Airborne laser point cloud strip mine feature ground feature extraction method and system

Abstract

The invention relates to the technical field of mining area management, and provides an airborne laser point cloud strip mine feature ground feature extraction method and system. The method comprises the following steps: carrying out voxelization downsampling on point cloud data, identifying step line candidate feature points based on a plane fitting residual error and normal vector angle dispersion, and carrying out fitting; performing spatial partitioning according to a distance relation between points and step lines, constructing a local neighborhood calculation context enhancement feature, and compressing to generate a coding feature; and through interpolation diffusion to non-sampling points, fusing the non-sampling points with hierarchical features to generate point-level semantic tags, and finally realizing accurate extraction of feature ground features such as steps and roads in the strip mine area. According to the invention, high-precision automatic extraction of the characteristic ground features of the strip mine area is realized, and technical support is provided for monitoring and management of the strip mine area.

Description

Technical Field

[0001] This invention relates to the field of mining area management technology, and in particular to a method and system for extracting feature features from open-pit mines using airborne laser point clouds. Background Technology

[0002] With the continuous expansion of open-pit mining, the terrain and landforms of open-pit mining areas are becoming increasingly complex and varied, leading to a rise in potential geological hazards and placing higher demands on the safety monitoring and management of these areas. Airborne LiDAR technology, as an efficient and high-precision three-dimensional topographic surveying method, can quickly acquire large-scale, high-density point cloud data, providing crucial technical support for the extraction and monitoring of terrain features in open-pit mining areas.

[0003] In the management and monitoring of open-pit mines, the accurate extraction of characteristic features such as steps, slopes, and transport roads is crucial for mine safety, disaster early warning, and production scheduling. Traditional open-pit mine feature extraction relies primarily on manual identification and measurement, which is labor-intensive and inefficient, failing to meet the real-time monitoring needs of large-scale mines. With the maturity of airborne laser point cloud data acquisition technology, how to automatically and accurately extract characteristic features from massive point cloud data has become a current research hotspot.

[0004] Currently, open-pit mine feature extraction technology suffers from the following shortcomings: Existing point cloud processing methods are computationally inefficient when handling large-scale open-pit mine point cloud data, often requiring extensive preprocessing and manual intervention, failing to achieve efficient and automated feature extraction. This is especially true when point cloud density is uneven and noise levels are high, making processing even more difficult. Open-pit mine terrain is complex and varied, and existing feature extraction algorithms lack sufficient accuracy in identifying key features such as step edges and slope change zones. Particularly under complex terrain conditions, traditional geometric feature-based methods struggle to accurately distinguish between real feature boundaries and noise points, leading to inaccurate extraction results. Existing technologies lack full utilization of the spatial context information of point cloud data; most methods only consider local geometric features, ignoring the spatial correlation between feature points, failing to effectively utilize global information to assist feature extraction, resulting in insufficient robustness of feature recognition in complex environments and difficulty adapting to the terrain characteristics of different mining areas. Summary of the Invention

[0005] This invention provides a method and system for extracting feature features from open-pit mines using airborne laser point clouds, which can solve the problems in the prior art.

[0006] A first aspect of the present invention provides a method for extracting feature features from open-pit mines using airborne laser point clouds, comprising:

[0007] Airborne laser point cloud data containing open-pit mine areas and surrounding features is acquired, and the airborne laser point cloud data is downsampled using voxelization to obtain downsampled point cloud data;

[0008] In the downsampled point cloud data, candidate feature points for step lines are identified in the local neighborhood of each target point based on a joint criterion of plane fitting residual and normal vector angle dispersion. Curve fitting is then performed on the candidate feature points for step lines, and step line fitting results are recursively generated.

[0009] The downsampled point cloud data is spatially correlated with the step line fitting result. The downsampled point cloud data is spatially divided into blocks according to the distance relationship from the point to the step line. A center point is selected from each spatial block, and a local neighborhood is constructed around each center point.

[0010] Within the local neighborhood, context-enhanced features are calculated based on the feature vectors of neighboring points. These context-enhanced features are then processed to obtain compressed features. Encoded features are then generated based on these compressed features and the feature vectors.

[0011] The encoded features are diffused to unsampled points through interpolation and fused with hierarchical features to generate point-level semantic labels. Target ground features are extracted based on the fitting results of the point-level semantic labels and the step line.

[0012] In the downsampled point cloud data, candidate feature points for step lines are identified within the local neighborhood of each target point based on a joint criterion of plane fitting residual and normal vector angle dispersion. Curve fitting is then performed on these candidate feature points to recursively generate step line fitting results, including:

[0013] A target plane is obtained by fitting within the local neighborhood of each target point, and the perpendicular distance from each point in the neighborhood to the target plane is calculated as the fitting residual.

[0014] For each point in the neighborhood, the normal vector is calculated by eigenvalue decomposition of the neighborhood covariance matrix. The normal vector is then projected onto the tangent plane of the target plane, and the angle between the projected vectors is calculated to determine the angular dispersion of the normal vector.

[0015] Points that simultaneously satisfy the condition that the fitting residual is greater than the first threshold and the normal vector angle dispersion is greater than the second threshold are marked as candidate feature points for the step line;

[0016] Using the candidate feature points of the step line as seed points, region growing is performed. Neighboring candidate feature points whose spatial distance is less than a distance threshold and whose normal vector angle is less than an angle threshold are merged into the same region to obtain multiple candidate feature point sequences.

[0017] For each candidate feature point sequence, the candidate feature points in the sequence are processed sequentially. Based on the alternating iteration of state prediction and observation update, step line fitting points are generated. All step line fitting points are connected to form the step line fitting result.

[0018] For each candidate feature point sequence, the candidate feature points in the sequence are processed sequentially. Based on the alternating iteration of state prediction and observation update, step line fitting points are generated. All step line fitting points are connected to form the step line fitting result, including:

[0019] Predict the prior state vector of the next candidate feature point based on the state vector of the current candidate feature point in the candidate feature point sequence;

[0020] Using the actual position coordinates of the next candidate feature point in the candidate feature point sequence as the observation vector, calculate the residual vector between the observation vector and the position coordinate components in the prior state vector;

[0021] Calculate the magnitude of the residual vector, adjust the prediction error covariance matrix based on the magnitude, and calculate the Kalman gain matrix based on the adjusted prediction error covariance matrix and the observation noise covariance matrix.

[0022] The prior state vector is corrected based on the Kalman gain matrix and the residual vector to obtain the posterior state vector. Based on the angle between the tangent direction component in the posterior state vector and the tangent direction component in the adjacent posterior state vector, curvature control points are inserted at the corresponding positions of the posterior state vector.

[0023] The position coordinate components in the posterior state vector are extracted as step line fitting points. The posterior state vector is used as the current state vector for the next iteration. All candidate feature points are processed sequentially, and all step line fitting points are connected with the curvature control points in the processing order to form the step line fitting result.

[0024] The downsampled point cloud data is spatially correlated with the step line fitting result. Based on the distance relationship from the points to the step line, the downsampled point cloud data is spatially divided into blocks. A center point is selected from each spatial block, and a local neighborhood is constructed around each center point, including:

[0025] For each point in the downsampled point cloud data, calculate the shortest distance from the point to all step lines in the step line fitting result, and use the shortest distance as the step line distance attribute of the point;

[0026] Based on the step line distance attribute, the downsampled point cloud data is divided into multiple distance intervals, each distance interval corresponding to a distance range;

[0027] For the point cloud data within each distance interval, the point cloud data is divided into multiple spatial blocks using a spatial grid partitioning method;

[0028] For each spatial block, the average spatial coordinates of all points within the block are calculated as the block center coordinates. Using the block center coordinates as a reference point, the point closest to the block center coordinates is selected as the center point within the spatial block.

