An edge-side adaptive reconstruction and privacy protection method for a 3D scanning base station

By establishing a joint processing relationship between reconstruction sufficiency state and privacy identifiability state in a 3D scanning base station, and utilizing the stable geometry of adjacent non-privacy spatial units for registration constraints and contour suppression processing, the problem of separating end-side reconstruction and privacy determination is solved, achieving efficient 3D reconstruction and privacy protection.

CN123118920APending Publication Date: 2026-11-17HEFEI HOUTU TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611458391.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-18
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing 3D scanning base stations separate reconstruction decisions and privacy assessments at the edge, making it impossible to coordinate the determination of necessary observations, registration feature sources, and fusion data representations. This results in a difficulty in maintaining consistency between reconstruction continuity and privacy identifiability, especially in areas with occlusion or insufficient geometric information, which may expose sensitive information.

Method used

By establishing a joint processing relationship between reconstruction sufficiency state and privacy identifiability state in a 3D scanning base station, registration constraints are established using the stable geometric structure of adjacent non-privacy spatial units, sensitive observations are transformed into geometrically fused data that is decoupled from source frames, and contour suppression processing is performed before output to ensure privacy protection.

Benefits of technology

While maintaining the continuity and integrity of the 3D reconstruction model, it reduces the association between identity textures and acquisition sources, thereby reducing the execution load, storage consumption, and computing power consumption of the edge processor and improving privacy protection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118920A_ABST
    Figure CN123118920A_ABST
Patent Text Reader

Abstract

This invention relates to the field of 3D image processing and point cloud reconstruction technology, specifically to an edge-side adaptive reconstruction and privacy protection method for 3D scanning base stations. The method involves acquiring multi-view scanning data from the base station, registering it to a unified 3D coordinate system, and dividing it into spatial units. Based on multi-view observation correlations and image sensitive region projections, the reconstruction sufficiency state and privacy identifiability state are determined, and a joint processing relationship between the two is established. For privacy-sensitive spatial units, registration constraints are established using the stable geometric structures of adjacent non-privacy spatial units, converting sensitive observations into geometrically fused data that decorrelates source frame correlations. Based on the relationship, partitioned surface fusion and state updates are performed to generate a 3D reconstruction model and remove texture mapping from sensitive meshes. This invention retains necessary surface geometry and local connectivity relationships while reducing the retention of sensitive textures and source correlations, and decreasing edge-side redundant registration, redundant fusion, and cache read / write operations.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of three-dimensional image processing and point cloud reconstruction technology, specifically to an edge-side adaptive reconstruction and privacy protection method for a three-dimensional scanning base station. Background Technology

[0002] Existing 3D scanning base stations typically use depth cameras, binocular imaging modules, or laser scanners to acquire color images, depth maps, and point cloud data of the target object. Based on pre-calibrated sensor intrinsic and extrinsic parameters, observations from different perspectives are transformed into a unified 3D coordinate system. After outlier removal, voxel downsampling, and surface normal estimation, the acquired data is coarsely registered using geometric key points and their local descriptive relationships. Then, the spatial pose of each observation frame is obtained through iterative nearest-point processing and pose map correction. The registered point cloud is used to generate a continuous surface through truncated symbolic distance field fusion, surface element fusion, or Poisson surface reconstruction. A 3D model is formed through hole repair, mesh simplification, and texture mapping. When privacy information is involved, the original scanning data is usually kept at the base station, and data flow is restricted by cache clearing, file encryption, and access control.

[0003] The existing technology has the following technical problems: the edge-side 3D reconstruction decision and privacy determination are separated, making it impossible to coordinate the determination of necessary observations, registration feature sources and fusion data representation for the same spatial region; the unified reconstruction process will continuously retain the original frames of occluded areas or areas with insufficient geometric information, which may contain data that can identify faces, human contours, documents and display content. However, directly occluding, blurring or deleting sensitive areas will destroy the feature correspondence between adjacent viewpoints, causing pose estimation to lose local constraints and forming misalignments or holes in the fusion surface; since the degree of reconstruction completeness and privacy recognizability are not mapped to the same 3D spatial unit, the base station can hardly use non-sensitive geometric structures to replace sensitive registration features, and it is also difficult to remove the association between geometric records and source frame textures while retaining the necessary surface geometry. As a result, it is difficult to keep the reconstruction continuity and privacy data minimization consistent during edge-side processing. Summary of the Invention

[0004] The purpose of this invention is to provide an edge-side adaptive reconstruction and privacy protection method for a three-dimensional scanning base station, which can solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An edge-side adaptive reconstruction and privacy protection method for a 3D scanning base station includes: acquiring multi-view scanning data of the 3D scanning base station, performing global initial registration of the scanning data to a unified 3D coordinate system, and dividing the data into 3D spatial units; The reconstruction sufficiency state of each three-dimensional spatial unit is determined based on the multi-view observation association, and the privacy identifiability state is determined based on the projection of the image sensitive area onto the unified three-dimensional coordinate system. The privacy identifiability state distinguishes between texture identifiability attributes and geometric contour identifiability attributes, and establishes a joint processing relationship between the two states. Based on the joint processing relationship, the observation retention method and the registration feature sources and fusion expressions associated with each other are allocated. For privacy-sensitive spatial units, registration constraints are established based on the stable geometric structures in adjacent non-privacy spatial units, and the registered sensitive observations are converted into geometric fusion data that is decoupled from the source frames. When the identifiable attribute of the geometric contour reaches a preset recognition threshold, contour suppression processing is performed on the geometric fusion data of the corresponding privacy-sensitive spatial unit to weaken the fine geometric contours that can be used for identity recognition. Based on the joint processing relationship, partition surface fusion is performed to update the two states and output the three-dimensional reconstruction model. When the three-dimensional reconstruction model needs to be transmitted externally, external transmission desensitization verification is performed to verify the geometric contour identifiable attributes of each privacy-sensitive spatial unit. For spatial units whose geometric contour identifiable attributes exceed the threshold, contour suppression desensitization is performed again. Only after the verification is passed can the model data be output externally.

[0006] Preferably, the determination of the reconstruction adequacy state includes: projecting the effective depth region of each observation frame onto the three-dimensional spatial unit according to the scanning pose, and establishing the coverage association between the three-dimensional spatial unit and the observation viewpoint; Effective observation distribution is obtained within each three-dimensional spatial unit, and integrity markers are formed based on the surface normal consistency and registration correspondence of adjacent observations; The coverage association is combined with the integrity flag to form the reconstruction adequacy state.

[0007] Preferably, the determination of the privacy identifiability status includes: obtaining the privacy-sensitive region from the color image of the multi-view scan data through semantic segmentation, and back-projecting the privacy-sensitive region onto the unified three-dimensional coordinate system using the depth map and scan pose synchronized with the color image; Back-projection points that overlap across frame projections are merged into a sensitive point set and a continuous association is established. Texture association attributes and geometric contour association attributes are recorded respectively. The geometric contour association attributes characterize the outer boundary shape, local curvature, and cross-view surface size consistency of the sensitive point set. When the geometric contour association attributes exceed a preset threshold, it is determined that the spatial unit has a geometric contour identity recognition risk. The privacy identifiability state is formed according to the three-dimensional spatial units covered by the sensitive point set; if the geometric contour association attribute exceeds the preset threshold, the privacy identifiability state of the spatial unit remains sensitive and triggers geometric contour suppression processing, which does not allow the retention of fine geometric contours that can identify the identity.

[0008] Preferably, the establishment of the joint processing relationship includes: pairing the reconstruction adequacy state and the privacy identifiability state according to the spatial index of the three-dimensional spatial unit to generate a processing identifier that includes observation preservation relationship and the registration source and fusion expression associated with each other; When the processing identifier corresponds to the privacy-sensitive spatial unit, the registration source is pointed to the adjacent non-privacy spatial unit and the fusion expression is limited to the geometric expression of the decoupled source frame; When the processing identifier corresponds to a non-privacy space unit, the observation retention relationship is determined based on the reconstruction sufficiency state.

[0009] Preferably, the establishment of the coverage association includes: establishing an observation support map by projecting the outlines of each observation frame in the unified three-dimensional coordinate system and the three-dimensional spatial unit, and connecting three-dimensional spatial units with shared boundaries and covered by the same observation viewpoint as support edges; When a new observation frame is added, the nodes that intersect with the new projection contour and the corresponding support edges are updated, and the integrity marker is reconstructed based on the updated observation support map.

[0010] Preferably, the establishment of the sensitive point set includes: extracting effective depth points and depth discontinuity edges intersecting with the region contour within the privacy-sensitive region, back-projecting the effective depth points into a core sensitive point set, and registering the region to be confirmed along the observation ray corresponding to the depth missing position; The area to be confirmed is cross-projected using subsequent observation perspectives, and the supplementary point set obtained from the cross-projection is combined with the core sensitive point set to form the sensitive point set.

[0011] Preferably, the establishment of the stable geometry includes: extracting the geometric boundary of cross-frame repeated observations along the shared boundary between the privacy-sensitive spatial cell and the adjacent non-privacy spatial cell, and selecting surface patches from the adjacent non-privacy spatial cells that maintain topological adjacency with the geometric boundary; Local registration constraints are formed using geometric boundaries and surface patches, and texture descriptors falling into the privacy-sensitive spatial unit are excluded from the registration correspondence set.

