Cave point cloud data acquisition and three-dimensional model construction method and system

Abstract

This invention discloses a method and system for acquiring point cloud data and constructing 3D models of karst caves, specifically relating to the field of 3D modeling and edge computing of karst caves. The method includes acquiring continuous point cloud data, pose data, and edge computing status data of a moving vehicle within a natural karst cave; performing temporal alignment and segmented registration on the continuous point cloud data and pose data; outputting a local model and a set of unobserved regions; reading the local model and the set of unobserved regions; performing spatial extension analysis and connectivity deduction on each unobserved region; and outputting a set of candidate structures corresponding to each unobserved region. By generating candidate structures for unobserved regions and calculating disambiguation values, the identification results of key unobserved regions are directly incorporated into subsequent supplementary data acquisition and model update processes, thereby improving the problem of difficulty in confirming key connectivity relationships in natural karst caves.

Description

Technical Field

[0001] This invention relates to the field of 3D modeling and edge computing technology for karst caves, and more specifically, to a method and system for acquiring point cloud data and constructing 3D models of karst caves. Background Technology

[0002] In the field of point cloud acquisition and 3D model construction of natural caves, the main goal of existing technologies is to obtain complete, continuous, and usable 3D spatial models for subsequent analysis in areas that are difficult for personnel to access. To achieve this goal, the usual practice is to have a mobile vehicle equipped with laser scanning or depth sensing equipment enter the cave and continuously collect point cloud data along a predetermined route. Then, the cave model is gradually generated by combining processes such as pose estimation, local stitching, global reconstruction, and coverage checking. In some solutions, the mobile vehicle is also used as an edge computing node to perform local processing on the collected data inside the cave before uploading it to the backend, so as to reduce transmission pressure and improve on-site processing efficiency. This type of approach can complete basic modeling tasks in scenarios where the path is relatively clear, the environment is relatively stable, and the data can be fully transmitted back. Taking the deep exploration scenario of natural karst caves as an example, the site often has narrow passages, bifurcated cavities, large changes in top elevation, limited communication distance, limited computing power of the carrier and the inability to rely on real-time human intervention. The system must rely on edge computing capabilities to collect, process and determine which areas need to be observed and which data can be directly used for model updates during the entry process. Under these conditions, although existing technologies can continuously collect new point clouds and expand the surface coverage, and edge computing nodes can also complete a certain degree of local stitching and filtering, in actual results, a situation often occurs where the model surface appears to have collected a lot of data and can be stitched together locally, but the key areas that determine the true connectivity of the cavity are still not clearly collected, such as whether the bifurcation is connected, whether there is a connecting cavity behind the slit, and whether two adjacent cavities are truly connected. This ultimately leads to inconsistent reconstruction results in different batches, or even if offline processing continues, the key spatial relationships cannot be confirmed. The reason for this problem is that even if edge computing is introduced into existing methods, the processing flow is mainly organized around point cloud compression, local registration and real-time transmission. The acquisition and modeling process is still mainly organized based on whether more points are collected and whether the coverage area is expanded. However, it does not further determine which unobserved areas will directly affect the correct understanding of the current spatial structure, nor does it take this impact as the basis for subsequent acquisition and model updates in the edge computing stage. Therefore, the technical problem to be solved by this application is how to identify key unobserved areas that play a decisive role in determining spatial connectivity in natural caves, based on edge computing conditions, under the circumstances where the computing power and communication capabilities of the carrier are limited and real-time human intervention is not possible, and to guide point cloud acquisition and 3D model construction accordingly. Summary of the Invention

[0003] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method and system for acquiring cave point cloud data and constructing a three-dimensional model. By performing candidate structure deduction and discrimination degree calculation on unobserved areas, and combining edge computing status to guide supplementary acquisition of target areas and incremental model confirmation, the problems mentioned in the background art are solved.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for acquiring point cloud data of karst caves and constructing a 3D model, comprising: S1. Acquire continuous point cloud data, pose data and edge computing status data of the mobile carrier in the natural cave, perform temporal alignment and segmented registration on the continuous point cloud data and pose data, and output the local model and the set of unobserved areas. S2. Read the local model and the set of unobserved regions, perform spatial extension analysis and connectivity deduction on each unobserved region, and output the candidate structure set corresponding to each unobserved region. S3. Read each candidate structure set, calculate the degree of distinction of each unobserved region for different candidate structures, and output the ambiguity cancellation value corresponding to each unobserved region. S4. Read the ambiguity resolution values, local model and edge computing status data, perform target area filtering and acquisition path generation, and output the target area and supplementary acquisition instructions. S5. Control the mobile carrier to perform supplementary acquisition on the target area according to the supplementary acquisition command, and perform incremental fusion and connectivity confirmation of the supplementary point cloud data with the local model to output the karst cave 3D model.

[0005] In a preferred embodiment, in S1, temporal alignment and segmented registration are performed on the continuous point cloud data and pose data, including: S1-1. Read the acquisition time of each point cloud frame in the continuous point cloud data, the recording time of each pose record in the pose data, and the status record generated by the edge computing node at each processing time. Among them, the status record is used to characterize the local point cloud processing progress and resource consumption at the corresponding processing time, and includes at least the number of point cloud frames to be processed, the number of registered point cloud frames, storage consumption, and the currently executed task; for each point cloud frame, the absolute value of the time difference between its acquisition time and the time of each pose record and the processing time of each status record is calculated, and the pose record and status record with the smallest absolute value of time difference are respectively used as the corresponding pose and the corresponding edge computing state of the point cloud frame, and the temporal alignment group is output. S1-2. Read the timing alignment group, calculate the change sequence between consecutive adjacent point cloud frames according to the displacement increment, rotation increment between the corresponding poses of adjacent point cloud frames and the proportion of overlapping points after registration of adjacent point cloud frames, and divide the point cloud frames with the same change direction and continuous spatial position in the change sequence into the same local segment, and output multiple local point cloud segments.

[0006] In a preferred embodiment, S1 further includes: S1-3. Read each local point cloud segment, perform frame-by-frame registration and cumulative fusion on the point cloud frames in each local point cloud segment according to the order of acquisition to obtain a local model, and perform visible surface projection and occlusion spatial search on the local model according to the acquisition line of sight, and output the set of unobserved areas not covered by the acquired point cloud.

[0007] In a preferred embodiment, in S2, spatial extension analysis and connectivity deduction are performed on each unobserved region, including: S2-1. Read the boundary point cloud around each unobserved region in the local model, extract the opening boundary of each unobserved region in reverse along the acquisition line, and continuously calculate the cross-sectional width change, boundary normal direction and center position offset of adjacent positions of the opening boundary, and output the extension direction sequence and cross-sectional change sequence corresponding to each unobserved region. S2-2. Read the extension direction sequence and cross-sectional change sequence, extrapolate the opening boundary segment by segment according to the extension direction sequence and generate the corresponding spatial cross-section segment by segment according to the cross-sectional change sequence, connect the continuously generated spatial cross-sections in sequence to form multiple test structures, and output the test structure set corresponding to each unobserved area. S2-3. Read each set of trial structures and local models. For each trial structure, calculate its continuity with the opening boundary, its occupancy conflict with the acquired space, and its connectivity with adjacent known cavities. Retain the trial structures that satisfy the conditions of boundary continuity, no spatial overlap, and connectivity path. Output the candidate structure set corresponding to each unobserved area.

[0008] In a preferred embodiment, S3 includes: S3-1. Read the candidate structures, opening boundaries, extension direction sequences, and cross-sectional change sequences corresponding to the same unobserved area. For each candidate structure, calculate the sum of the distances between corresponding points on the boundary, the sum of the angle differences between the structure's extension direction and the extension direction sequence, and the sum of the differences between the structure's cross-sectional dimensions and the cross-sectional change sequence. Output the boundary deviation values ​​corresponding to each candidate structure. S3-2. Read the boundary deviation values ​​of each candidate structure, local model and the candidate structure. For each candidate structure, calculate the overlapping volume of its occupied space with the acquired space, the shortest connected path length between it and the adjacent known cavity, and the occlusion discrepancy area after projection along the original acquisition line of sight. Add the boundary deviation value, overlapping volume, shortest connected path length and occlusion discrepancy area and output the total difference value corresponding to each candidate structure.

[0009] In a preferred embodiment, S3 further includes: S3-3. Read the total difference values ​​corresponding to the same unobserved region, sort them in ascending order of total difference value, and subtract the total difference values ​​of adjacent candidate structures one by one after sorting. Output the sorting result and adjacent difference value sequence corresponding to the unobserved region. S3-4. Read the sorting results and the adjacent difference sequence, take the maximum difference in the adjacent difference sequence, and take the candidate structures on both sides of the maximum difference as the preferred candidate structure and the second-best candidate structure, respectively. Then calculate the difference between the total difference of the second-best candidate structure and the total difference of the preferred candidate structure, and use this difference as the ambiguity elimination value for the corresponding unobserved region.