[0029] Using the center point as the center of a sphere, the neighborhood radius is determined based on the point cloud density of the spatial block, and all points whose distance from the center point is less than the neighborhood radius constitute a local neighborhood.

[0030] Within the local neighborhood, context-enhanced features are calculated based on the feature vectors of neighboring points. These context-enhanced features are then processed to obtain compressed features. Encoded features are generated based on the compressed features and the feature vectors, including:

[0031] For neighborhood points within a local neighborhood, a linear transformation is used to map the feature vectors of the neighborhood points into query vectors, key vectors, and value vectors, respectively.

[0032] Calculate the relevance score between the query vector and the key vector, calculate the distance decay weight based on the spatial distance between neighboring points and fuse it with the relevance score to obtain a spatially perceived relevance score, and obtain context-enhanced features based on the spatially perceived relevance score and the value vector;

[0033] The context enhancement features are grouped by channel, and the channels are divided into multiple subgroups. Each subgroup is expanded by channel expansion, and different expansion factors are used for different subgroups to obtain multi-scale expanded features.

[0034] The multi-scale extended features are selectively activated by applying a gating mechanism to obtain gated modulation features. Each subgroup of the gated modulation features is then compressed into channels, and the compressed subgroup features are concatenated to obtain compressed features.

[0035] Calculate the channel attention weights between the compressed feature and the feature vector, perform channel weighting on the compressed feature based on the channel attention weights, and perform residual fusion between the weighted compressed feature and the feature vector to obtain the encoded feature.

[0036] Calculate the relevance score between the query vector and the key vector, calculate the distance decay weight based on the spatial distance between neighboring points and fuse it with the relevance score to obtain a spatially perceived relevance score, and obtain context-enhanced features based on the spatially perceived relevance score and the value vector, including:

[0037] Calculate the inner product of the query vector and the key vector, and scale the inner product according to the feature dimension of the query vector to obtain the basic relevance score;

[0038] Calculate the Euclidean distance between the center point and each neighboring point within the local neighborhood, and calculate the distance decay weight based on the Euclidean distance, wherein the Euclidean distance is inversely proportional to the distance decay weight;

[0039] The spatial modulation correlation score is determined based on the basic correlation score and the distance attenuation weight, and the spatial modulation correlation score is then biased to obtain the spatial perception correlation score.

[0040] The temperature parameter is calculated based on the numerical distribution characteristics of the spatial perception correlation score, and the spatial perception correlation score is scaled using the temperature parameter to obtain the adjusted spatial perception correlation score.

[0041] The adjusted spatial perception relevance score is mapped to a positive distribution using an exponential function. The mapping result is normalized in the neighborhood point dimension to obtain normalized relevance weights. The normalized relevance weights are then weighted and aggregated with the value vector. Position encoding compensation is applied to the aggregation result to obtain context-enhanced features.

[0042] The encoded features are diffused to unsampled points through interpolation and fused with hierarchical features to generate point-level semantic labels. Target features are extracted based on the fitting result between the point-level semantic labels and the step line, including:

[0043] For unsampled points within a spatial block, search for the nearest sampled center points. Based on the spatial distance between the unsampled point and each sampled center point, perform distance-weighted interpolation on the encoding features of each sampled center point to obtain the interpolated encoding features of the unsampled point.

[0044] The interpolation coding features are fused with the hierarchical features of the corresponding spatial location across hierarchical levels to obtain fused features. Based on the fused features, semantic prediction is performed on the unsampled points to determine the point-level semantic labels of the unsampled points.

[0045] Spatial connectivity analysis is performed on the generated point-level semantic labels, and points with the same point-level semantic labels and spatial connectivity are clustered to obtain candidate land cover regions;

[0046] For each candidate feature area, the spatial envelope range of the candidate feature area is calculated, and its relative positional relationship with the fitting result of the step line is used to determine the spatial functional zone to which the candidate feature area belongs;

[0047] Candidate land cover areas that meet the spatial functional zoning constraints and whose number of points meets the regional scale requirements are marked as valid target land cover areas, and the classification of the valid target land cover areas is determined based on point-level semantic labels.

[0048] A second aspect of the present invention provides an airborne laser point cloud feature extraction system for open-pit mines, comprising:

[0049] The first unit is used to acquire airborne laser point cloud data containing open-pit mine areas and surrounding features, and to perform voxel downsampling on the airborne laser point cloud data to obtain downsampled point cloud data.

[0050] The second unit is used to identify candidate feature points of the step line in the local neighborhood of each target point in the downsampled point cloud data based on the joint criterion of plane fitting residual and normal vector angle dispersion, and to perform curve fitting on the candidate feature points of the step line to recursively generate the step line fitting result;

[0051] The third unit is used to spatially correlate the downsampled point cloud data with the step line fitting result, divide the downsampled point cloud data into spatial blocks according to the distance relationship from the point to the step line, select a center point for each spatial block, and construct a local neighborhood around each center point.

[0052] The fourth unit is used to calculate context-enhanced features based on the feature vectors of neighboring points within the local neighborhood, process the context-enhanced features to obtain compressed features, and generate encoded features based on the compressed features and the feature vectors.

[0053] The fifth unit is used to diffuse the encoded features to unsampled points through interpolation and fuse them with hierarchical features to generate point-level semantic labels, and extract target ground features based on the fitting results of the point-level semantic labels and the step line.

[0054] A third aspect of the present invention provides an electronic device, comprising:

[0055] processor;

[0056] Memory used to store processor-executable instructions;

[0057] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0058] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0059] This invention identifies candidate feature points for step lines using a joint criterion based on the plane fitting residual and the dispersion of the normal vector angle. This enables more accurate capture of step line features in the complex topography of open-pit mines, improving the accuracy and robustness of feature extraction. The invention utilizes the step line fitting results to spatially partition point cloud data, fully considering the topographic features of open-pit mines. This allows subsequent processing to better adapt to the unique topographic structure of the mining area, improving processing efficiency. The invention introduces a context-enhanced feature computation mechanism based on local neighborhoods, effectively capturing local structural information in point cloud data and significantly improving feature representation capabilities. Through the design of compressed features and feature vector generation encoding features, this invention achieves efficient encoding of multi-level features in point cloud data, which is beneficial for subsequent feature identification. This invention solves the problem of information loss after point cloud downsampling through feature diffusion and hierarchical feature fusion, ensuring the integrity and accuracy of the final extraction results. This invention forms a complete technical route, from downsampling and feature extraction to semantic annotation, achieving high-precision automatic extraction of feature features in open-pit mines, providing technical support for open-pit mine monitoring and management. Attached Figure Description

[0060] Figure 1 This is a schematic flowchart of the airborne laser point cloud open-pit mine feature extraction method according to an embodiment of the present invention.

[0061] Figure 2 This is a schematic diagram of the process for generating step line fitting results according to an embodiment of the present invention. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0064] Figure 1 This is a schematic flowchart of the airborne laser point cloud method for extracting feature features in open-pit mines, according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0065] Airborne laser point cloud data containing open-pit mine areas and surrounding features is acquired, and the airborne laser point cloud data is downsampled using voxelization to obtain downsampled point cloud data;

[0066] In the downsampled point cloud data, candidate feature points for step lines are identified in the local neighborhood of each target point based on a joint criterion of plane fitting residual and normal vector angle dispersion. Curve fitting is then performed on the candidate feature points for step lines, and step line fitting results are recursively generated.

[0067] The downsampled point cloud data is spatially correlated with the step line fitting result. The downsampled point cloud data is spatially divided into blocks according to the distance relationship from the point to the step line. A center point is selected from each spatial block, and a local neighborhood is constructed around each center point.

[0068] Within the local neighborhood, context-enhanced features are calculated based on the feature vectors of neighboring points. These context-enhanced features are then processed to obtain compressed features. Encoded features are then generated based on these compressed features and the feature vectors.