[0012] Preferably, the partitioned surface fusion includes: establishing a reconstruction task chain associated with three-dimensional spatial units based on the processing identifier; For non-privacy space units that are not sufficiently reconstructed, new observations are incorporated into the local incremental registration task based on the global initial registration and then surface fusion is performed. For fully reconstructed non-privacy spatial units, registered observations are converted into fused surfaces and task associations for duplicate observations are removed; For the privacy-sensitive spatial unit, the geometric fusion data is written into a fusion task that does not contain texture mapping relationships; Write the output of each fusion task back to the corresponding 3D spatial unit.

[0013] Preferably, after the local registration constraint is completed, the depth points in the privacy-sensitive spatial unit are transformed to the unified three-dimensional coordinate system, and the geometric records containing only spatial position and surface orientation are accumulated according to the three-dimensional spatial unit. After writing the geometric record into the geometric fusion data, the index relationship between the source frame and the texture region of the source frame accessed from the geometric record is deleted, and controlled erasure is performed on the privacy-sensitive original data in the source frame; if the geometric contour identifiable attribute of the privacy-sensitive spatial unit reaches the preset recognition threshold, contour suppression processing is performed on the geometric record to weaken the local curvature and contour details that can be used for identity recognition, and desensitized geometric fusion data is obtained. The written geometric fusion data is used as the subsequent fusion input for the privacy-sensitive spatial unit.

[0014] Preferably, after the output of each fusion task is written back to the corresponding three-dimensional spatial unit, the reconstruction sufficiency state is updated according to the coverage relationship of the written-back surface to the observation view, and the privacy identifiability state is updated according to the spatial overlap relationship between the written-back surface and the privacy-sensitive spatial unit. Only the three-dimensional spatial units whose reconstruction adequacy state or privacy identifiability state has changed are rewritten into the reconstruction task chain; After the reconstruction task chain is completed, the fused surfaces of each 3D spatial unit are stitched together to form a 3D reconstruction model, and the mapping relationship between the privacy-sensitive spatial unit and the texture atlas is removed. For privacy-sensitive spatial units whose geometric contour recognition attributes reach the threshold, mesh contour suppression is performed. The cache of the original scan data of the corresponding privacy-sensitive area in local storage is cleared. When the 3D reconstruction model needs to be transmitted externally, desensitization verification is performed to verify the desensitization status of the geometric contours of all privacy-sensitive spatial units. Only after the verification is passed can the model be output externally.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By establishing a joint processing relationship between the reconstruction sufficiency state and the privacy identifiability state using the same 3D spatial index, the observation retention, registration source, and fusion representation of each spatial region can be determined by the same state combination. Privacy-sensitive spatial units no longer use identity-related texture description information for registration. Instead, local constraints are established by the geometric boundaries and topologically adjacent surface patches of repeated cross-frame observations in adjacent non-privacy spatial units, ensuring that sensitive regions can maintain spatial connectivity with surrounding surfaces even after exiting feature matching. Sensitive depth points that have completed registration are converted into geometric records containing only spatial location and surface orientation. The index relationship between the geometric records and the source frame and texture region is removed, allowing the output model to retain the surface morphology required for reconstruction while reducing the persistent retention of identity textures, acquisition source associations, and sensitive observations in the derived model.

[0016] 2. The observation support map establishes a connection between the projection contour of the observation frame and the 3D spatial unit. When a new observation enters, only the nodes intersecting with the projection contour and the corresponding support edges are updated. With the reconstruction task chain organized by spatial unit, the registration and surface fusion processing can be limited to the areas where the observation relationship or joint state changes, reducing repeated registration, repeated fusion, and cache read / write in stable areas. Privacy-sensitive areas are back-projected into a sensitive point set through synchronous depth map, and spatial coverage is supplemented by depth discontinuity edges, unconfirmed observation rays, and subsequent viewpoint cross-projections, so that the edges of 2D areas or depth-deficient locations can be included in the privacy state update. After the fusion task is completed, only the state-changed areas are reprocessed, and the mapping relationship between privacy-sensitive spatial units and texture atlas is removed, thereby reducing the execution load, storage usage, model update time, and computational energy consumption of the edge processor. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the overall process of adaptive reconstruction and privacy protection at the 3D scanning base station end. Figure 2 To reconstruct the sufficiency measurement map; Figure 3 A graph of privacy-identifiable metrics; Figure 4 This is a joint determination boundary margin diagram. Detailed Implementation

[0018] Please refer to Figure 1 In one embodiment, an edge-side adaptive reconstruction and privacy protection method for a three-dimensional scanning base station includes: acquiring multi-view scanning data of the three-dimensional scanning base station, performing global initial registration of the scanning data to a unified three-dimensional coordinate system, and dividing the three-dimensional spatial units; The reconstruction sufficiency state of each three-dimensional spatial unit is determined based on the multi-view observation association, and the privacy identifiability state is determined based on the projection of the image sensitive area onto the unified three-dimensional coordinate system. The privacy identifiability state distinguishes between texture identifiability attributes and geometric contour identifiability attributes, and establishes a joint processing relationship between the two states. Based on the joint processing relationship, the observation retention method and the registration feature sources and fusion expressions associated with each other are allocated. For privacy-sensitive spatial units, registration constraints are established based on the stable geometric structures in adjacent non-privacy spatial units, and the registered sensitive observations are converted into geometric fusion data that is decoupled from the source frames. When the identifiable attribute of the geometric contour reaches a preset recognition threshold, contour suppression processing is performed on the geometric fusion data of the corresponding privacy-sensitive spatial unit to weaken the fine geometric contours that can be used for identity recognition. Based on the joint processing relationship, partition surface fusion is performed to update the two states and output the three-dimensional reconstruction model. When the three-dimensional reconstruction model needs to be transmitted externally, external transmission desensitization verification is performed to verify the geometric contour identifiable attributes of each privacy-sensitive spatial unit. For spatial units whose geometric contour identifiable attributes exceed the threshold, contour suppression desensitization is performed again. Only after the verification is passed can the model data be output externally.

[0019] In this embodiment, the 3D scanning base station is deployed in a scenario where digital modeling of the target object needs to be completed at the acquisition location. The base station is equipped with an image acquisition interface, a depth data receiving interface, a timing synchronization processor, an edge computing processor, and a local storage area. Color images, depth maps, and point clouds are written to the same observation frame record according to the acquisition time. The observation frame record includes a frame identifier, sensor source, acquisition time, initial camera pose value, and data buffer address. The edge computing processor only reads the record to be processed from the local storage area and writes the fusion result back to the local spatial index. After the original image completes the necessary projection and registration, it enters the controlled cache cleanup queue. The processing chain from data entering the base station to the formation of the 3D model does not depend on sending the original scanning data to the remote device.

[0020] Specifically, the multi-view scanning data consists of continuous observations that change the viewpoint around the target object. Color images and depth maps are paired by the acquisition time and sensor trigger number. Existing point cloud data is associated with the corresponding color image by the acquisition frame identifier in the point record. If there are multiple sensor observations in the same trigger period, the timing synchronization processor establishes the master observation frame according to the time sequence and registers other observations in the same acquisition window as companion frames. The companion frames retain their respective coordinate transformation entries instead of directly stitching data. During subsequent registration, the initial relative pose is derived from the calibration relationship between the master observation frame and the companion frames, thereby avoiding the loss of the observation direction and the source relationship of privacy areas due to merging data from different sources in advance.

[0021] Furthermore, the unified three-dimensional coordinate system is selected from the base station coordinate system or the first observation frame coordinate system that meets the requirements of image clarity and depth integrity. Sensor calibration data is used to back-project depth pixels into sensor coordinate points, and then convert them into unified three-dimensional coordinate points through the rotation and translation relationships of the corresponding observation frames. Before the conversion, the depth map is marked with no return value, flying points, and mixed depths that cross object boundaries. Isolated points in the point cloud are removed according to the neighborhood connectivity. For observations that repeatedly fall into the same local position, the source index is retained instead of being averaged in advance. The processed spatial points still carry the observation direction, original pixel position, and registration status for reconstruction sufficiency judgment and privacy projection.

[0022] In this embodiment, the three-dimensional spatial unit adopts a block structure that is fixedly associated with a unified three-dimensional coordinate system. Each block has a spatial index, boundary range, adjacent block index, observation reference table, geometric fusion area, and status record area. Spatial points are written into the observation reference table of the corresponding block according to the range of their coordinates. Spatial points located near the block boundary simultaneously register boundary references with adjacent blocks without copying the original data. The geometric fusion area stores surface distance records, surface direction records, and fusion source status formed by multiple observations. The status record area stores reconstruction sufficiency status, privacy identifiability status, and joint processing identifier, so that reconstruction judgment, privacy judgment, and subsequent processing at the same spatial location can be directly associated through the spatial index.

[0023] In a preferred embodiment, the three-dimensional spatial unit can also adopt a hierarchical block structure dynamically generated based on the target surface distribution. The edge computing processor first registers all observations using coarser spatial blocks, and then creates sub-blocks for blocks containing surface points and with large observation changes. The parent block only saves the range and summary state of the sub-blocks, while the specific geometric records are stored in the leaf-level blocks. When a privacy-sensitive area crosses the parent-child boundary, it is propagated downward to the relevant leaf-level blocks through spatial enclosing relationships. After the reconstruction task is completed, adjacent leaf-level blocks with stable states and small geometric changes can be merged into a unified fusion area, so that the spatial division is adjusted according to the surface distribution changes of the scanned object, while keeping the spatial index traceable during state updates.