[0010] In a preferred embodiment, in S4, the ambiguity resolution values, local model, and edge computing status data are read, and target region filtering and acquisition path generation are performed, including: S4-1. Read the ambiguity resolution value, spatial location of each unobserved region in the local model, and the number of unprocessed point cloud frames, registered point cloud frames, and storage usage in the edge computing status data for each unobserved region. Calculate the path length to the current position of the mobile carrier, the passable space width of the corresponding region, and the newly added point cloud volume after supplementary acquisition for each unobserved region. Combine and sort the ambiguity resolution value, path length, passable space width, and newly added point cloud volume, and output the target region. S4-2. Read the target area and local model, extract the center points of the passable space sequentially from the current position of the mobile carrier to the target area, calculate the change in steering angle and the width of the passage gap for the line segments connecting adjacent center points, and output the target path point sequence. S4-3. Read the target path point sequence and edge computing status data, generate motion control commands and corresponding point cloud acquisition commands for each path point according to the order of the target path point sequence, and combine the motion control commands and point cloud acquisition commands to form supplementary acquisition commands.

[0011] In a preferred embodiment, S5 includes: S5-1. Read the supplementary acquisition command, control the mobile carrier to reach each acquisition position in the target area along the path corresponding to the supplementary acquisition command, and acquire supplementary point cloud data, current position pose data and acquisition line of sight data at each acquisition position, and output the supplementary acquisition data set. S5-2. Read the supplementary acquisition data group and local model, transform each supplementary point cloud data to the coordinate system of the local model according to the corresponding current position and pose data, and perform overlapping region registration between the current supplementary point cloud and the previously fused point cloud according to the acquisition order, and then accumulate and superimpose them to output the supplementary fusion model. S5-3. Read the supplementary fusion model, the candidate structure set corresponding to the target region, and the unobserved region set. For each candidate structure, calculate the sum of the distances between its boundary points and the corresponding boundary points in the supplementary fusion model, the overlapping volume of its occupied space with the existing space in the supplementary fusion model, and the length of the connected path between it and the adjacent known cavity. Add the sum of the distances, the overlapping volume, and the length of the connected path to obtain the confirmation difference for each candidate structure. Take the candidate structure with the smallest confirmation difference as the confirmed structure and output the connectivity confirmation result.

[0012] In a preferred embodiment, S5 further includes: S5-4. Read the supplementary fusion model and connectivity confirmation results, write the spatial boundary corresponding to the confirmed structure into the supplementary fusion model, delete the unobserved area corresponding to the confirmed structure, update the connection position, connection direction and spatial occupancy range between each cavity area, and output the three-dimensional model of the cave. The three-dimensional model of the karst cave includes a set of cave boundary point clouds, a set of cavity connection relationships, and a set of confirmed regional spatial occupancy in a unified coordinate system.

[0013] A system for acquiring point cloud data of karst caves and constructing 3D models, including: The temporal modeling module is used to acquire continuous point cloud data, pose data and edge computing status data of the mobile carrier in the natural cave, perform temporal alignment and segment registration on the continuous point cloud data and pose data, and output local models and sets of unobserved regions. The structural deduction module is used to read the local model and the set of unobserved regions, perform spatial extension analysis and connectivity deduction on each unobserved region, and output the candidate structure set corresponding to each unobserved region. The ambiguity discrimination module is used to read each candidate structure set, calculate the degree of discrimination of each unobserved region against different candidate structures, and output the ambiguity cancellation value corresponding to each unobserved region. The region guidance module is used to read various ambiguity resolution values, local model and edge computing status data, perform target region filtering and acquisition path generation, and output target region and supplementary acquisition instructions; The fusion confirmation module is used to control the mobile carrier to perform supplementary acquisition on the target area according to the supplementary acquisition command, and to perform incremental fusion and connectivity confirmation of the supplementary point cloud data with the local model, and output the 3D model of the cave.

[0014] The technical effects and advantages of this invention are as follows: 1. By generating candidate structures for unobserved areas and calculating ambiguity elimination values, the identification results of key unobserved areas are directly introduced into the subsequent supplementary data collection and model update process, thereby improving the problem of difficulty in confirming key connectivity relationships in natural caves; 2. By performing temporal alignment, local segmentation and intra-segment fusion on continuous point cloud data, pose data and edge computing state data, the impact of temporal mismatch and cross-regional mixing on local models can be relatively reduced, thereby improving the targeting of unobserved area extraction; 3. By generating multiple trial structures based on the opening boundary, extension direction sequence, and cross-sectional change sequence, and by combining local models to screen candidate structures, the coherence and interpretability of the spatial morphology inference behind the unobserved area can be relatively enhanced. 4. By incorporating boundary deviation, spatial overlap, connectivity path, and occlusion consistency into the candidate structure differentiation process, misjudgments based solely on local coverage results can be relatively suppressed, thereby improving the stability of structure discrimination in unobserved areas. 5. By combining ambiguity resolution values, access conditions, and edge computing status data to filter target areas and generate supplementary acquisition instructions, the effectiveness of supplementary acquisition can be relatively improved under conditions of limited computing power and storage, alleviating the processing burden caused by invalid acquisition. 6. By merging the supplementary point cloud data with the incremental local model and confirming the candidate structure accordingly, and then updating the cave boundary point cloud set, cavity connection relationship set, and confirmed area spatial occupancy set, the ability of the karst cave 3D model to represent the real spatial connectivity relationship can be relatively improved. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method steps of the present invention.

[0016] Figure 2 This is a schematic diagram of the system modules of the present invention. Detailed Implementation

[0017] 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.

[0018] Refer to the instruction manual appendix Figure 1-2 The method for acquiring point cloud data and constructing a 3D model of a karst cave according to the present invention includes: S1. Acquire continuous point cloud data, pose data and edge computing status data of the mobile carrier in the natural cave, perform temporal alignment and segmented registration on the continuous point cloud data and pose data, and output the local model and the set of unobserved areas. In this embodiment, continuous point cloud data, pose data, and edge computing state data are first organized into unified input data that is correspondable, segmentable, and fusionable, used to form a local model and extract the set of unobserved regions. The processing principle is as follows: first, the correspondence between point cloud frames and pose and state records is established under the same time reference; then, local segmentation is completed based on the continuity of spatial changes during continuous acquisition; subsequently, frame-by-frame registration and cumulative fusion are performed within each local area; and the correspondence between observed surfaces and their subsequent unobserved spatial locations is used to identify unobserved regions. This process avoids subsequent spatial extension analysis being directly based on temporal mismatches, local mixing, or unclear unobserved ranges. The implementation process includes the following steps: First, the acquisition time of each point cloud frame in the continuous point cloud data, the recording time of each pose record in the pose data, and the status records generated by the edge computing nodes at each processing time are uniformly read. Among them, the status records are continuously written by the edge computing nodes in the order of processing during the local point cloud processing, and are used to characterize the local point cloud processing progress and resource consumption at the corresponding processing time. It includes at least the number of point cloud frames to be processed, the number of registered point cloud frames, storage consumption, and the currently executed point cloud processing step. The currently executed point cloud processing step is used to characterize which processing state the edge computing node is in at the corresponding time: point cloud caching, frame-by-frame registration, intra-segment fusion, or model update. For any point cloud frame, the absolute value of the time difference between its acquisition time and the time of all pose records is calculated, and the pose record corresponding to the minimum value is taken as the corresponding pose of the point cloud frame. Then, the absolute value of the time difference between the acquisition time of the point cloud frame and the processing time of all status records is calculated, and the status record corresponding to the minimum value is taken as the corresponding edge computing state of the point cloud frame. If the minimum time difference corresponds to multiple records, the record whose time is earlier than the acquisition time of the point cloud frame and is closest to the acquisition time is selected first. If there is no record earlier than the acquisition time, the record whose time is later than the acquisition time and is closest to the acquisition time is selected. After the correspondence is completed, the point cloud frame, the corresponding pose, and the corresponding edge calculation state are combined and written into the temporal alignment group for subsequent local partitioning to continue reading. If no usable pose record or usable state record is found before or after the acquisition time of a point cloud frame, the point cloud frame is written into the abnormal record area and removed from the current round of processing sequence to prevent temporally mismatched point clouds from entering the subsequent registration chain. Subsequently, based on temporal alignment groups, the continuous acquisition process is locally divided, so that point cloud frames with consistent spatial variation trends and continuous positions are grouped into the same local range, avoiding the direct merging of point clouds that cross bifurcation points, sharp turning points, or obvious structural abrupt changes. Specifically, according to the acquisition order of point cloud frames, the displacement increment, rotation increment, and overlap ratio are calculated for any two adjacent frames. The displacement increment is obtained from the position coordinate difference between the corresponding poses of the later frame and the corresponding poses of the earlier frame; the rotation increment is obtained from the attitude angle difference between the corresponding poses of the later frame and the corresponding poses of the earlier frame; the overlap ratio is obtained by transforming the point clouds of adjacent frames to the same coordinate system according to their respective poses, counting the number of points in the later frame that fall within the nearest neighbor search radius of the earlier frame, and then dividing by the total number of points in the later frame. The nearest neighbor search radius is twice the average distance from each point in the current point cloud frame to its nearest neighbor, and the average distance is calculated from the nearest neighbor distance of all points in the current point cloud frame. Distance statistics are obtained; based on the aforementioned three quantities, a change sequence between consecutive adjacent point cloud frames is formed; when the displacement vector direction between several consecutive frames remains in the same direction, the attitude angle change maintains the same increasing or decreasing trend, and the geometric center line connecting adjacent point cloud frames is still within the extension range of the current displacement direction, the several consecutive frames are divided into the same local point cloud segment; when the displacement direction reverses, the attitude angle change changes from continuous increase to continuous decrease or from continuous decrease to continuous increase, or the overlap ratio changes from continuous stability to continuous decrease, the position where the change occurs is determined as the new local starting position; after completing the traversal, multiple local point cloud segments are output and written into the local segment index table for subsequent segment-by-segment fusion; if the overlap ratio between two adjacent frames cannot be stably calculated due to local water mist, dripping water echoes, or floating particles, the displacement increment and rotation increment are retained as the division basis, and the adjacent frame pair is marked as a low-confidence registration pair for priority verification during subsequent fusion; After completing the local segmentation, frame-by-frame registration and cumulative fusion are performed on the point cloud frames within each local point cloud segment according to the acquisition sequence to form a local model. Within the local model, a set of unobserved regions not covered by the acquired point clouds is identified. During processing, the current point cloud frame is first transformed to the intra-segment coordinate system of its local point cloud segment based on its corresponding pose. Then, the overlapping area between the current point cloud frame and the fused point cloud from the previous moment is used as the matching range to establish nearest neighbor pairs, and the sum of Euclidean distances between all matching point pairs is calculated. Based on this, the translation and rotation of the current point cloud frame are iteratively updated, causing the sum of Euclidean distances to decrease successively. When the change in the sum of Euclidean distances between two adjacent iterations is less than the change in the previous iteration, the process continues. When the distance sum is one percent of the sum of the distances, or when the number of iterations reaches twenty, the current frame registration is stopped, and the current point cloud frame after registration is written into the fused point cloud to obtain the corresponding local model. After the local model is formed, visible surface projection is performed on each observed surface point in the local model according to the corresponding acquisition line direction. Specifically, a search ray is constructed from each observed surface point along the opposite direction of its corresponding acquisition line direction. The continuous space behind the ray that does not fall into the acquired point cloud is searched within the bounded space of the local model. The space that satisfies the condition that there is an observed surface at the front end, no point cloud occupants in the continuous space behind, and the continuous space is connected to the boundary of the surrounding cavity is determined as the unobserved space. Next, unobserved spaces that are spatially adjacent and have connected openings are merged to obtain an unobserved region set. The opening position, spatial range, and corresponding acquisition line of sight of each unobserved region are written into the unobserved region record table for use in subsequent spatial extension analysis and candidate structure generation. If, during the frame-by-frame fusion process, consecutive gaps and boundary tears are found in the overlapping area of ​​adjacent fused point clouds, the previous fusion result of the local point cloud segment is reverted, and the current point cloud frame and the next frame are reclassified into a new local point cloud segment. Intra-segment registration and fusion are performed again to prevent erroneous fusion results from continuing to be passed to the unobserved region extraction process. Through the above processing, continuous point cloud data, pose data, and edge computing state data can be established under the same temporal correspondence. Then, the continuous acquisition process is divided into multiple local point cloud segments with consistent spatial variation trends. Furthermore, local models and sets of unobserved regions are formed within each local point cloud segment, thus providing a unified, stable, and directly callable input foundation for subsequent opening boundary extraction, candidate structure deduction, and ambiguity elimination value calculation. After this processing, on the one hand, the interference of temporally mismatched point clouds, cross-structure abrupt point clouds, and erroneously fused point clouds on the subsequent generation of candidate structures can be reduced. On the other hand, it can also make the unobserved regions have clear opening locations, spatial ranges, and line-of-sight sources, which facilitates subsequent calculations based on the real connectivity relationships. In practical applications, for example, a tracked mobile carrier equipped with a lidar and inertial measurement unit is sent into a natural cave branch that is difficult for personnel to enter. The mobile carrier moves forward along the narrow passage and continuously collects point cloud data. The edge computing node synchronously records the number of point cloud frames to be processed, the number of registered point cloud frames, the storage usage, and the point cloud processing step currently being executed at each processing moment. The system first maps each frame of point cloud to the pose and state record that is closest in time. Then, it completes local segmentation based on the displacement changes, attitude changes, and overlap ratio between adjacent point cloud frames. Subsequently, it performs frame-by-frame registration and cumulative fusion on each local point cloud segment and identifies the set of unobserved areas behind the bifurcation, behind the slit, and behind the position of the top elevation change in the local model. This provides a clear spatial basis for subsequent determination of whether these unobserved areas lead to independent cavities or are connected to known cavities.

