Method for generating spatial smart object representation oriented to occluded areas

Abstract

The application provides a spatial intelligent object representation method for occlusion area traceable generation. Spatial intelligent perception feature data recording scene geometry and dynamic evolution trend are generated based on spatial environment sensing data stream of the scene, and deep gradient component and optical flow displacement vector are extracted therefrom to determine the occlusion area in the scene and generate a corresponding occlusion mask graph; the occlusion mask graph is used to limit a target reconstruction area in the spatial intelligent perception feature data, and context features around the target reconstruction area are extracted to perform multi-scale geometric completion and semantic migration filling processing to determine object geometric features and semantic consistency features; cross-modal alignment processing and physical rationality correction processing are performed on the object geometric features and semantic consistency features to generate corrected object attribute features; scene uncertainty modeling evaluation is performed based on the corrected object attribute features to generate spatial intelligent object representation for the occlusion area traceable.

Description

Technical Field

[0001] This application relates to the field of spatial intelligence and artificial intelligence computing, and more specifically, to a method for representing spatial intelligent objects that can be traced and generated for occluded regions. Background Technology

[0002] With the rapid development of spatial intelligent agent navigation, augmented reality (AR), and digital twin technologies, systems are placing extremely high demands on their ability to perceive and understand spatial objects in complex environments. Especially in dynamic interactive scenarios with occlusion, achieving reliable reconstruction and semantic completion of invisible areas is crucial for ensuring safe robot path planning and immersive virtual-real interaction.

[0003] In existing spatial object representation schemes, 3D reconstruction techniques based on depth estimation or Neural Radiation Field (NeRF) are commonly used. This approach first acquires point cloud or image data of the scene using multimodal sensors; then, it uses neural networks to geometrically represent the visible regions, and fills in missing areas using simple heuristic algorithms or generative adversarial networks; finally, it outputs the reconstructed scene results using voxel meshes or point cloud topology.

[0004] However, this traditional generative approach has significant technical drawbacks. Since the reconstruction of occluded regions often relies on implicit inference from the model, the generation process is similar to a "black box," lacking traceable generation logic. This makes it impossible for users to determine the source and reliability of the completed content. Furthermore, existing solutions often exhibit a disconnect between geometric shape and semantic attributes when handling cross-modal data alignment, and lack quantitative assessment of the uncertainty of the generated results. This leads to low physical plausibility of spatial object representations, making it difficult to meet the application requirements of highly reliable spatial intelligent systems. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides a spatial intelligent object representation method for traceable generation of occluded regions, thereby at least alleviating the aforementioned technical problems.

[0006] A spatial intelligent object representation method for traceable generation of occluded regions includes:

[0007] Step 1: Generate spatial intelligent perception feature data that records the scene's geometric structure and dynamic evolution trend based on the spatial environment sensing data stream for the scene, and extract the depth gradient component and optical flow displacement vector from it to determine the occlusion area in the scene and generate the corresponding occlusion mask map.

[0008] Step 2: Use the occlusion mask image to define the target reconstruction region in the spatial intelligent perception feature data, and extract the context features around the target reconstruction region to perform multi-scale geometric completion and semantic transfer filling processing to determine the object's geometric features and semantic consistency features.

[0009] Step 3: Perform cross-modal alignment and physical rationality correction on the geometric features and semantic consistency features of the object to generate corrected object attribute features;

[0010] Step 4: Perform scene uncertainty modeling and evaluation based on the corrected object attribute features to generate a traceable spatial intelligent object representation for occluded areas.

[0011] Optionally, generating the spatial intelligent perception feature data in step 1 specifically includes:

[0012] The color image component, depth information component, and point cloud data component contained in the spatial environment sensing data stream are obtained.

[0013] The color image components, depth information components, and point cloud data components are timestamped using a unified time reference to generate a time-synchronized sensor stream.

[0014] The temporal synchronous sensing stream is reprojected and mapped to a spatial coordinate system using a spatial calibration matrix to generate the spatial intelligent sensing feature data.

[0015] Optionally, generating the occlusion mask image in step 1 specifically includes:

[0016] Local depth gradient calculation is performed on the spatial intelligent sensing feature data to obtain the depth gradient components that characterize abrupt changes in spatial structure;

[0017] Perform pixel trajectory tracking processing between consecutive frames on the spatial intelligent perception feature data to obtain the optical flow displacement vector characterizing the dynamic evolution trend;

[0018] Based on the depth gradient components and the optical flow displacement vector, feature-deficient regions are identified, and semantic attribute labeling is performed on the feature-deficient regions to generate the occlusion mask map.

[0019] Optionally, step 2 includes performing the multi-scale geometric completion process, comprising:

[0020] Based on the occlusion mask map, the spatial boundary of the target reconstruction region determined in the spatial intelligent perception feature data is used to perform geometric extrapolation using a multi-scale voxel generation algorithm to obtain a rough geometric shell.

[0021] Fine-grained feature refinement is performed on the coarse geometric shell using surface curvature constraints to determine the object geometry that reflects the complete structure of the occluded region.

[0022] Optionally, the semantic transfer imputation process performed in step 2 specifically includes:

[0023] Analyze the feature distribution pattern of known semantic categories in the spatial intelligent perception feature data surrounding the target reconstruction area, and calculate the semantic correlation between the feature distribution pattern and the boundary of the space to be reconstructed;

[0024] Based on the semantic relevance, the corresponding contextual semantic weights are transferred to the geometric feature space to determine the semantic consistency features that are consistent with the surrounding environment attributes.

[0025] Optionally, step 2 further includes constructing a traceable generation chain, specifically including:

[0026] Real-time recording of input sensor source identifiers, model version operators, and intermediate feature mapping states during the execution of the multi-scale geometric completion and semantic transfer filling processes;

[0027] The input sensing source identifier, the model version operator, and the intermediate feature mapping state are encapsulated into an audit payload with logical dependencies to construct a traceable generation chain with a unique traceability identifier that runs through the perception layer to the representation layer.

[0028] Optionally, generating the corrected object attribute features in step 3 specifically includes:

[0029] The traceable generation chain is retrieved, coordinate information is extracted from the geometric features of the object, and cross-modal feature alignment is performed with the category attributes extracted from the semantic consistency features to eliminate the matching deviation between spatial location and semantic attributes;

[0030] The physical rationality correction process is performed on the aligned cross-modal features using preset gravity support constraints and object contact logic to generate the corrected object attribute features.

[0031] Optionally, step 4 involves performing scenario uncertainty modeling and evaluation, specifically including:

[0032] The modified object attribute features are subjected to multi-sample probability estimation using the Monte Carlo sampling algorithm to generate a confidence map representing the credibility of the prediction results.

[0033] Based on the confidence map, the uncertainty distribution of the occluded area is determined, and then the corrected object attribute features are associated and encapsulated with the uncertainty distribution to generate the traceable spatial intelligent object representation for the occluded area.

[0034] Optionally, the method further includes a dynamic update action:

[0035] Real-time monitoring of the residual values ​​between the completed content in the spatial intelligent object representation generated for the occluded area and the actual observation values ​​in the subsequently received spatial environment sensing data stream;

[0036] In response to the residual value exceeding a preset deviation threshold, the logic breakpoint where the deviation occurred in the traceable generation chain is retrieved, and incremental fine-tuning of the spatial intelligent object representation for the traceable occluded area is performed based on the actual observation value.

[0037] Optionally, the semantic transfer filling process is implemented by retrieving a pre-established spatial semantic knowledge graph, specifically including:

[0038] Retrieve prior models of object categories in the spatial semantic knowledge graph that match the semantic context and spatial attributes of the target reconstruction region, and construct global semantic topological constraints for semantic transfer filling based on the prior models of object categories;

[0039] Numerical correction is performed on the semantic relevance based on the global semantic topology constraints, and weight ratio correction is performed on the context semantic weight based on the global semantic topology constraints;

[0040] The known region semantic features that adapt to the target reconstruction region are determined based on the corrected semantic relevance and the corrected context semantic weight.

