Multi-level self-adaptive point cloud registration method and device for strip mine and medium

Abstract

The invention relates to a multi-level self-adaptive point cloud registration method and device for a strip mine, and a medium, and the method comprises the steps: carrying out the preliminary down-sampling of two input frames of original point clouds, extracting the feature similarity information of a second layer of point clouds, carrying out the self-adaptive screening of a region point set with the highest registration value based on the similarity, and further constructing refined third and fourth layers of point clouds, feature-driven hierarchical registration representation is realized, and the problems of information redundancy and structure loss in a traditional fixed downsampling strategy are effectively avoided. Meanwhile, an angle information modeling mechanism is introduced into a feature extraction module, and rotation invariance is enhanced by encoding a relative included angle relationship between local neighborhoods, so that the method can still stably extract structural features and realize accurate registration when the attitude change of the sensor is severe. According to the method, the accuracy and robustness of registration are remarkably improved while the registration efficiency is guaranteed, and the method is suitable for scenes such as strip mines with high requirements for real-time sensing and registration robustness.

Description

Technical Field

[0001] This invention relates to a point cloud registration method, specifically a multi-level adaptive point cloud registration method, equipment, and medium for open-pit mines. Background Technology

[0002] Point cloud registration technology, as a core method for realizing multi-view 3D information fusion and scene reconstruction, has important applications in fields such as autonomous driving, robot navigation, and surveying and modeling. Current mainstream methods improve registration efficiency and stability at multiple scales by constructing a multi-layered point cloud pyramid structure and performing layered feature matching from coarse to fine. However, in large-scale scenarios such as open-pit mines, traditional point cloud registration methods face severe challenges, necessitating more efficient and robust adaptive registration techniques.

[0003] The open-pit mine environment has the following typical characteristics, making traditional registration methods difficult to apply:

[0004] 1. Massive data scale: A single scan can acquire billions of point clouds, and the open-pit mine covers a wide area and has a large spatial span, far exceeding the data volume of urban roads and indoor scenes, making traditional methods computationally inefficient;

[0005] 2. High environmental dynamism: The movement of mine trucks, real-time changes in the shape of ore piles, blasting operations, etc., cause significant differences in point clouds in the time dimension, making static registration methods prone to failure;

[0006] 3. Complex geometric structure: The stepped mining face, irregular ore piles and exposed rock strata form a multi-layered mixed geometric feature, making it difficult to effectively extract key features from the fixed pyramid structure;

[0007] 4. Severe obstruction and noise: Dust, equipment obstruction, and slope terrain cause point cloud loss and noise interference, affecting registration accuracy;

[0008] 5. High real-time requirements: Open-pit mine safety monitoring (such as slope displacement early warning) requires real-time or near-real-time point cloud registration capabilities, which traditional methods cannot meet.

[0009] Current mainstream methods mostly employ a fixed-layer point cloud pyramid structure, relying on pre-defined downsampling strategies to construct multi-scale levels and perform feature matching at different scales. However, in large-scale scenarios such as open-pit mines, these methods suffer from the following key problems:

[0010] First, the static pyramid structure lacks adaptability.

[0011] 1) Feature loss problem: Fixed downsampling strategies cannot distinguish between high-value structural areas (such as mining faces and equipment outlines) and redundant backgrounds (such as flat roads), and tend to retain redundant information while ignoring key geometric structures, resulting in the loss of key geometric information in the coarse registration stage.

[0012] 2) Unreasonable allocation of computing resources: The complexity varies greatly in different areas of open-pit mines (such as dense equipment areas and open goaf areas), and fixed-layer pyramids cannot be adaptively adjusted, resulting in wasted computing resources or local registration failures.

[0013] 3) Poor rotation invariance: The movement of large equipment in open-pit mines (such as electric shovels and drilling rigs) causes significant rotational changes in point clouds. Existing feature extraction methods lack adaptability to spatial rotational changes, resulting in inconsistent feature representations under different viewing conditions, which seriously affects the matching stability and overall accuracy between point clouds.

[0014] Second, the dynamic scene registration capability is insufficient: traditional methods rely on static scene assumptions and cannot adapt to dynamic factors such as changes in the shape of the mine pile and vehicle movement, resulting in registration drift between historical point clouds and real-time scanning data; traditional algorithms consume significantly more computation time when processing point clouds of billions, making it difficult to meet the real-time monitoring needs of open-pit mines.

