A three-dimensional point cloud terrain segmentation method for unstructured planetary surfaces

Abstract

The application discloses a three-dimensional point cloud terrain segmentation method for unstructured planetary surface, and belongs to the technical field of deep space exploration and mobile robot autonomous navigation. The method comprises the following steps: performing abnormal point and negative obstacle sensitive preprocessing on the obtained original point cloud; performing grid division on the effective point cloud, decoupling the slope and roughness features of the terrain based on principal component analysis; processing sparse data through adaptive density checking, and performing initial classification on the grid according to geometric features; performing double-mode elevation updating, performing time sequence probability fusion on the candidate ground area, and introducing a forced downward correction strategy for the candidate obstacle area to correct under-segmentation; performing space-time joint interpolation on the hollow area and the sparse area to generate a complete terrain, and outputting a semantic segmentation result in combination with multi-dimensional features. The application is especially suitable for the environment perception of a planetary surface exploration vehicle on the moon, Mars and other planets, and effectively improves the precision of environment perception and the safety of navigation under rugged terrain.

Description

Technical Field

[0001] This invention belongs to the field of deep space exploration and autonomous navigation technology for mobile robots. Specifically, it relates to a three-dimensional point cloud terrain segmentation method for unstructured planetary surfaces. It is applicable to environmental perception and segmentation of three-dimensional lidar point clouds of planetary rovers, and can also be extended to the field of autonomous exploration and navigation in other complex lunar environments. Background Technology

[0002] With the rapid development of deep space exploration technology, planetary rovers, represented by lunar and Martian rovers, have become indispensable platforms for exploring the surfaces of extraterrestrial bodies. When rovers perform missions on extremely complex and unstructured planetary surfaces (such as the lunar surface, which is full of craters, rocks, and steep slopes), their autonomous navigation and path planning capabilities highly depend on accurate perception of the surrounding environment. Compared with visual cameras that rely on illumination conditions, 3D lidar is considered a highly promising core sensor for planetary surface exploration due to its advantages such as high ranging accuracy, long detection range, and insensitivity to ambient light.

[0003] In existing technologies, constructing a rasterized elevation map based on 3D LiDAR point clouds and performing terrain segmentation is the mainstream method for environmental perception of ground mobile robots. Its general process includes: point cloud preprocessing (such as noise reduction and filtering), 2D raster division, and line-of-sight detection and obstacle avoidance. This type of method performs well in structured ground environments.

[0004] However, when applied directly to extreme, unstructured environments such as planetary surfaces, existing technologies reveal significant limitations. First, they suffer from serious deficiencies in negative obstacle detection. The numerous impact craters on planetary surfaces are typical and high-risk negative obstacles (terrain depressions). To filter out "subsurface" noise points generated by sensor multipath reflections and other factors, existing algorithms typically employ simple line-of-sight rejection or height threshold filtering, treating points located below historical ground models as noise and discarding them. This leads to the erroneous discarding of true point cloud data from the bottom of craters, creating data holes in elevation maps that prevent rovers from detecting the presence of deep craters, posing a serious safety hazard.

[0005] Secondly, the ability to distinguish sloping terrain is insufficient. Existing methods rely heavily on the height variance of point clouds within a grid to assess terrain ruggedness. However, on planetary surfaces, there are often slopes with steep inclines but relatively flat surfaces. The point clouds of such terrain also have significant vertical differences, resulting in an inflated height variance value, which is then misclassified by existing algorithms as impassable, rough obstacle areas. This misclassification significantly reduces the traversable area of ​​the rover, leading to frequent path planning failures or unnecessary detours.

[0006] Furthermore, it exhibits poor robustness to long-distance sparse point clouds and dynamic occlusion environments. The density of lidar point clouds decreases sharply with increasing detection distance, and geometric features directly based on sparse point cloud computing are prone to distortion. Simultaneously, during the movement of the rover, dynamic occlusion by obstacles may cause the map area to remain unupdated for extended periods, resulting in "undersegmentation," where the terrain height in certain areas of the map is overestimated, failing to reflect the true terrain after obstacles are removed or wheels have dented the ground.