[0024] Specifically, the initial registration of multi-view scan data prioritizes geometric key points in non-sensitive candidate regions. Before a privacy-identifiable state is formed, the edge computing processor first masks pixels of known sensitive categories through image semantic results, then extracts curvature change points, depth boundaries, and local plane intersections from the remaining depth region. It establishes correspondences between candidate points with matching local geometric distributions in adjacent observations, estimates rigid body transformations between observation frames based on these correspondences, and sends the transformed overlapping points to iterative nearest-point processing to correct poses. If candidate points are mainly concentrated in sensitive regions, the fine registration of related observations is temporarily suspended and the initial pose values ​​are retained. Local constraints are established after adjacent non-privacy spatial units form stable geometric structures.

[0025] The above registration is a global reference registration, used to establish a transformation reference for a unified three-dimensional coordinate system and the observation distribution of the initial three-dimensional spatial units. When supplementing new observations for spatial units that have not been fully reconstructed, a local incremental registration is performed. It uses the coordinate system obtained from the global reference registration and the fused stable surface as constraints, and only performs pose correction and feature matching within the local spatial units where the state has changed. There is no need to repeat the global registration for all observation frames. The two are registration processes at different levels and stages, which together constitute the incremental adaptive reconstruction process.

[0026] In this embodiment, the multi-view observation association is not represented solely by frame order, but is composed of the projection coverage relationship from the observation frame to the three-dimensional spatial unit, the registration correspondence relationship between observation frames, and the shared boundary relationship between adjacent three-dimensional spatial units. The edge computing processor records the set of spatial units covered by each observation frame, records the viewing direction that can be observed on the surface of each spatial unit, and establishes registration edges for observation frames with common spatial points or common surface boundaries. Subsequent new observations only modify the local associations that intersect with their covered set. The reconstruction sufficiency state is generated by the viewing coverage, observation continuity, and pose consistency in the local associations, without needing to repeatedly perform global comparisons on all scan frames.

[0027] The privacy identifiability status is based on the actual coverage of the sensitive areas of the image in a unified three-dimensional coordinate system. The edge computing processor performs semantic segmentation on the color image to obtain the face, human body shape, document surface, display area, and pre-registered sensitive object areas. After aligning the segmented areas with the synchronous depth map, it back-projects them into three-dimensional sensitive points. Locations lacking depth retain the corresponding observation rays and image contours. Sensitive points generated from multiple perspectives are merged into three-dimensional spatial units through spatial proximity and surface continuity relationships. The status recording area simultaneously distinguishes texture identifiability attributes, geometric contour identifiability attributes, and cross-frame continuous correlation attributes. When the geometric contour identifiability attribute reaches a preset identifiability threshold, the geometric contour of the spatial unit is determined to have an identity recognition risk. In addition to removing the texture mapping, contour suppression processing is also required for the geometric fusion data of the spatial unit. Fine local geometric contours are smoothed and blurred to weaken the local curvature and contour details that can be used for identity recognition, thus avoiding the possibility of recovering identity information through stable geometric contours or continuous observation relationships after only deleting the color texture.

[0028] Furthermore, the joint processing relationship adopts processing records with three-dimensional spatial units as keys. Each processing record references the reconstruction sufficiency state and the privacy identifiability state, and saves the types of observations that can be retained, the source of features that can participate in registration, the data format written into the geometric fusion area, and the dependencies between the tasks to be performed. The processing record does not directly replace the two states with a single score, but retains the category and formation basis of each state. When any formation basis changes due to new observations, cache clearing, or sensitive area expansion, only the processing record of the corresponding spatial unit is reconstructed, so that the privacy judgment can limit the source of registration data and the reconstruction state can determine the minimum geometric observations that still need to be retained under the restricted conditions.

[0029] For 3D spatial units whose privacy-identifiable states are not marked as sensitive and whose reconstruction is insufficient, the processing record retains observation references that can provide new perspectives or supplement occlusion surfaces, and adds new observations to the local registration and surface fusion tasks. For 3D spatial units whose privacy-identifiable states are not marked as sensitive and whose reconstruction is sufficient, the processing record uses stable fusion surfaces as subsequent inputs and releases duplicate observation references. For privacy-sensitive spatial units, the processing record prohibits identity-related texture descriptors from entering the registration correspondence set and limits the surface fusion input to geometric records that have had their source frame inverse index removed. For spatial units that have both insufficient observations and privacy-sensitive attributes, the processing record selects geometric boundaries and topologically adjacent surface patches from adjacent non-privacy spatial units to form constraints. When reliable constraints cannot be formed, the unreconstructable label is retained and speculative surfaces are not generated.

[0030] In this embodiment, the stable geometry in adjacent non-privacy space cells consists of surface boundaries observed repeatedly across frames, local planar patches, and their topological adjacency relationship. The edge computing processor searches for geometric boundaries that can be located in multiple registered observations along the boundaries of privacy-sensitive space cells. Then, it selects surface patches with continuous normal changes and sufficient observation support from the outside of the boundaries. The geometric boundaries define the spatial seams between sensitive and non-sensitive regions. The surface patches provide the directional constraints required for local pose solving. Local registration only uses the geometric correspondence within the boundaries and surface patches and does not read the color texture of sensitive regions. The obtained pose transforms the sensitive depth point and maintains its connection position with adjacent surfaces.

[0031] When the registered sensitive observations are converted into geometrically fused data that is decoupled from the source frame, the edge computing processor aggregates sensitive depth points according to three-dimensional spatial units. It only writes the spatial position, surface orientation, and fusion support status in the unified three-dimensional coordinates, and does not write the original pixel coordinates, color texture coordinates, or the index that can access the source frame in reverse. After the geometric fusion area performs position consistency checks on multiple sensitive observations of the same local surface, it forms a fusion record. Forming a fusion record means decoupling the link between the temporary sensitive point and the source frame buffer address. If subsequent state updates require reprocessing the same spatial position, the surface relationship is restored from the geometric fusion record and adjacent non-sensitive constraints, without re-activating the cleaned identity texture.

[0032] Partitioned surface fusion uses three-dimensional spatial units as processing boundaries. It converts non-sensitive observations or geometric fusion data that have been decoupled from the source frame into local distance records and surface orientation records. Records from different perspectives and with compatible spatial positions are accumulated according to the observation support relationship. Records with obvious conflicts are retained as branches to be confirmed and await subsequent perspective judgment. After local fusion is completed, the surface position and normal continuity information are exchanged at the spatial unit boundary. Local surfaces with compatible boundaries are stitched into a continuous mesh. When the boundary lacks support, the open boundary marker is retained. Texture mapping only references the color areas that are allowed to be retained in non-privacy spatial units. No mapping to the source texture is established for the mesh patches corresponding to privacy-sensitive spatial units.

[0033] In this embodiment, after each local fusion result is written back, the multi-view coverage and surface consistency of the corresponding spatial unit are recalculated, and the spatial overlap between the fused surface and the range of the 3D sensitive points is rechecked. Spatial units whose states have changed enter the next round of reconstruction task chain, while spatial units whose states have not changed maintain the current fused surface and are not processed repeatedly. When there are no spatial units to be processed in the task chain, the edge computing processor connects the local mesh according to the spatial index, checks whether the privacy-sensitive mesh patch still has a texture atlas mapping or a source frame inverse index, clears data links that do not conform to the processing record, generates a 3D reconstruction model, and writes the model's spatial block index, privacy state summary, and cache cleanup record into the local model file.

[0034] In a preferred embodiment, determining the reconstruction adequacy state includes: projecting the effective depth region of each observation frame onto a three-dimensional spatial cell based on the scan pose, and establishing a coverage association between the three-dimensional spatial cell and the observation viewpoint; obtaining the effective observation distribution within each three-dimensional spatial cell, and forming an integrity marker based on the surface normal consistency relationship and registration correspondence relationship of adjacent observations; and combining the coverage association and the integrity marker into a reconstruction adequacy state.

[0035] Specifically, the scanning pose is composed of the rotation and translation relationships from the observation frame coordinate system to the unified three-dimensional coordinate system. The edge computing processor traverses the effective depth pixels in the depth map, back-projects the pixels into spatial points according to the camera imaging relationship, and then uses the scanning pose to transform the spatial points and determine the three-dimensional spatial unit they fall into. At the same time, the direction from the camera optical center to the spatial point is registered as the observation direction. Areas in the depth map that are occluded by the foreground are not registered as covered. The far-end depth points reached after passing through existing surfaces are assigned to the corresponding surface layers according to the visibility check, so as to avoid mistaking the front and rear surfaces in the same projection direction as the same spatial coverage.

[0036] The effective observation distribution of each three-dimensional spatial unit is generated through local voxel occupancy and surface adjacency. The end-side computing processor arranges the points falling into the spatial unit according to local coordinates, finds adjacent occupancy positions along the estimated surface direction, forms a surface patch with continuous occupancy positions, and retains isolated occupancy positions as points to be confirmed. When a surface patch crosses the boundary of a spatial unit, it is connected to the adjacent spatial unit through boundary reference. If the same surface patch is repeatedly observed from different perspectives, the number of independent observation directions and the coverage range are recorded. If it is repeatedly collected only from close observation directions, it is merged into the same perspective support without increasing the coverage category.