[0015] Third, the multi-sensor fusion capability is limited: open-pit mines typically use multiple sensors such as LiDAR, vision, and IMU to collect data, but existing methods are mostly designed for single sensors and lack effective multi-modal data fusion mechanisms, resulting in a decrease in registration accuracy in sparse or occluded point cloud areas.

[0016] Therefore, there is an urgent need for a point cloud registration method that can dynamically select structural regions based on feature distribution and has good rotation invariance, so as to balance computational efficiency and registration accuracy and meet the actual needs of high robust 3D perception in ultra-large-scale dynamic scenarios such as open-pit mines. Summary of the Invention

[0017] The purpose of this invention is to provide a multi-level adaptive point cloud registration method for open-pit mines, which can dynamically select structural regions based on feature distribution, has good rotation invariance, and can simultaneously balance computational efficiency and registration accuracy, meeting the practical needs of high robust 3D perception in ultra-large-scale dynamic scenarios such as open-pit mines.

[0018] To achieve the above objectives, this invention provides a multi-level adaptive point cloud registration method for open-pit mines, comprising the following steps:

[0019] S1. Input the source frame point cloud and the target frame point cloud, and perform two-layer downsampling on the source frame point cloud and the target frame point cloud to obtain the first layer point cloud and the second layer point cloud;

[0020] S2. An improved kernel convolutional network based on the introduction of angle information is used to extract features from the first layer point cloud and the second layer point cloud respectively, so as to obtain point cloud features with rotation invariance.

[0021] S3. Based on the feature similarity between the two frames of the second-layer point cloud, select the local regions with high similarity, construct the third-layer point cloud, and extract its features.

[0022] S4. Repeat step S3 to further filter out local regions with high similarity, construct the fourth layer of point cloud and extract its features;

[0023] S5. Enhance the interaction of features between the second, third and fourth layer point clouds and fuse multi-level contextual information;

[0024] S6. Perform coarse registration based on the third layer point cloud to estimate the initial pose transformation;

[0025] S7. Perform fine registration based on the second layer point cloud, optimize the initial pose, and obtain the final registration result.

[0026] This invention presents a multi-level adaptive point cloud registration method for open-pit mines. By constructing a multi-level feature extraction and matching enhancement system, it significantly improves the registration accuracy and efficiency of point clouds in complex environments such as open-pit mines compared to existing technologies. In terms of feature modeling, an improved kernel-point convolutional structure incorporating angle information effectively enhances rotation invariance and feature representation capabilities, significantly improving feature extraction quality and matching robustness. Simultaneously, a progressive local region filtering mechanism improves the focus and accuracy of feature matching while reducing unnecessary computational overhead. A multi-scale upsampling and downsampling fusion structure expands the network's ability to model geometric semantic information at different resolutions, jointly optimizing the coarse and fine registration processes. This reduces the method's dependence on initial pose estimation and allows for higher registration accuracy while maintaining high efficiency.

[0027] In some embodiments of the present invention, the downsampling in step S1 utilizes a voxel grid filtering method to generate point cloud layers of different resolutions and establish inter-layer neighborhood index relationships. Specifically, two layers of voxel downsampling are performed on the input source frame point cloud and the target frame point cloud respectively to generate the first layer and the second layer point cloud; at the same time, the neighborhood relationship of each layer and the cross-layer downsampling and upsampling indexes are constructed to lay the foundation for subsequent multi-scale feature extraction.

[0028] In some embodiments of the present invention, the extraction based on the improved kernel-point convolutional network in step S2 includes:

[0029] S21. Construct a local neighborhood structure and calculate the angle information between neighboring points; specifically, construct a local neighborhood structure suitable for convolution operations for the point cloud from which feature information needs to be extracted, and calculate the angle information between neighboring points based on the structure.

[0030] S22. Initialize spatially distributed kernel points based on spherical neighborhood search and calculate the angle information of the kernel point center; specifically, based on the spherical neighborhood search strategy, construct a local point set within a fixed radius for each point and initialize several spatially distributed kernel points, each kernel point having a predefined position coordinate in three-dimensional space; at the same time, calculate the center point of the kernel point and treat the kernel point as a neighborhood point to calculate the angle information;

[0031] S23. Perform weighted feature aggregation based on the Euclidean distance and angle difference between neighboring points and kernel points to generate rotation-invariant features. Specifically, for each center point in the point cloud from which feature information needs to be extracted, perform weighted calculation using the Euclidean distance and angle difference between its neighboring points and each kernel point. Through weighted feature aggregation, kernel point convolution operation is implemented to obtain new local descriptive features.