[0069] The encoded features are diffused to unsampled points through interpolation and fused with hierarchical features to generate point-level semantic labels. Target ground features are extracted based on the fitting results of the point-level semantic labels and the step line.

[0070] In one optional implementation, candidate feature points for step lines are identified in the local neighborhood of each target point in the downsampled point cloud data based on a joint criterion of plane fitting residual and normal vector angle dispersion. Curve fitting is then performed on these candidate feature points to recursively generate step line fitting results, including:

[0071] A target plane is obtained by fitting within the local neighborhood of each target point, and the perpendicular distance from each point in the neighborhood to the target plane is calculated as the fitting residual.

[0072] For each point in the neighborhood, the normal vector is calculated by eigenvalue decomposition of the neighborhood covariance matrix. The normal vector is then projected onto the tangent plane of the target plane, and the angle between the projected vectors is calculated to determine the angular dispersion of the normal vector.

[0073] Points that simultaneously satisfy the condition that the fitting residual is greater than the first threshold and the normal vector angle dispersion is greater than the second threshold are marked as candidate feature points for the step line;

[0074] Using the candidate feature points of the step line as seed points, region growing is performed. Neighboring candidate feature points whose spatial distance is less than a distance threshold and whose normal vector angle is less than an angle threshold are merged into the same region to obtain multiple candidate feature point sequences.

[0075] For each candidate feature point sequence, the candidate feature points in the sequence are processed sequentially. Based on the alternating iteration of state prediction and observation update, step line fitting points are generated. All step line fitting points are connected to form the step line fitting result.

[0076] Figure 2 This is a schematic diagram illustrating the process of generating step line fitting results according to an embodiment of the present invention. Figure 2 As shown, firstly, point cloud data of the indoor environment is acquired. Downsampling is then used to reduce the data volume and improve subsequent processing efficiency. A commonly used downsampling method is voxel grid filtering, which divides the 3D space into equal-sized cubic grids. Points within each grid are replaced by their centroids, thus significantly reducing the number of points while preserving spatial structural features.

[0077] In the downsampled point cloud, local neighborhood analysis is performed for each target point to identify candidate feature points for the step line. Local neighborhoods are typically determined using either K-nearest neighbors or spherical neighborhoods. For example, the 30 nearest points around the target point or all points within a 3-centimeter radius are selected as its local neighborhood.

[0078] For the local neighborhood of each target point, a plane fitting is first performed. Principal component analysis is used to calculate the covariance matrix of the neighborhood point set, and eigenvalue decomposition yields three eigenvectors and their corresponding eigenvalues. The eigenvector with the smallest eigenvalue is the plane normal vector, which is used to determine the equation of the target plane. The perpendicular distance from each point in the neighborhood to this plane is calculated as the plane fitting residual. If the distance from a point to the plane exceeds 2 millimeters, it indicates that the point is located in a geometrically discontinuous region.

[0079] Simultaneously, the normal vector dispersion of points within the neighborhood is calculated. For each point in the neighborhood, the covariance matrix of its small local neighborhood (e.g., the 15 nearest points) is constructed, and the normal vector is obtained through eigenvalue decomposition. All normal vectors are projected onto the tangent plane of the target plane (i.e., the component of the normal vector in the normal direction of the target plane is removed), and the angle between these projected vectors is calculated. The normal vector angle dispersion is defined as a statistical measure of the angle between the projected vectors, such as the standard deviation or the maximum angle. When the normal vector angle dispersion is greater than 15 degrees, it indicates that the region has edge features.

[0080] Points that simultaneously meet two conditions are marked as candidate feature points for the step line: the plane fitting residual is greater than the first threshold (e.g., 2 mm) and the normal vector angular dispersion is greater than the second threshold (e.g., 15 degrees). These points are usually located at spatial geometric discontinuities and have obvious directional changes, which conforms to the geometric characteristics of the step edge.

[0081] Next, spatial clustering is performed on the labeled candidate feature points using a distance-based region growing algorithm. Starting with any unprocessed candidate feature point as a seed point, adjacent candidate points with a spatial distance of less than 5 mm and an angle between their normal vectors of less than 30 degrees are merged into the same region. This process is iterated until no new points can be added, and then a new seed point is selected to start the next region growing, ultimately obtaining multiple spatially continuous candidate feature point sequences.

[0082] Curve fitting is performed on each candidate feature point sequence to generate a smooth and coherent step line. A recursive Bayesian filtering method is then used to transform the curve fitting problem into a state estimation problem. First, the starting state (position and orientation) is initialized. Then, the points in the sequence are processed sequentially: in the prediction phase, the position and orientation of the next point are estimated based on the current state; in the update phase, the prediction results are corrected by incorporating information from actual observation points. This alternating iteration ensures the smoothness of the fitted curve and its fidelity to the original data.

[0083] Finally, all fitted points are connected in sequence to form a step line, which can be used for applications such as 3D reconstruction of indoor scenes and robot navigation. The accuracy of the fitting results can be verified by comparing with manually labeled data. Experiments show that this method has a stable step line recognition capability in various indoor environments and is particularly suitable for processing real point cloud data containing noise.

[0084] In one optional implementation, each candidate feature point sequence is processed sequentially, and step line fitting points are generated based on alternating iterations of state prediction and observation update. All step line fitting points are then connected to form a step line fitting result, including:

[0085] Predict the prior state vector of the next candidate feature point based on the state vector of the current candidate feature point in the candidate feature point sequence;

[0086] Using the actual position coordinates of the next candidate feature point in the candidate feature point sequence as the observation vector, calculate the residual vector between the observation vector and the position coordinate components in the prior state vector;

[0087] Calculate the magnitude of the residual vector, adjust the prediction error covariance matrix based on the magnitude, and calculate the Kalman gain matrix based on the adjusted prediction error covariance matrix and the observation noise covariance matrix.

[0088] The prior state vector is corrected based on the Kalman gain matrix and the residual vector to obtain the posterior state vector. Based on the angle between the tangent direction component in the posterior state vector and the tangent direction component in the adjacent posterior state vector, curvature control points are inserted at the corresponding positions of the posterior state vector.

[0089] The position coordinate components in the posterior state vector are extracted as step line fitting points. The posterior state vector is used as the current state vector for the next iteration. All candidate feature points are processed sequentially, and all step line fitting points are connected with the curvature control points in the processing order to form the step line fitting result.

[0090] This invention processes each candidate feature point sequence sequentially, generating step line fitting points based on alternating state prediction and observation updates. All step line fitting points are then connected to form the final step line fitting result. This method utilizes the Kalman filter concept to achieve accurate fitting of the step line.

[0091] The candidate feature point sequence contains multiple candidate feature points, each with location coordinate information, typically represented by two-dimensional planar coordinates (x, y). During the step line fitting process, each candidate feature point in the sequence needs to be processed sequentially. Each time a candidate feature point is processed, an iterative process of state prediction and observation update is performed. The state vector contains location coordinate components and tangent direction components. The location coordinate components represent the position of the feature point, and the tangent direction component represents the tangent direction of the step line at that feature point.

[0092] When processing candidate feature points, the prior state vector of the next candidate feature point is first predicted based on the current candidate feature point's state vector. The current state vector contains position coordinate components (x, y) and tangent direction components (dx, dy), and a linear prediction model is used for the prediction process. When predicting the prior state vector of the next candidate feature point, the predicted position coordinates of the next candidate feature point can be obtained by adding the position coordinate components from the current state vector to the product of the tangent direction components and a preset step size, while keeping the tangent direction components unchanged. The preset step size can be adjusted according to the distribution density of candidate feature points; the default value can be set to the average distance between two adjacent candidate feature points, typically ranging from 0.1 to 10 pixels.