[0041] The semantic features of the known region are directionally transferred to the target reconstructed region to complete the semantic transfer filling process.

[0042] Technical advantages of the technical solution provided in this application

[0043] 1. Addressing the issues of fuzzy boundary recognition and lack of dynamic evolution patterns in occlusion detection.

[0044] This application generates spatial intelligent perception feature data recording geometric structure and dynamic trends based on spatial environment sensing data stream in step 1, and extracts depth gradient components and optical flow displacement vectors to generate an occlusion mask map. This step solves the technical defect of inaccurate occlusion region definition caused by traditional solutions relying solely on static images. Since depth gradients can reflect abrupt changes in spatial structure, and optical flow displacement can capture the motion continuity of dynamic objects, the combination of the two enables the generated occlusion mask map to accurately locate "perceptual voids" in the spatiotemporal dimension, providing highly physically relevant boundary guidance for subsequent reconstruction tasks and significantly improving the robustness of occlusion region identification.

[0045] 2. Addressing the issues of the missing content and untraceable generation logic.

[0046] This application utilizes an occlusion mask map in step 2 to define the target reconstruction region and extracts surrounding contextual features to perform multi-scale geometric completion and semantic transfer filling. This step solves the "discontinuity" phenomenon in traditional generation algorithms where the completed content is inconsistent with the surrounding environment. Compared to traditional blind content filling, this application uses the contextual features of known regions as source material and ensures the dual coherence of the generated content at both the geometric and semantic levels through multi-scale feature mapping. Furthermore, since the reconstruction action is based on explicit mask map guidance and contextual feature transfer, this deterministic processing chain provides underlying logical support for the traceability of the final result, effectively alleviating the problem of uninterpretable generated results in traditional schemes.

[0047] 3. Addressing the issues of cross-modal feature migration and generation results violating physical principles.

[0048] This application performs cross-modal alignment and physical plausibility correction on the geometric and semantic consistency features of the object in step 3 to generate corrected object attribute features. This step solves the technical bottleneck of spatial misalignment between "object shape" and "object attributes" in traditional solutions. Compared with traditional pure algorithm fitting that does not consider external constraints, this application uses physical correction mechanisms (such as gravity constraints and support surface detection) to intervene in the generated features a second time, ensuring that the corrected object attribute features are not only visually complete but also conform to objective interaction laws in physical logic, significantly improving the realism of spatial intelligent object representation.

[0049] 4. Addressing the issues of insufficient reliability of the results and lack of evaluation criteria.

[0050] This application performs scene uncertainty modeling and evaluation based on the modified object attribute features in step 4. This step addresses the technical deficiency of traditional solutions that only provide the result without giving confidence levels. By quantifying the uncertainty distribution of the generated results, the method in this application can clearly inform the backend system (such as robot navigation or AR rendering engine) which areas are highly reliable and which areas have potential deviations, thus providing key risk measurement indicators for system decision-making. Combined with the aforementioned traceability characteristics, the final generated traceable spatial intelligent object representation for occluded areas achieves a complete technical closed loop of "perception-reconstruction-verification-traceability". Attached Figure Description

[0051] Figure 1 This application provides a method for representing spatial intelligent objects that can be traced and generated in an occluded area.

[0052] Figure 2 This application provides a spatial intelligent object representation device that can be traced and generated for occlusion areas.

[0053] Figure 3 This is an electronic device according to an embodiment of the present application. Detailed Implementation

[0054] like Figure 1 As shown in the figure, this application provides a method for representing spatial intelligent objects that can be traced and generated in an occluded region, comprising:

[0055] Step 1: Generate spatial intelligent perception feature data that records the scene's geometric structure and dynamic evolution trend based on the spatial environment sensing data stream for the scene, and extract the depth gradient component and optical flow displacement vector from it to determine the occlusion area in the scene and generate the corresponding occlusion mask map.

[0056] Step 2: Use the occlusion mask image to define the target reconstruction region in the spatial intelligent perception feature data, and extract the context features around the target reconstruction region to perform multi-scale geometric completion and semantic transfer filling processing to determine the object's geometric features and semantic consistency features.

[0057] Step 3: Perform cross-modal alignment and physical rationality correction on the geometric features and semantic consistency features of the object to generate corrected object attribute features;

[0058] Step 4: Perform scene uncertainty modeling and evaluation based on the corrected object attribute features to generate a traceable spatial intelligent object representation for occluded areas.

[0059] Optionally, generating the spatial intelligent perception feature data in step 1 specifically includes:

[0060] Step 11: Obtain the color image component, depth information component, and point cloud data component contained in the spatial environment sensing data stream;

[0061] Step 12: Perform timestamp alignment processing on the color image component, the depth information component, and the point cloud data component using a unified time reference to generate a time-synchronized sensor stream.

[0062] Step 13: Perform reprojection and spatial coordinate system mapping processing on the temporal synchronous sensing stream using a spatial calibration matrix to generate the spatial intelligent sensing feature data.

[0063] Preferably, the specific implementation process of obtaining the color image component, depth information component and point cloud data component contained in the spatial environment sensing data stream in step 11 is as follows: first, the spatial environment sensing data stream is decomposed into components and effective information is extracted to obtain the color image component, depth information component and point cloud data component contained in the spatial environment sensing data stream. The spatial environment sensing data stream originates from real-time environmental data collected by multi-source sensing units deployed on mobile carriers, which may include indoor mobile robots, augmented reality vision terminals, and other devices with spatial perception capabilities. The color image component, acquired by a color vision camera, contains texture, edge, color gradient, and semantic contour information of objects within the scene, providing a texture and semantic basis for subsequent scene semantic recognition. The depth information component, acquired by a structured light depth sensor, contains information on the straight-line distance between each pixel and the sensing unit, local surface normals, and depth continuity, providing a close-range dense distance basis for subsequent spatial structure abrupt change recognition. The point cloud data component, acquired by a 3D LiDAR, contains the 3D coordinates, reflection intensity, and point cloud density information of spatial sampling points within the scene, providing a mid-to-long-range sparse spatial structure basis for subsequent spatial structure representation. The color image component, depth information component, and point cloud data component complement each other, avoiding the problems of missing close-range details or distortion of long-range structures that exist in single-modal data. This provides a multi-dimensional basic data source for subsequent time synchronization and spatial registration processing, and also provides multi-dimensional basic data support for the accurate identification of occluded areas.

[0064] Preferably, in step 12, when performing timestamp alignment processing on the color image component, the depth information component, and the point cloud data component using a unified time reference to generate a time-synchronized sensor stream, firstly, a common clock reference signal is configured for all sensing units participating in environmental data acquisition, establishing a globally unified time reference and eliminating frequency deviations and cumulative timekeeping errors of the clocks of different sensing units; subsequently, addressing the sampling frequency differences between the color vision camera, structured light depth sensor, and 3D LiDAR, a sliding window time interpolation algorithm is used. Using the timestamp closest to the current processing node in the acquisition time sequence as a reference, the acquisition frame sequences of the color image component, the depth information component, and the point cloud data component are interpolated and corrected in the time dimension, ensuring a one-to-one matching relationship between the frame sequences of each component in the time dimension, with the time deviation between frames controlled within the millisecond range (e.g., 10). (Within milliseconds); Simultaneously, combining the carrier attitude change data collected by the inertial measurement unit, the time dimension compensation for attitude offset is performed on each component frame sequence collected under dynamic motion conditions to eliminate inter-frame temporal misalignment caused by rapid carrier movement; Finally, the validity of each component frame sequence after interpolation correction and attitude compensation is verified in the time dimension, invalid frames with time deviations exceeding the preset range are removed, and frame sequences with high time matching are retained and integrated to generate a time-synchronized sensor stream. The time-synchronized sensor stream eliminates the time misalignment problem between different modal data, ensuring that during subsequent spatial coordinate system mapping processing, each component data at the same time node corresponds to the same spatial state in the scene, avoiding spatial position matching deviation caused by time asynchrony, and providing a time-consistent data source for subsequent unified mapping of the spatial coordinate system.