[0032] In some embodiments of the present invention, the adaptive downsampling in steps S3 and S4 includes:

[0033] First, calculate the cosine similarity matrix between the point cloud features of the two frames;

[0034] Secondly, similarity inference is performed using the Softmax activation function to filter out valid points that meet the bidirectional matching conditions;

[0035] Finally, points that do not meet the criteria are removed to obtain a higher-level sparse point cloud.

[0036] In some embodiments of the present invention, the interactive enhancement and fusion of multi-layer point cloud features in step S5 includes the following steps:

[0037] S51. Establish a cross-level feature propagation index through KNN search;

[0038] S52. Perform self-attention and cross-attention enhancement on each layer of features;

[0039] S53. Upsample low-resolution features to a high-resolution point cloud;

[0040] S54. Downsample the high-resolution features to the sparse point cloud;

[0041] S55. Features at different scales are spliced ​​together and fused using MLP to form a unified representation.

[0042] The present invention also discloses an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described point cloud registration method.

[0043] The present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described point cloud registration method. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating a multi-level adaptive point cloud registration method for open-pit mines according to an embodiment of the present invention.

[0045] Figure 2 This is a diagram of the kernel convolution model based on rotation-invariant feature extraction in this invention;

[0046] Figure 3 This is a flowchart of adaptive downsampling in an embodiment of the present invention;

[0047] Figure 4 This is a diagram of the context feature interaction network model in an embodiment of the present invention. Detailed Implementation

[0048] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0049] The following is for reference. Figures 1-4 A multi-level adaptive point cloud registration method for open-pit mines according to embodiments of the present invention is described, comprising the following steps:

[0050] S1. Input the source frame point cloud and the target frame point cloud, and perform two-layer downsampling on the source frame point cloud and the target frame point cloud to obtain the first layer point cloud and the second layer point cloud;

[0051] S2. An improved kernel convolutional network based on the introduction of angle information is used to extract features from the first layer point cloud and the second layer point cloud respectively, so as to obtain point cloud features with rotation invariance.

[0052] S3. Based on the feature similarity between the two frames of the second-layer point cloud, select the local regions with high similarity, construct the third-layer point cloud, and extract its features.

[0053] S4. Repeat step S3 to further filter out local regions with high similarity, construct the fourth layer of point cloud and extract its features;

[0054] S5. Enhance the interaction of features between the second, third and fourth layer point clouds and fuse multi-level contextual information;

[0055] S6. Perform coarse registration based on the third layer point cloud to estimate the initial pose transformation;

[0056] S7. Perform fine registration based on the second layer point cloud, optimize the initial pose, and obtain the final registration result.

[0057] This invention employs an adaptive hierarchical construction mechanism driven by feature similarity. By dynamically analyzing the local feature distribution of the initial point cloud, it automatically identifies regions with high registration value and constructs a more refined high-level point set accordingly. This replaces the traditional fixed pyramid downsampling method, reducing redundant point interference from the source and preserving key structural information, thereby improving registration quality without increasing computational complexity. Simultaneously, a rotation-invariant feature extraction mechanism based on angle information modeling is introduced. By encoding the relative angles and spatial structural relationships between neighboring points, the stability of point cloud features under changing viewing angles is enhanced. This mechanism effectively combats feature mismatch caused by sensor attitude changes, ensuring the robustness and continuity of the registration process. Ultimately, this method significantly improves registration accuracy and robustness while maintaining registration efficiency, making it suitable for scenarios such as open-pit mines where real-time perception and registration robustness requirements are high.

[0058] Assume the source frame point cloud is S, the target frame point cloud is P, and S = {s} i ∈R 3 {i = 1, ..., M}, P = {p} i ∈R 3 Let M be the number of point clouds in the original frame and N be the number of point clouds in the target frame.

[0059] Given that a single frame of point cloud data can contain hundreds of thousands of points in a large scene, and considering device limitations and efficiency issues, we first perform downsampling processing on the source frame point cloud S and the target frame point cloud P. For example, we can use a sampleable voxel grid filtering method to downsample, generate point cloud layers of different resolutions, and establish inter-layer neighborhood index relationships.

[0060] Specifically, two point clouds are generated from the source frame point cloud S and the target frame point cloud P through voxel downsampling, resulting in the first point cloud layer: S 1 P 1 First layer point cloud: S 2 P 2 We used radius search and KNN methods to construct the self-neighborhood, downsampled neighborhood and upsampled index of each layer, respectively, to support subsequent convolution and feature fusion.