[0037] The surface normal consistency relationship is formed by the angle relationship between the local normals of adjacent observations in a unified three-dimensional coordinate system. The end-side computing processor estimates the local surface orientation for each registered observation, unifies the normals of the same surface patch to the visible side of the scanning device, and then calculates the directional proximity between the normals of different observations. Observations with conflicting orientations are registered as geometric branches and their pose correspondence is traced. The registration correspondence is recorded by the spatial residual after the transformation of local geometric key points, the similarity of the neighborhood shape of the corresponding points, and the closure of the shared boundary. Only observations that have both surface continuity and registration correspondence support are included in the integrity label.

[0038] Furthermore, the sufficiency state reconstruction can be formed by quantization records of the coverage component, the continuous distribution component, the normal consistent component, and the pose corresponding component. The quantization records are calculated using the following formula: ; in, Spatial index representing a three-dimensional spatial unit. This represents the reconstruction sufficiency measure for spatial index u. This represents the normalized value of the effective surface area covered by independent observation directions. This represents the normalized value of the continuous occupancy of a local voxel along the surface direction. A normalized value representing the degree of uniformity of the normals across the observed surface. This represents the normalized value indicating that the registration correspondence satisfies the local residual distribution. and The values ​​represent the combined weights of the four components, with the sum of the values ​​being 1. The plus sign indicates that each component is added together according to its corresponding weight, and the multiplication sign indicates the contribution ratio of the component value to the reconstruction adequacy value.

[0039] When calculating the coverage component, the visible area of ​​the surface patches formed in the spatial unit is used as the statistical benchmark. The area that can be projected onto the surface patch without obstruction from each independent observation direction is deduplicated and accumulated, and the ratio of the accumulated area to the statistical benchmark is limited to between 0 and 1. The distribution continuity component is obtained by the ratio of the number of continuously occupied positions inside the surface patch to the total number of candidate surface positions. The normal consistency component is mapped from the average value of the proximity of the cross-observation normal direction to between 0 and 1. The pose correspondence component is obtained by the ratio of the number of correspondences that satisfy the local residual distribution to the total number of candidate correspondences. When there is a lack of candidate correspondences, the state is kept in the unconfirmed state and not directly assigned a sufficient state.

[0040] For example, if the coverage component of a spatial cell is 0.8, the continuous distribution component is 0.6, the uniform normal component is 0.9, and the pose corresponding component is 0.7, and the four combined weights are 0.30, 0.25, 0.20, and 0.25 respectively, the reconstruction sufficiency value is 0.745. If the sufficient state boundary value formed by the local benchmark observation is 0.75, the spatial cell retains the insufficient reconstruction mark and enters the supplementary observation task. After the new viewpoint is written, the coverage component becomes 0.9, the continuous distribution component becomes 0.8, the uniform normal component remains 0.9, and the pose corresponding component becomes 0.8, with a corresponding value of 0.85. If there is no uncovered continuous surface area, the state recording area converts the spatial cell to a fully reconstructed mark and stops retaining repeated observations.

[0041] In this embodiment, the sufficient state boundary value is not directly determined by hardware performance, but is generated by the base station after completing a set of benchmark scans on the calibration object. The spatial units generated by the benchmark scan are manually or labeled according to whether the surface is closed, whether the cross-view position is consistent, and whether the boundary is continuous. The edge computing processor statistically analyzes the value distribution of sufficient and insufficient samples and records the state boundary value at the boundary between the two. During the actual scan, the state is still corrected by combining the uncovered continuous area and the geometric branches to be confirmed. The quantization record is used to sort the spatial units to be processed and trigger state transitions. A single value is not used to replace the spatial coverage and surface connection relationship.

[0042] Please refer to Figure 2 , Figure 2 The horizontal axis on the left represents the components of the reconstructed evidence, in order of importance: sufficient coverage, continuous distribution, consistent normal direction, and pose correspondence. The vertical axis represents the weighting and current average contribution, with values ​​ranging from 0 to 1. The diagonal bars represent... Four weighted terms, represented by dotted bars. After weighting, the formed Average contribution. The horizontal axis on the right represents the status record number, and the vertical axis represents the overall reconstruction value. The black line represents... The gray dashed line represents the sufficient boundary, which is a process that gradually changes with new observations and state write-backs. .Depend on Figure 2 It is evident that sufficient coverage and pose correspondence account for a relatively high proportion in the weight configuration, indicating that the edge processing prioritizes ensuring that the effective surface is covered by independent viewpoints and requires that the local registration correspondence fall within an acceptable residual distribution; during state progression, It gradually increased from about 0.48 to about 0.85, and then surpassed [a certain value] in the latter part. This indicates that the addition of new perspectives, surface continuity, and normal consistency together drive the spatial unit from a state of needing supplementary observations to a state of full reconstruction.

[0043] In another preferred embodiment, the establishment of coverage association includes: establishing an observation support map by projecting the contours of each observation frame in a unified three-dimensional coordinate system and three-dimensional spatial units, and connecting three-dimensional spatial units with shared boundaries and covered by the same observation viewpoint as support edges; when adding a new observation frame, updating the nodes intersecting with the new projection contour and the corresponding support edges, and reconstructing the integrity markers based on the updated observation support map.

[0044] The observation frame nodes in the observation support diagram save the scan pose and projection contour. The spatial nodes correspond to the three-dimensional spatial units and save the effective surface range. The covering edge between the observation frame node and the spatial node represents the visible projection of the observation frame onto the inner surface of the spatial unit. The support edge between two spatial nodes indicates that they have a shared boundary and at least one common observation view. When establishing the support edge, the boundary point set and observation direction corresponding to the common observation are recorded. If the shared boundary is observed by different and non-overlapping viewpoints, only the spatial adjacency relationship is saved and no observation support relationship is established to prevent the boundary lacking common observation from being misjudged as a continuous surface.

[0045] When a new observation frame enters the observation support map, the edge computing processor queries intersecting spatial nodes based on the spatial enclosing range of the new projection contour. It only performs visibility judgment, cover edge establishment, and boundary point update on the nodes obtained from the query. Then, it visits adjacent spatial nodes along the new or changed support edge and recalculates the relevant integrity markers. Spatial nodes that do not intersect with the new projection contour and whose support edges have not changed remain in their original state. When an observation frame is removed by the cache cleanup process, the corresponding node in the graph is first marked as inaccessible, and then its geometric contribution is checked to see if it has been converted into a fusion record. After the conversion is completed, the cover edge is deleted. Spatial nodes that are still supported by other observations do not re-enter the registration task.

[0046] Preferably, the integrity markers include five types of internal records: closed surfaces, open boundaries, occlusions to be observed, registration conflicts, and depth to be confirmed. Internal records can coexist in the same spatial unit, and the reconstruction sufficiency state is formed based on the combination of internal records. A closed surface indicates that the surface patch in the spatial unit can be continuously stitched with the boundary of the adjacent unit. An open boundary indicates that there are unconnected surface edges. Occlusions to be observed indicate that the observation ray is blocked by the foreground surface. Registration conflicts indicate that different observations form incompatible surface branches. Depth to be confirmed indicates that the image area has texture but lacks corresponding depth. The task scheduler selects a processing path to supplement the viewpoint, re-register, or maintain the unknown area based on the internal records.

[0047] In a preferred embodiment, determining the privacy identifiability state includes: obtaining privacy-sensitive regions from color images of multi-view scan data through semantic segmentation; back-projecting the privacy-sensitive regions onto a unified three-dimensional coordinate system using a depth map synchronized with the color images and scan pose; merging the back-projection points that overlap across frames into a sensitive point set and establishing a continuous association, recording texture association attributes and geometric contour association attributes respectively; and forming a privacy identifiability state according to the three-dimensional spatial units covered by the sensitive point set.

[0048] Texture association attributes are generated from the local color distribution within the sensitive area, the continuous occurrence of texture key points, and the stability of patterns within the area. The attribute record only saves the normalized recognizability and category label, and does not save the descriptive vector that can restore the original image. Geometric contour association attributes are generated from the outer boundary shape of the sensitive point set, the local curvature distribution, and the consistency of surface size across viewing angles. When the geometric contour association attribute exceeds the preset threshold, even if the colored texture has been removed, the privacy recognizability status remains in the sensitive category, and geometric contour suppression processing is triggered. Privacy protection cannot be achieved solely by texture erasure.

[0049] In this embodiment, before the color image enters the semantic segmentation process, distortion correction is completed according to the sensor calibration record, and the color components affected by exposure changes are converted into image representations with brightness and chroma separation. The semantic segmentation processor uses the encoding path to extract local texture and contour features, uses the decoding path to restore the features to the original image position and outputs the category confidence record of each pixel. The edge computing processor forms connected regions according to pixel categories, merges similar regions separated by short-distance occlusion, deletes isolated responses that do not have depth support and cannot appear continuously in adjacent frames, and forms privacy-sensitive regions and their image contours from the retained regions.