[0019] S2. Read the local model and the set of unobserved regions, perform spatial extension analysis and connectivity deduction on each unobserved region, and output the candidate structure set corresponding to each unobserved region. In the implementation of S2, the purpose of performing spatial extension analysis and connectivity deduction on the unobserved region is to further determine the spatial structure that may extend from the opening boundary of each unobserved region, based on the spatial range of the observed surface and the unobserved region given by the local model, and to filter out the trial results that are inconsistent with the known spatial structure or cannot be established, so as to form a candidate structure set that can be used for subsequent ambiguity elimination value calculation. The basic principle is to first extract the opening boundary at the junction of the unobserved region and the local model, and then obtain the extension direction sequence and cross-sectional change sequence from the local geometric changes of the opening boundary. Based on these two sequences, multiple sets of feasible trial structures are generated. Finally, each trial structure is compared with the collected space and adjacent known cavities in the local model, retaining those with valid geometric connections, non-conflicting spatial occupancy, and achievable connectivity. Through this process, the unobserved region, which originally only showed missing space, can be transformed into a candidate set of structures with clear geometric extension trends and connectivity possibilities. This implementation process includes the following steps: First, the boundary point clouds around each unobserved region in the local model are read to determine the boundary morphology between each unobserved region and the observed space, and based on this, the extension direction sequence and cross-sectional change sequence required for subsequent extrapolation are generated. The input includes the boundary point clouds around the unobserved regions in the local model, the spatial range of the corresponding unobserved regions, and the direction of the acquisition line of sight recorded when the unobserved regions were formed. During processing, the boundary point clouds directly adjacent to the unobserved region are extracted from the position where the unobserved region meets the observed surface along the opposite direction of the acquisition line of sight, and then reordered according to the connectivity order of the boundary point clouds in the circumferential direction of the opening to form an opening boundary point sequence. Subsequently, the midpoint of the adjacent point pair in the opening boundary point sequence is used as the local sampling position, and the cross-sectional width change, boundary normal direction, and center position offset of each local sampling position are calculated respectively. Among them, the cross-sectional width is obtained by the distance between the opposite boundary points on both sides of the opening boundary at the current sampling position, and the cross-sectional width change is the difference between the cross-sectional width of the current sampling position and the cross-sectional width of the previous sampling position. The boundary normal direction is obtained by the change in the angle between the boundary normal vector at the current sampling position and the boundary normal vector at the previous sampling position. The center position offset is obtained by the position difference vector between the boundary center point corresponding to the current sampling position and the boundary center point corresponding to the previous sampling position. The center position offset vector is projected in the opposite direction of the acquisition line of sight, and the direction with the largest projection component is taken as the local extension direction of the current sampling position. The local extension directions are combined in the order of the opening boundaries to obtain the extension direction sequence. The cross-sectional width changes corresponding to each sampling position are combined in the same order to obtain the cross-sectional change sequence. The extension direction sequence and the cross-sectional change sequence are written into the structural extrapolation record table of the corresponding unobserved area for subsequent extrapolation to generate test structures. If there are missing boundary points at a local sampling position, making it impossible to calculate the cross-sectional width directly, the average cross-sectional width of the two nearest available sampling positions before and after the local sampling position is used as the substitute value, and an interpolation mark is written into the structural extrapolation record table to reduce its priority when screening test structures later. Subsequently, the opening boundary is extrapolated segment by segment based on the extension direction sequence and the cross-sectional change sequence. Multiple trial structures are generated through various extension combinations to address the problem that a single extrapolation result is insufficient to cover various real spatial morphologies such as bifurcation, bending, and cavity expansion of natural caves. The input includes the opening boundary, extension direction sequence, and cross-sectional change sequence corresponding to each unobserved area. During processing, the opening boundary is used as the starting cross-section, and the process is advanced segment by segment into the unobserved area according to the direction corresponding to the extension direction sequence. Each advancement generates a new spatial cross-section. The center position of the new spatial cross-section is obtained by moving the center position of the previous spatial cross-section along the current extension direction by one extrapolation step. The extrapolation step is three times the average point spacing of the current opening boundary. This value is derived from the statistical results of the boundary point cloud to ensure that adjacent extrapolated cross-sections remain continuous without being excessively sparse. The size of the new spatial cross-section is obtained by superimposing the size of the previous spatial cross-section with the width change value corresponding to the current cross-sectional change sequence. To generate multiple trial structures, instead of using a single continuous extrapolation method for all spatial cross-sections, three types of cross-sectional update results are constructed simultaneously at each extrapolation position: The first type maintains the current extension direction and updates the dimensions according to the current cross-sectional change value. The second type adds a steering component with the same steering amount as the current boundary normal direction along the current extension direction before updating the dimensions. The third type subtracts a steering component with the same steering amount as the current boundary normal direction along the current extension direction before updating the dimensions. Thus, within the same unobserved region, different extrapolation positions will form different continuous spatial cross-section chains due to different steering component values. Spatial cross-sections continuously generated by the same type of steering combination are sequentially connected to form a set of trial structures. After connecting all spatial cross-section chains corresponding to all steering combinations, a set of trial structures is obtained, and the cross-section sequence, center trajectory, and spatial occupancy range corresponding to each trial structure are written into the trial structure record table for subsequent comparison with the local model. If, during the extrapolation process, it is found that the spatial cross-section generated at a certain continuous extrapolation position has exceeded the outer spatial range of the unobserved region, the extrapolation chain is stopped, and the spatial cross-section generated before the stop is taken as the end cross-section of the trial structure to avoid unbounded extension of the trial structure. After obtaining the set of trial structures, each trial structure is compared with the local model one by one to retain the truly possible structural results and form a candidate structure set. The input includes the set of trial structures, the local model, the opening boundary, the acquired spatial range, and the known cavities adjacent to the unobserved area in space. During processing, the continuity between each trial structure and the opening boundary is calculated first. Specifically, the boundary points of the starting section of the trial structure and the opening boundary points are established as nearest neighbor pairs. The sum of the distances between all corresponding point pairs is calculated, and the angle difference between the normal of the starting section and the average normal of the opening boundary is calculated. The sum of the distances and the angle difference are used together as the continuity result. Subsequently, the occupancy conflict between the trial structure and the acquired space is calculated. Specifically, the number of voxels that overlap with the point cloud voxels of the local model within the space occupied by the trial structure is counted, and the number of overlapping voxels is divided by the total number of voxels of the trial structure to obtain the occupancy overlap ratio. The voxel side length is twice the average point spacing of the local model, and this value is derived from the local model points. Cloud density statistics are then analyzed. Next, the connectivity between the test structure and adjacent known cavities is calculated. Specifically, starting from the center point of the end section of the test structure, an adjacency search is performed along the cavity space in the local model. If a continuous cavity path that does not cross the space occupied by the acquired space reaches the adjacent known cavity, it is recorded as a connected path, and the cumulative length of the path is written into the connectivity results. If no path exists, it is recorded as no connected path. After comparing all test structures, the test structure with the smallest sum of distances between the starting section and the opening boundary, the smallest difference in normal angles, zero overlap ratio, and a connected path is retained as a candidate structure. The candidate structure set is written into the candidate structure record table of the corresponding unobserved area for subsequent discrimination calculation. If all test structures have space conflicts or no connected paths, the test structure with the best continuity and the smallest overlap ratio is retained as a temporary structure, and a supplementary acquisition mark is written into the candidate structure record table so that the unobserved area is given priority in the subsequent supplementary acquisition judgment process. Through the above processing, starting from the unobserved area in the local model, the opening boundary can be extracted and extended direction sequence and cross-sectional change sequence can be formed. Then, multiple test structures can be generated based on various continuous turning combinations, thus making up for the deficiency that a single path extrapolation cannot cover multiple structural possibilities under the conditions of bending, expansion and bifurcation inside natural caves. Furthermore, by combining the continuity of opening boundary connection, the collected spatial occupancy conflict situation and the connectivity and reachability relationship of adjacent known cavities, after screening the test structures, a set of candidate structures with clear geometric origin and connectivity basis can be output, which establishes a clear input basis for subsequent ambiguity elimination value calculation. In practical applications: For example, after a mobile carrier completes a round of local modeling in a natural cave, an unobserved area is identified behind a bifurcation. The system first calculates the boundary normal direction and cross-sectional width change from the boundary point cloud around the opening of the unobserved area. Then, starting from the opening boundary, it constructs three types of continuous spatial cross-sectional chains: straight outward extrapolation, left-deflection outward extrapolation, and right-deflection outward extrapolation, thus forming multiple trial structures. Subsequently, these trial structures are compared with the collected space in the local model and the known cavity on the other side of the bifurcation. Trial structures that overlap with the collected wall and those that cannot reach the known cavity are eliminated. Finally, trial structures that smoothly connect with the opening boundary, do not conflict with the collected space internally, and can establish reachability with adjacent cavities are retained as the candidate structure set. This provides a basis for subsequent determination of whether the unobserved area corresponds to an independent cavity or a bifurcation connected cavity.