[0093] After prediction, the actual position coordinates of the next candidate feature point in the candidate feature point sequence are used as the observation vector. The observation vector contains only position information and not direction information. The residual vector between the observation vector and the position coordinate components of the prior state vector is calculated. The residual vector represents the difference between the predicted and actual positions. The residual vector is calculated by subtracting the predicted position coordinates from the actual position coordinates.

[0094] Next, the magnitude of the residual vector, i.e., the Euclidean norm of the residual vector, is calculated, reflecting the degree of deviation between the predicted and actual positions. The prediction error covariance matrix is ​​adjusted based on the magnitude of the residual vector. The adjustment mechanism is as follows: when the magnitude of the residual vector exceeds a preset threshold, the elements of the prediction error covariance matrix are proportionally amplified, making the system more sensitive to the observed data; when the magnitude of the residual vector is less than the preset threshold, the prediction error covariance matrix remains unchanged. The preset threshold can be set to 2 to 5 pixel units, and the specific value can be adjusted according to the characteristics of the step lines in the application scenario. The amplification ratio can be set as the ratio of the residual vector magnitude to the preset threshold, but cannot exceed the maximum amplification factor. The default value of the maximum amplification factor is 5, which can be adjusted between 2 and 10.

[0095] The Kalman gain matrix is ​​calculated based on the adjusted prediction error covariance matrix and the observation noise covariance matrix. The observation noise covariance matrix reflects the uncertainty of the observation data and can be preset according to the accuracy characteristics of the acquisition equipment. It is usually set as a diagonal matrix, with diagonal elements ranging from 0.1 to the square of 5 pixel units. The calculation of the Kalman gain matrix follows the standard Kalman filtering algorithm, which involves multiplying the adjusted prediction error covariance matrix by the transpose of the observation matrix, and then multiplying this by a combination of the observation matrix, the prediction error covariance matrix, and the observation noise covariance matrix to obtain the final Kalman gain matrix.

[0096] After calculating the Kalman gain matrix, the prior state vector is corrected based on the Kalman gain matrix and the residual vector to obtain the posterior state vector. The correction formula is the prior state vector plus the product of the Kalman gain matrix and the residual vector. The posterior state vector contains the corrected position coordinate components and tangent direction components. The position coordinate components are closer to the actual observation position, and the tangent direction components are also adjusted to adapt to the local shape of the step line.

[0097] After obtaining the posterior state vector, it is necessary to calculate the angle between the tangent direction components of this posterior state vector and the tangent direction components of the adjacent posterior state vectors. The angle is calculated using the vector dot product method, that is, the dot product of the two tangent direction vectors is divided by the product of their magnitudes, and then the inverse cosine is taken to obtain the radian value. When the calculated angle exceeds a preset angle threshold, it indicates that there is a significant directional change in the step line at this point, and a curvature control point needs to be inserted at the corresponding position in the posterior state vector. The preset angle threshold can be set between 15 and 45 degrees, with a default value of 30 degrees. The coordinates of the curvature control point are the same as the position coordinate components in the corresponding posterior state vector, but they are marked as control point type in the system to distinguish them from ordinary step line fitting points.

[0098] The position coordinate components are extracted from the posterior state vector as the step line fitting points. The posterior state vector is then used as the current state vector for the next iteration. All candidate feature points in the candidate feature point sequence are processed sequentially until the last candidate feature point is processed. After processing all candidate feature points, a series of step line fitting points and curvature control points are obtained.

[0099] Connect all the fitted points of the step line with the curvature control points in the processing order to form the fitted result of the step line. Piecewise linear interpolation or cubic spline interpolation methods can be used for the connection to generate a smooth transition between adjacent points, especially to ensure the continuity and smoothness of the curve at the curvature control points. When using cubic spline interpolation, the curvature control points act as control points for the spline curve, affecting the shape of the curve near those points, so that the fitted result better fits the geometric characteristics of the step.

[0100] A specific example is as follows: Suppose there is a sequence containing 5 candidate feature points with coordinates (10,20), (15,21), (20,25), (25,35), and (30,36). The initial state vector is set to [(10,20),(1,0.2)], where (10,20) are the initial position coordinate components, and (1,0.2) are the initial tangent direction components. The prediction error covariance matrix is ​​initially set as a diagonal matrix with diagonal elements of 1.0, the observation noise covariance matrix is ​​set as a diagonal matrix with diagonal elements of 0.5, the preset angle threshold is set to 30 degrees, and the preset residual magnitude threshold is set to 3 pixels.

[0101] After processing the first candidate feature point (10,20), the prior state vector for the next point is predicted using the initial state vector as [(15,21),(1,0.2)]. The observation vector is (15,21), and the residual vector is calculated as (0,0), with a residual magnitude of 0. Keeping the prediction error covariance matrix unchanged, the Kalman gain matrix is ​​calculated. Since the residual is zero, the posterior state vector is the same as the prior state vector, i.e., [(15,21),(1,0.2)]. The position coordinate component (15,21) is extracted as the first step line fitting point.

[0102] When processing the second candidate feature point, the prior state vector for predicting the next point is [(20,22),(1,0.2)]. The observation vector is (20,25), and the calculated residual vector is (0,3), with a residual magnitude of 3. Since the residual magnitude equals the preset threshold, the element values ​​of the prediction error covariance matrix are adjusted to 1.0×(3 / 3)=1.0. The Kalman gain matrix is ​​calculated, yielding the posterior state vector [(20,25),(0.95,0.5)]. The position coordinate component (20,25) is extracted as the second step line fitting point.

[0103] When processing the third candidate feature point, the prior state vector for predicting the next point is [(24.75,27.5),(0.95,0.5)]. The observation vector is (25,35), and the calculated residual vector is (0.25,7.5), with a residual magnitude of 7.5. Since the residual magnitude is greater than the preset threshold, the element values ​​of the prediction error covariance matrix are adjusted to 1.0×(7.5 / 3)=2.5, but not exceeding the maximum amplification factor of 5. The Kalman gain matrix is ​​calculated, yielding the posterior state vector [(25,35),(0.8,0.9)]. The calculated tangent direction angle is approximately 22 degrees, less than the preset angle threshold, so no curvature control point is inserted. The position coordinate component (25,35) is extracted as the fitting point for the third step line.

[0104] When processing the fourth candidate feature point, the prior state vector for predicting the next point is [(29,39.5),(0.8,0.9)]. The observation vector is (30,36), and the calculated residual vector is (1,-3.5), with a residual magnitude of approximately 3.6. Since the residual magnitude is greater than the preset threshold, the element values ​​of the prediction error covariance matrix are adjusted to 1.0×(3.6 / 3)=1.2. The Kalman gain matrix is ​​calculated, yielding the posterior state vector [(30,36),(0.9,0.1)]. The tangent direction angle is calculated to be approximately 50 degrees, which is greater than the preset angle threshold. A curvature control point is inserted at coordinates (30,36). The position coordinate component (30,36) is extracted as the fourth step line fitting point.

[0105] Finally, all the step line fitting points [(15,21), (20,25), (25,35), (30,36)] and curvature control points [(30,36)] are connected in the order of processing to form the step line fitting result. At the curvature control points, the fitted curve will have a smoother transition, better adapting to the geometric characteristics of the step line.

[0106] In one optional implementation, the downsampled point cloud data is spatially correlated with the step line fitting result. The downsampled point cloud data is then spatially divided into blocks based on the distance relationship from the points to the step line. A center point is selected from each spatial block, and a local neighborhood is constructed around each center point, including:

[0107] For each point in the downsampled point cloud data, calculate the shortest distance from the point to all step lines in the step line fitting result, and use the shortest distance as the step line distance attribute of the point;

[0108] Based on the step line distance attribute, the downsampled point cloud data is divided into multiple distance intervals, each distance interval corresponding to a distance range;

[0109] For the point cloud data within each distance interval, the point cloud data is divided into multiple spatial blocks using a spatial grid partitioning method;

[0110] For each spatial block, the average spatial coordinates of all points within the block are calculated as the block center coordinates. Using the block center coordinates as a reference point, the point closest to the block center coordinates is selected as the center point within the spatial block.