[0007] In summary, there is an urgent need for a three-dimensional point cloud terrain segmentation method specifically designed for the extremely unstructured environment of planetary surfaces. This method would address issues such as missed detection of negative obstacles, misjudgment of slopes, and improve the robustness of processing sparse and dynamic data, thereby ensuring the autonomous navigation safety and mission execution efficiency of planetary rovers. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a three-dimensional point cloud terrain segmentation method for unstructured planetary surfaces, a gradient consistency preservation method for negative obstacles on planetary surfaces, and a slope-roughness joint criterion based on PCA decoupling. Through dual-mode elevation updates and spatiotemporal interpolation, it achieves crater detection and accurate slope identification.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] A method for 3D point cloud terrain segmentation for unstructured planetary surfaces includes:

[0011] Step 1: Obtain the original 3D lidar point cloud data of the planet's surface;

[0012] Step 2: Perform preprocessing on the original 3D LiDAR point cloud data to detect outliers and negative obstacles, thereby obtaining effective point cloud data;

[0013] Step 3: Divide the obtained effective point cloud data into grids, and calculate the slope and roughness of the point cloud in each grid based on the principal component analysis method.

[0014] Step 4: Perform an adaptive point density check on each grid cell, and mark the grid cells as sparse holes or non-sparse grid cells based on the check results. Then, perform an initial classification of non-sparse grid cells based on the slope and roughness.

[0015] Step 5: Calculate the current observed elevation and confidence level of each grid cell, and execute the dual-mode elevation update strategy based on the initial classification results to obtain the updated elevation map and confidence level.

[0016] Step 6: Perform spatiotemporal joint interpolation on the unupdated sparse and void regions to generate a complete elevation map and use it as a reference for the next frame, while synchronously updating the corresponding confidence level.

[0017] Step 7: Based on the complete elevation map and the slope and roughness, perform multi-dimensional feature discrimination on the original 3D lidar point cloud data, and finally output a multi-dimensional semantic segmentation result containing flat ground, slope, obstacles and negative obstacles.

[0018] Furthermore, step 2 specifically includes:

[0019] The 3D coordinates of each point in the current point cloud are compared with the elevation of the corresponding location in the previous frame of the historical elevation map. If a point is above the ground represented by the previous frame of the historical elevation map, it is retained. If a point is below the ground represented by the previous frame of the historical elevation map, gradient consistency detection is triggered.

[0020] The gradient consistency detection is as follows: calculate the vertical gradient of the line connecting the point to several nearest neighbor points in its three-dimensional spatial neighborhood; if the statistically obtained gradient distribution conforms to the preset continuous descent physical law, then the point is marked as a feature point representing a negative obstacle and retained; if the point exhibits an isolated abrupt change and has no neighborhood support conforming to the physical law, then it is determined to be a discrete noise point and removed.

[0021] Furthermore, step 3 specifically includes:

[0022] Calculate the covariance matrix for the three-dimensional coordinates of all points within each grid, and perform eigenvalue decomposition on the covariance matrix to obtain three eigenvalues ​​sorted by size;

[0023] The roughness is characterized by calculating the proportion of the smallest eigenvalue in the sum of the three eigenvalues; the smaller the proportion, the smoother the surface.

[0024] The slope is characterized by calculating the angle between the eigenvector corresponding to the minimum eigenvalue and the preset gravity direction vector.

[0025] Furthermore, step 4 specifically includes:

[0026] Based on the angular resolution parameters of the 3D LiDAR and the distance from the center of each grid to the sensor, calculate the theoretical expected number of points for that grid.

[0027] If the actual number of point clouds within a grid is less than the product of the theoretically expected number of points and a preset ratio threshold, then the grid is determined to be a sparse and hollow grid, and the initial classification step is skipped.

[0028] If the actual number of point clouds is greater than or equal to the product, then based on the slope and roughness values ​​of the grid calculated in step 3, it is initially classified as a candidate ground region or a candidate obstacle region. The candidate ground region includes flat ground and continuous terrain regions with different slopes.

[0029] Furthermore, the dual-mode elevation update strategy in step 5 specifically includes:

[0030] The current observation elevation of each grid cell is calculated based on the minimum height weighted by the number of points, and the current observation confidence is calculated based on the number of point clouds in that grid cell.