[0020] S3. Read each candidate structure set, calculate the degree of distinction of each unobserved region for different candidate structures, and output the ambiguity cancellation value corresponding to each unobserved region. In this embodiment, the purpose of calculating the distinguishability of candidate structures corresponding to the same unobserved region is to further determine which type of candidate structure better matches the actual spatial extension result of the unobserved region, based on the candidate structures already derived from the opening boundary, extension direction sequence, and cross-sectional change sequence. This distinguishability is then quantified into an ambiguous resolution value that can be directly used for subsequent target region screening. The processing principle is as follows: first, the boundary deviation of each candidate structure is calculated around the opening boundary and extension trend; then, the spatial validity of each candidate structure is calculated by combining the collected space in the local model and adjacent known cavities; subsequently, all candidate structures within the same unobserved region are sorted and subjected to difference analysis; finally, an ambiguous resolution value that characterizes the degree of separation between the preferred candidate structure and its competing structures is output. This implementation process includes the following steps: First, the deviation between each candidate structure and the original opening shape of the unobserved region is calculated through boundary geometric consistency to ensure that subsequent comparisons are based on the same geometric starting point. Inputs include each candidate structure, opening boundary, extension direction sequence, and cross-sectional change sequence corresponding to the same unobserved region. During processing, for any candidate structure, its initial cross-sectional boundary points are extracted and established in a circumferential order with the boundary points in the opening boundary. If the number of boundary points differs, equidistant interpolation is performed along the boundary arc length on the side with fewer points to ensure a consistent number of points. Then, the Euclidean distance between corresponding points is calculated and accumulated to obtain the sum of distances between corresponding boundary points. Subsequently, the extension direction of the structure is constructed based on adjacent trajectory points of the candidate structure's center trajectory, and the angle between the direction vectors at corresponding positions in the extension direction sequence is calculated. All angles are accumulated to obtain the sum of angle differences between the structure's extension direction and the extension direction sequence. Further, the cross-sectional dimensions of each extrapolated position of the candidate structure are read and compared with the corresponding positions in the cross-sectional change sequence. The absolute values ​​of the width changes are subtracted one by one and then summed to obtain the sum of the differences between the structural cross-sectional dimensions and the cross-sectional change sequence. To ensure a consistent calculation method for the subsequent total difference, in this embodiment, the three types of results are uniformly converted to a voxel count method: the sum of the distances between corresponding points on the boundary is divided by the average point spacing of the local model and rounded to obtain the number of boundary deviation points; the sum of the angle differences is divided by the average normal rotation angle of adjacent opening boundaries and rounded to obtain the number of direction deviation points; the sum of the cross-sectional dimension differences is divided by the average cross-sectional width change of the opening boundaries and rounded to obtain the number of cross-sectional deviation points. The number of boundary deviation points, the number of direction deviation points, and the number of cross-sectional deviation points are then added together to output the boundary deviation value corresponding to the candidate structure and written into the candidate structure deviation record table for subsequent spatial validity calculation. If there are interpolation marks at local positions in the extension direction sequence or cross-sectional change sequence, the angle and dimension differences at the corresponding positions are still included in the calculation, but the interpolation source mark is simultaneously written into the candidate structure deviation record table for subsequent verification. Subsequently, the validity of each candidate structure in the local model is calculated based on spatial occupancy and occlusion consistency to avoid retaining candidate structures that appear geometrically reasonable but actually contradict the acquired space based solely on the opening boundary fitting results. The input includes each candidate structure, the local model, and the corresponding boundary deviation value for each candidate structure. During processing, for any candidate structure, its spatial occupancy range is first discretized into a set of voxels. The voxel side length is twice the average point spacing of the local model. This value is derived from the statistical results of the local model's point cloud density and is used to ensure that the candidate structure occupancy judgment maintains the same resolution as the local model density. Then, the voxel set of the candidate structure and the acquired spatial voxels are statistically analyzed. The number of overlapping voxels in the set is used as the overlapping volume of the occupied space and the acquired space. Then, starting from the center point of the candidate structure's end section, an adjacency search is performed within the cavity space voxels in the local model. The search rule is to expand only to adjacent cavity voxels that are coplanar, share edges, or share angles with the current voxel, and to accumulate the path length from the starting point to each expanded voxel. When the set of cavity voxels corresponding to an adjacent known cavity is reached for the first time, the current accumulated path length is used as the shortest connected path length between the candidate structure and the adjacent known cavity. If all reachable cavity voxels have been traversed without reaching an adjacent known cavity, the diagonal of the circumscribed space of the currently unobserved region is then... The line length is used as the alternative path length and marked as disconnected. Further, the candidate structure is projected onto the projection plane of the local model along the original acquisition line of sight to obtain the candidate structure projection area. Then, the observed occlusion areas of the local model in the same line of sight are projected onto the same projection plane, and the number of projection grids not covered by the observed occlusion areas in the candidate structure projection area is counted to obtain the occlusion discrepancy area. To eliminate aperture breaks caused by the direct addition of different physical quantities, in this embodiment, the shortest connected path length is divided by the voxel side length and rounded to obtain the path voxel number; the occlusion discrepancy area is divided by the area of ​​a single grid on the projection plane and rounded to obtain the occlusion discrepancy grid. The overlapping volume itself is represented by the number of overlapping voxels, which is on the same order of magnitude as the boundary deviation value, the number of path voxels, and the number of occlusion mismatch grids. Then, the boundary deviation value, the number of overlapping voxels, the number of path voxels, and the number of occlusion mismatch grids are added together to output the total difference value corresponding to the candidate structure, and written into the candidate structure total difference value record table for subsequent sorting. If the adjacent known cavities in the local model have not yet formed a complete cavity voxel set, the number of overlapping voxels and the number of occlusion mismatch grids are still calculated, and the number of path voxels converted from the diagonal length of the outer space of the unobserved area is used as a substitute value to ensure that the candidate structure can continue to participate in the sorting, but a pending sampling mark is written at the same time. Furthermore, the total differences of all candidate structures within the same unobserved region are sorted and adjacent differences are calculated to construct the basis for subsequent ambiguity elimination values. The input includes the total differences corresponding to each unobserved region. During processing, all candidate structures are first rearranged in ascending order of total difference to obtain a sorting result, and the candidate structure numbers in the sorting result are written sequentially into a sorting record table. Subsequently, the differences between adjacent candidate structures in the sorting result are calculated one by one, specifically by subtracting the total difference of the previous candidate structure from the total difference of the subsequent candidate structure, resulting in an adjacent difference sequence. This adjacent difference sequence is then written into the same unobserved region. A domain distinction record table is used to subsequently determine the preferred and second-best candidate structures. To avoid instability in subsequent discrimination due to an insufficient number of candidate structures or identical differences, when only one candidate structure exists in the same unobserved region, that candidate structure is directly written to the first position of the sorting result, and the adjacent difference sequences are marked as empty. When there are two or more candidate structures and two or more identical maximum differences appear in the adjacent difference sequences, the pair of adjacent candidate structures with higher sorting positions is selected as the subsequent representative structure pair to ensure that the discrimination result primarily reflects the degree of separation between the candidate structure with the lowest total difference and its nearest competing structure. Finally, based on the sorting results and the adjacent difference sequence, the preferred candidate structure, the second-best candidate structure, and the corresponding ambiguity elimination value are determined to transform the competitive strength between candidate structures into numerical results that can be directly used for target region filtering. The input includes the sorting results and the adjacent difference sequence. During processing, the maximum difference in the adjacent difference sequence is read first, and the corresponding position of the maximum difference in the sorting results is located. Then, the first candidate structure to the right of the maximum difference is determined as the representative structure to the right of the maximum difference. At the same time, to avoid directly mistaking the adjacent structure to the left of the maximum difference as the preferred candidate structure when the maximum difference appears in the later part of the sorting, in this embodiment, all candidate structures to the left of the maximum difference are read first, and the candidate structure with the smallest total difference is selected as the representative structure to the left of the maximum difference. This representative structure to the left of the maximum difference is taken as the preferred candidate structure, and the maximum difference is... The structure on the right is considered the second-best candidate structure. Then, the difference between the total difference of the second-best candidate structure and the total difference of the first-best candidate structure is calculated, and this difference is used as the disambiguation value for the corresponding unobserved region. If the adjacent difference sequence is empty, the unique candidate structure is determined as the first-best candidate structure, and the disambiguation value is recorded as the difference between the total difference of the unique candidate structure and zero, indicating that there are no competing candidate structures in the unobserved region. After the calculation is completed, the first-best candidate structure number, the second-best candidate structure number, and the disambiguation value are written into the unobserved region differentiation result table for subsequent target region screening. If the total difference of all candidate structures in the ranking result is the same, the first-ranked candidate structure is considered the first-best candidate structure, the second-ranked candidate structure is considered the second-best candidate structure, and the disambiguation value is recorded as zero, indicating that the unobserved region is still in a highly ambiguous state. Through the above processing, we can first calculate the degree of deviation of each candidate structure from the original boundary morphology and spatial extension trend, starting from the opening boundary, extension direction sequence, and cross-sectional change sequence. Then, combined with the acquired space, adjacent known cavities, and original acquisition line of sight in the local model, we can determine the validity of each candidate structure in terms of geometric connection, spatial occupation, and occlusion interpretation. After unifying all the results to the same counting caliber, we form a total difference value, avoiding the direct addition of results with different dimensions that would cause a break in the calculation chain. Furthermore, by using the ranking results and adjacent difference sequences, we can determine the degree of separation between the preferred candidate structure and the second-best candidate structure, and output an ambiguity elimination value that can directly characterize the strength of ambiguity in the unobserved area, thus establishing a clear, continuous, and executable input basis for subsequent target area screening. In practical applications: For example, in an unobserved area behind a bifurcated cavity in a natural cave, the system has generated three sets of candidate structures, corresponding to three spatial interpretations: straight connectivity, left-bending connectivity, and right-expanding termination. During processing, the system first calculates the boundary distances between the three candidate structures and the opening boundaries, the angle differences between them and the extension direction sequence, and the dimensional differences between them and the cross-sectional change sequence. These results are then converted into boundary deviation values ​​with a unified counting aperture. Subsequently, the system counts the number of overlapping voxels between each candidate structure and the collected space, and the distance between each candidate structure and adjacent known cavities. The path voxels and the number of occlusion mismatch grids along the original acquisition line are used to form a total difference. If the total difference between the straight connected structure and the left-bending connected structure is close, while the total difference between the right-expanding terminating structure is significantly larger, the system will determine the optimal structure on the side with the lower total difference as the priority candidate structure after sorting, and determine the structure that competes with it most closely as the second-best candidate structure. Then, the difference between the two total differences will be output as the ambiguity cancellation value. This will determine whether the unobserved area currently has a sufficiently clear structure distinction result, or whether it still needs to enter the subsequent supplementary acquisition process.