[0111] Using the center point as the center of a sphere, the neighborhood radius is determined based on the point cloud density of the spatial block, and all points whose distance from the center point is less than the neighborhood radius constitute a local neighborhood.

[0112] Spatially associating the downsampled point cloud data with the fitted result of the step line, and spatially dividing the downsampled point cloud data into blocks based on the distance relationship from points to the step line, is a crucial step in achieving local feature extraction. This process first requires calculating the shortest distance from each point in the point cloud data to the step line, then spatially dividing the data based on these distance values, and finally constructing local neighborhoods to support subsequent feature extraction.

[0113] In practical implementation, after acquiring downsampled point cloud data, the first step is to calculate the shortest distance from each point to the step line. For each point in the downsampled point cloud, the distance from that point to all fitted step lines is calculated, and the minimum value is taken as the step line distance attribute of that point. Specifically, if the 3D coordinates of a point in the downsampled point cloud are (x, y, z), and a fitted step line can be represented as a straight line segment in space with endpoint coordinates of (x1, y1, z1) and (x2, y2, z2), then the distance from that point to this line segment can be calculated using the point-to-line segment distance formula. The minimum value of the distance from that point to all step lines is taken as the step line distance attribute of that point, denoted as d_step.

[0114] Based on the calculated step line distance attributes, the downsampled point cloud data is divided into multiple distance intervals. For example, the point cloud can be divided into a near distance interval (0-0.2 meters), a medium distance interval (0.2-0.5 meters), and a far distance interval (0.5-1.0 meters). The division of distance intervals can be adjusted according to the actual application scenario and the distribution characteristics of the point cloud. In some embodiments, the distance intervals can be set at equal intervals, such as setting an interval every 0.1 meters; in other embodiments, a non-uniform division method can be adopted, with smaller distance intervals set for areas close to the step line to improve accuracy.

[0115] For point cloud data within each distance interval, the data is further divided into multiple spatial blocks using a spatial grid partitioning method. Specifically, a grid structure can be established in three-dimensional space, and the grid size can be dynamically adjusted according to the point cloud density. For example, a smaller grid size, such as 0.1 m × 0.1 m × 0.1 m, can be used in areas with high point cloud density; while a larger grid size, such as 0.2 m × 0.2 m × 0.2 m, can be used in areas with low point cloud density. For each distance interval, a suitable grid size is determined based on its spatial distribution characteristics. Then, all points within that interval are assigned to the corresponding grid according to their three-dimensional coordinates, thus forming multiple spatial blocks.

[0116] After dividing the space into blocks, a center point needs to be selected for each block. First, the average spatial coordinates of all points within each block are calculated, and this average is used as the center coordinates of that block. Let there be n points within a block, with coordinates (x_i, y_i, z_i), i=1,2,...,n. Then the block center coordinates (x_c, y_c, z_c) are calculated as follows: x_c = Σx_i / n, y_c = Σy_i / n, z_c = Σz_i / n. After obtaining the block center coordinates, the point within the block closest to these coordinates is selected as the center point. Specifically, the Euclidean distance from each point within the block to the block center coordinates is calculated, and the point with the smallest distance is selected as the center point of that block.

[0117] Using the selected center point as the sphere's center, a local neighborhood is constructed. The radius of this neighborhood needs to consider the point cloud density of the spatial block. In areas with high point cloud density, a smaller neighborhood radius can be used; in areas with low point cloud density, a larger neighborhood radius can be used to ensure that the neighborhood contains a sufficient number of points. In practice, the neighborhood radius can be calculated based on the number and spatial distribution of points within the spatial block. For example, the neighborhood radius can be set to 2-3 times the average distance between points within the spatial block, or it can be set to a fixed value empirically, such as 0.15 meters. After determining the neighborhood radius, the distance from all points within the spatial block to the center point is calculated. The set of all points whose distance is less than the neighborhood radius is defined as the local neighborhood of that center point.

[0118] In practical applications, when processing point cloud data of indoor scenes containing multiple steps, the first step is to calculate the distance from each point to the nearest step line using the steps described above. For example, 0-0.1 meters is defined as the first distance interval, 0.1-0.3 meters as the second distance interval, and so on. Then, a grid is applied to each distance interval, such as using a grid size of 0.05m × 0.05m × 0.05m in the first distance interval and 0.1m × 0.1m × 0.1m in the second distance interval. For each resulting spatial block, the mean coordinates of the points within the block are calculated, and the point closest to this mean is selected as the center point. Finally, the neighborhood radius is determined based on the point density of each spatial block, constructing a local neighborhood to provide a foundation for subsequent feature extraction and analysis.

[0119] Through the above steps, the spatial association between the downsampled point cloud data and the step line fitting results was completed, spatial segmentation based on distance relationship was realized, and a local neighborhood based on the center point was successfully constructed, laying the foundation for subsequent point cloud processing and feature extraction.

[0120] In one optional implementation, context-enhanced features are calculated based on feature vectors of neighboring points within the local neighborhood; compressed features are obtained after processing the context-enhanced features; and encoded features are generated based on the compressed features and the feature vectors, including:

[0121] For neighborhood points within a local neighborhood, a linear transformation is used to map the feature vectors of the neighborhood points into query vectors, key vectors, and value vectors, respectively.

[0122] Calculate the relevance score between the query vector and the key vector, calculate the distance decay weight based on the spatial distance between neighboring points and fuse it with the relevance score to obtain a spatially perceived relevance score, and obtain context-enhanced features based on the spatially perceived relevance score and the value vector;

[0123] The context enhancement features are grouped by channel, and the channels are divided into multiple subgroups. Each subgroup is expanded by channel expansion, and different expansion factors are used for different subgroups to obtain multi-scale expanded features.

[0124] The multi-scale extended features are selectively activated by applying a gating mechanism to obtain gated modulation features. Each subgroup of the gated modulation features is then compressed into channels, and the compressed subgroup features are concatenated to obtain compressed features.

[0125] Calculate the channel attention weights between the compressed feature and the feature vector, perform channel weighting on the compressed feature based on the channel attention weights, and perform residual fusion between the weighted compressed feature and the feature vector to obtain the encoded feature.

[0126] First, a local neighborhood containing the 3D point cloud data is obtained. Each point has a corresponding feature vector. The neighborhood points within this local neighborhood are processed by linear transformations, mapping their feature vectors to query vectors, key vectors, and value vectors. Specifically, assuming the feature vector dimension of the neighborhood points is D, three different linear transformation matrices can be used to map the feature vectors to the query space, key space, and value space, respectively. The linear transformation can be represented as the product of the feature vector and the weight matrix, yielding query vectors, key vectors, and value vectors of dimension D'.

[0127] Next, the relevance score between the query vector and the key vector is calculated. For each pair of points in the neighborhood, the dot product of their query vector and key vector is calculated to obtain the original relevance score. Simultaneously, a distance decay weight is calculated based on the spatial distance between neighboring points. A Gaussian function can be used to decay the distance, giving higher weights to points that are spatially close. The original relevance score is multiplied by the distance decay weight to obtain the spatially perceived relevance score, which comprehensively considers feature similarity and spatial relationship. Then, the spatially perceived relevance score is normalized, for example, using a softmax function, to ensure that the sum of the weights is 1. Finally, the normalized spatially perceived relevance score is weighted and summed with the value vector to obtain the context-enhanced feature.