[0061] In some embodiments of the present invention, step S2, which introduces kernel-point convolutional networks with angle information to extract features of the first and second layer point clouds, includes the following steps:

[0062] S21. Construct a local neighborhood structure and calculate the angle information between neighboring points: Construct a local neighborhood structure suitable for convolution operations for the point cloud from which feature information needs to be extracted, and calculate the angle information between neighboring points based on the structure.

[0063] S22. Initialize spatially distributed core points based on spherical neighborhood search and calculate the angle information of the core point center: Based on the spherical neighborhood search strategy, construct a local point set within a fixed radius for each point and initialize several spatially distributed core points. Each core point has predefined position coordinates in three-dimensional space; at the same time, calculate the center point of the core point and treat the core point as a neighborhood point to calculate the angle information.

[0064] S23. Perform weighted feature aggregation based on the Euclidean distance and angle difference between neighboring points and kernel points to generate rotation-invariant features: For each center point in the point cloud from which feature information needs to be extracted, perform weighted calculation using the Euclidean distance and angle difference between its neighboring points and each kernel point. Through weighted feature aggregation, kernel point convolution operation is achieved to obtain new local descriptive features.

[0065] Specifically, such as Figure 2 As shown, for each point x in the point cloud from which features are to be extracted... i First, based on the pre-obtained self-neighborhood And calculate the angle information between neighboring points.

[0066] Next, several fixed-distribution kernel points are initialized. The positions of these kernel points are fixed offsets relative to preset points, simulating the convolutional kernel structure in traditional images. Kernel points can also be considered as spatial templates, achieving spatially perceptual feature convergence by matching their distances to actual neighboring points. Simultaneously, the center point of each kernel point is calculated, and the kernel points are treated as neighboring points to calculate angular information.

[0067] Then, by using the features f of the neighborhood points j The convolutional feature representation f(p) of a point is obtained by weighting and summing the features based on its distance and angle from the kernel point. i ):

[0068]

[0069] Among them, w m (p j ) represents the weight of point i relative to kernel point j, calculated based on the Gaussian kernel distance weight; M represents the number of point clouds in the original frame.

[0070] After obtaining the initial geometric features, a residual convolution module is added to enhance the representation capability, and the number of channels is increased from the input dimension to twice the initial feature dimension to obtain the high-dimensional geometric features of the point cloud.

[0071] Finally, the features of the two point cloud layers sampled in step S2 are obtained. in, Features of the first layer of point cloud, This represents the characteristics of the second layer of point clouds.

[0072] In some embodiments of the present invention, the adaptive downsampling in steps S3 and S4 includes: first calculating the cosine similarity matrix between the features of two frames of point clouds; then performing similarity inference through the Softmax activation function to filter valid points that meet the bidirectional matching conditions; and finally, removing points that do not meet the conditions to obtain a higher-level sparse point cloud.

[0073] In other words, the feature similarity in step S3 is determined by the feature similarity of the two frames of the second-layer point cloud. and The cosine similarity between the reference points is represented and converted into a matching score using the Softmax function. In this matrix, each reference point selects its K most similar source points as candidate matches; then, the similarity list of these source points is checked in reverse to determine if the reference point is also considered one of its K most similar points. If they are neighbors, the reference point and its corresponding source point are considered to be within the overlapping region of the two frame point clouds. Finally, the reference points that meet the conditions are retained as the point cloud for the next layer to obtain a higher-level sparse point cloud.

[0074] Specifically, such as Figure 3 As shown, the point cloud features from two frames of the second layer are first utilized. and Solve for the cosine similarity matrix D between two point clouds. in, yes The transpose of the matrix;

[0075] Then, the Softmax activation function is used to infer the cosine similarity matrix D to obtain the similarity matrix M between features.

[0076] Next, the similarity matrix M is used to determine valid points. For each reference point x... i Select the k most similar source points in the i-th row of the similarity matrix M, with indices {j1,...,j...}. k Further check whether bidirectional nearest neighbor consistency is satisfied: that is, determine x. i Does it also appear among the top k most similar target points selected in the column vector of the similarity matrix corresponding to the source point?