[0021] S4. Read the ambiguity resolution values, local model and edge computing status data, perform target area filtering and acquisition path generation, and output the target area and supplementary acquisition instructions. In this embodiment, the purpose of reading the disambiguation values, local model data, and edge computing status data corresponding to each unobserved region and performing target region filtering and acquisition path generation is to determine the priority target regions for supplementary acquisition based on the already obtained results of distinguishing the strength of each unobserved region, combined with the current position of the mobile vehicle, the traffic conditions in the local model, and the current processing capacity of the edge computing nodes, and to generate supplementary acquisition instructions corresponding to the target regions. The processing principle is as follows: first, the supplementary acquisition benefits and arrival costs are calculated for each unobserved region, and the edge computing status data is converted into constraints on the current data processing capacity. Then, target regions are filtered under the premise of satisfying the current processing capacity constraints. Subsequently, a sequence of target path points to reach the target region is generated based on the traffic space in the local model. Finally, mobile control instructions and point cloud acquisition instructions are generated based on the target path point sequence and edge computing status data to keep the supplementary acquisition action consistent with the current data processing rhythm of the edge computing nodes. This implementation process includes the following steps: First, by jointly calculating the disambiguation value, spatial location, path cost, and incremental data of each unobserved region, the target region for the current round of supplementary data collection is selected. This ensures that the target region can prioritize connectivity determination without exceeding the data processing capacity of the edge computing node. The input includes the disambiguation value corresponding to each unobserved region, the spatial location of each unobserved region in the local model, the current location of the mobile carrier, and the number of point cloud frames to be processed, the number of registered point cloud frames, and the storage usage in the edge computing status data. During processing, the current position of the mobile carrier and the opening positions of each unobserved area are first used as the start and end points in the local model. Adjacency search is performed within the passable space of the cavity, and the distance between the centers of each adjacent passable voxel on the search path is accumulated to obtain the path length from the current position of the mobile carrier to each unobserved area. The passable voxel is obtained by subtracting the space occupied by the collected space from the cavity space in the local model. The voxel side length is three times the average point spacing of the local model. This value comes from the point cloud density statistics of the local model. Subsequently, multiple continuous cross-sections are captured along the path forward at the opening location of the corresponding unobserved area. The effective passage width within each cross-section that does not overlap with the space already collected is counted, and the minimum value is taken as the passable space width corresponding to the unobserved area. Further, to calculate the newly added point cloud volume after supplementary collection, a set of candidate supplementary collection viewpoints is constructed around the opening location of the unobserved area. The set of candidate supplementary collection viewpoints consists of three locations at distances of one, two, and three times the length of the mobile carrier body from the opening location, and all three locations are located in the current path direction. Then, ray projection is performed on the visible space of the unobserved area from each candidate supplementary collection viewpoint along the collection line of sight. The number of rays that can first fall into the surface not covered by the already collected point cloud is counted, and this number is multiplied by the average sampling density of the current sensor within a unit viewpoint to obtain the predicted value of the newly added point cloud volume corresponding to the candidate supplementary collection viewpoint. The maximum predicted value of the newly added point cloud volume among the three candidate supplementary collection viewpoints is taken as the newly added point cloud volume corresponding to the unobserved area after supplementary collection. Then, the number of point cloud frames to be processed in the edge computing status data is added to the number of registered point cloud frames to obtain the total number of frames in the current processing queue; the storage usage is divided by the total storage capacity of the edge computing nodes to obtain the current storage usage ratio; then, the total number of frames in the current processing queue is multiplied by the average number of points per frame to obtain the current number of points to be processed, and the average number of registered frames completed by the edge computing nodes in the past thirty processing times is used as the current unit cycle processing capacity. This average value comes from the statistical results of the historical status records of the edge computing nodes; if the newly added point cloud volume corresponding to an unobserved area is greater than the remaining processable points after subtracting the current number of points to be processed from the current unit cycle processing capacity, then the unobserved area is recorded as an overloaded area and will not participate in the target area selection in this round; the remaining unobserved areas are sorted in the following order: first Unobserved areas are sorted by disambiguation value from smallest to largest, prioritizing those with lower discrimination and stronger ambiguity. If disambiguation values ​​are the same, they are then sorted by path length from smallest to largest. If path lengths are the same, they are then sorted by passable space width from largest to smallest. If passable space widths are the same, they are then sorted by newly added point cloud volume from smallest to largest. The unobserved area at the top of the list is designated as the target area, and its number, opening location, candidate supplementary sampling viewpoints, and corresponding predicted newly added point cloud volume are written into the target area record table for subsequent path generation. If all unobserved areas are marked as overloaded areas, the unobserved area with the smallest newly added point cloud volume is selected as the target area, and a load reduction sampling marker is written into the target area record table to reduce the number of sampling locations later. Subsequently, a sequence of target path points is generated based on the accessible space in the target area and the local model. This enables the mobile vehicle to stably reach the corresponding re-sampling position in the target area along the accessible space and provides a sequential path basis for the generation of subsequent mobile control commands. The input includes the target area, the local model, and the current position of the mobile vehicle. During processing, the set of accessible space voxels between the current position of the mobile vehicle and the candidate re-sampling viewpoints in the target area is first read from the local model. Then, taking the current position of the mobile vehicle as the starting point and the viewpoint with the largest predicted value of newly added point cloud volume among the candidate re-sampling viewpoints in the target area as the ending point, a shortest path search is performed within the set of accessible space voxels to obtain a target access path composed of continuous accessible voxels. Then, the center point of the accessible space is extracted along the target access path at fixed intervals. The fixed interval is half the length of the mobile vehicle's fuselage. This value comes from the structural size constraints of the mobile vehicle and is used to ensure that the adjacent path points can reflect the path bending changes without being too dense and increasing the control burden. Every fixed interval, at the current... The center position equidistant from the already acquired spatial boundary is calculated within the cross-section of the passage voxel, and this center position is taken as a target path point. Next, the change in turning angle and the clearance width are calculated for each line segment connecting adjacent target path points. The change in turning angle is obtained by the angle between the direction vector of the current line segment and the direction vector of the previous line segment, and the clearance width is obtained by the minimum distance between the two already acquired spatial boundaries on either side of the cross-section perpendicular to the path direction at the midpoint of the current line segment. After all target path points are extracted, the target path point coordinates, the lengths of adjacent line segments, the corresponding change in turning angle, and the clearance width are written into the target path point sequence table in chronological order for subsequent supplementary acquisition commands. If a candidate supplementary acquisition viewpoint in the target area is unreachable during the path search, the path search is restarted by switching to the next candidate supplementary acquisition viewpoint based on the newly added point cloud prediction value from largest to smallest. If all candidate supplementary acquisition viewpoints are unreachable, the process returns to the target area record table and a new unobserved area is selected as the next highest-ranked target area. After obtaining the target path point sequence, the path following action and point cloud acquisition action are linked and configured according to the edge computing status data to form a supplementary acquisition command, so that the movement rhythm, acquisition rhythm and edge processing rhythm of the mobile carrier are consistent. The input includes the target path point sequence and edge computing status data. During processing, the sequential number, path point coordinates, adjacent line segment length, corresponding turning angle change and passage gap width of each path point in the target path point sequence are read first. Then, the total number of frames in the current processing queue, the current storage occupancy ratio and the current unit cycle processing capacity in the edge computing status data are read. For each target path point, the forward distance command is generated according to the adjacent line segment length, and the turning angle command is generated according to the turning angle change. The two are combined to form the movement control command for the corresponding path point. Then, the point cloud acquisition is determined according to the total number of frames in the current processing queue and the current unit cycle processing capacity. The specific rule is: when the total number of frames in the current processing queue is less than or equal to the number of frames that can be processed corresponding to the current unit cycle processing capacity, point cloud acquisition is performed once at the path point. When the total number of frames in the current processing queue exceeds the number of frames that can be processed per unit cycle, skip the point cloud acquisition of the current path point and only perform movement control. When the current storage occupancy rate is greater than 80%, adjust the acquisition interval of subsequent path points from once per point to once every other path point. The 80% allocation is due to edge computing node storage management rules to ensure that remaining write space is retained during supplementary acquisition. Furthermore, adjust the movement speed and acquisition line of sight based on the width of the passage gap at each path point: when the passage gap width is less than twice the width of the mobile carrier, the movement speed is half the rated speed; when the passage gap width is greater than or equal to twice the width of the mobile carrier... The movement speed is set to the rated speed; the line of sight for data acquisition is always pointed in the direction of the line connecting the opening of the target area and the current path point to ensure that the supplementary data acquisition focuses on covering the space behind the opening of the unobserved area; after generating all the movement control commands and point cloud acquisition commands for all path points, the two are combined in the order of the path points to form a supplementary acquisition command, which is written into the supplementary acquisition task table for subsequent mobile carriers to execute; if it is found that the total number of frames in the current processing queue continues to increase during the command generation process and does not decrease for three consecutive processing times, a delayed acquisition mark is written into the supplementary acquisition task table, and the point cloud acquisition frequency of all subsequent path points is reduced to once every two path points to prevent the supplementary acquisition results from not being processed in time due to point cloud backlog at the edge computing nodes; Through the above processing, the ambiguity elimination value, spatial location, accessibility conditions, and incremental data of the unobserved area can be uniformly incorporated into the target area selection process, and the edge computing status data can be substantially transformed into data processing capability constraints for this round of supplementary acquisition, thereby avoiding the selection of target areas based solely on spatial distance or coverage area. Furthermore, by combining the accessible space in the local model to generate a sequence of target path points, and dynamically configuring movement control commands and point cloud acquisition commands based on the current processing load of the edge computing nodes, the supplementary acquisition actions can be matched with the edge processing capabilities, reducing the risk of data accumulation and command mismatch after supplementary acquisition. In practical applications: For example, after a mobile vehicle completes a round of local modeling and candidate structure differentiation in a natural cave, the system finds that one of the three unobserved regions has the lowest ambiguity resolution value, but this region is far away and the expected increase in point cloud volume is large. Another region is closer but has a slightly higher ambiguity resolution value. At this time, the system first eliminates regions where the expected number of supplementary acquisition points exceeds the current remaining processing capacity based on the current number of points to be processed by the edge computing node, the processing capacity per unit cycle, and the storage occupancy ratio. Then, it sorts the remaining regions according to the ambiguity resolution value, path length, passable space width, and increase in point cloud volume to determine the target region. Subsequently, it extracts the target path point sequence from the current position to the supplementary acquisition viewpoint of the target region in the local model, and generates supplementary acquisition instructions based on the change in turning angle of each path point, the passable gap width, and the current edge processing load. This enables the mobile vehicle to enter the target region and complete supplementary acquisition at a pace that the current edge computing node can handle.