[0031] For grids initially classified as candidate ground, a temporal probability fusion mode is adopted: the current observation confidence and the existing historical confidence of the grid are used as weights to perform a weighted average of the current observation elevation and the historical elevation estimates to obtain a new elevation estimate, and the historical confidence is cumulatively updated based on the current observation confidence.

[0032] For grids initially classified as candidate obstacle regions, a forced downward correction check mode is adopted: the current observed elevation is compared with the historical elevation estimate of the grid; if the current observed elevation is less than the historical elevation estimate, the weighted fusion logic is ignored, the elevation of the grid is directly corrected to the current observed elevation, and its confidence level is increased within the preset upper limit; otherwise, the historical elevation estimate and confidence level remain unchanged.

[0033] Furthermore, step 6 specifically includes:

[0034] A spiral search strategy expanding outward from the center is adopted to locate all sparse, hollow graticles that have not been updated in step 5, as well as graticles that are themselves hollow.

[0035] For each sparse, hollow raster to be interpolated, search for the raster in its surrounding neighborhood that has already been updated in elevation, and calculate the weighted average of the elevations of these neighboring raster lattices.

[0036] Using the existing confidence level of the sparse hole grid as the fusion weight, the weighted average elevation of the neighborhood and the historical elevation estimate of the sparse hole grid are weighted and summed to obtain the interpolated elevation value, thereby filling the sparse holes.

[0037] Furthermore, the specific calculation method for the weighted average in the aforementioned temporal probability fusion mode is as follows:

[0038] The new elevation estimate equals the product of the current observed elevation and the current observation confidence level, plus the product of the historical elevation estimate and the historical confidence level, and then divided by the sum of the current observation confidence level and the historical confidence level.

[0039] Furthermore, step 7 specifically includes:

[0040] Based on the generated complete elevation map, the relative elevation difference between the elevation of each point in the original point cloud and the estimated ground elevation of its corresponding location in the complete elevation map is calculated; a joint determination is then made by combining the roughness and slope of the raster to which each point belongs.

[0041] If the roughness of the grid is lower than the first preset threshold and the slope is less than the second preset threshold, then the point is determined to be a flat surface.

[0042] If the roughness of the grid is lower than the first preset threshold and the slope is between the second preset threshold and the third preset threshold, then the point is determined to be a passable slope.

[0043] If the roughness of the grid is higher than the first preset threshold, or the relative height difference is greater than the fourth preset threshold, then the point is determined to be a positive obstacle.

[0044] For the negative obstacle feature points marked and retained in step 2, if their relative height difference is negative and the absolute value exceeds the fifth preset threshold, then the point is determined to be a negative obstacle.

[0045] In a second aspect, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method for three-dimensional point cloud terrain segmentation for unstructured planetary surfaces.

[0046] Thirdly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned method for three-dimensional point cloud terrain segmentation for unstructured planetary surfaces.

[0047] The beneficial effects of this invention are as follows:

[0048] This invention effectively solves the problem of missing detection of negative obstacles such as meteorite craters, significantly improving detection safety. It abandons the traditional approach of simply removing subsurface points and innovatively proposes a preprocessing method based on depth comparison and gradient consistency detection of the previous frame's historical elevation map. This method can intelligently separate noise from continuously descending negative obstacle features, thus effectively filtering out noise while fully preserving the crucial safety features of the crater walls and floor point clouds. This ensures that the generated elevation map accurately reflects the terrain depressions, fundamentally avoiding the risk of the rover falling into deep craters.

[0049] This invention achieves precise decoupling of terrain geometric features, significantly improving the accuracy of slope terrain identification. By introducing a joint criterion of slope and roughness based on principal component analysis, the invention decouples the two features of terrain inclination and surface ruggedness. This enables the algorithm to accurately identify traversable slopes that are "slope-heavy but have a smooth surface," avoiding misclassification as impassable obstacles. This maximizes the effective traversable area of ​​the rovers and provides better terrain information for global path planning.