[0050] The sensitive category table used for semantic segmentation is pre-registered by the data processing rules of the scanning task. The category table records whether the category is mainly based on texture recognition, whether it is mainly based on geometric contour recognition, and whether it needs to be continuously tracked across frames. The face and document surfaces retain both texture and contour markers. The display area retains texture markers and defines the projection range with the boundary of the bearing plane. The human body shape retains contour markers and establishes an overall association for continuous limb areas. The edge computing processor does not directly write the semantic category name into the final 3D model, but converts the category into a privacy status flag and saves it in the local status record area for joint processing to select the data expression method.

[0051] Specifically, each effective depth pixel in the privacy-sensitive area is back-projected through camera calibration parameters. During calculation, the difference between the pixel's lateral position and the principal point's lateral position is multiplied by the depth value and divided by the lateral focal length scale to obtain the lateral component of the sensor coordinates. The difference between the pixel's vertical position and the principal point's vertical position is multiplied by the depth value and divided by the vertical focal length scale to obtain the vertical component. The depth value is used as the forward component. Then, the rotation and translation relationships of the scanning pose are used to transform it to a unified three-dimensional coordinate system. The transformed sensitive points retain the area marker and observation direction, and the color pixel values ​​are no longer used as the subsequent registration input.

[0052] Cross-frame projection overlap determination takes the sensitive point surface in a unified 3D coordinate system as the object. The edge computing processor projects the sensitive point of the current frame onto the local surface coordinates of the existing sensitive point set, checks the spatial proximity, surface orientation proximity, and semantic category compatibility. Points that meet the compatibility requirements are included in the same sensitive point set and an anonymous reference of the current observation frame is added to the persistent association table. Points that do not meet the compatibility requirements are established as a new sensitive point set. When two sensitive point sets are connected by a continuous surface and have the same category in subsequent views, their 3D coverage can be merged, but the anonymous reference of the source frame is still retained separately according to the observation until the geometric fusion data is formed and then uniformly removed.

[0053] Texture association attributes are generated from the local color distribution within the sensitive area, the persistent occurrence of texture key points, and the stability of patterns within the area. The attribute record only saves the normalized identifiability and category label, and does not save the descriptive vector that can restore the original image. Geometric contour association attributes are generated from the outer boundary shape of the sensitive point set, the local curvature distribution, and the consistency of surface size across viewing angles. The edge computing processor writes both into the spatial unit state. If the color texture has been removed but the geometric contour can still be associated with multiple observations, the privacy identifiability state still maintains the sensitive category, and the geometric processing restrictions are not lifted due to the decrease of a single texture attribute.

[0054] In another preferred embodiment, the establishment of the sensitive point set includes: extracting effective depth points and depth discontinuities intersecting with the region contour within the privacy-sensitive region; back-projecting the effective depth points into a core sensitive point set; and registering the region to be confirmed along the observation ray corresponding to the depth missing position; using subsequent observation perspectives to perform cross-projection on the region to be confirmed; and combining the supplementary point set obtained from the cross-projection with the core sensitive point set to form the sensitive point set.

[0055] Depth discontinuity edges are obtained by comparing the depth changes on both sides of the privacy-sensitive region contour. The edge computing processor reads the effective depth samples on the inner and outer sides along the image contour, registers the depth changes belonging to the same continuous surface as sensitive surface boundaries, registers the depth changes from the foreground to the background as occlusion boundaries, and registers the segments with no depth on the inner side of the contour but stable depth on the outer side as depth-deficient boundaries. The sensitive surface boundaries are back-projected with the effective depth points and limit the range of the core sensitive point set. The occlusion boundaries retain the front-to-back layer relationship, and the depth-deficient boundaries generate an observation ray to be confirmed from the camera optical center through the corresponding contour pixels.

[0056] When writing subsequent observation perspectives, the edge computing processor transforms the observation ray to be confirmed to the subsequent observation coordinate system, queries the image position and corresponding depth through which the ray passes, and if the semantic region of the subsequent perspective is compatible with the original privacy-sensitive region category and the back-projection point falls into the spatial neighborhood of the ray to be confirmed, then the back-projection point is registered as a supplementary point; if the subsequent depth shows that the ray position is occluded by another surface, then the pending confirmation state is maintained and the occluding surface is recorded; if multiple subsequent perspectives show that the outer side of the original outline is free space, then the pending confirmation region is reduced. The cross-projection process only modifies the spatial units that intersect with the observation ray to be confirmed, and does not expand the sensitive range without observation support.

[0057] Furthermore, the privacy-identifiable state can be quantized based on texture-identifiable components, geometric contour-identifiable components, and cross-frame correlation components. The quantization record is calculated using the following formula: ; Here, u still represents the spatial index of a three-dimensional spatial unit. This represents the privacy-identifiable value for spatial index u. This represents the normalized value that indicates the consistency of texture key points and region patterns across multiple frames. This represents a normalized value indicating that the 3D contour and local curvature remain consistent across multiple viewpoints. This represents a normalized value indicating the continuous correlation of a set of sensitive points across frames. and Each of the three components represents a combined weight with a sum of 1. The plus sign indicates the accumulation of components, and the multiplication sign indicates the contribution ratio of the corresponding component to the privacy identifiability value. The symbols maintain a different meaning boundary from those in the reconstruction adequacy quantification record.

[0058] Texture-identifiable components are normalized by the ratio of the number of continuously occurring texture keypoints to the number of candidate keypoints in sensitive regions, the cross-frame similarity order of regional patterns, and category confidence records. Geometric contour-identifiable components are normalized by the cross-viewpoint overlap of the outer boundary of the sensitive point set, the stability of the local curvature order, and the consistency of surface scale. Cross-frame correlation components are obtained by the ratio of the number of times the sensitive point set is continuously matched by different observations to the number of times it is visible. When any component corresponds to a sensitive category that should not be retained, a mandatory sensitivity flag can be directly written. The combined value is used to process queue sorting and general sensitive category conversion, without overwriting the mandatory sensitivity flag.

[0059] For example, if the texture identifiable component of a certain 3D spatial unit is 0.9, the geometric contour identifiable component is 0.7, and the cross-frame association component is 0.8, and the corresponding combination weights are 0.40, 0.35, and 0.25 respectively, the privacy identifiable value is 0.805. If the sensitive state boundary value formed by the local data rules is 0.70, the spatial unit enters the privacy sensitive processing path, the registration source is redirected to the adjacent non-privacy spatial unit, and texture mapping is prohibited. When the texture data is removed, if the texture identifiable component decreases while the geometric contour identifiable component is still higher than its component boundary, the sensitive flag is forced to continue to be retained, and the geometric record still adopts the fusion expression of removing the source frame association.

[0060] Please refer to Figure 3 , Figure 3 This describes the formation process of the corresponding privacy-identifiable metrics. The left-hand bar chart uses texture identification, contour identification, and persistent association as the horizontal axis, and weights and contributions as the vertical axis. The dotted bar chart reflects the proportion of the three types of privacy evidence in the combined judgment. The average contribution in the current session state records. The right-hand broken line section uses the state record number as the horizontal axis and... With the vertical axis as the pivot, the black broken line represents the change in privacy identifiability as a function of multi-view projection, sensitive point set merging, and cross-frame correlation updates; the gray dashed line represents the sensitive boundary. .from Figure 3 It can be seen from this that Overall, it remains However, the curve shows an upward and fluctuating trend as the state progresses, indicating that even after the texture data is removed, the geometric contour and persistent correlation still have a lasting impact on privacy judgment; if subsequent observations make... or The contribution continues to increase, and even if the color texture is no longer preserved, the spatial unit may still maintain a sensitive processing path.

[0061] In this embodiment, the determination of the spatial unit covered by the sensitive point set adopts parallel recording of surface overlap and observation ray overlap. Sensitive points with actual depth are written into the surface overlap record according to their spatial position, and areas lacking depth are written into the ray overlap record according to the spatial range traversed by the observation ray. When a spatial unit has both types of records, the sensitive geometry is preferentially limited according to the surface overlap range. When there is only a ray overlap record, the unconfirmed state is maintained and the corresponding observation texture is restricted from entering the registration. After the depth is supplemented by the subsequent viewpoint, the ray overlap is converted into surface overlap, so that the edge of the two-dimensional sensitive area and the missing depth position can continuously participate in the three-dimensional privacy state update.

[0062] In a preferred embodiment, the establishment of the joint processing relationship includes: pairing the reconstruction adequacy state and the privacy identifiability state according to the spatial index of the three-dimensional spatial unit to generate a processing identifier that includes the observation preservation relationship and the registration source and fusion expression associated with each other; when the processing identifier corresponds to a privacy-sensitive spatial unit, pointing the registration source to the adjacent non-privacy spatial unit and limiting the fusion expression to the geometric expression that decouples the source frame; when the processing identifier corresponds to a non-privacy spatial unit, determining the observation preservation relationship according to the reconstruction adequacy state.

[0063] The joint processing relationship is implemented through a state pairing table. Each record in the state pairing table uses the spatial index and state version as keys. The reconstruction sufficiency field references the overriding association and integrity markers, the privacy identifiability field references the sensitive point set and attribute records, and the processing identifier field references the observation retention table, the registration source table, and the fusion expression table. The edge computing processor reads two states with the same version under the same spatial index and generates a processing identifier. If the versions of the two states are different, the spatial unit corresponding to the earlier state is first written into the state update queue, and the processing identifier is generated after the update is completed. This avoids using reconstruction information and privacy information from different times to form processing decisions.