[0022] S5. Control the mobile carrier to perform supplementary acquisition on the target area according to the supplementary acquisition command, and perform incremental fusion and connectivity confirmation of the supplementary point cloud data with the local model to output the three-dimensional model of the cave. In this embodiment, the purpose of supplementary acquisition, incremental fusion, and connectivity confirmation is to transform the previously selected target area into actual supplementary acquisition results, and to use the supplementary acquisition results to confirm the candidate structures corresponding to the target area. Finally, the confirmed spatial boundaries and connectivity relationships are written into the model to form a 3D model of the karst cave that can be used for subsequent analysis and scheduling. The processing principle is as follows: First, supplementary acquisition of the target area is completed according to the supplementary acquisition instructions to obtain supplementary point cloud data, current position pose data, and acquisition line-of-sight direction data. Then, incremental fusion is performed between the supplementary point cloud data and the local model in a unified coordinate system. Subsequently, the supplementary fusion model is used to calculate the confirmation difference for each candidate structure corresponding to the target area and select the confirmed structure. Finally, the confirmed structure is written into the supplementary fusion model, the confirmed unobserved areas are deleted, and the connectivity relationships and spatial occupancy results between cavities are updated. This implementation process includes the following steps: First, supplementary acquisition data sets for the target area are obtained by executing supplementary acquisition commands, so that subsequent incremental fusion is based on the actual supplementary acquisition results. The input is the supplementary acquisition command, which includes at least the target path point sequence, the movement control command corresponding to each acquisition position, and the point cloud acquisition command. During processing, the mobile carrier moves sequentially according to the target path point sequence. After reaching each acquisition position, a position stability judgment is performed. Specifically, the position coordinates and attitude angles at three consecutive sampling times are read. When the maximum difference between the position coordinates at the three sampling times is less than one percent of the length of the mobile carrier's fuselage and the maximum difference between the attitude angles is less than one degree, the current position is determined to be stable. Among them, one percent and one degree are determined by... Due to the constraints on the pose acquisition accuracy of the mobile carrier; after the position stabilizes, a point cloud acquisition is performed at the current acquisition position, and the corresponding current position pose data and acquisition gaze direction data are read. The supplementary point cloud data, current position pose data and acquisition gaze direction data are combined and written into the supplementary acquisition data group for subsequent fusion reading; if the current position does not meet the stabilization condition for three consecutive sampling times, the point cloud acquisition is delayed and the current position is maintained; if the stabilization condition is still not met after three delays, the current position pose data and acquisition gaze direction data are retained, and the acquisition position is written into the low stability position recording area, so that the priority of the supplementary point cloud at this position is reduced during subsequent fusion. Subsequently, a supplementary fusion model is formed through unified coordinate transformation and successive superposition, enabling the supplementary point cloud data to maintain the same spatial reference as the local model. The inputs are the supplementary acquired data sets and the local model. During processing, the current position and pose data corresponding to each set of supplementary point cloud data are first read, and the supplementary point cloud data is transformed from the sensor coordinate system to the coordinate system where the local model is located. Then, according to the acquisition order, the overlapping area between the current supplementary point cloud and the previously fused point cloud is extracted, nearest neighbor pairs are established within the overlapping area, and the sum of Euclidean distances between all corresponding point pairs is calculated. Then, the translation and rotation of the current supplementary point cloud are iteratively updated, and the sum of Euclidean distances is successively reduced. When the sum of the Euclidean distances between two adjacent iterations is less than one percent of the sum of the Euclidean distances in the previous iteration, or when the number of iterations reaches twenty, the current supplementary point cloud registration is stopped, and the registered current supplementary point cloud is accumulated and written into the fused point cloud to obtain the supplementary fusion model. After all supplementary point clouds have been written, the supplementary fusion model is written into the model update cache for subsequent candidate structure confirmation. If a group of supplementary point clouds cannot establish a sufficient number of nearest neighbor pairs in the overlapping area, its coordinate transformation results are retained but not included in the current round of superposition, and the group of supplementary point clouds is written into the point cloud area to be reviewed to avoid low-quality supplementary sampling results directly affecting structure confirmation. Next, the candidate structures corresponding to the target region are confirmed one by one through the supplementary fusion model, transforming the unobserved regions from the inference results into confirmed structures. The inputs are the supplementary fusion model, the set of candidate structures corresponding to the target region, and the set of unobserved regions. During processing, for any candidate structure, the boundary points of the candidate structure are first extracted, and the corresponding boundary point closest to the candidate structure is searched in the supplementary fusion model. The Euclidean distance between the corresponding boundary points is calculated one by one and accumulated to obtain the boundary point distance sum. Then, the spatial occupancy range of the candidate structure is discretized into a set of voxels, and the number of voxels that overlap with the existing spatial voxel set of the supplementary fusion model is counted to obtain the overlapping volume. The voxel side length is twice the average point spacing of the local model, and this value comes from the statistical results of the point cloud density of the local model. Subsequently, starting from the center point of the end section of the candidate structure, an adjacency search is performed within the cavity space voxel corresponding to the supplementary fusion model. The shortest path length to the adjacent known cavity is accumulated. If the cavity is not reached... For adjacent known cavities, the diagonal length of the outer space of the target region is used as the alternative path length and marked as disconnected. Then, to maintain consistency in the measurement of the confirmation difference, the distance between boundary points is divided by the average point spacing of the local model and rounded to obtain the number of boundary deviation points. The shortest path length is divided by the voxel side length and rounded to obtain the number of path voxels. The overlapping volume is directly represented by the number of overlapping voxels. The number of boundary deviation points, the number of overlapping voxels, and the number of path voxels are then added together to obtain the confirmation difference corresponding to the candidate structure. After calculating all candidate structures, the candidate structure with the smallest confirmation difference is selected as the confirmed structure, and the confirmed structure number, confirmation difference, and disconnected marker are written into the connectivity confirmation result table for subsequent model updates. If two or more candidate structures have the same confirmation difference, the candidate structure with the smallest number of boundary deviation points is selected as the confirmed structure. If the number of boundary deviation points is still the same, the candidate structure with the higher ranking position is selected as the confirmed structure. Finally, by writing the confirmed structure into the supplementary fusion model and writing back the update results, a directly usable 3D model of the karst cave is formed. The inputs are the supplementary fusion model and the connectivity confirmation results. During processing, the spatial boundary points, center trajectory, and spatial occupancy range corresponding to the confirmed structure are first read, and the spatial boundary points are written into the cave boundary point cloud set in the supplementary fusion model. Then, based on the shortest connectivity path results between the confirmed structure and adjacent known cavities, the connection start point, connection end point, and connection path direction corresponding to the target area are updated in the cavity connection relationship table. Subsequently, the spatial occupancy range corresponding to the confirmed structure is written into the existing... Confirm the spatial occupancy set of the confirmed area and delete the unobserved area corresponding to the confirmed structure from the unobserved area record table; after completing the write-back, output the 3D model of the cave; the 3D model of the cave includes the cave boundary point cloud set, the cavity connection relationship set, and the confirmed area spatial occupancy set under a unified coordinate system; if there are no connectivity markers in the connectivity confirmation results, the spatial boundary of the confirmed structure is still written into the cave boundary point cloud set and the confirmed area spatial occupancy set, but the area corresponding to the confirmed structure is marked as an unconnected terminated cavity in the cavity connection relationship set, so as to be prioritized for verification during subsequent re-sampling; Through the above processing, the supplementary acquisition results of the target area can be incrementally fused with the local model in a unified coordinate system. Based on the supplementary fusion model, the candidate structures corresponding to the target area are confirmed, transforming the candidate structures that were originally in the deduction stage into confirmed structures. Furthermore, by writing back the spatial boundaries, cavity connection relationships, and spatial occupancy ranges corresponding to the confirmed structures, a well-defined and stable 3D model of the karst cave can be formed. This allows subsequent judgment of unobserved areas, path planning, and structural analysis to be based on the updated model. In practical applications: for example, after a mobile carrier arrives at the target area behind a certain fork in the road according to the supplementary acquisition command, in three... The system acquires supplementary point cloud data, current position pose data, and line-of-sight direction data at the acquisition location. The supplementary point cloud data is then transformed to the coordinate system of the local model and superimposed to obtain a supplementary fusion model. Subsequently, the system calculates the number of boundary deviation points, the number of overlapping voxels, and the number of path voxels for multiple candidate structures corresponding to the target area, and selects the candidate structure with the smallest confirmation difference as the confirmed structure. Finally, the spatial boundary of the confirmed structure is written into the cave boundary point cloud set, the connectivity results between it and adjacent cavities are written into the cavity connection relationship set, and its spatial occupancy range is written into the confirmed area spatial occupancy set, thereby obtaining the updated 3D model of the cave.