[0050] This method enhances adaptability to sparse point clouds and dynamic environments, as well as map accuracy. Through an adaptive density checking mechanism, it dynamically identifies and processes sparse grids caused by distance, avoiding feature calculation distortion. Combined with a dual-mode elevation update strategy (temporal probabilistic fusion and forced downward correction), it not only smoothly fuses multi-frame observation information to improve estimation stability but also forces map elevation correction when terrain lower than historical records is detected (e.g., after obstacles are removed). This effectively solves the "undersegmentation" problem caused by dynamic occlusion, ensuring the timeliness and accuracy of the terrain model. Attached Figure Description

[0051] Figure 1 This is a schematic diagram illustrating the principle of a three-dimensional point cloud terrain segmentation method for unstructured planetary surfaces according to the present invention. Detailed Implementation

[0052] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0053] like Figure 1 As shown, the present invention provides a method for three-dimensional point cloud terrain segmentation for unstructured planetary surfaces, which specifically includes the following steps:

[0054] Step 1: Obtain the raw 3D lidar point cloud data Points_row of the planet's surface.

[0055] Step 2: Perform outlier and negative obstacle sensitivity preprocessing on the original 3D LiDAR point cloud data; perform depth comparison using the previous frame's historical elevation map, and perform gradient consistency detection on points located below the previous frame's historical ground to remove discrete noise and retain negative obstacle feature points representing craters, obtaining valid point cloud data Points_valid. Specifically, the negative obstacle sensitivity preprocessing operation is as follows:

[0056] The core of this approach lies in addressing the flaw in traditional line-of-sight detection methods, which often inadvertently delete data from the bottom of craters. In unstructured environments on planetary surfaces, negative obstacles (such as craters) are high-risk areas. LiDAR data often contains false subsurface points generated by multipath reflections, and traditional methods typically simply discard points located below the surface as noise. This method introduces a complete elevation map generated in the previous frame as a historical reference (Map_history). First, the height of the current point is compared with the height of the corresponding location in the previous frame's historical reference (Map_history) using depth comparison. If the current point is above the historical surface, it is retained directly; if it is below, gradient consistency detection is triggered. The nearest neighbor set of the current point in its spatial neighborhood is searched, and the vertical gradient (Gradient_vertical) of the line connecting the current point and its neighbors is calculated. If the statistically displayed gradient distribution satisfies the physical slope constraint of a continuous descent, it indicates that the current point belongs to a continuously descending crater wall or bottom, and it is marked as a negative obstacle feature point and retained; if the current point exhibits an isolated abrupt change relative to its surrounding points and lacks neighborhood support, it is determined to be discrete noise and discarded. The final output valid point cloud data Points_valid removes noise while fully preserving the pit bottom data that is crucial for safety.

[0057] Step 3: Divide the acquired valid point cloud data (Points_valid) into raster cells, calculate the point cloud covariance matrix within each cell, and perform principal component analysis to obtain the slope (Slope_alpha) and roughness (Roughness_rho) representing the terrain's geometric features. The specific operations for principal component analysis and geometric feature extraction are as follows:

[0058] Due to the complex terrain of planetary surfaces, a single height variance cannot effectively distinguish between steep slopes and vertical obstacles (both have significant height differences). Principal component analysis (PCA) can be used to decouple terrain features. First, the effective point cloud is projected onto a two-dimensional grid, and the covariance matrix of the point cloud coordinates within each grid is calculated. Eigenvalue decomposition is then performed on this matrix, yielding three eigenvalues: Eigen_val_1, Eigen_val_2, and Eigen_val_3 (arranged in descending order). The core of roughness (Roughness_rho) calculation lies in quantifying the surface's dispersion, specifically represented by the proportion of the smallest eigenvalue (Eigen_val_3) in the sum of all eigenvalues. A smaller value indicates a surface closer to a smooth plane, and this index replaces the traditional Z-axis height variance. The core of slope (Slope_alpha) calculation lies in analyzing the ground's tilt direction, specifically represented by the angle between the eigenvector corresponding to the smallest eigenvalue (i.e., the normal vector of the local plane) and the gravity direction vector (Vector_gravity). Using these two indicators, the algorithm can accurately identify "steep but smooth" passable slope areas.