[0065] Preferably, in step 13, the spatial calibration matrix is ​​used to perform reprojection and spatial coordinate system mapping processing on the temporal synchronous sensor stream to generate spatial intelligent perception feature data. This includes: First, retrieving the spatial calibration matrix obtained in advance through calibration experiments. The spatial calibration matrix records the rotation and translation parameters between the coordinate systems of the color vision camera, structured light depth sensor, and 3D LiDAR and the preset world coordinate system. The rows and columns of the spatial calibration matrix correspond to the coordinate axis transformation relationship of different sensing unit coordinate systems, and the elements at the intersection of rows and columns correspond to the rotation and translation amounts between different coordinate axes, used to achieve linear transformation between different coordinate systems. Subsequently, based on the rotation and translation parameters in the spatial calibration matrix, reprojection operations are performed on the pixel coordinate system corresponding to the color image component, the camera coordinate system corresponding to the depth information component, and the radar coordinate system corresponding to the point cloud data component in the temporal synchronous sensor stream, respectively, mapping the independent coordinate systems corresponding to the three components to the preset world coordinate system. At the same time, combined with the real-time attitude data collected by the inertial measurement unit, the coordinate system offset generated during the carrier movement is corrected in real time to ensure that the mapped coordinate information corresponds to the actual spatial position in the scene. Next, spatial dimension matching and verification are performed on the component data after coordinate system one-mapping, correcting minor deviations generated during the coordinate mapping process. This ensures that the coordinate information corresponding to the same spatial location within the scene remains consistent across the color image component, depth information component, and point cloud data component. Finally, the color image component, depth information component, and point cloud data component that have completed spatial coordinate system one-mapping and matching verification are structurally integrated, and a unified spatiotemporal index label is added to each component data, generating spatial intelligent perception feature data with a unified spatiotemporal reference. This spatial intelligent perception feature data eliminates the spatial misalignment problem between different modal data, enabling multi-dimensional environmental data to form a stable correspondence within the same spatiotemporal framework. This provides unified and reliable basic feature data for subsequent steps such as depth gradient component and optical flow displacement vector extraction, occlusion region identification, and mask map generation, avoiding misjudgment of occluded regions due to spatiotemporal misalignment.

[0066] Optionally, generating the occlusion mask image in step 1 specifically includes:

[0067] Step 14: Perform local depth gradient calculation on the spatial intelligent perception feature data to obtain the depth gradient component characterizing abrupt changes in spatial structure;

[0068] Step 15: Perform pixel trajectory tracking processing on the spatial intelligent perception feature data between consecutive frames to obtain the optical flow displacement vector that represents the dynamic evolution trend;

[0069] Step 16: Identify feature-deficient regions based on the depth gradient components and the optical flow displacement vector, and perform semantic attribute labeling on the feature-deficient regions to generate the occlusion mask map.

[0070] Preferably, the specific implementation process of performing local depth gradient calculation on the spatial intelligent perception feature data in step 14 to obtain the depth gradient component representing the abrupt change in spatial structure is as follows: taking the spatial intelligent perception feature data that has completed spatiotemporal alignment and coordinate mapping as the processing object, firstly, a dense depth map corresponding to the depth information component is extracted from the spatial intelligent perception feature data. Each pixel in the dense depth map corresponds to the distance value between the spatial position in the scene and the sensing unit. Then, taking each pixel in the dense depth map as the center, a neighborhood window of a preset size is selected, and the gradient change rate is calculated on the distance value in the neighborhood window to obtain the depth gradient value corresponding to each pixel. Finally, the depth gradient values ​​corresponding to all pixels in the dense depth map are integrated to form a depth gradient component covering the entire scene. Depth gradient components are used to characterize abrupt changes in spatial structure within a scene. The distance values ​​corresponding to continuous and complete object surfaces in a scene exhibit a smooth transition, with the corresponding depth gradient values ​​in a lower range. However, at the boundaries of object occlusion, there will be a jump in distance values, and the corresponding depth gradient values ​​will exceed the preset gradient threshold. This region is marked as the first type of occlusion candidate region, providing a spatial structure-level judgment basis for the comprehensive identification of subsequent feature-deficient regions, while avoiding misjudgment problems caused by relying solely on texture features to identify occlusion.

[0071] Preferably, step 15, which involves performing inter-frame pixel trajectory tracking processing on the spatial intelligent perception feature data to obtain the optical flow displacement vector representing the dynamic evolution trend, includes: using the spatial intelligent perception feature data as the processing object, firstly extracting temporally continuous color image components from the spatial intelligent perception feature data; simultaneously, combining the unified spatiotemporal index marker carried in the spatial intelligent perception feature data, matching the depth information component and point cloud data component that are temporally synchronized with the color image components; then, performing pixel-level trajectory tracking processing on the temporally continuous multi-frame color image components, first compensating for the global pixel displacement caused by the motion of the sensing carrier itself, then calculating the relative displacement direction and displacement magnitude of each pixel between consecutive frames to obtain the optical flow vector corresponding to each pixel; finally, integrating the optical flow vectors corresponding to all pixels to form an optical flow displacement vector covering the entire scene. Optical flow displacement vectors are used to characterize the dynamic evolution trend of objects in a scene. The optical flow vectors corresponding to the static background and the surface of unoccluded objects in the scene conform to the continuous and smooth motion law. However, the pixels in the occluded area will have abnormal states such as interruption, abrupt change, or mismatch with the global motion law. These abnormal areas are marked as second-type occlusion candidate areas, which complement the information of the first-type occlusion candidate areas. This can effectively distinguish between abrupt changes in the object's own structure and depth jumps caused by occlusion, further reducing the probability of misjudgment of occluded areas and providing a dynamic motion level judgment basis for the comprehensive identification of subsequent feature-missing areas.

[0072] Preferably, step 16, which identifies feature-deficient regions based on the depth gradient component and the optical flow displacement vector, specifically includes: taking the first type of occlusion candidate region corresponding to the depth gradient component and the second type of occlusion candidate region corresponding to the optical flow displacement vector as processing objects, firstly performing spatial position matching and fusion processing on the two types of occlusion candidate regions, taking the spatial intersection of the two types of occlusion candidate regions as the initial occlusion region, and simultaneously correcting the initial occlusion region by combining multi-viewpoint matching difference information, eliminating misjudged regions caused by sensing noise and missing object surface texture, thus obtaining feature-deficient regions; subsequently, classifying the occlusion degree of the feature-deficient regions, dividing them into partially occluded regions and completely occluded regions according to the completeness of depth and texture information within the feature-deficient regions. Feature-deficient regions are spatial regions within the scene where sensing data cannot be completely collected due to object occlusion, and are the core target regions for subsequent occlusion mask map generation, while also providing precise spatial boundaries for defining the target reconstruction region in subsequent steps.