[0023] Furthermore, the present invention also includes a system for acquiring cave point cloud data and constructing a 3D model, the system comprising: The temporal modeling module is used to acquire continuous point cloud data, pose data and edge computing status data of the mobile carrier in the natural cave, perform temporal alignment and segment registration on the continuous point cloud data and pose data, and output local models and sets of unobserved regions. The structural deduction module is used to read the local model and the set of unobserved regions, perform spatial extension analysis and connectivity deduction on each unobserved region, and output the candidate structure set corresponding to each unobserved region. The ambiguity discrimination module is used to read each candidate structure set, calculate the degree of discrimination of each unobserved region against different candidate structures, and output the ambiguity cancellation value corresponding to each unobserved region. The region guidance module is used to read various ambiguity resolution values, local model and edge computing status data, perform target region filtering and acquisition path generation, and output target region and supplementary acquisition instructions; The fusion confirmation module is used to control the mobile carrier to perform supplementary acquisition on the target area according to the supplementary acquisition command, and to perform incremental fusion and connectivity confirmation of the supplementary point cloud data with the local model, and output the 3D model of the cave.

[0024] The core idea of ​​this scheme is to allow a mobile platform to collect continuous point cloud data while moving within a natural cave. This data is then combined with pose data to achieve local alignment and segmented fusion, forming a local model. From this model, unobserved areas with visible boundaries but unclear backgrounds are identified. Instead of treating these unobserved areas as ordinary gaps, the system deduces multiple possible spatial structures based on the shape, extension direction, and cross-sectional changes of the opening boundaries. These possible structures are then compared one by one to determine their distinguishability, and this result is used to prioritize additional data collection. After identifying the target area, the system combines the passable space in the local model with the current processing capabilities of edge computing nodes to generate supplementary data collection paths and instructions. Finally, the newly collected point cloud is fused with the original local model, and the results confirm the actual structure corresponding to the target area. The confirmed spatial boundaries, cavity connections, and spatial occupancy results are then written into the model to obtain the final 3D model of the cave. When applying this scheme, for example, during deep exploration of natural caves, a mobile platform enters a bifurcation area along a narrow passage. Although a considerable amount of point cloud data has been collected, there may still be a section of space behind the bifurcation that remains unclear. The system will first determine whether this unobserved area is more likely to continue forward, turn left into another cavity, or simply a dead corner behind a closed cavity wall. If it's still unclear, the system won't blindly continue scanning the entire cave. Instead, it will prioritize selecting the area most influencing the determination of connectivity and then move to a more suitable location along a passable route to collect additional data. After this additional data collection, the system merges the newly collected point cloud data with the existing model to confirm whether the area is connected and where it extends. Finally, the results are updated into the 3D model. The advantage of this approach is that it doesn't simply pursue a larger number of point clouds, but prioritizes clearly capturing the key locations that truly affect the determination of the cave's spatial structure.