[0128] The context-enhanced features are processed by channel grouping, dividing the channel dimension into multiple subgroups. Assuming the channel dimension of the context-enhanced features is C, it can be evenly divided into N subgroups, each containing C / N channels. Channel expansion is then performed on each subgroup, with different expansion factors for different subgroups. For example, the first subgroup is expanded by a factor of 1 (preserving the original size), the second subgroup by a factor of 2, the third subgroup by a factor of 4, and so on. Channel expansion can be achieved through transposed convolution or interpolation, resulting in multi-scale expanded features.

[0129] A gating mechanism is applied to the multi-scale extended features for selective activation, enhancing useful information and suppressing redundant information. Specifically, gating weights can be generated using the sigmoid function, and each channel of the multi-scale extended features is weighted to obtain the gated modulation features. Each subgroup of the gated modulation features is then compressed, using methods such as average pooling or max pooling to reduce the number of channels, compressing each subgroup to 1 / M of the original number of channels (M being a preset compression ratio). The compressed subgroup features are then concatenated to obtain the compressed features.

[0130] Finally, channel attention weights are calculated between the compressed features and the original feature vectors. Channel statistics are extracted using global average pooling, and channel weights are generated through a two-layer fully connected network to capture the dependencies between channels. The compressed features are then weighted according to these channel attention weights to enhance the contribution of important channels. The weighted compressed features are then residually fused with the original feature vectors, for example, through addition or concatenation operations, to obtain the encoded features. Residual connections help preserve original feature information while incorporating contextual enhancement information, thus improving feature representation capabilities.

[0131] In point cloud segmentation applications, the encoded features extracted using the above methods can better represent local geometric structures and semantic information, effectively improving segmentation accuracy. In point cloud recognition tasks, this method considers the impact of spatial distance on feature aggregation, enhancing the spatial discriminative ability of features and thus improving recognition accuracy. In autonomous driving scenarios, this method can better distinguish between different categories of point clouds, such as roads, vehicles, and pedestrians, improving the reliability and safety of the perception system.

[0132] In one optional implementation, a relevance score is calculated between the query vector and the key vector; a distance decay weight is calculated based on the spatial distance between neighboring points and fused with the relevance score to obtain a spatially perceived relevance score; and context-enhanced features are obtained based on the spatially perceived relevance score and the value vector, including:

[0133] Calculate the inner product of the query vector and the key vector, and scale the inner product according to the feature dimension of the query vector to obtain the basic relevance score;

[0134] Calculate the Euclidean distance between the center point and each neighboring point within the local neighborhood, and calculate the distance decay weight based on the Euclidean distance, wherein the Euclidean distance is inversely proportional to the distance decay weight;

[0135] The spatial modulation correlation score is determined based on the basic correlation score and the distance attenuation weight, and the spatial modulation correlation score is then biased to obtain the spatial perception correlation score.

[0136] The temperature parameter is calculated based on the numerical distribution characteristics of the spatial perception correlation score, and the spatial perception correlation score is scaled using the temperature parameter to obtain the adjusted spatial perception correlation score.

[0137] The adjusted spatial perception relevance score is mapped to a positive distribution using an exponential function. The mapping result is normalized in the neighborhood point dimension to obtain normalized relevance weights. The normalized relevance weights are then weighted and aggregated with the value vector. Position encoding compensation is applied to the aggregation result to obtain context-enhanced features.

[0138] In spatially perceptual feature enhancement methods, a spatial distance factor is introduced when calculating the correlation between the query vector and the key vector in order to enhance the spatial perception of features.

[0139] First, feature data containing the center point and its local neighboring points is obtained, where the center point feature vector serves as the query vector, and the neighboring point feature vectors serve as the key and value vectors, respectively. This feature data can come from images, point clouds, or other data sources containing spatial information.

[0140] The basic relevance score between the query vector and the key vector is calculated. Specifically, for the query vector q and the key vector k, their similarity in the feature space is measured by calculating their inner product q×k. Considering the influence of vector dimension on the inner product value, scaling is required, typically by dividing by the square root of the query vector's feature dimension d, i.e., basic relevance score s = q×k / sqrt(d). This scaling stabilizes the gradient and avoids the gradient vanishing problem caused by excessively large inner product values ​​in high-dimensional feature spaces.

[0141] Next, we calculate the Euclidean distance between the center point and each neighboring point in the local neighborhood. Assuming the center point's coordinates are p_0 and a neighboring point's coordinates are p_i, the Euclidean distance between them is... Based on this distance, a distance decay weight w is calculated, such that the weight of more distant neighboring points is smaller. This can be achieved using a Gaussian decay function w = exp(-d_i² / σ²), where σ is a parameter controlling the decay rate.

[0142] The spatial modulation correlation score is obtained by fusing the baseline correlation score with the distance attenuation weight. The fusion method can be multiplicative modulation: s' = s × w, or weighted combination: s' = α × s + (1-α) × w, where α is a coefficient balancing the importance of the two. A bias correction is applied to the spatial modulation correlation score by introducing a learnable bias term b, resulting in the spatial perception correlation score s'' = s' + b. This bias correction adjusts the correlation baseline and enhances the model's expressive power.

[0143] The temperature parameter τ is calculated based on the numerical distribution characteristics of the spatial perception relevance scores. τ can be determined based on the statistical characteristics of the scores in the current batch of data, for example, by calculating the standard deviation στ of the scores and setting τ = β × στ, where β is an adjustable coefficient. The spatial perception relevance scores are then scaled using the temperature parameter: s''' = s'' / τ. This adjustment controls the degree of focus of subsequent attention weights.

[0144] The adjusted spatial perception relevance score is mapped to a positive distribution using an exponential function: a = exp(s'''). The mapping result is then normalized along the neighborhood point dimension to obtain the normalized relevance weight: α = a / Σa. This step ensures that the sum of all attention weights is 1, meeting the requirements of a weighted average.

[0145] Finally, the normalized relevance weights are weighted and aggregated with the value vector: o = Σ(α × v), where v represents the value vector. To preserve location information, location encoding compensation is applied to the aggregation result: o' = o + PE, where PE is the location code. This compensation mechanism enhances the model's ability to perceive spatial location, especially when dealing with data with complex spatial structures.

[0146] One application of this method is in point cloud scene understanding, where calculating the spatial perception correlation between the center point and surrounding points can better capture local geometric features. Compared to traditional methods, introducing spatial distance attenuation weights allows the model to adaptively adjust attention allocation based on the physical distance between points, thereby generating more spatially perceptive feature representations.

[0147] Through the above steps, feature similarity and spatial distance information are effectively integrated to generate enhanced features with context awareness, which can be applied to various scenarios that require simultaneous understanding of feature semantics and spatial structure.

[0148] In one optional implementation, the encoded features are diffused to unsampled points through interpolation and fused with hierarchical features to generate point-level semantic labels. Target features are then extracted based on the fitting result between the point-level semantic labels and the step line, including:

[0149] For unsampled points within a spatial block, search for the nearest sampled center points. Based on the spatial distance between the unsampled point and each sampled center point, perform distance-weighted interpolation on the encoding features of each sampled center point to obtain the interpolated encoding features of the unsampled point.

[0150] The interpolation coding features are fused with the hierarchical features of the corresponding spatial location across hierarchical levels to obtain fused features. Based on the fused features, semantic prediction is performed on the unsampled points to determine the point-level semantic labels of the unsampled points.

[0151] Spatial connectivity analysis is performed on the generated point-level semantic labels, and points with the same point-level semantic labels and spatial connectivity are clustered to obtain candidate land cover regions;

[0152] For each candidate feature area, the spatial envelope range of the candidate feature area is calculated, and its relative positional relationship with the fitting result of the step line is used to determine the spatial functional zone to which the candidate feature area belongs;

[0153] Candidate land cover areas that meet the spatial functional zoning constraints and whose number of points meets the regional scale requirements are marked as valid target land cover areas, and the classification of the valid target land cover areas is determined based on point-level semantic labels.