[0073] Preferably, step 16, which involves performing semantic attribute labeling on the feature-deficient regions to generate the occlusion mask image, specifically includes: taking the feature-deficient regions as the processing object, firstly, calling a pre-trained semantic segmentation network to perform full-scene semantic segmentation processing on the color image component and depth information component in the spatial intelligent perception feature data, obtaining the semantic categories corresponding to each spatial region in the scene, including scene object types such as pedestrians, walls, furniture, vehicles, and ground; then, performing spatial position matching between the feature-deficient regions and the full-scene semantic segmentation results, labeling each feature-deficient region with corresponding semantic attributes, and clarifying the object category to which the feature-deficient regions belong; finally, using the spatial size of the scene as a benchmark, generating a single-channel mask matrix that matches the spatial size of the scene, assigning the matrix positions corresponding to the feature-deficient regions to preset occlusion identifier values, assigning the matrix positions corresponding to the non-occluded regions to background identifier values, and simultaneously adding corresponding semantic attribute labels and occlusion degree labels to each feature-deficient region, integrating them to form an occlusion mask image. The occlusion mask map fully records the spatial location, boundary range, degree of occlusion, and semantic category of the occluded area in the scene. It provides accurate spatial boundaries and semantic priors for the limitation of the target reconstruction area, multi-scale geometric completion, and semantic transfer filling in subsequent steps, avoiding problems such as spatial range deviation and semantic attribute mismatch in the subsequent completion process.

[0074] Optionally, step 2 includes performing the multi-scale geometric completion process, comprising:

[0075] Step 21: Based on the occlusion mask map, the spatial boundary of the target reconstruction region determined in the spatial intelligent perception feature data is used to perform geometric extrapolation using a multi-scale voxel generation algorithm to obtain a rough geometric shell.

[0076] Step 22: Perform fine-grained feature refinement on the coarse geometric shell using surface curvature constraints to determine the object geometry that reflects the complete structure of the occluded area.

[0077] Preferably, in step 21, when determining the spatial edge to be reconstructed based on the occlusion mask image in the spatial intelligent perception feature data, the occlusion mask image and the spatial intelligent perception feature data are used as processing objects. First, the set of pixel coordinates corresponding to the regions marked with occlusion identifier values ​​is extracted from the occlusion mask image. Combining the unified spatiotemporal index and spatial coordinate system mapping parameters carried by the spatial intelligent perception feature data, the two-dimensional pixel coordinates in the occlusion mask image are converted into three-dimensional spatial coordinates in the world coordinate system. Then, the closed boundary contour of the occlusion region in three-dimensional space is fitted and generated to obtain the spatial boundary to be reconstructed. At the same time, the semantic attribute markers attached to the occlusion mask image are combined to clarify the object category to which the spatial boundary to be reconstructed belongs, providing prior information on the corresponding object geometric shape for subsequent geometric extrapolation. The spatial boundary to be reconstructed precisely defines the three-dimensional spatial range that needs to be geometrically completed, avoiding the overflow or loss of spatial range during the completion process. At the same time, the matched prior information on the object geometric shape can constrain the overall direction of the geometric extrapolation, so that the extrapolation result conforms to the conventional spatial structure characteristics of the corresponding object category.

[0078] Preferably, in step 21, when using a multi-scale voxel generation algorithm to perform geometric extrapolation to obtain a rough geometric shell, the spatial boundary to be reconstructed is taken as the processing object, and the multi-scale voxel generation algorithm is used to perform geometric extrapolation processing of the three-dimensional space. First, at the macro scale, based on the geometric boundary and depth gradient information of the visible area around the spatial boundary to be reconstructed, voxel dilation processing is performed on the voxel units at the corresponding positions in the three-dimensional space, and the spatial envelope of the occluded area is constructed along the geometric extension direction of the visible area to form the basic geometric contour corresponding to the occluded area. Then, at the meso scale, the temporal synchronization sensing information of consecutive frames in the spatial intelligent perception feature data is retrieved, and the main structure of the basic geometric contour is supplemented using cross-frame geometric consistency constraints and local motion field prediction results to restore the main geometric structure of the object corresponding to the occluded area. Finally, at the micro scale, local feature upsampling processing is performed on the geometric contour after the main structure is supplemented to fill the small holes and structural gaps in the geometric contour, forming a rough geometric shell that completely covers the entire range of the spatial boundary to be reconstructed. The multi-scale voxel generation algorithm, through a hierarchical and progressive extrapolation process from macro to micro, takes into account both the overall morphological rationality and local structural integrity of the extrapolated geometric structure. Unlike the overall morphological distortion problem that is prone to occur in traditional single voxel completion methods, the generated coarse geometric shell provides a complete and morphologically compliant basic geometric framework for subsequent fine-grained feature refinement.

[0079] Preferably, in step 22, when performing fine-grained feature refinement on the coarse geometric shell using surface curvature constraints to determine the object geometric features reflecting the complete structure of the occluded region, the coarse geometric shell is used as the processing object, and fine-grained feature refinement is performed based on surface curvature constraints. First, the surface normal vector and curvature value corresponding to each voxel unit in the coarse geometric shell are extracted. At the same time, the curvature distribution law of the surface of the object in the visible area around the boundary of the space to be reconstructed is extracted. Combined with the surface features of the object category to which the occluded region belongs, surface curvature constraint conditions matching the target object are constructed. Then, based on the surface curvature constraint conditions, the smoothness of the outer surface of the coarse geometric shell is adjusted and the details are corrected, so that the surface curvature change of the completed region and the surface curvature change of the surrounding visible region maintain a continuous and smooth transition, eliminating surface abrupt changes and structural noise generated during the geometric extrapolation process. Finally, the topological connectivity of the corrected geometric structure is checked to ensure that the geometric structure of the completed region and the object geometric structure of the surrounding visible region maintain topological connectivity, and there is no structural break or spatial misalignment. Finally, the object geometric features reflecting the complete structure of the occluded region are obtained. Surface curvature constraints ensure that the completed geometry maintains a high degree of consistency with the surface features of the original visible area of ​​the object, avoiding the rigid splicing and morphological distortion of the geometry. At the same time, topological connectivity verification ensures that the completed geometry conforms to the spatial structural logic of the real object. The resulting object geometric features provide a precise three-dimensional geometric foundation for the determination of subsequent semantic consistency features and cross-modal feature alignment processing.

[0080] Optionally, the semantic transfer imputation process performed in step 2 specifically includes:

[0081] Step 23: Analyze the feature distribution pattern of known semantic categories in the spatial intelligent perception feature data around the target reconstruction area, and calculate the semantic correlation between the feature distribution pattern and the boundary of the space to be reconstructed;

[0082] Step 24: Transfer the corresponding contextual semantic weights to the geometric feature space based on the semantic relevance to determine the semantic consistency features that are consistent with the surrounding environment attributes.

[0083] Preferably, in step 23, when analyzing the feature distribution pattern of known semantic categories in the spatial intelligent perception feature data surrounding the target reconstruction region and calculating the semantic correlation between the feature distribution pattern and the boundary of the space to be reconstructed, the boundary of the space to be reconstructed corresponding to the target reconstruction region is used as a benchmark. First, a visible spatial region extending outward from the boundary of the space to be reconstructed within a preset range is delineated from the spatial intelligent perception feature data as a context reference region. The context reference region contains complete and collectable color image components, depth information components, and point cloud data components, and has clear semantic and spatial structure information. Subsequently, semantic parsing processing is performed on the spatial intelligent perception feature data within the context reference region to extract the known semantic categories corresponding to each spatial location within the region. Semantic categories, along with the corresponding texture gradient features, depth continuity features, spatial size ratio features, and adjacency relationship features for each known semantic category, are statistically analyzed and integrated to determine the numerical distribution range and variation patterns of various features, forming the feature distribution patterns of known semantic categories within the context reference region. Combined with semantic attribute markers added to the occlusion mask image, the target semantic category to which the spatial boundary to be reconstructed belongs is determined. Using the target semantic category as the core, the feature similarity, spatial adjacency density, and positional correlation between each known semantic category and the target semantic category within the context reference region are calculated, ultimately integrating the semantic correlation between the spatial boundary to be reconstructed and the surrounding known semantic categories. The delineation of the context reference region provides reliable prior information about the visible area for semantic analysis; the extraction of feature distribution patterns clarifies the semantic feature logic and spatial correlation rules of the surrounding environment; and the semantic correlation provides a quantitative basis for subsequent semantic weight allocation and transfer, avoiding the problem of semantic attributes being disconnected from the surrounding environment due to indiscriminate semantic filling.