[0077] If the two-way condition is not met, it is considered not to belong to the valid region and can be removed as an outlier; if the two-way condition is met, then x is considered to be... i This region is considered valid and can be retained for further feature aggregation, local matching, or weighted registration. The third point cloud is obtained through adaptive downsampling: S 3 P 3 And extract its features in accordance with the method shown in step S2:

[0078] Similarly, by referring to the above steps, adaptive downsampling of the third-layer point cloud is used to obtain the fourth-layer point cloud S. 4 P 4 and its corresponding features

[0079] In some embodiments of the present invention, the interactive enhancement and fusion of multi-layer point cloud features in step S5 includes the following steps:

[0080] S51. Establish cross-level feature propagation index through KNN search: Perform KNN search in the original point cloud to obtain the nearest neighbor index, which is used to upsample the sparse features to the high-resolution point cloud layer for nearest neighbor interpolation backpropagation; at the same time, perform KNN search from the original point cloud to the sparse point set to obtain the index and coordinate information required for sparse downsampling fusion.

[0081] S52. Perform self-attention and cross-attention enhancement on the features of each layer: Perform self-attention enhancement on the point cloud features of the second, third and fourth layers respectively, and then further perform cross-attention enhancement on the source frame point cloud and the target frame point cloud of each layer.

[0082] S53. Upsample low-resolution features to high-resolution point cloud: For low-resolution features, upsample them to the corresponding positions in dense point cloud using the nearest neighbor index obtained in S51 to obtain preliminary upsampled feature representation; at the same time, upsample the low-resolution features using the upsampling index to enhance the spatial response capability of the foreground region.

[0083] S54. Downsample high-resolution features to sparse point cloud: For dense features, downsample the index required for downsampling fusion obtained in S51 to the previous point set to obtain the fused sparse representation; at the same time, further downsample it to the sparse point set to generate new sparse features to support cross-scale modeling.

[0084] S55. Features at different scales are concatenated and fused using an MLP to form a unified representation: Upsampled and downsampled features at the same scale are concatenated by feature dimension, fusing geometric and semantic information from different scales to form a unified representation. Finally, the fused features are dimensionally compressed and reorganized using an MLP network to improve their expressive power and be used as input for subsequent modules.

[0085] Specifically, such as Figure 4 As shown, step S5 mainly utilizes the point cloud and feature information of the second, third and fourth layers to interactively enhance the point cloud features of the third layer.

[0086] In this process, to align the second-layer point cloud with the third-layer point cloud, for each point in the third layer, its K nearest neighbors are found in the second-layer point set, and their spatial coordinates and indices are extracted. Using the features of these neighboring points F2, the feature representation F of the current downsampled point is jointly calculated through distance-weighted interpolation. 3_down This achieves downsampling alignment. The formula is expressed as:

[0087] F 3_down =F3+MLP(down_sampling(F2));

[0088] Here, F3 represents the features of the third layer point cloud before upsampling and downsampling, MLP represents the multilayer perceptron, and down_sampling() represents the feature downsampling operation through distance-weighted interpolation.

[0089] Correspondingly, to align the fourth-layer point cloud with the third-layer point cloud, the nearest neighbor F4 of each point in the third-layer point cloud is found in the fourth-layer point set. The corresponding low-resolution features are then assigned values ​​or subjected to distance-weighted propagation to complete the feature upsampling alignment F. 3_up Its formula is expressed as:

[0090] F 3_up =F3+MLP(down_sampling(F4));

[0091] Here, F3 represents the features of the third layer point cloud before upsampling and downsampling, MLP represents the multilayer perceptron, and up_sampling() represents the feature upsampling operation.

[0092] Next, the aligned third-layer point cloud features F are obtained by upsampling and downsampling these two samples respectively. 3_up and F 3_down The points are stitched together and fused using an MLP network to obtain the final interactive point cloud features. The formula is expressed as:

[0093]

[0094] Here, concat(x,y) means concatenating the features along the feature channel dimension.

[0095] Features after interaction Perform self-attention modeling to capture the contextual dependencies between points;

[0096] Based on the features of the source / target point clouds after self-attention modeling, similarity scores are calculated between point pairs. A threshold is applied to these similarity scores to obtain a set of significantly matched point pairs, guiding further cross-attention interactions between the source and target point clouds, resulting in the final features of the third layer of source and target point clouds. and

[0097] Understandably, in step S5, the interactive enhancement of features first calculates the K-nearest neighbor index of each layer of point cloud based on point-to-point distance to construct a local structure for subsequent feature enhancement and fusion; then, self-attention, cross-attention, upsampling and downsampling fusion, and MLP feature compression are performed sequentially to fuse multi-scale geometric and semantic information; initial matching scores and compatibility scores are obtained through feature matching; salient points are selected based on matching confidence, and the expression of salient regions is enhanced through consistency attention; finally, cross-attention guided by salient regions is used to further interact features between the source / target point clouds to improve the discriminativeness and robustness of matching.