[0154] After completing the semantic encoding of spatial blocks and the fitting of step lines, the encoded features need to be diffused to unsampled points through interpolation and fused with hierarchical features to generate point-level semantic labels, ultimately extracting the target features. This process includes four main steps: encoded feature interpolation and diffusion, feature fusion and semantic prediction, candidate feature region clustering, and target feature extraction.

[0155] First, interpolation diffusion of encoded features is performed on the unsampled points within the spatial block. Due to the sparse sampling strategy, only some sampled center points have encoded features, while a large number of unsampled points lack feature representation. For each unsampled point, the K nearest sampled center points in its spatial neighborhood are searched, with K typically set between 8 and 12. The interpolated encoded features of the unsampled points are calculated using a distance-based weighted average method. Specifically, let the position of the unsampled point p be (x_p, y_p, z_p), and its k-th nearest sampled center point be c_k, with position (x_k, y_k, z_k). Then, the spatial Euclidean distance d_k = sqrt((x_p-x_k)^2 + (y_p-y_k)^2 + (z_p-z_k)^2) is calculated. The weight w_k = 1 / d_k^2 is defined according to the distance and normalized to obtain the final weight w_k' = w_k / ∑w_i. The interpolated coding feature of the unsampled point p is F_p = ∑(w_k' × F_k), where F_k is the coding feature of the k-th nearest sampled center point. Using this interpolation method, all unsampled points obtain coding feature representations consistent with the feature space of the sampled center points.

[0156] Secondly, the interpolated encoded features and hierarchical features are fused to generate point-level semantic labels. In the preceding steps, hierarchical features of different scales have been obtained through a multi-layer feature extraction network. For each unsampled point, cross-hierarchical feature fusion is performed using a combination of feature concatenation and an attention mechanism, based on the hierarchical features corresponding to its spatial location index. Let the interpolated encoded feature of the unsampled point p be F_p, and the corresponding hierarchical feature be H_p. Then, the fused feature G_p = Attention(Concat(F_p, H_p)), where Attention represents the attention mechanism used to adjust the weights of features from different sources. A multi-layer perceptron (MLP) is used to perform semantic prediction on the fused feature G_p, outputting a classification probability vector. The category with the highest probability is taken as the semantic label for that point. Semantic categories include common land cover types such as buildings, vegetation, ground, and walls. In this way, the semantic label propagation from sparse sampling to dense point clouds is completed, achieving fine-grained point-level semantic segmentation.

[0157] Then, spatial connectivity analysis and candidate feature region clustering are performed based on point-level semantic labels. A region growing algorithm is used to cluster points with the same semantic labels. A spatial search radius *r* is set, and for each point, its neighboring points within *r* are searched. If a neighboring point has the same semantic label as the current point, it is assigned to the same cluster. The entire point cloud is traversed using either breadth-first search (BFS) or depth-first search (DFS) algorithms until all points have been visited and assigned to their respective clusters. Each cluster constitutes a candidate feature region and is assigned a unique region identifier. To improve clustering efficiency, spatial indexing structures such as octrees or KD-trees can be used to accelerate the neighborhood search process.

[0158] Finally, combining the step line fitting results and spatial functional zoning constraints, effective target features are extracted. For each candidate feature region, its three-dimensional spatial envelope (i.e., minimum circumscribed cube) is calculated, and the relative positional relationship between this envelope and the fitted step line is analyzed. If the envelope of a candidate feature region spatially intersects with a certain step line or the distance is less than a preset threshold, the functional zoning to which the candidate feature belongs is determined according to the spatial functional zoning criteria defined by the step line, such as a building area, green area, or road area. Further zoning-specific filtering rules are set, including the minimum number of points in the region (N_min), the maximum number of points in the region (N_max), and region shape indicators. Only candidate feature regions that meet the functional zoning constraints and whose number of points satisfies N_min ≤ N ≤ N_max are retained and marked as effective target features. The final classification of each effective target feature is determined based on the category with the highest frequency in the point-level semantic labels.

[0159] The above process completes the entire workflow from point cloud semantic segmentation to target feature extraction, achieving accurate identification and classification of target features such as buildings and vegetation in the scene. The extracted target features not only possess precise 3D geometric shapes but also contain semantic information and functional attributes, providing a foundation for subsequent scene understanding and 3D reconstruction. This method is particularly suitable for feature extraction tasks in complex urban environments and can effectively handle challenging scenarios such as undulating terrain and dense buildings.

[0160] The airborne laser point cloud open-pit mine feature extraction system of this invention includes:

[0161] The first unit is used to acquire airborne laser point cloud data containing open-pit mine areas and surrounding features, and to perform voxel downsampling on the airborne laser point cloud data to obtain downsampled point cloud data.

[0162] The second unit is used to identify candidate feature points of the step line in the local neighborhood of each target point in the downsampled point cloud data based on the joint criterion of plane fitting residual and normal vector angle dispersion, and to perform curve fitting on the candidate feature points of the step line to recursively generate the step line fitting result;

[0163] The third unit is used to spatially correlate the downsampled point cloud data with the step line fitting result, divide the downsampled point cloud data into spatial blocks according to the distance relationship from the point to the step line, select a center point for each spatial block, and construct a local neighborhood around each center point.

[0164] The fourth unit is used to calculate context-enhanced features based on the feature vectors of neighboring points within the local neighborhood, process the context-enhanced features to obtain compressed features, and generate encoded features based on the compressed features and the feature vectors.

[0165] The fifth unit is used to diffuse the encoded features to unsampled points through interpolation and fuse them with hierarchical features to generate point-level semantic labels, and extract target ground features based on the fitting results of the point-level semantic labels and the step line.

[0166] A third aspect of the present invention provides an electronic device, comprising:

[0167] processor;

[0168] Memory used to store processor-executable instructions;

[0169] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0170] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0171] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0172] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for extracting feature features from open-pit mines using airborne laser point clouds, characterized in that, include: Airborne laser point cloud data containing open-pit mine areas and surrounding features is acquired, and the airborne laser point cloud data is downsampled using voxelization to obtain downsampled point cloud data; In the downsampled point cloud data, candidate feature points for step lines are identified in the local neighborhood of each target point based on a joint criterion of plane fitting residual and normal vector angle dispersion. Curve fitting is then performed on the candidate feature points for step lines to recursively generate step line fitting results. The downsampled point cloud data is spatially correlated with the step line fitting result. The downsampled point cloud data is spatially divided into blocks according to the distance relationship between the points and the step line. A center point is selected from each spatial block, and a local neighborhood is constructed around each center point. Within the local neighborhood, context-enhanced features are calculated based on the feature vectors of neighboring points. These context-enhanced features are then processed to obtain compressed features. Finally, encoded features are generated based on the compressed features and the feature vectors. The encoded features are diffused to unsampled points through interpolation and fused with hierarchical features to generate point-level semantic labels. Target ground features are extracted based on the fitting results of the point-level semantic labels and the step line.

2. The method according to claim 1, characterized in that, In the downsampled point cloud data, candidate feature points for step lines are identified within the local neighborhood of each target point based on a joint criterion of plane fitting residual and normal vector angle dispersion. Curve fitting is then performed on these candidate feature points to recursively generate step line fitting results, including: A target plane is obtained by fitting within the local neighborhood of each target point, and the perpendicular distance from each point in the neighborhood to the target plane is calculated as the fitting residual. For each point in the neighborhood, the normal vector is calculated by eigenvalue decomposition of the neighborhood covariance matrix. The normal vector is then projected onto the tangent plane of the target plane, and the angle between the projected vectors is calculated to determine the angular dispersion of the normal vector. Points that simultaneously satisfy the condition that the fitting residual is greater than the first threshold and the normal vector angle dispersion is greater than the second threshold are marked as candidate feature points for the step line; Using the candidate feature points of the step line as seed points, region growing is performed. Neighboring candidate feature points whose spatial distance is less than a distance threshold and whose normal vector angle is less than an angle threshold are merged into the same region to obtain multiple candidate feature point sequences. For each candidate feature point sequence, the candidate feature points in the sequence are processed sequentially. Based on the alternating iteration of state prediction and observation update, step line fitting points are generated. All step line fitting points are connected to form the step line fitting result.