[0064] The observation retention table does not directly copy the scan data, but records the still accessible anonymous references to observations, the data retention categories, and the release conditions. Incompletely reconstructed non-privacy spatial units retain observation references that can add independent perspectives or supplement open boundaries. Completely reconstructed non-privacy spatial units only retain the fused surface and its observation support count. Privacy-sensitive spatial units temporarily store necessary depth references and prohibit color texture references from entering subsequent tasks before the geometric fusion data is formed. Depth reference release is triggered after the geometric fusion data is written. Spatial units that are to be confirmed due to occlusion only retain the observation rays and anonymous status, and do not retain the sensitive pixel content corresponding to the rays.

[0065] The registration source table records the geometric boundaries, surface patches, and their respective non-privacy space indexes that are allowed to be used in the current spatial cell. The fusion representation table records the geometric fusion data or non-reconstructable markers of the current spatial cell using the original non-sensitive observations, stable fusion surfaces, and decoupled source frames. The two are associated through the same processing identifier. When the privacy state changes from non-sensitive to sensitive, the edge computing processor discards the entries in the original registration source table that point to the texture features inside the spatial cell, queries adjacent non-privacy spatial cells along the shared boundary, and reconstructs the registration source table. At the same time, the fusion representation is switched to geometric records. The change in reconstruction state only adjusts the observation retention table and task priority relationship, and does not remove privacy processing restrictions.

[0066] Please refer to Figure 4 , Figure 4 The horizontal axis represents the state record number, and the vertical axis represents the boundary distance. The dashed zero-value reference line is used to distinguish whether the state value has crossed the corresponding boundary. The solid black line represents the reconstructed boundary distance. ,Right now relative to sufficient boundary The degree of slack; the gray dotted line represents the privacy and security distance. Privacy boundaries Compared to A safe space is still preserved. Figure 4 middle A negative value in the initial segment indicates that while early spatial units have observational support, at least one of the following—coverage, continuity, normal consistency, or pose correspondence—is insufficient to support adequate reconstruction; as state records increase, As the line gradually approaches and crosses the zero line, it indicates that the local reconstruction conditions have changed from insufficient to satisfied. Conversely, It remained positive throughout, but the safety distance narrowed somewhat in the latter part, indicating that although the privacy identifiability value did not exceed... However, it is still necessary to maintain texture mapping constraints and source frame association unchecks. Figure 4 Placing reconstruction sufficiency and privacy identifiability within the same criterion explains why collaborative processing relationships cannot be judged solely on... or Single-item decision: Spatial units are only suitable for regular fusion when the reconstruction margin meets the requirements and the privacy security distance has not been exhausted; otherwise, they should be transferred to supplementary observation, sensitive geometric transformation or verification processing.

[0067] In this embodiment, the state pairing table adopts a conservative conversion logic for conflict states. If the reconstructed state shows sufficient surfaces but the privacy state is still pending confirmation, the existing geometric surfaces are maintained and texture mapping is paused. If the privacy state shows sensitive information but the reconstructed state lacks viewpoint support, local constraints are established using adjacent non-sensitive structures and open boundary markers are retained. If a registration conflict occurs in the reconstructed state and the conflicting branch overlaps with the sensitive point set, the texture description information within the conflicting region is stopped, and the pose is re-solved using only the geometric correspondence outside the overlapping range. If the sensitive range shrinks after the state update, the original sensitive spatial unit still needs to pass the state version check and source frame index check before it can be transferred to the non-sensitive fusion path.

[0068] In another preferred embodiment, the establishment of a stable geometry includes: extracting the geometric boundary of cross-frame repeated observations along the shared boundary between the privacy-sensitive spatial cell and the adjacent non-privacy spatial cell, and selecting surface patches that are topologically adjacent to the geometric boundary from the adjacent non-privacy spatial cells; using the geometric boundary and the surface patches to form a local registration constraint, and excluding texture descriptors that fall into the privacy-sensitive spatial cell from the registration correspondence set.

[0069] The extraction of geometric boundaries is based on the depth and surface orientation changes near the shared boundary. The edge computing processor transforms the boundary candidates from multiple registered observations to a unified three-dimensional coordinate system. Candidates that are spatially adjacent, have compatible tangent directions, and are located at the outer edge of the same sensitive range are merged into boundary trajectories. Each local segment in the boundary trajectory stores the number of observation sources and their location discreteness. Only segments that appear repeatedly in multiple independent views and remain spatially adjacent to the outer edge of the sensitive point set enter the stable geometric structure. Boundary candidates that appear in a single frame or move with the view are registered as occlusion boundaries and do not participate in local registration constraints.

[0070] Surface patches that maintain topological adjacency with geometric boundaries are obtained through surface connectivity search outside the boundary trajectory. The search starts from the boundary point in the non-privacy space cell and extends along the local mesh or voxel adjacency relationship to the surface position that is continuous in the normal direction. The extension stops when encountering privacy-sensitive space indexes, open boundaries, or geometric branches. The obtained surface patches must be provided with consistent spatial support by at least two observation directions, and their connection order with the geometric boundary is saved in the state record area. When the local surface is resampled or the space cell is repartitioned, the adjacency relationship is restored according to the connection order, without relying on texture features that may have been deleted.

[0071] Local registration constraints are formed by the point-to-line correspondence of the boundary trajectory and the point-to-face correspondence of the surface patch. The edge computing processor matches the non-sensitive boundary points in the observation to be registered to the boundary trajectory and the non-sensitive surface points to the topologically adjacent surface patches. Based on the spatial distance and surface orientation differences of the corresponding points, the local pose correction is constructed. The rotation and translation relationships of the observation to be registered are iteratively updated. When multiple feasible correspondences occur, the corresponding branches that are consistent with the connection order of the shared boundary are retained first. During the iteration process, each update reprojects the privacy-sensitive region and deletes the texture descriptors that fall into the sensitive spatial unit until the pose change and geometric correspondence residuals enter the local reference distribution.

[0072] The exclusion of texture descriptors is achieved through dual checks of two-dimensional sensitive regions and three-dimensional sensitive spatial indexes. When extracting key points of image texture, the edge computing processor deletes candidates that fall into two-dimensional sensitive regions. After establishing cross-frame correspondence, the remaining key points are back-projected to a unified three-dimensional coordinate system, and then correspondences that fall into privacy-sensitive spatial units or the range of observation rays to be confirmed are deleted. Key points near the boundary that have both sensitive and non-sensitive projections do not enter the registration correspondence set. Geometric boundaries and surface patches are corresponded using spatial positions and surface orientations without color components, so that local pose solving does not depend on image content that may represent identity.

[0073] In a further embodiment, after the local registration constraint is completed, the depth points in the privacy-sensitive spatial unit are converted to a unified three-dimensional coordinate system, and geometric records containing only spatial position and surface orientation are accumulated according to the three-dimensional spatial unit; after the geometric records are written into the geometric fusion data, the index relationship of accessing the source frame and the texture region of the source frame from the geometric records is deleted, and the written geometric fusion data is used as the subsequent fusion input of the privacy-sensitive spatial unit.

[0074] The geometric record is saved by combining the internal coordinates of the spatial unit with the unified three-dimensional coordinate transformation relationship. The spatial position is written to the local position record with the origin of the spatial unit as a reference. The surface direction is recorded according to the unified orientation writing direction. The observation support status record is recorded to see whether multiple anonymous observations form consistent support for the same local surface, but the source frame address corresponding to the anonymous observation is not recorded. Before writing, the edge computing processor checks whether the sensitive depth point is on the visible surface, whether it is continuous with the adjacent non-sensitive boundary, and whether it belongs to the registration conflict branch. Depth points that meet the first two conditions and do not belong to the conflict branch are entered into the geometric record, and the remaining depth points remain in the pending confirmation state.

[0075] The writing of geometric fusion data adopts a processing method that separates the temporary recording area and the formal fusion area. The transformed geometric records first enter the temporary recording area and are clustered according to spatial location. The position records and orientation records within the same surface cluster are written to the formal fusion area after cross-observation consistency checks. After the formal fusion area returns the write confirmation flag, the index manager finds the source frame buffer index and texture region index along the anonymous observation references of the temporary records and unlinks them. Then, the anonymous observation references in the temporary records are cleared. If the writing process is not completed, the temporary links are maintained and the unconfirmed records in the formal fusion area are rolled back to avoid the intermediate state where the source frame links have been deleted but the geometric surface has not yet been formed.

[0076] The geometric fusion data in the formal fusion zone serves as the only data source for sensitive spatial units that can be included in subsequent fusion tasks. Subsequent reconstruction sufficiency updates read surface position, orientation, and observation support status from the geometric fusion data. Privacy and identifiability updates read the sensitive range from the spatial unit status without reading the source image. If a new observation covers the same sensitive spatial unit again, the new depth point is merged with the data in the formal fusion zone after local registration and geometric record conversion. New color textures are not mapped using texture atlases. After merging, new source frame links are processed according to the same write confirmation and index de-indexing process.

