pre-classification to obtain the corresponding feature matrix. Then, the obtained feature matrices are concatenated into wavelet feature matrices according to the order of the pre-divided sub-matrices; Step 5. Classify the traditional curvature feature matrix according to the obtained wavelet feature matrix point cloud category, and fuse the traditional curvature feature matrix and wavelet feature matrix to construct the fused point cloud feature matrix of each frame of laser point cloud; Step 6. Perform pose transformation on the fused point cloud feature matrix generated for each frame of laser point cloud. After all the fused point cloud feature matrices have been transformed, the laser factor map is obtained and tightly coupled with the previously acquired IMU data for optimization. Finally, the tightly coupled optimization result is fed into the input pose graph optimization to generate the global map.

2. The SLAM method fusing wavelet transform features and curvature features according to claim 1, characterized in that, Step 3 specifically involves: Step 3.

1. Convert the 3D point cloud into a 2D matrix using polar coordinate projection. ; in and As a two-dimensional matrix The units of rows and columns, Represents the polar radius. Represents the polar angle, and the height in the three-dimensional point cloud coordinate system. As The content is simultaneously removed from the two-dimensional matrix. Invalid data with a negative value; Will For coordinates with negative values ​​or no fill content, set all coordinate data to 0. Step 3.

2. Convert the two-dimensional matrix The matrix is ​​split into multiple sub-matrices using a row-first, column-second traversal method. Then, based on the point cloud density and rate of change characteristics in each sub-matrix, each sub-matrix is ​​pre-divided into four categories.

3. The SLAM method fusing wavelet transform features and curvature features according to claim 2, characterized in that, Step 3.1 specifically involves: The coordinates of the 3D point cloud scanned by the lidar Convert to two-dimensional matrix coordinates and retain height ,in Representing the three-dimensional point cloud along , , Coordinates in three directions; definition Fill the two-dimensional matrix with content. After filling the two-dimensional matrix, find the content in the two-dimensional matrix. For coordinates with values ​​less than 0 and no fill, set all data in those coordinates to 0 to obtain a two-dimensional matrix. That is, the polar coordinate matrix, as follows: Z= ； in The number of rows and columns is The unit parameter of the behavior is the polar radius. The unit parameter of the column is polar angle. ; The unit distance of the row is the maximum detection range of the lidar and The ratio, with the column unit scale being 360° and The ratio; Representation matrix The Middle row and number Column elements, , .

4. The SLAM method fusing wavelet transform features and curvature features according to claim 2, characterized in that, Step 3.2 specifically involves: Step 3.2.

1. Determine the point cloud density in each submatrix; First, according to Size division of two-dimensional matrix traverse the two-dimensional matrix in row-major order. Multiple submatrices are obtained. The size of the submatrix is Two-dimensional matrix The size is , ; Step 3.2.

2. Then, based on the point cloud density and rate of change, adjust each sub-matrix. Perform pre-classification; I. First, determine the submatrix The point cloud density is determined as follows: If submatrix If the number of valid points in a dataset exceeds 80% of all data, it is considered dense; otherwise, it is considered sparse. II. Further determine each submatrix The rate of change characteristics of the midpoint cloud are determined as follows: From each submatrix Random selection There are 10 valid points, and the points adjacent to each valid point are defined as surrounding number points; Calculate the mean of the number points surrounding each valid point, and obtain the result through calculation. The mean of the surrounding numerical points corresponding to each valid point; The average of the surrounding number points is further calculated and compared with... The ratio is calculated using the average of the valid points; If the ratio of the two is greater than the preset rate of change threshold, it is determined that the change is large; otherwise, it is determined that the change is stable. Based on the above judgment principles I and II, each submatrix is... Based on their characteristics, they are divided into four categories: If the submatrix is ​​sparse and its change is stationary, then it is marked as ; If the submatrix is ​​dense and changes smoothly, it is marked as ; If the submatrix is ​​sparse and varies greatly, it is marked as ; If the submatrix is ​​dense and varies greatly, it is marked as ; Among them, subscript Represents from a two-dimensional matrix The first division in the middle Submatrix .