3. The method according to claim 2, characterized in that, For each candidate feature point sequence, the candidate feature points in the sequence are processed sequentially. Based on the alternating iteration of state prediction and observation update, step line fitting points are generated. All step line fitting points are connected to form the step line fitting result, including: Predict the prior state vector of the next candidate feature point based on the state vector of the current candidate feature point in the candidate feature point sequence; Using the actual position coordinates of the next candidate feature point in the candidate feature point sequence as the observation vector, calculate the residual vector between the observation vector and the position coordinate components in the prior state vector; Calculate the magnitude of the residual vector, adjust the prediction error covariance matrix based on the magnitude, and calculate the Kalman gain matrix based on the adjusted prediction error covariance matrix and the observation noise covariance matrix. The prior state vector is corrected based on the Kalman gain matrix and the residual vector to obtain the posterior state vector. Based on the angle between the tangent direction component in the posterior state vector and the tangent direction component in the adjacent posterior state vector, curvature control points are inserted at the corresponding positions of the posterior state vector. The position coordinate components in the posterior state vector are extracted as step line fitting points. The posterior state vector is used as the current state vector for the next iteration. All candidate feature points are processed sequentially, and all step line fitting points are connected with the curvature control points in the processing order to form the step line fitting result.

4. The method according to claim 1, characterized in that, The downsampled point cloud data is spatially correlated with the step line fitting result. Based on the distance relationship from the points to the step line, the downsampled point cloud data is spatially divided into blocks. A center point is selected from each spatial block, and a local neighborhood is constructed around each center point, including: For each point in the downsampled point cloud data, calculate the shortest distance from the point to all step lines in the step line fitting result, and use the shortest distance as the step line distance attribute of the point; Based on the step line distance attribute, the downsampled point cloud data is divided into multiple distance intervals, each distance interval corresponding to a distance range; For point cloud data within each distance interval, the point cloud data is divided into multiple spatial blocks using a spatial grid partitioning method; For each spatial block, the average spatial coordinates of all points within the block are calculated as the block center coordinates. Using the block center coordinates as a reference point, the point closest to the block center coordinates is selected as the center point within the spatial block. Using the center point as the center of a sphere, the neighborhood radius is determined based on the point cloud density of the spatial block, and all points whose distance from the center point is less than the neighborhood radius constitute a local neighborhood.

5. The method according to claim 1, characterized in that, Within the local neighborhood, context-enhanced features are calculated based on the feature vectors of neighboring points. These context-enhanced features are then processed to obtain compressed features. Encoded features are generated based on the compressed features and the feature vectors, including: For neighborhood points within a local neighborhood, a linear transformation is used to map the feature vectors of the neighborhood points into query vectors, key vectors, and value vectors, respectively. Calculate the relevance score between the query vector and the key vector, calculate the distance decay weight based on the spatial distance between neighboring points and fuse it with the relevance score to obtain a spatially perceived relevance score, and obtain context-enhanced features based on the spatially perceived relevance score and the value vector; The context enhancement features are grouped by channel, and the channels are divided into multiple subgroups. Each subgroup is expanded by channel expansion, and different expansion factors are used for different subgroups to obtain multi-scale expanded features. The multi-scale extended features are selectively activated by applying a gating mechanism to obtain gated modulation features. Each subgroup of the gated modulation features is then compressed into channels, and the compressed subgroup features are concatenated to obtain compressed features. Calculate the channel attention weights between the compressed feature and the feature vector, perform channel weighting on the compressed feature based on the channel attention weights, and perform residual fusion between the weighted compressed feature and the feature vector to obtain the encoded feature.

6. The method according to claim 5, characterized in that, Calculate the relevance score between the query vector and the key vector, calculate the distance decay weight based on the spatial distance between neighboring points and fuse it with the relevance score to obtain a spatially perceived relevance score, and obtain context-enhanced features based on the spatially perceived relevance score and the value vector, including: Calculate the inner product of the query vector and the key vector, and scale the inner product according to the feature dimension of the query vector to obtain the basic relevance score; Calculate the Euclidean distance between the center point and each neighboring point within the local neighborhood, and calculate the distance decay weight based on the Euclidean distance, wherein the Euclidean distance is inversely proportional to the distance decay weight; The spatial modulation correlation score is determined based on the basic correlation score and the distance attenuation weight, and the spatial modulation correlation score is then biased to obtain the spatial perception correlation score. The temperature parameter is calculated based on the numerical distribution characteristics of the spatial perception correlation score, and the spatial perception correlation score is scaled using the temperature parameter to obtain the adjusted spatial perception correlation score. The adjusted spatial perception relevance score is mapped to a positive distribution using an exponential function. The mapping result is normalized in the neighborhood point dimension to obtain normalized relevance weights. The normalized relevance weights are then weighted and aggregated with the value vector. Position encoding compensation is applied to the aggregation result to obtain context-enhanced features.

7. The method according to claim 1, characterized in that, The encoded features are diffused to unsampled points through interpolation and fused with hierarchical features to generate point-level semantic labels. Target features are extracted based on the fitting result between the point-level semantic labels and the step line, including: For unsampled points within a spatial block, search for the nearest sampled center points. Based on the spatial distance between the unsampled point and each sampled center point, perform distance-weighted interpolation on the encoding features of each sampled center point to obtain the interpolated encoding features of the unsampled point. The interpolation coding features are fused with the hierarchical features of the corresponding spatial location across hierarchical levels to obtain fused features. Based on the fused features, semantic prediction is performed on the unsampled points to determine the point-level semantic labels of the unsampled points. Spatial connectivity analysis is performed on the generated point-level semantic labels, and points with the same point-level semantic labels and spatial connectivity are clustered to obtain candidate land cover regions; For each candidate feature area, the spatial envelope range of the candidate feature area is calculated, and its relative positional relationship with the fitting result of the step line is used to determine the spatial functional zone to which the candidate feature area belongs; Candidate land cover areas that meet the spatial functional zoning constraints and whose number of points meets the regional scale requirements are marked as valid target land cover areas, and the classification of the valid target land cover areas is determined based on point-level semantic labels.

8. An airborne laser point cloud open-pit mine feature extraction system, used to implement the method as described in any one of claims 1-7, characterized in that, include: The first unit is used to acquire airborne laser point cloud data containing open-pit mine areas and surrounding features, and to perform voxel downsampling on the airborne laser point cloud data to obtain downsampled point cloud data. The second unit is used to identify candidate feature points of the step line in the local neighborhood of each target point in the downsampled point cloud data based on the joint criterion of plane fitting residual and normal vector angle dispersion, and to perform curve fitting on the candidate feature points of the step line to recursively generate the step line fitting result; The third unit is used to spatially correlate the downsampled point cloud data with the step line fitting result, divide the downsampled point cloud data into spatial blocks according to the distance relationship between the points and the step line, select a center point for each spatial block, and construct a local neighborhood around each center point; The fourth unit is used to calculate context-enhanced features based on the feature vectors of neighboring points within the local neighborhood, process the context-enhanced features to obtain compressed features, and generate encoded features based on the compressed features and the feature vectors. The fifth unit is used to diffuse the encoded features to unsampled points through interpolation and fuse them with the hierarchical features to generate point-level semantic labels, and extract target ground features based on the fitting result of the point-level semantic labels and the step line.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images