[0059] Step 4: Perform adaptive point density check on each grid cell, calculate the expected number of points (Count_expected), and determine whether the grid is sparse based on the set ratio threshold (Ratio_threshold). Sparse grid cells are then interpolated, while non-sparse grid cells are initially classified according to the slope-roughness criterion. The adaptive density check operation is as follows:

[0060] Considering that the data density of LiDAR decreases exponentially with increasing detection distance, sparse point clouds at long distances can lead to distortion in geometric feature calculations (e.g., insufficient point count can cause variance calculation failure). Therefore, a fixed point count threshold for the entire image cannot be used; instead, a dynamic expected point count needs to be calculated. Based on the LiDAR's angular resolution parameter `Resolution_angular` and the distance `Distance_grid` from the grid center to the sensor, the theoretically expected number of points to be hit at that location is calculated using a physical model. If the actual number of points within the grid is less than `Count_expected` multiplied by a preset ratio threshold `Ratio_threshold`, the grid is determined to be a sparse, hollow grid. For sparse, hollow grids, to avoid misclassification, the classification step is skipped and left for subsequent interpolation steps. For non-sparse, hollow grids with sufficient points, initial classification is performed based on the geometric features calculated in step 3: if the roughness is low and the slope is small, it is marked as a candidate ground; otherwise, it is marked as a candidate obstacle.

[0061] Step 5: Calculate the current observed elevation and confidence level of each grid cell. Based on the initial classification results from Step 4, perform dual-mode elevation update and temporal probability fusion on the candidate ground grid cells. The specific operation of dual-mode elevation update is as follows:

[0062] To address the uncertainties introduced by location drift and dynamic occlusion in unstructured terrain, a differentiated update strategy is adopted. The first step is to calculate instantaneous observations. First, the weighted minimum height of the point cloud within the grid is calculated as the current observation elevation, Height_current. Simultaneously, the current observation confidence_obs is calculated based on the number of point clouds within the grid. The calculation logic is: the more points, the higher the confidence; when the number of points reaches a saturation threshold, the confidence is normalized to 1. The second step is to perform a dual-mode update. For candidate ground grids, a confidence-based temporal weighted fusion mechanism is used. The core is to use confidence as a weight to achieve a balance between current observations and historical memory. This weighted averaging algorithm can effectively utilize high-confidence observations to correct the map, while maintaining historical state when observation quality is poor. At the same time, historical confidence needs to be cumulatively updated so that it increases with the number of observations. For candidate obstacle grids, the core is to prevent under-segmentation (i.e., obstacle occlusion leading to overestimation of ground height). It is checked whether Height_current is significantly lower than Height_history. If this condition is met, it means that terrain deeper than the historical record has been detected (for example, after the planetary rover moves, the bottom of the pit that was originally hidden by rocks becomes visible, or the ground subsidence caused by the wheels sinking). In this case, the above fusion logic must be ignored, a forced downward update must be performed, the map elevation must be directly corrected to Height_current, and the confidence of the grid must be increased within the preset upper limit to ensure the accuracy of the terrain model.

[0063] Step 6: Perform spatiotemporal joint interpolation on the sparse, hollow grid cells not updated in Step 5 using a spiral search to generate a complete elevation map (Map_full), and feed it back to Step 2 as a historical reference for the next frame. The specific operations of spatial interpolation and closed-loop feedback are as follows:

[0064] In unstructured terrain surveying, data gaps are unavoidable. To construct a continuous and complete terrain model to support closed-loop feedback, it is necessary to fill in the sparsely populated graticules marked in step 4, the sparsely populated graticules that have not been updated in step 5, and graticules that are themselves void. First, a spiral search is performed centered on the sensor to locate all sparsely populated graticules that have not been updated. For each sparsely populated graticule, the surrounding graticules that have been updated in elevation are searched, and the weighted average elevation of the neighborhood, Neighbor_avg, is calculated. Then, the confidence_val obtained in step 5 is used as a weighting factor to perform interpolation updates. The specific calculation logic is as follows: when the graticule confidence is low, the interpolation result mainly depends on the average height of the surrounding neighborhood, using spatial smoothing to fill the gaps; when the graticule confidence is high, the interpolation result mainly retains its historical elevation estimate, using temporal information to maintain map stability. This method achieves a dynamic balance between spatial and temporal information. Finally, to reflect the aging of data over time, confidence decay processing is performed on the entire map. The final generated complete elevation map, Map_full, is fed back to the system's input and used as Map_history for the next frame's processing, forming a closed loop in time sequence.