[0077] After the geometric fusion data is formally written into the formal fusion area and the source frame index relationship is removed, the edge computing processor marks the corresponding original sensitive source frame as cleanable and performs sensitive original data erasure according to the local cache management strategy. The erased content includes the color texture pixels, original pixel coordinates, source frame identifier, and related information that can be traced back to the acquisition source in the source frame. For privacy-sensitive spatial units whose geometric contour recognition attributes reach the threshold, contour suppression is further performed on their geometric fusion records to weaken the local geometric details that can identify the identity and output geometric records that only retain the general shape and do not retain identity features. The original non-sensitive source frames also enter the cache cleanup queue after the corresponding spatial unit is fully reconstructed and the repeated observations are de-associated.

[0078] This invention also includes a desensitization verification process before the model is transmitted externally: when the 3D reconstruction model needs to be transmitted to the cloud or external devices, the edge computing processor reads the privacy identifiability status of each 3D spatial unit and traverses all privacy-sensitive spatial units; if the identifiability attribute of the geometric contour of the spatial unit is higher than the threshold, contour suppression desensitization processing is performed on the geometric mesh of the unit to remove fine geometric contour features that can be used for identity recognition; only after the desensitization verification is completed is the model data allowed to be output and transmitted externally, blocking the privacy leakage channel brought by geometric fusion data; if external transmission is not performed, only local cache cleanup is performed.

[0079] Thus, privacy protection covers the entire process from data acquisition and input, intermediate processing, local model storage, to external transmission and output: raw sensitive data only exists temporarily during geometric transformation and is erased after geometric record generation and index removal are completed; intermediate fused data is processed hierarchically according to the identifiable attributes of geometric contours, and contour suppression is performed on high-risk geometric contours, without retaining identity textures, identifiable geometric contours, or associations with the data acquisition source; finally, the output model removes texture mapping from sensitive areas and performs additional geometric contour desensitization verification before external transmission, achieving the goal of minimizing the retention of sensitive information throughout the entire process, thus reducing the continuous retention of identity textures, identifiable geometric contours, data acquisition source associations, and sensitive observations in the derived model.

[0080] In a preferred embodiment, partitioned surface fusion includes: establishing a reconstruction task chain associated by three-dimensional spatial units based on processing identifiers; for non-privacy spatial units with insufficient reconstruction, incorporating new observations into the registration task and then performing surface fusion; for non-privacy spatial units with sufficient reconstruction, converting registered observations into fused surfaces and removing the task association of duplicate observations; for privacy-sensitive spatial units, writing geometric fusion data into a fusion task that does not contain texture mapping relationships; and writing the output of each fusion task back to the corresponding three-dimensional spatial unit.

[0081] The reconstruction task chain consists of task nodes and spatial dependency edges. Task nodes store the corresponding spatial index, input data reference, processing type, state version, and output position. Spatial dependency edges indicate that the shared boundary of adjacent spatial units can only be performed after the previous fusion result is formed. The edge computing processor first generates local registration nodes, geometric transformation nodes, surface fusion nodes, and state write-back nodes according to the processing identifier, and then connects the task nodes according to the input-output relationship. When the state version of the same spatial unit changes, the old version of the task node that has not yet been executed is discarded. The output of the completed task with unchanged input continues to be used as the input of the subsequent node, and the cleaned source frame is not reread.

[0082] For non-privacy spatial units with insufficient reconstruction, the task chain queries the observation support map to find new observations that can be supplemented by independent perspectives or open boundaries. It establishes a local geometric correspondence between the new observations and existing stable surfaces. After completing the pose correction, it writes the new point cloud into the surface fusion node. The surface fusion node updates the local distance record according to the consistency of spatial position and orientation. The source state of the surface branches formed by the new observations is retained and waits for confirmation from adjacent perspectives. When the new surface crosses the boundary of the spatial unit, a boundary update task is generated and adjacent spatial units are notified to re-check the shared boundary. Local surfaces not covered by the new projection maintain the original fusion record.

[0083] For fully reconstructed non-privacy spatial units, the task chain no longer retains the original point set of repeated observations. Instead, it converts the registered and consistent observations into stable fusion surfaces. The stable fusion surfaces store spatial location, orientation, color fusion references, and observation support counts. After the conversion is confirmed, the association between repeated observations and subsequent tasks is removed. For privacy-sensitive spatial units, the surface fusion node only receives geometric fusion data from the formal fusion region and writes a prohibited texture mapping flag in the task metadata. The boundary stitching node can read the spatial location and orientation of adjacent non-privacy surfaces, but must not extend the adjacent texture coordinates to sensitive mesh patches.

[0084] The fusion task employs a spatial cell-level write-back lock to ensure that the state and surface are updated synchronously. Before the task is executed, the current state version of the spatial cell is read and saved as the input version. After the execution is completed, the state version is read again. If the two versions are consistent, the local fused surface, open boundary records, and observation support states are written to the corresponding spatial cell and the surface version is incremented. If the two versions are inconsistent, the write-back is abandoned and the task node is reconstructed according to the new processing identifier. After the write-back is completed, the input data reference is released, non-sensitive duplicate observations are placed in the release queue, and sensitive geometric temporary records are placed in the index cleanup queue according to the confirmation flag returned by the formal fusion area.

[0085] In a further embodiment, after the output of each fusion task is written back to the corresponding 3D spatial unit, the reconstruction adequacy state is updated according to the coverage relationship of the written-back surface to the observation viewpoint, and the privacy identifiability state is updated according to the spatial overlap relationship between the written-back surface and the privacy-sensitive spatial unit; only the 3D spatial units whose reconstruction adequacy state or privacy identifiability state has changed are rewritten into the reconstruction task chain; after the reconstruction task chain is completed, the fusion surfaces of each 3D spatial unit are stitched together to form a 3D reconstruction model, and the mapping relationship between the privacy-sensitive spatial unit and the texture atlas is removed.

[0086] The write-back of the surface coverage relationship to the observation viewpoint is updated through reverse visibility checks. The end-side computational processor traces back from the local surface position along the registered observation direction to the corresponding observation plane, checks whether the ray is blocked by other fused surfaces, and registers the surface range that is not blocked and falls into the effective depth region as supported coverage. The surface originally formed by the unconfirmed point is converted into a continuous surface record after obtaining multiple observation supports. The range that was originally judged as covered but is occluded in the new pose has its corresponding coverage edge removed. The updated coverage association, continuous surface record and registration consistency record re-form the reconstruction sufficiency state.

[0087] The spatial overlap relationship between the write-back surface and the privacy-sensitive spatial unit is checked jointly by the spatial index and the surface enclosure range. When a local surface falls into the sensitive spatial index, it inherits the corresponding privacy status flag. When a local surface crosses the boundary between sensitive and non-sensitive spaces, it is divided into grid pieces along the outer edge of the sensitive point set and the mapping permission is saved separately. When the observed ray intersects with the write-back surface, the surface near the intersection point is transferred to the privacy status pending review. After subsequent cross-projection confirms that the intersection point belongs to a sensitive surface, a formal sensitivity flag is written. When it is confirmed that it does not belong to a sensitive surface, the ray overlap record is canceled, but the status version change record is retained.

[0088] Before rewriting the 3D spatial unit into the reconstruction task chain, the state category, formation basis, and data references before and after the update are compared. If the category remains unchanged and only the observation support count is increased, the state record is updated without creating a new registration task. If the state category changes, the open boundary range changes, or the privacy mapping permission changes, a new task node is generated, and the directly adjacent spatial units are notified to check the shared boundary along the support edge of the observation support map. The notification does not continue to cross spatial units whose state has not changed, thereby keeping the loop processing range within the local area affected by this write-back. The task chain ends when there are no unexecuted nodes, pending write-back confirmations, or unresolved version conflicts.

[0089] During local surface fusion stitching, the edge computing processor reads the boundary vertices and surface orientations of adjacent spatial units according to the spatial index. It merges boundary vertices that have the same spatial position and continuous orientation, retains boundaries that are close in position but belong to different surface layers as independent mesh edges, and keeps the position where open boundaries meet non-reconstructable regions as closed marker edges without performing surface inference. When sensitive mesh patches and non-sensitive mesh patches share geometric boundaries, only spatial vertices are merged, without merging texture coordinates or source image references. The resulting overall mesh inherits the privacy mapping permission of each local surface.

[0090] Texture atlas generation only traverses non-privacy meshes with texture mapping permissions. From color observations that are still accessible, it selects regions with minimal viewpoint occlusion and small projection distortion and writes them into the atlas. Atlas entries store the mapping from non-sensitive meshes to textured regions. **A "no texture" flag is written to the corresponding location of privacy-sensitive meshes.** Simultaneously, for privacy-sensitive meshes with geometrically identifiable attributes exceeding a threshold, vertex smoothing and local curvature suppression are performed to eliminate identity-related geometric details. Double verification is performed during model export / external transmission checks: both verifying the absence of texture coordinates and verifying that the geometric contours of sensitive regions have undergone contour suppression and desensitization. If the verification fails, model output is prohibited. **If a previous model version established a texture mapping for the same spatial location, after the status is updated to sensitive, the old atlas entries and mesh mappings are deleted based on the spatial index, and then the atlas index for the current model version is generated.**

[0091] In a preferred embodiment, when the scanning process is interrupted due to device pause or task switching, the edge computing processor saves the state version of each spatial unit, the observation support map, the formal fusion region, and the unfinished task nodes, but does not save the sensitive source frame links that have completed geometric transformation. During the recovery process, the processor first checks the confirmation flag and the state matching table version written to the formal fusion region, rolls back the unconfirmed fusion records and cancels the task nodes with inconsistent input versions, and then reconstructs the task chain from the still accessible non-sensitive observations, geometric fusion data, and shared boundary relationships, so that the recovery process continues to comply with the registration source constraints and fusion expression constraints formed before the interruption.