5. The SLAM method fusing wavelet transform features and curvature features according to claim 1, characterized in that, In step 4, db3 wavelet feature extraction is performed on the sub-matrix pre-classified in step 3, specifically as follows: The four matrix attribute categories of the pre-classified submatrix are defined as follows: If the submatrix is ​​sparse and its change is stationary, then it is marked as ; If the submatrix is ​​dense and changes smoothly, it is marked as ; If the submatrix is ​​sparse and varies greatly, it is marked as ; If the submatrix is ​​dense and varies greatly, it is marked as ; Define the low-pass filter for db3 as follows: The high-pass filter is ; By convolving the high-pass filter and the low-pass filter, the following four feature judgment matrices are obtained: ， ， ， ； in Used for detecting the target matrix, i.e., the submatrix. The corner point, Used for detecting target matrix Horizontal edge point, Used for detecting target matrix The vertical edge point, Used for detecting target matrix Plane point; Fill the edges of the above four feature judgment matrices with 0. Size, of which Size of the submatrix; The attributes obtained after pre-classification are submatrix and the padded Convolution yields the feature matrix ; The attributes obtained after pre-classification are submatrix and the padded Convolution yields the feature matrix ; The attributes obtained after pre-classification are submatrix and the padded Convolution yields the feature matrix ; The attributes obtained after pre-classification are submatrix and the padded Convolution yields the feature matrix ; The characteristic matrix , , , These represent the object's face, horizontal edge, vertical edge, and details, respectively. Then, the calculated feature matrix results are processed according to the pre-segmented two-dimensional matrix. By concatenating the individual sub-matrices in the order they were obtained, the complete wavelet feature matrix is ​​obtained. , The expression for a matrix is: 。 6. The SLAM method fusing wavelet transform features and curvature features according to claim 1, characterized in that, In step 5, the wavelet feature matrix and the traditional curvature feature matrix are downsampled before feature fusion; For wavelet characteristic matrix Perform downsampling, sampling every other row and column, from the original Become The size becomes the original. The wavelet feature matrix after downsampling is obtained. The expression is as follows: = ； Traditional curvature feature matrix With wavelet feature matrix The same downsampling method is used to obtain a curvature feature matrix of the same size as the one extracted by wavelet transform. The expression is as follows: = 。 7. The SLAM method fusing wavelet transform features and curvature features according to claim 6, characterized in that, In step 5, the curvature feature matrix is... The classification process is as follows: The wavelet feature matrix has been obtained through step 4. Each point cloud belongs to a specific category among the four categories LL, LH, HL, and HH, which correspond to surfaces, horizontal edges, vertical edges, and detail features, respectively. Curvature feature matrix The point cloud categories are based on the wavelet feature matrix. The same subscript is mapped to the corresponding position for category classification and judgment. The features of the point cloud values ​​belong to specific categories among LL, LH, HL, and HH.

8. The SLAM method fusing wavelet transform features and curvature features according to claim 7, characterized in that, In step 5, the specific process of feature fusion is as follows: For curvature characteristic matrix After partitioning, different fusion methods are selected based on the different characteristics of each sub-matrix. The new feature matrix after fusion based on wavelet transform feature extraction and curvature feature extraction is denoted as... ; The LL class embodies the planar features of the point cloud, and the selection matrix is ​​used. and Smaller numbers are added to the new feature matrix ; LH and HL classes are in an intermediate state in terms of numerical value, and are calculated using matrices. and The average value is added to the new feature matrix. ; HH class reflects corner features, selection matrix and Larger numbers are added to the new feature matrix ; Curvature feature matrix With wavelet characteristic matrix After all point cloud values ​​are blended, the blended value is obtained. .

9. A computer device comprising a processor and one or more memories, wherein executable code is stored in the memories, characterized in that, When the processor executes the executable code, it implements the steps of the SLAM method for fusing wavelet transform features and curvature features as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it is used to implement the steps of the SLAM method for fusing wavelet transform features and curvature features as described in any one of claims 1 to 8.

Images