[0084] Preferably, the semantic correlation between the feature distribution pattern and the boundary of the space to be reconstructed in step 23 can be implemented in a manner that is not limited to the following: First, along the closed contour of the boundary of the space to be reconstructed, extract the boundary points directly adjacent to the context reference region as semantic anchors. Each semantic anchor corresponds to a known semantic category and a corresponding feature value within the context reference region. Then, for each semantic anchor, calculate the feature matching degree between its corresponding known semantic category and the target semantic category of the region to be reconstructed. The feature matching degree is obtained by comparing the differences between the two semantic categories in texture gradient features, depth continuity features, and spatial structure features. The smaller the difference, the higher the feature matching degree. Then, combine the distribution density of the semantic anchors on the boundary of the space to be reconstructed, and assign a corresponding weighting coefficient to the feature matching degree of each semantic anchor. The denser the distribution of semantic anchors on the boundary contour, the higher the corresponding weighting coefficient. Finally, perform a weighted summation on the feature matching degrees of all semantic anchors to obtain the final semantic correlation degree. The extraction of semantic anchors accurately captures the semantic connection boundary between the region to be reconstructed and the visible region. The weighted processing based on distribution density conforms to the actual adjacency of the boundary of the space to be reconstructed, avoiding the edge semantic deviation caused by uniform weighting, so that the generated semantic correlation can accurately reflect the influence of the surrounding semantic environment on the region to be reconstructed.

[0085] Preferably, in step 24, when transferring the corresponding contextual semantic weights to the geometric feature space based on the semantic relevance to determine the semantic consistency features consistent with the surrounding environment attributes, the calculated semantic relevance is used as the processing object. First, according to the numerical distribution of the semantic relevance, corresponding contextual semantic weights are assigned to the semantic features corresponding to different known semantic categories within the context reference area. The higher the semantic relevance value of the known semantic category, the higher the assigned contextual semantic weight. At the same time, a pre-established spatial semantic knowledge graph is retrieved, and the object category prior model that matches the target semantic category is searched. Based on the common semantic features and spatial topological rules of similar objects recorded in the object category prior model, global constraint correction is performed on the initially assigned contextual semantic weights to eliminate the problem of unreasonable weight allocation caused by local contextual feature bias. Subsequently, the corrected contextual semantic weights are weighted and fused with the semantic features of the corresponding known semantic categories to generate semantic filling features that adapt to the spatial boundary to be reconstructed and the target semantic category. These semantic filling features are then mapped to the 3D geometric feature space corresponding to the object's geometric features, ensuring a one-to-one mapping between each semantic label in the semantic filling features and the 3D spatial coordinates in the object's geometric features, eliminating spatial misalignment between semantic attributes and geometric structure. Finally, by minimizing the semantic consistency loss function, iterative optimization is performed on the mapped semantic features, constraining the semantic labels of the region to be reconstructed to have a continuous spatial transition and consistent category matching with the semantic labels of the surrounding context reference regions, avoiding semantic drift or cross-object semantic leakage. This ultimately yields semantically consistent features that are consistent with the surrounding environment. The differentiated allocation of contextual semantic weights ensures the rationality of semantic transfer; the mapping to the geometric feature space solves the problem of the disconnect between geometric shape and semantic attributes in traditional schemes; and the iterative optimization of the semantic consistency loss function further ensures the global consistency of the semantic attributes of the completed region, providing a precise semantic foundation for subsequent cross-modal feature alignment matching that matches the geometric structure.

[0086] Preferably, in a scenario, when determining the specific implementation of the semantic consistency feature in step 24, the semantic consistency feature is further optimized in terms of temporal dimension by combining temporally continuous spatial intelligent perception feature data. Specifically, firstly, historical semantic features of the same object corresponding to the target reconstruction region are extracted from the spatial intelligent perception feature data of multiple consecutive frames before the current frame, and the temporal variation pattern and stability of the historical semantic features are statistically analyzed. Then, the historical semantic features are matched and verified with the semantic consistency feature generated in the current frame. If a temporal semantic category jump is detected between the two, a corresponding temporal weight is assigned based on the temporal stability of the historical semantic features in consecutive frames. The semantic consistency feature is then smoothed and corrected using the temporal weight, so that the generated semantic consistency feature remains stable between temporally continuous frames. This temporal optimization process adapts to the temporal change requirements brought about by viewpoint switching and object movement in dynamic scenes, avoids frequent jumps in semantic attributes caused by brief changes in viewpoint in dynamic occlusion scenes, and further improves the stability and reliability of semantic consistency features in dynamic interactive scenes.

[0087] Optionally, step 2 further includes constructing a traceable generation chain, specifically including:

[0088] Step 25: Record the input sensor source identifier, model version operator, and intermediate feature mapping status in real time when performing the multi-scale geometric completion and semantic transfer filling processes;

[0089] Step 26: Encapsulate the input sensor source identifier, the model version operator, and the intermediate feature mapping state into an audit payload with logical dependencies, so as to construct the traceable generation chain with a unique traceability identifier that runs through the perception layer to the representation layer.

[0090] Preferably, the specific implementation process of real-time recording of the entire process information of multi-scale geometric completion and semantic transfer filling in step 25 is as follows: taking the complete execution link of multi-scale geometric completion and semantic transfer filling as the recording object, a full-node recording mechanism is synchronously triggered when the processing flow starts. Along the execution order of the processing flow, synchronous information collection and retention are performed on the input, operation, and output of each processing node. First, the input sensor source identifier is collected and recorded. The input sensor source identifier comes from the unified spatiotemporal index marker carried by the spatial intelligent perception feature data. Specifically, it includes the sensor unit type, acquisition frame number, acquisition timestamp, and spatiotemporal synchronization verification result corresponding to the color image component, depth information component, and point cloud data component participating in multi-scale geometric completion and semantic transfer filling. At the same time, the data call path and weight allocation information of each sensor data in the processing process are recorded to form a complete input source traceability record. The complete record of the input sensor source identifier clarifies the original data source of the occluded area completion content, providing the most basic data source basis for the traceability verification of the subsequent generated results, and at the same time providing the core information of the input layer for the encapsulation of the audit payload.

[0091] Preferably, in step 25, the specific technical implementation of recording the model version operator is carried out by taking all the computational units called in the multi-scale geometric completion and semantic transfer filling processes as the processing objects. For each computational operator used in the multi-scale voxel generation algorithm, surface curvature constraint refinement, semantic correlation calculation, and context semantic weight transfer processing, the function definition, version number, core computational logic, running parameter configuration, and weight initialization information of the operator are recorded. At the same time, the execution order of each operator in the processing flow, the input and output interface relationship, and the computational error and convergence status information after the operator is executed are recorded. For the model structures such as semantic segmentation networks and feature fusion networks used in the processing flow, the training version of the model, the network structure configuration, the source of the pre-training dataset, and the network layers and weight update information called in this processing are recorded simultaneously to form a complete model operator traceability record. The full process record of the model version operator clarifies the execution basis of all computational logic in the occluded region completion process, avoids the problem of unknown computational process in traditional black-box generation schemes, provides the core basis at the computational level for the logical traceability of subsequent generation results, and provides the core information of the inference layer for the encapsulation of audit payloads.