[0098] In some embodiments of the present invention, step S6 utilizes the final features of the third-layer source point cloud and the target point cloud. and Calculate the matching score between points, and transfer the points s from the source point cloud. i The point p with the highest matching score in the target point cloud j These are considered as corresponding points. Using this rule, corresponding points are found for all points in the source point cloud, and the coarse pose T is solved using Singular Value Decomposition (SVD). c .

[0099] In some embodiments of the present invention, step S7 utilizes coarse pose T c Using prior values ​​in conjunction with the second-layer point cloud, the fine pose T is solved through the Iterative Closest Point (ICP) algorithm. r .

[0100] In summary, the multi-level adaptive point cloud registration method for open-pit mines according to embodiments of the present invention significantly improves the registration accuracy and efficiency of point clouds in complex environments such as open-pit mines compared to existing technologies by constructing a multi-level feature extraction and matching enhancement system. In terms of feature modeling, an improved kernel-point convolutional structure incorporating angle information effectively enhances rotation invariance and feature representation capabilities, significantly improving feature extraction quality and matching robustness. Simultaneously, a progressive local region filtering mechanism improves the focus and accuracy of feature matching while reducing unnecessary computational overhead. The multi-scale upsampling and downsampling fusion structure expands the network's ability to model geometric and semantic information at different resolutions, jointly optimizing the coarse and fine registration processes. This weakens the dependence on initial pose estimation and allows for higher registration accuracy while maintaining high efficiency.

[0101] The present invention also discloses an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described point cloud registration method.

[0102] The present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described point cloud registration method.

[0103] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A multi-level adaptive point cloud registration method for open-pit mines, characterized in that, Includes the following steps: S1. Input the source frame point cloud and the target frame point cloud, and perform two-layer downsampling on the source frame point cloud and the target frame point cloud to obtain the first layer point cloud and the second layer point cloud; S2. An improved kernel convolutional network based on the introduction of angle information is used to extract features from the first layer point cloud and the second layer point cloud respectively, so as to obtain point cloud features with rotation invariance. S3. Based on the feature similarity between the two frames of the second-layer point cloud, select the local regions with high similarity, construct the third-layer point cloud, and extract its features. S4. Repeat step S3 to further filter out local regions with high similarity, construct the fourth layer of point cloud and extract its features; S5. Enhance the interaction of features between the second, third and fourth layer point clouds and fuse multi-level contextual information; S6. Perform coarse registration based on the third layer point cloud to estimate the initial pose transformation; S7. Perform fine registration based on the second layer point cloud, optimize the initial pose, and obtain the final registration result.

2. The multi-level adaptive point cloud registration method for open-pit mines according to claim 1, characterized in that, The downsampling in step S1 uses a voxel grid filtering method to generate point cloud layers of different resolutions and establish inter-layer neighborhood index relationships.

3. The multi-level adaptive point cloud registration method for open-pit mines according to claim 1, characterized in that, The improved kernel-point convolutional network in step S2 includes: S21. Construct a local neighborhood structure and calculate the angle information between neighborhood points; S22. Initialize the spatial distribution of kernel points based on spherical neighborhood search, and calculate the angle information of the kernel point center; S23. Perform weighted feature aggregation based on the Euclidean distance and angle difference between neighboring points and the core point to generate rotation-invariant features.

4. The multi-level adaptive point cloud registration method for open-pit mines according to claim 1, characterized in that, The interactive enhancement and fusion of multi-layer point cloud features in step S5 includes the following steps: S51. Establish a cross-level feature propagation index through KNN search; S52. Perform self-attention and cross-attention enhancement on each layer of features; S53. Upsample low-resolution features to a high-resolution point cloud; S54. Downsample the high-resolution features to the sparse point cloud; S55. Features at different scales are spliced ​​together and fused using MLP to form a unified representation.

5. The multi-level adaptive point cloud registration method for open-pit mines according to claim 1, characterized in that, The adaptive downsampling in steps S3 and S4 includes: Calculate the cosine similarity matrix between the point cloud features of two frames; Similarity inference is performed using the Softmax activation function to filter valid points that meet the bidirectional matching conditions; Points that do not meet the criteria are removed to obtain a higher-level sparse point cloud.

6. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the point cloud registration method according to any one of claims 1-5.

7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the point cloud registration method according to any one of claims 1-5.

Images