[0065] Step 7: Based on the complete elevation map and the slope and roughness, perform multi-dimensional feature discrimination on the original 3D LiDAR point cloud data, and finally output a multi-dimensional semantic segmentation network containing flat ground, slope, obstacles, and negative obstacles. The multi-dimensional feature discrimination operation is as follows:

[0066] The core lies in utilizing decoupled geometric features and reconstructed complete terrain to construct multi-condition semantic decision-making logic. First, the relative height difference (Height_diff) between the height of each point in the original point cloud and the corresponding ground height in Map_full is calculated. The output is based on the following logic: if the raster roughness (Roughness_rho) is less than a preset threshold and the slope (Slope_alpha) is less than a preset flat ground angle threshold, it is determined to be flat ground, where the energy consumption for passage is lowest; if the roughness is still very small, but the slope is between the flat ground angle threshold and the maximum climbing angle, it is determined to be a passable slope, which is a key advantage of this method compared to traditional elevation segmentation; if the roughness is large, or the relative height difference (Height_diff) is greater than the obstacle height threshold, it is determined to be a positive obstacle; for feature points retained in step 2, if their relative height difference (Height_diff) is negative and its absolute value exceeds a certain threshold, it is ultimately confirmed as a negative obstacle.

[0067] In a second aspect, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method for three-dimensional point cloud terrain segmentation for unstructured planetary surfaces.

[0068] Thirdly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned method for three-dimensional point cloud terrain segmentation for unstructured planetary surfaces.

[0069] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for three-dimensional point cloud terrain segmentation for unstructured planetary surfaces, characterized in that, include: Step 1: Obtain the original 3D lidar point cloud data of the planet's surface; Step 2: Perform preprocessing on the original 3D LiDAR point cloud data to detect outliers and negative obstacles, thereby obtaining effective point cloud data; Step 3: Divide the obtained effective point cloud data into grids, and calculate the slope and roughness of the point cloud in each grid based on the principal component analysis method. Step 4: Perform an adaptive point density check on each grid cell, and mark the grid cells as sparse holes or non-sparse grid cells based on the check results. Then, perform an initial classification of non-sparse grid cells based on the slope and roughness. Step 5: Calculate the current observed elevation and confidence level of each grid cell, and execute the dual-mode elevation update strategy based on the initial classification results to obtain the updated elevation map and confidence level. Step 6: Perform spatiotemporal joint interpolation on the unupdated sparse and void regions to generate a complete elevation map and use it as a reference for the next frame, while synchronously updating the corresponding confidence level. Step 7: Based on the complete elevation map and the slope and roughness, perform multi-dimensional feature discrimination on the original 3D lidar point cloud data, and finally output a multi-dimensional semantic segmentation result containing flat ground, slope, obstacles and negative obstacles.

2. The method for three-dimensional point cloud terrain segmentation for unstructured planetary surfaces according to claim 1, characterized in that, Step 2 specifically includes: The 3D coordinates of each point in the current point cloud are compared with the elevation of the corresponding location in the previous frame of the historical elevation map. If a point is above the ground represented by the previous frame of the historical elevation map, it is retained. If a point is below the ground represented by the previous frame of the historical elevation map, gradient consistency detection is triggered. The gradient consistency detection is as follows: calculate the vertical gradient of the line connecting the point to several nearest neighbor points in its three-dimensional spatial neighborhood; if the statistically obtained gradient distribution conforms to the preset continuous descent physical law, then the point is marked as a feature point representing a negative obstacle and retained; if the point exhibits an isolated abrupt change and has no neighborhood support conforming to the physical law, then it is determined to be a discrete noise point and removed.

3. The method for three-dimensional point cloud terrain segmentation for unstructured planetary surfaces according to claim 1, characterized in that, Step 3 specifically includes: Calculate the covariance matrix for the three-dimensional coordinates of all points within each grid, and perform eigenvalue decomposition on the covariance matrix to obtain three eigenvalues ​​sorted by size; The roughness is characterized by calculating the proportion of the smallest eigenvalue in the sum of the three eigenvalues; the smaller the proportion, the smoother the surface. The slope is characterized by calculating the angle between the eigenvector corresponding to the minimum eigenvalue and the preset gravity direction vector.