[0025] The above description is merely a preferred embodiment of the present invention and is 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 acquiring point cloud data of karst caves and constructing a 3D model, characterized in that, include: S1. Acquire continuous point cloud data, pose data and edge computing status data of the mobile carrier in the natural cave, perform temporal alignment and segmented registration on the continuous point cloud data and pose data, and output the local model and the set of unobserved areas. S2. Read the local model and the set of unobserved regions, perform spatial extension analysis and connectivity deduction on each unobserved region, and output the candidate structure set corresponding to each unobserved region. S3. Read each candidate structure set, calculate the degree of distinction of each unobserved region for different candidate structures, and output the ambiguity cancellation value corresponding to each unobserved region. S4. Read the ambiguity resolution values, local model and edge computing status data, perform target area filtering and acquisition path generation, and output the target area and supplementary acquisition instructions. S5. Control the mobile carrier to perform supplementary acquisition on the target area according to the supplementary acquisition command, and perform incremental fusion and connectivity confirmation of the supplementary point cloud data with the local model to output the karst cave 3D model.

2. The method for acquiring point cloud data and constructing a three-dimensional model of a karst cave according to claim 1, characterized in that: In S1, temporal alignment and segmented registration are performed on continuous point cloud data and pose data, including: S1-1. Read the acquisition time of each point cloud frame in the continuous point cloud data, the recording time of each pose record in the pose data, and the status record generated by the edge computing node at each processing time. Among them, the status record is used to characterize the local point cloud processing progress and resource consumption at the corresponding processing time, and includes at least the number of point cloud frames to be processed, the number of registered point cloud frames, storage consumption, and the currently executed task; for each point cloud frame, the absolute value of the time difference between its acquisition time and the time of each pose record and the processing time of each status record is calculated, and the pose record and status record with the smallest absolute value of time difference are respectively used as the corresponding pose and the corresponding edge computing state of the point cloud frame, and the temporal alignment group is output. S1-2. Read the timing alignment group, calculate the change sequence between consecutive adjacent point cloud frames according to the displacement increment, rotation increment between the corresponding poses of adjacent point cloud frames and the proportion of overlapping points after registration of adjacent point cloud frames, and divide the point cloud frames with the same change direction and continuous spatial position in the change sequence into the same local segment, and output multiple local point cloud segments.

3. The method for acquiring point cloud data and constructing a three-dimensional model of a karst cave according to claim 2, characterized in that: S1 also includes: S1-3. Read each local point cloud segment, perform frame-by-frame registration and cumulative fusion on the point cloud frames in each local point cloud segment according to the order of acquisition to obtain a local model, and perform visible surface projection and occlusion spatial search on the local model according to the acquisition line of sight, and output the set of unobserved areas not covered by the acquired point cloud.

4. The method for acquiring point cloud data and constructing a three-dimensional model of a karst cave according to claim 3, characterized in that: In S2, spatial extension analysis and connectivity deduction are performed on each unobserved region, including: S2-1. Read the boundary point cloud around each unobserved region in the local model, extract the opening boundary of each unobserved region in reverse along the acquisition line, and continuously calculate the cross-sectional width change, boundary normal direction and center position offset of adjacent positions of the opening boundary, and output the extension direction sequence and cross-sectional change sequence corresponding to each unobserved region. S2-2. Read the extension direction sequence and cross-sectional change sequence, extrapolate the opening boundary segment by segment according to the extension direction sequence and generate the corresponding spatial cross-section segment by segment according to the cross-sectional change sequence, connect the continuously generated spatial cross-sections in sequence to form multiple test structures, and output the test structure set corresponding to each unobserved area. S2-3. Read each set of trial structures and local models. For each trial structure, calculate its continuity with the opening boundary, its occupancy conflict with the acquired space, and its connectivity with adjacent known cavities. Retain the trial structures that satisfy the conditions of boundary continuity, no spatial overlap, and connectivity path. Output the candidate structure set corresponding to each unobserved area.

5. The method for acquiring point cloud data and constructing a three-dimensional model of a karst cave according to claim 4, characterized in that: In S3, the following are included: S3-1. Read the candidate structures, opening boundaries, extension direction sequences, and cross-sectional change sequences corresponding to the same unobserved area. For each candidate structure, calculate the sum of the distances between corresponding points on the boundary, the sum of the angle differences between the structure's extension direction and the extension direction sequence, and the sum of the differences between the structure's cross-sectional dimensions and the cross-sectional change sequence. Output the boundary deviation values ​​corresponding to each candidate structure. S3-2. Read the boundary deviation values ​​of each candidate structure, local model and the candidate structure. For each candidate structure, calculate the overlapping volume of its occupied space with the acquired space, the shortest connected path length between it and the adjacent known cavity, and the occlusion discrepancy area after projection along the original acquisition line of sight. Add the boundary deviation value, overlapping volume, shortest connected path length and occlusion discrepancy area and output the total difference value corresponding to each candidate structure.

6. The method for acquiring point cloud data and constructing a three-dimensional model of a karst cave according to claim 5, characterized in that: S3 also includes: S3-3. Read the total difference values ​​corresponding to the same unobserved region, sort them in ascending order of total difference value, and subtract the total difference values ​​of adjacent candidate structures one by one after sorting. Output the sorting result and adjacent difference value sequence corresponding to the unobserved region. S3-4. Read the sorting results and the adjacent difference sequence, take the maximum difference in the adjacent difference sequence, and take the candidate structures on both sides of the maximum difference as the preferred candidate structure and the second-best candidate structure, respectively. Then calculate the difference between the total difference of the second-best candidate structure and the total difference of the preferred candidate structure, and use this difference as the ambiguity elimination value for the corresponding unobserved region.

7. The method for acquiring point cloud data and constructing a three-dimensional model of a karst cave according to claim 6, characterized in that: In S4, the ambiguity resolution values, local model, and edge computing status data are read, and target area filtering and acquisition path generation are performed, including: S4-1. Read the ambiguity resolution value, spatial location of each unobserved region in the local model, and the number of unprocessed point cloud frames, registered point cloud frames, and storage usage in the edge computing status data for each unobserved region. Calculate the path length to the current position of the mobile carrier, the passable space width of the corresponding region, and the newly added point cloud volume after supplementary acquisition for each unobserved region. Combine and sort the ambiguity resolution value, path length, passable space width, and newly added point cloud volume, and output the target region. S4-2. Read the target area and local model, extract the center points of the passable space sequentially from the current position of the mobile carrier to the target area, calculate the change in steering angle and the width of the passage gap for the line segments connecting adjacent center points, and output the target path point sequence. S4-3. Read the target path point sequence and edge computing status data, generate motion control commands and corresponding point cloud acquisition commands for each path point according to the order of the target path point sequence, and combine the motion control commands and point cloud acquisition commands to form supplementary acquisition commands.

8. The method for acquiring point cloud data and constructing a three-dimensional model of a karst cave according to claim 7, characterized in that: S5 includes: S5-1. Read the supplementary acquisition command, control the mobile carrier to reach each acquisition position in the target area along the path corresponding to the supplementary acquisition command, and acquire supplementary point cloud data, current position pose data and acquisition line of sight data at each acquisition position, and output the supplementary acquisition data set. S5-2. Read the supplementary acquisition data group and local model, transform each supplementary point cloud data to the coordinate system of the local model according to the corresponding current position and pose data, and perform overlapping region registration between the current supplementary point cloud and the previously fused point cloud according to the acquisition order, and then accumulate and superimpose them to output the supplementary fusion model. S5-3. Read the supplementary fusion model, the candidate structure set corresponding to the target region, and the unobserved region set. For each candidate structure, calculate the sum of the distances between its boundary points and the corresponding boundary points in the supplementary fusion model, the overlapping volume of its occupied space with the existing space in the supplementary fusion model, and the length of the connected path between it and the adjacent known cavity. Add the sum of the distances, the overlapping volume, and the length of the connected path to obtain the confirmation difference for each candidate structure. Take the candidate structure with the smallest confirmation difference as the confirmed structure and output the connectivity confirmation result.

9. The method for acquiring point cloud data and constructing a three-dimensional model of a karst cave according to claim 8, characterized in that: S5 also includes: S5-4. Read the supplementary fusion model and connectivity confirmation results, write the spatial boundary corresponding to the confirmed structure into the supplementary fusion model, delete the unobserved area corresponding to the confirmed structure, update the connection position, connection direction and spatial occupancy range between each cavity area, and output the three-dimensional model of the cave. The three-dimensional model of the karst cave includes a set of cave boundary point clouds, a set of cavity connection relationships, and a set of confirmed regional spatial occupancy in a unified coordinate system.

10. A system for acquiring point cloud data of karst caves and constructing 3D models, comprising the method for acquiring point cloud data of karst caves and constructing 3D models according to claim 1, characterized in that, include: The temporal modeling module is used to acquire continuous point cloud data, pose data and edge computing status data of the mobile carrier in the natural cave, perform temporal alignment and segment registration on the continuous point cloud data and pose data, and output local models and sets of unobserved regions. The structural deduction module is used to read the local model and the set of unobserved regions, perform spatial extension analysis and connectivity deduction on each unobserved region, and output the candidate structure set corresponding to each unobserved region. The ambiguity discrimination module is used to read each candidate structure set, calculate the degree of discrimination of each unobserved region against different candidate structures, and output the ambiguity cancellation value corresponding to each unobserved region. The region guidance module is used to read various ambiguity resolution values, local model and edge computing status data, perform target region filtering and acquisition path generation, and output target region and supplementary acquisition instructions; The fusion confirmation module is used to control the mobile carrier to perform supplementary acquisition on the target area according to the supplementary acquisition command, and to perform incremental fusion and connectivity confirmation of the supplementary point cloud data with the local model, and output the 3D model of the cave.

Images