[0092] Preferably, the specific implementation process of recording intermediate feature mapping states in step 25 is as follows: taking the intermediate processing results output by each operation node in the multi-scale geometric completion and semantic transfer filling process as the processing object, according to the execution sequence of the processing flow, the basic geometric contour generated by macro-scale voxel dilation, the main geometric structure generated by meso-scale structure supplementation, and the detailed geometric features generated by micro-scale upsampling are recorded in sequence during the multi-scale geometric completion process. At the same time, the distribution law of context features, the calculation result of semantic correlation, the result of context semantic weight allocation, and the result of semantic filling feature generation are recorded in the semantic transfer filling process. For each intermediate feature mapping state, the feature dimension, numerical distribution range, confidence score, mapping relationship with the previous input features, and input matching relationship with the subsequent operation nodes are recorded synchronously. At the same time, key indicator data such as geometric consistency error and semantic consistency loss generated during the processing are recorded to form a complete intermediate process traceability record. The full-time record of intermediate feature mapping states completely restores the entire mapping process from the original sensing data to the final object's geometric features and semantic consistency features. This fills the gap in the untraceable intermediate links of the generation process in traditional schemes, providing a full-link process basis for error location and logical verification of subsequent generation results. At the same time, it provides the core information of constraint layer and process verification for the encapsulation of audit payloads.

[0093] Preferably, when encapsulating the recorded information into an audit payload with logical dependencies in step 26, the following implementation methods can also be used, but are not limited to: taking the input sensor source identifier, model version operator, and intermediate feature mapping state of the complete record as the processing objects, first, according to the actual execution logic of multi-scale geometric completion and semantic transfer filling processing, sorting out the logical dependencies between the recorded information, clarifying that the input sensor source identifier is the prerequisite for all operations, the model version operator is the execution rule for feature mapping processing, and the intermediate feature mapping state is the progressive processing result after the operation rule is applied to the basic data, and the finally generated object geometric features and semantic consistency features are complete. The final output of the link processing is then structured and integrated according to a four-layer progressive structure: input layer, inference layer, constraint layer, and output layer. The input layer encapsulates information related to the input sensor source identifier; the inference layer encapsulates information related to the model version operator; the constraint layer encapsulates information related to consistency constraints and error indicators in the intermediate feature mapping states; and the output layer encapsulates the geometric features, semantic consistency features, and confidence information of the finally generated object. Finally, a temporal signature and logical association marker are added to the structured and integrated information, forming an inseparable logical dependency between the information at each level. This encapsulation forms an audit payload with complete logical dependencies. The structured encapsulation of the audit payload integrates the scattered end-to-end record information into an auditable data body with a complete logical link, ensuring that each step of the generation process has a clear logical connection. This avoids the problem of broken traceability links caused by scattered record information and provides a core data carrier for the final construction of a traceable generation chain.

[0094] Preferably, in step 26, when constructing the traceable generation chain with a unique traceability identifier that runs through the perception layer to the representation layer, the encapsulated audit payload is used as the processing object. First, based on the full-link content, collection timestamp, and processing node information of the audit payload, a unique and tamper-proof traceability identifier is generated through hash operation. This unique traceability identifier forms a one-to-one binding relationship with the audit payload, and the unique traceability identifier contains an index address that can quickly locate information at each level. Then, the audit payload bound with the unique traceability identifier is associated and stored with the corresponding generated object geometric features and semantic consistency features, so that the final generated spatial object feature data can directly retrieve the corresponding full-link audit payload through the unique traceability identifier, forming a complete and continuous traceable chain from the original sensor data input of the perception layer, to the model operator operation of the inference layer, to the intermediate feature mapping of the constraint layer, and finally to the object feature output of the representation layer, thus completing the construction of the traceable generation chain. After the traceable generation chain is constructed, the logical continuity and version consistency of each node in the generation chain are verified in real time through time-series signatures and logical continuity verification rules. If a logical break or version conflict is detected, source tracing and repair and link reconstruction will be triggered to ensure the integrity and effectiveness of the traceable generation chain. This traceable generation chain differs from traditional black-box generation schemes, realizing the traceability, verifiability, and auditability of the entire process of occlusion area completion generation. It provides a traceable logical basis for subsequent cross-modal feature alignment and physical rationality correction, and provides a precise link index for deviation location and incremental fine-tuning in the subsequent dynamic update process. It also provides complete supporting data for the credibility assessment and security audit of spatial intelligent object representation results.

[0095] Optionally, generating the corrected object attribute features in step 3 specifically includes:

[0096] Step 31: Retrieve the traceable generation chain, extract coordinate information from the geometric features of the object, and perform cross-modal feature alignment with the category attributes extracted from the semantic consistency features to eliminate the matching deviation between spatial location and semantic attributes;

[0097] Step 32: Perform physical rationality correction processing on the aligned cross-modal features using preset gravity support constraints and object contact logic to generate the corrected object attribute features.

[0098] Preferably, the specific implementation process of retrieving the traceable generation chain and extracting coordinate information from the object's geometric features in step 31 is as follows: taking the object's geometric features and semantic consistency features corresponding to the target reconstruction region as processing benchmarks, firstly, by using the unique traceability identifier associated with the object's geometric features and semantic consistency features, the complete traceable generation chain corresponding to the completion process of the target reconstruction region is retrieved; then, the spatial coordinate system mapping parameters corresponding to the original sensing data are extracted from the input layer of the traceable generation chain, the operator execution logic and feature mapping rules corresponding to the geometric completion and semantic filling processing are extracted from the inference layer of the traceable generation chain, the verification indicators of geometric consistency and semantic consistency are extracted from the constraint layer of the traceable generation chain, and the generation sequence and spatial matching benchmarks corresponding to the object's geometric features and semantic consistency features are extracted from the output layer of the traceable generation chain; then, based on the spatial matching benchmarks extracted from the traceable generation chain, the three-dimensional spatial coordinate information, surface normal vector information, and geometric topological association information corresponding to each voxel unit are extracted from the object's geometric features and integrated to form a complete geometric coordinate dataset, while simultaneously extracting the semantic category attributes, semantic weight distribution, and adjacency semantic association information corresponding to each spatial location from the semantic consistency features and integrating them to form a complete semantic category attribute set. The retrieval of traceable generation chains and the extraction of benchmark information provide a unified spatial benchmark and generation logic basis for subsequent cross-modal feature alignment, avoiding the problem of inconsistent matching benchmarks between geometric and semantic features due to differences in generation chains, while ensuring the traceability and verifiability of the entire alignment process.

[0099] Preferably, in step 31, when performing cross-modal feature alignment to eliminate the matching deviation between spatial location and semantic attributes, the geometric coordinate dataset and the semantic category attribute set are used as processing objects. First, a unified three-dimensional feature embedding space is constructed. Each three-dimensional spatial coordinate in the geometric coordinate dataset is used as a position anchor point in the three-dimensional feature embedding space. The semantic information corresponding to the semantic category attribute set is mapped to the embedding dimension corresponding to the position anchor point, so that each spatial position anchor point simultaneously carries the geometric structure information and semantic attribute information of the corresponding position. Subsequently, based on the feature mapping rules recorded in the traceable generation chain, spatial position matching verification is performed on the mapped features, comparing the spatial coordinate information and semantic attribute information corresponding to each spatial position anchor point one by one. The system identifies regions where geometric coordinates and semantic categories are spatially misaligned or have mismatched boundaries. For these regions, it performs correction processing by combining the geometric topological relationships and semantic adjacency of the context neighborhood. This adjusts the spatial mapping of semantic attribute information, ensuring the semantic category boundary perfectly aligns with the geometric structure boundary, while eliminating mismatches and leaks of semantic attributes across different geometric structures. Finally, by minimizing the cross-modal alignment loss function, iterative optimization is performed on the corrected features, establishing a precise one-to-one match between geometric coordinates and semantic category attributes in three-dimensional space. This completes the cross-modal feature alignment process, resulting in aligned features with a precise match between geometric structure and semantic attributes. Cross-modal feature alignment solves the core problem of the disconnect between geometric form and semantic attributes in traditional solutions, ensuring a complete match between the completed object's geometric structure and semantic attributes in three-dimensional space. This provides a foundation of unified structural and attribute data for subsequent physical plausibility correction.