[0092] In another preferred embodiment, when the scanned object contains a local motion region, the edge computing processor separates the motion surface from the static reference surface through the geometric correspondence residual in continuous observation. The static non-privacy surface is used to establish a unified three-dimensional coordinate system and a stable geometric structure. The motion region establishes an independent spatial state version according to the acquisition time period. The motion surface that belongs to the privacy-sensitive range is still converted into geometric fusion data that is decoupled from the source frame. The motion surfaces of different acquisition time periods are not directly accumulated into the same fusion surface. When the model is output, a state version is selected or the time-layered recording is maintained according to the scanning task. The texture atlas mapping continues to be restricted by the privacy flag of each state version.

Claims

1. A method for end-side adaptive reconstruction and privacy protection of a three-dimensional scanning base station, characterized in that, include: Acquire multi-view scanning data from a 3D scanning base station, perform global initial registration of the scanning data to a unified 3D coordinate system, and divide the data into 3D spatial units; The reconstruction sufficiency state of each three-dimensional spatial unit is determined based on the correlation of multi-view observations, and the privacy identifiability state is determined based on the projection of the image sensitive area onto the unified three-dimensional coordinate system. The privacy identifiability state distinguishes between texture identifiability attributes and geometric contour identifiability attributes, and establishes a joint processing relationship between the two states. Based on the joint processing relationship, the observation retention method and the registration feature sources and fusion expressions associated with each other are allocated. For privacy-sensitive spatial units, registration constraints are established based on the stable geometric structures in adjacent non-privacy spatial units, and the registered sensitive observations are converted into geometric fusion data that is decoupled from the source frames. When the identifiable attribute of the geometric contour reaches a preset recognition threshold, contour suppression processing is performed on the geometric fusion data of the corresponding privacy-sensitive spatial unit to weaken the fine geometric contours that can be used for identity recognition. Based on the joint processing relationship, partition surface fusion is performed to update the two states and output the three-dimensional reconstruction model. When the three-dimensional reconstruction model needs to be transmitted externally, external transmission desensitization verification is performed to verify the geometric contour identifiable attributes of each privacy-sensitive spatial unit. For spatial units whose geometric contour identifiable attributes exceed the threshold, contour suppression desensitization is performed again. Only after the verification is passed can the model data be output externally.

2. The end-side adaptive reconstruction and privacy protection method of a three-dimensional scanning base station according to claim 1, characterized in that, The determination of the reconstruction adequacy state includes: projecting the effective depth region of each observation frame onto the three-dimensional spatial unit according to the scanning pose, and establishing the coverage association between the three-dimensional spatial unit and the observation viewpoint; Effective observation distribution is obtained within each three-dimensional spatial unit, and integrity markers are formed based on the surface normal consistency and registration correspondence of adjacent observations; The coverage association is combined with the integrity flag to form the reconstruction adequacy state.

3. The end-side adaptive reconstruction and privacy protection method of a three-dimensional scanning base station according to claim 1, characterized in that, The determination of the privacy identifiability status includes: obtaining the privacy-sensitive region from the color image of the multi-view scan data through semantic segmentation, and back-projecting the privacy-sensitive region onto the unified three-dimensional coordinate system using the depth map and scan pose synchronized with the color image; Back-projection points that overlap across frame projections are merged into a sensitive point set and a continuous association is established. Texture association attributes and geometric contour association attributes are recorded respectively. The geometric contour association attributes characterize the outer boundary shape, local curvature, and cross-view surface size consistency of the sensitive point set. When the geometric contour association attributes exceed a preset threshold, it is determined that the spatial unit has a geometric contour identity recognition risk. The privacy identifiability state is formed according to the three-dimensional spatial units covered by the sensitive point set; if the geometric contour association attribute exceeds the preset threshold, the privacy identifiability state of the spatial unit remains sensitive and triggers geometric contour suppression processing, which does not allow the retention of fine geometric contours that can identify the identity.

4. The end-side adaptive reconstruction and privacy protection method of a three-dimensional scanning base station according to claim 1, characterized in that, The establishment of the joint processing relationship includes: pairing the reconstruction adequacy state with the privacy identifiability state according to the spatial index of the three-dimensional spatial unit, and generating a processing identifier that includes the observation preservation relationship and the registration source and fusion expression associated with each other; When the processing identifier corresponds to the privacy-sensitive spatial unit, the registration source is pointed to the adjacent non-privacy spatial unit and the fusion expression is limited to the geometric expression of the decoupled source frame; When the processing identifier corresponds to a non-privacy space unit, the observation retention relationship is determined based on the reconstruction sufficiency state.

5. The end-side adaptive reconstruction and privacy protection method of a three-dimensional scanning base station according to claim 2, characterized in that, The establishment of the coverage association includes: establishing an observation support map by projecting the outlines of each observation frame in the unified three-dimensional coordinate system and the three-dimensional spatial unit, and connecting three-dimensional spatial units with shared boundaries and covered by the same observation viewpoint as support edges; When a new observation frame is added, the nodes that intersect with the new projection contour and the corresponding support edges are updated, and the integrity marker is reconstructed based on the updated observation support map.

6. The end-side adaptive reconstruction and privacy protection method of a three-dimensional scanning base station according to claim 3, characterized in that, The establishment of the sensitive point set includes: extracting the effective depth points and depth discontinuity edges intersecting with the region contour within the privacy-sensitive region, back-projecting the effective depth points into a core sensitive point set, and registering the region to be confirmed along the observation ray corresponding to the depth missing position. The area to be confirmed is cross-projected using subsequent observation perspectives, and the supplementary point set obtained from the cross-projection is combined with the core sensitive point set to form the sensitive point set.

7. The end-side adaptive reconstruction and privacy protection method of a three-dimensional scanning base station according to claim 4, characterized in that, The establishment of the stable geometry includes: extracting the geometric boundary of cross-frame repeated observations along the shared boundary between the privacy-sensitive spatial cell and the adjacent non-privacy spatial cell, and selecting surface patches that maintain topological adjacency with the geometric boundary from the adjacent non-privacy spatial cells; Local registration constraints are constructed using geometric boundaries and surface patches, and texture descriptors falling into the privacy-sensitive spatial unit are excluded from the registration correspondence set.

8. The end-side adaptive reconstruction and privacy protection method for a three-dimensional scanning base station according to claim 7, characterized in that, The partitioned surface fusion includes: establishing a reconstruction task chain associated with three-dimensional spatial units based on the processing identifier; For non-privacy space units that are not sufficiently reconstructed, new observations are incorporated into the local incremental registration task based on the global initial registration and then surface fusion is performed. For fully reconstructed non-privacy spatial units, registered observations are converted into fused surfaces and task associations for duplicate observations are removed; For the privacy-sensitive spatial unit, the geometric fusion data is written into a fusion task that does not contain texture mapping relationships; Write the output of each fusion task back to the corresponding 3D spatial unit.

9. The end-side adaptive reconstruction and privacy protection method for a three-dimensional scanning base station according to claim 8, characterized in that, After the local registration constraint is completed, the depth points in the privacy-sensitive spatial unit are transformed to the unified three-dimensional coordinate system, and the geometric records containing only spatial position and surface orientation are accumulated according to the three-dimensional spatial unit. After writing the geometric record into the geometric fusion data, the index relationship between the source frame and the texture region of the source frame accessed from the geometric record is deleted, and controlled erasure is performed on the privacy-sensitive original data in the source frame; if the geometric contour identifiable attribute of the privacy-sensitive spatial unit reaches the preset recognition threshold, contour suppression processing is performed on the geometric record to weaken the local curvature and contour details that can be used for identity recognition, and desensitized geometric fusion data is obtained. The written geometric fusion data is used as the subsequent fusion input for the privacy-sensitive spatial unit.

10. The end-side adaptive reconstruction and privacy protection method for a three-dimensional scanning base station according to claim 8, characterized in that, After the output of each fusion task is written back to the corresponding three-dimensional spatial unit, the reconstruction sufficiency state is updated according to the coverage relationship of the written-back surface to the observation view, and the privacy identifiability state is updated according to the spatial overlap relationship between the written-back surface and the privacy-sensitive spatial unit. Only the three-dimensional spatial units whose reconstruction adequacy state or privacy identifiability state has changed are rewritten into the reconstruction task chain; After the reconstruction task chain is completed, the fused surfaces of each three-dimensional spatial unit are stitched together to form a three-dimensional reconstruction model, and the mapping relationship between the privacy-sensitive spatial unit and the texture atlas is removed; for privacy-sensitive spatial units whose geometric contour recognition attributes reach the threshold, mesh contour suppression is performed; and cache cleanup is performed on the original scan data of the corresponding privacy-sensitive area in local storage. When a 3D reconstruction model needs to be transmitted externally, a desensitization verification is performed to check the desensitization status of the geometric contours of all privacy-sensitive spatial units. Only after the verification is passed can the model be output externally.