4. The method for three-dimensional point cloud terrain segmentation for unstructured planetary surfaces according to claim 1, characterized in that, Step 4 specifically includes: Based on the angular resolution parameters of the 3D LiDAR and the distance from the center of each grid to the sensor, calculate the theoretical expected number of points for that grid. If the actual number of point clouds within a grid is less than the product of the theoretically expected number of points and a preset ratio threshold, then the grid is determined to be a sparse and hollow grid, and the initial classification step is skipped. If the actual number of point clouds is greater than or equal to the product, then based on the slope and roughness values ​​of the grid calculated in step 3, it is initially classified as a candidate ground region or a candidate obstacle region. The candidate ground region includes flat ground and continuous terrain regions with different slopes.

5. A method for three-dimensional point cloud terrain segmentation for unstructured planetary surfaces according to claim 1, characterized in that, The dual-mode elevation update strategy in step 5 specifically includes: The current observation elevation of each grid cell is calculated based on the minimum height weighted by the number of points, and the current observation confidence is calculated based on the number of point clouds in that grid cell. For grids initially classified as candidate ground, a temporal probability fusion mode is adopted: the current observation confidence and the existing historical confidence of the grid are used as weights to perform a weighted average of the current observation elevation and the historical elevation estimates to obtain a new elevation estimate, and the historical confidence is cumulatively updated based on the current observation confidence. For grids initially classified as candidate obstacle regions, a forced downward correction check mode is adopted: the current observed elevation is compared with the historical elevation estimate of the grid; if the current observed elevation is less than the historical elevation estimate, the weighted fusion logic is ignored, the elevation of the grid is directly corrected to the current observed elevation, and its confidence level is increased within the preset upper limit; otherwise, the historical elevation estimate and confidence level remain unchanged.

6. The method for three-dimensional point cloud terrain segmentation for unstructured planetary surfaces according to claim 1, characterized in that, Step 6 specifically includes: A spiral search strategy expanding outward from the center is adopted to locate all sparse, hollow graticles that have not been updated in step 5, as well as graticles that are themselves hollow. For each sparse, hollow raster to be interpolated, search for the raster in its surrounding neighborhood that has already been updated in elevation, and calculate the weighted average of the elevations of these neighboring raster lattices. Using the existing confidence level of the sparse hole grid as the fusion weight, the weighted average elevation of the neighborhood and the historical elevation estimate of the sparse hole grid are weighted and summed to obtain the interpolated elevation value, thereby filling the sparse holes.

7. A method for three-dimensional point cloud terrain segmentation for unstructured planetary surfaces according to claim 5, characterized in that, The specific calculation method for the weighted average in the aforementioned temporal probability fusion mode is as follows: The new elevation estimate equals the product of the current observed elevation and the current observation confidence level, plus the product of the historical elevation estimate and the historical confidence level, and then divided by the sum of the current observation confidence level and the historical confidence level.

8. A method for three-dimensional point cloud terrain segmentation for unstructured planetary surfaces according to claim 1, characterized in that, Step 7 specifically includes: Based on the generated complete elevation map, the relative elevation difference between the elevation of each point in the original point cloud and the estimated ground elevation of its corresponding location in the complete elevation map is calculated; a joint determination is then made by combining the roughness and slope of the raster to which each point belongs. If the roughness of the grid is lower than the first preset threshold and the slope is less than the second preset threshold, then the point is determined to be a flat surface. If the roughness of the grid is lower than the first preset threshold and the slope is between the second preset threshold and the third preset threshold, then the point is determined to be a passable slope. If the roughness of the grid is higher than the first preset threshold, or the relative height difference is greater than the fourth preset threshold, then the point is determined to be a positive obstacle. For the negative obstacle feature points marked and retained in step 2, if their relative height difference is negative and the absolute value exceeds the fifth preset threshold, then the point is determined to be a negative obstacle.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When one or more programs are executed by the one or more processors, the one or more processors implement the three-dimensional point cloud terrain segmentation method for unstructured planetary surfaces as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, enable the processor to implement the three-dimensional point cloud terrain segmentation method for unstructured planetary surfaces as described in any one of claims 1-8.

Images