[0100] Preferably, the definition of the gravity support constraint and object contact logic in step 31 includes: firstly, combining the application environment such as indoor navigation and AR interaction to which the scene belongs, constructing a physical constraint system that conforms to the physical operation rules of the real world. The gravity support constraint is constructed based on the gravity direction of the scene and the force balance rules of the object, clarifying the core constraint rules that the object in the scene must have a stable support contact surface, the suspended structure must conform to the mechanical structural characteristics of the corresponding object category, and the center of gravity of the object must be within the coverage of the support surface. The object contact logic is constructed based on the spatial interaction rules of objects in the real world, clarifying the core constraint rules that the contact boundary between objects of different semantic categories must maintain geometric fit, the spatial occupation range of different objects must not overlap, and the adjacency relationship between objects must conform to the conventional interaction logic of the corresponding semantic category. At the same time, the gravity support constraint and object contact logic are associated and matched with the semantic categories in the scene, and corresponding constraint thresholds and verification rules are configured for objects of different semantic categories such as pedestrians, walls, furniture, vehicles, and ground, forming a physical rationality verification system that can adapt to different object types. This physical constraint system aligns with the core requirements of spatial object physical rationality in application scenarios such as space intelligent agent navigation and augmented reality interaction, avoiding the generation of object structures that do not conform to real physical rules, and providing clear verification basis and correction standards for subsequent physical rationality correction processing.

[0101] Preferably, in step 32, when performing physical rationality correction processing on the aligned cross-modal features using preset gravity support constraints and object contact logic to generate the corrected object attribute features, the aligned features that have completed cross-modal alignment are taken as the processing objects. First, a first-stage mechanical rationality correction is performed based on gravity support constraints. The geometric structure information, semantic category attributes, and spatial position information of the target object are extracted from the aligned features. The center of gravity position, support contact surface range, and force distribution state of the target object are calculated. Unreasonable structural regions with no effective support, center of gravity exceeding the support range, and suspended structures that do not conform to the mechanical characteristics of the corresponding object category are identified. Subsequently, for the identified unreasonable structural regions, the geometric structure is locally smoothed by combining the conventional structural features of the corresponding semantic category object and the support relationship of neighboring objects. At the same time, the semantic attribute information of the corresponding position is corrected to ensure the adjustment. The resulting structure conforms to gravity support constraints, and the geometric structure and semantic information maintain a precise match. Then, based on object contact logic, a second stage of spatial interaction rationality correction is performed. This involves overlapping detection of the spatial occupancy of all objects within the scene, identifying abnormal regions where spatial overlap, non-fitting contact boundaries, or adjacency relationships do not conform to semantic logic are observed. For these abnormal regions, the boundaries and spatial positions of the geometric structure are fine-tuned in conjunction with object contact logic, while simultaneously correcting the semantic boundary information at the corresponding positions. This ensures no spatial overlap between objects and that contact relationships conform to real-world physical interaction rules. Finally, the features that have undergone two stages of correction undergo a second verification of topological connectivity and semantic consistency, ensuring that the corrected features simultaneously meet the requirements of accurate geometric structure, matching semantic attributes, and compliance with physical rules. This ultimately generates the corrected object attribute features. The physical rationality correction process further guarantees the rationality of the spatial object representation from the perspective of real physical rules, avoiding problems such as object suspension and spatial penetration that occur in traditional generation schemes, which are inconsistent with real-world scenarios. This allows the generated spatial object representation to be directly adapted to application scenarios with high requirements for spatial perception reliability, such as spatial intelligent agent navigation, augmented reality interaction, and digital twin modeling.

[0102] Optionally, step 4 involves performing scenario uncertainty modeling and evaluation, specifically including:

[0103] Step 41: Perform multi-sample probability estimation on the corrected object attribute features using the Monte Carlo sampling algorithm to generate a confidence map representing the credibility of the prediction results;

[0104] Step 42: Determine the uncertainty distribution of the occluded area based on the confidence map, and then associate and encapsulate the corrected object attribute features with the uncertainty distribution to generate the traceable spatial intelligent object representation for the occluded area.

[0105] Preferably, the specific implementation process of performing multi-sample probability estimation processing on the corrected object attribute features using the Monte Carlo sampling algorithm in step 41 is as follows: Taking the corrected object attribute features as the processing object, the corrected object attribute features are first structurally decomposed to extract the three-dimensional spatial coordinate information, geometric structure information, semantic category attribute information, and the multimodal input data weight distribution information recorded in the corresponding traceable generation chain, forming a feature dataset for probability estimation; then, based on the designed dual-channel probability estimation framework, a deterministic prediction branch and an uncertain estimation branch are constructed respectively. The deterministic prediction branch is used to output the optimal geometric structure and semantic attributes corresponding to each spatial location within the target reconstruction area. The optimal prediction result is evaluated through an uncertainty estimation branch, which quantifies the reliability of the best prediction result. Then, for the uncertainty estimation branch, a Monte Carlo sampling algorithm is used to perform multi-sample random sampling. Within a preset sampling range, multiple forward inference calculations with random deactivation are performed on the feature dataset to obtain multiple sets of prediction results for each spatial location within the target reconstruction region. Finally, variance statistics and probability distribution fitting are performed on the multiple sets of prediction results for each spatial location to calculate the dispersion and probability distribution interval of the prediction result for each spatial location, obtaining a confidence score for each spatial location. The confidence scores of all spatial locations are integrated to generate a confidence map covering the entire target reconstruction region. This process quantifies the reliability of the prediction result through multiple sampling inferences, unlike traditional schemes that only output a single prediction result. It provides a quantitative data foundation for the subsequent reliability evaluation of spatial object representations and also provides a reliability reference for decision-making in higher-level applications such as spatial intelligent agent navigation and augmented reality interaction.

[0106] Preferably, the execution logic of the Monte Carlo sampling algorithm in step 41 is as follows: First, based on the occlusion degree of the target reconstruction region, semantic category, and the proportion of input data sources recorded in the traceable generation chain, an appropriate number of sampling times and random deactivation probability are set. For completely occluded regions and regions with a higher proportion of semantic prior information in the proportion of input data sources, a higher number of sampling times are set. For partially occluded regions and regions with a higher proportion of measured sensor data in the proportion of input data sources, a relatively lower number of sampling times are set, ensuring the accuracy of probability estimation while controlling the consumption of computational resources. Subsequently, in each sampling process, random deactivation processing is performed on the feature channels of the feature dataset, randomly blocking the information input of some feature channels to simulate... The prediction results vary under different input data perturbations. After each sampling, the output shows the predicted geometric structure and semantic attribute values ​​for each spatial location within the target reconstruction area. After completing all preset sampling cycles, the 3D coordinate variance is calculated for all geometric structure prediction values ​​at the same spatial location, and the category probability distribution is statistically analyzed for all semantic attribute prediction values ​​at the same spatial location. The geometric structure variance and semantic category probability distribution are then weighted and fused to obtain the final confidence score for that spatial location. The confidence score is negatively correlated with the dispersion of the prediction results; the lower the dispersion, the higher the confidence score, indicating higher reliability of the prediction result at that location, and vice versa. This adaptive sampling strategy aligns with the differences in information completeness at different locations within the occluded area, avoiding computational redundancy or insufficient estimation accuracy issues caused by a uniform sampling strategy. This allows the generated confidence map to more accurately reflect the actual reliability of the prediction results at different locations.

[0107] Preferably, in step 42, when determining the uncertainty distribution of the occluded region based on the confidence map, the generated confidence map is used as the processing object. First, the confidence map is matched one-to-one with the three-dimensional spatial coordinates of the target reconstruction region by combining the occlusion mask map, and the two-dimensional confidence score is mapped to each voxel unit in the three-dimensional space to obtain a voxel-level confidence dataset in the three-dimensional space. Then, the spatial distribution of the voxel-level confidence dataset is statistically analyzed. According to the numerical range of the confidence score, the target reconstruction region is classified into high-confidence regions, medium-confidence regions, and low-confidence regions. At the same time, the spatial location, semantic category, and proportion of input data sources recorded in the traceable generation chain corresponding to different confidence regions are statistically analyzed to analyze the correlation between the confidence score and the input data type and the degree of occlusion. Then, combined with the intermediate feature mapping state of multi-scale geometric completion and semantic transfer filling processing recorded in the traceable generation chain, the source of prediction deviation for each confidence region is traced and marked to clarify the core influencing factors of uncertainty in different regions. Finally, an uncertainty distribution covering the entire spatial range of the occluded region and carrying the deviation source mark is formed. This process transforms the confidence information of the plane into a three-dimensional uncertainty distribution bound to the spatial object. At the same time, it completes the source association between confidence and generation link, so that the uncertainty distribution can not only reflect the confidence level of the prediction result, but also clarify the cause of the confidence difference, providing a precise positioning basis for the dynamic update and optimization of the subsequent object representation.

[0108] Preferably, the specific implementation process of associating and encapsulating the corrected object attribute features with the uncertainty distribution in step 42 to generate the traceable spatial intelligent object representation for occluded areas is as follows: Taking the corrected object attribute features, uncertainty distribution, and corresponding traceable generation chain as processing objects, firstly, the three-dimensional geometric structure information and semantic category attribute information in the corrected object attribute features are bound one-to-one with the confidence score and deviation source marker of the corresponding spatial location in the uncertainty distribution, so that each three-dimensional spatial voxel unit simultaneously carries four types of core information: geometric structure, semantic attributes, confidence quantification, and deviation source; then, a corresponding unique traceability identifier is attached to the bound three-dimensional feature data. This unique traceability identifier is consistent with the unique traceability identifier of the traceable generation chain. The full-link audit payload of the corresponding completion process can be directly retrieved through this unique traceability identifier; then, all the bound data is structurally encapsulated and integrated according to the hierarchical structure of spatial coordinate index, geometric structure module, semantic attribute module, confidence assessment module, and traceability identifier module. At the same time, a corresponding temporal marker and spatial registration reference information are attached to each module, and finally, a traceable spatial intelligent object representation for occluded areas is generated. This encapsulation process deeply integrates geometric structure, semantic attributes, credibility information, and traceability links, forming an integrated spatial object representation that combines "results + credibility interpretation + source traceability." Unlike traditional solutions that only output geometric reconstruction results, this approach enables the generated spatial object representation to possess interpretable, traceable, and verifiable characteristics, thus better adapting to the application requirements of highly reliable spatial intelligent systems.

[0109] Optionally, the method further includes a dynamic update action:

[0110] Step 51: Monitor in real time the residual values ​​between the completed content in the traceable spatial intelligent object representation facing the occluded area and the actual observation values ​​in the subsequently received spatial environment sensing data stream;

[0111] Step 52: In response to the residual value exceeding the preset deviation threshold, retrieve the logic breakpoint where the deviation is generated by the location of the traceable generation chain, and perform incremental fine-tuning of the spatial intelligent object representation of the traceable occluded area based on the real observation value.

[0112] Preferably, the residual value of real-time monitoring of the completed content in the traceable spatial intelligent object representation of the occluded area and the real observation value in the subsequently received spatial environment sensing data stream in step 51 includes: taking the generated traceable spatial intelligent object representation of the occluded area and the subsequently continuously received spatial environment sensing data stream as the processing objects, firstly, following the same processing flow as step 1, performing multimodal component extraction, timestamp alignment processing, and spatial coordinate system mapping processing on the subsequently received spatial environment sensing data stream to generate new frame spatial intelligent perception feature data that is consistent with the spatiotemporal reference of the original scene; then, based on the new frame spatial intelligent perception feature data, the visibility status of the original occluded area is detected in real time to identify whether the original occluded area is unoccluded and generates effective real observation values ​​under the new observation perspective. The real observation values ​​are the geometric structure information and semantic attribute information corresponding to the area directly collected by the sensor, including the texture semantic information corresponding to the color image component, the dense distance information corresponding to the depth information component, and the three-dimensional spatial structure information corresponding to the point cloud data component; then, the occluded area... The completed content corresponding to the original occluded region in the domain-traceable spatial intelligent object representation is spatially registered with the real observations in the new frame's spatial intelligent perception feature data. Based on a unified world coordinate system, the three-dimensional spatial coordinates of the completed content and the real observations are matched and aligned one-to-one to eliminate spatial position deviations caused by viewpoint changes and carrier motion. Finally, deviation calculations are performed between the registered completed content and the real observations, calculating coordinate deviations, surface curvature deviations, and topological deviations in the geometric structure dimension, as well as category matching deviations and semantic boundary deviations in the semantic attribute dimension. Weighted fusion processing is performed on the deviation values ​​of different dimensions to obtain a residual value that comprehensively reflects the degree of difference between the completed content and the real observations. The weights of the weighted fusion are adapted and adjusted according to the uncertainty distribution corresponding to the region, with higher weights assigned to the deviation dimensions corresponding to low-confidence regions, so that the residual value can more accurately reflect the actual degree of deviation of the completion result, providing a quantitative basis for subsequent update trigger judgments.

[0113] Preferably, in step 52, when the traceable generation chain is retrieved in response to the residual value exceeding a preset deviation threshold, a corresponding deviation threshold is first preset based on the application requirements of the scenario and the semantic category of the target object. The deviation threshold is divided into two categories: geometric deviation threshold and semantic deviation threshold. For scenarios with high requirements for geometric accuracy, such as spatial intelligent agent navigation and security monitoring, a stricter geometric deviation threshold is set. For scenarios with high requirements for semantic consistency, such as augmented reality interaction and digital twin visualization, a stricter semantic deviation threshold is set. Subsequently, the calculated residual value is compared with the preset deviation threshold. When the residual value exceeds the preset deviation threshold, subsequent deviations are triggered. For differential localization and incremental fine-tuning, when the residual value does not exceed the preset deviation threshold, only the confidence score of the corresponding region is updated based on the actual observation value, without modifying the structure and attributes. After the update process is triggered, the complete traceable generation chain corresponding to the completion content is retrieved through the unique traceability identifier bound in the traceable spatial intelligent object representation of the occluded region. The audit payload encapsulated in the traceable generation chain is obtained. The audit payload includes the input sensor source identifier of the input layer, the model version operator of the inference layer, the intermediate feature mapping state of the constraint layer, and the completion result generation record of the output layer, providing a complete end-to-end traceability basis for subsequent deviation logic breakpoint localization.

[0114] Preferably, the specific implementation process of locating the logic breakpoint in step 52 is as follows: taking the audit load corresponding to the retrieved traceable generation chain and the deviation region corresponding to the calculated residual value as the processing objects, the logic breakpoint is located using a reverse tracing method from the output layer to the input layer; firstly, based on the spatial location of the deviation region, the completion result generation record of the corresponding region is located in the output layer of the traceable generation chain, clarifying the generation logic and weight allocation corresponding to the completion content of the region; then, backtracking to the constraint layer of the traceable generation chain, comparing the intermediate feature mapping state in the completion process of the region, analyzing the feature differences between the intermediate features and the actual observation values, and determining whether the deviation is generated in the mapping, fusion, or constraint links of the intermediate features; then, backtracking to the inference layer of the traceable generation chain. The process involves verifying the model version operators, operational logic, and parameter configurations used during the completion process in that region. The reasonableness of the operator results is then validated against real observations to determine if the deviation originates in the inference stage of the model operators. Finally, the process traces back to the input layer of the traceable generation chain, checking the input sensor source identifiers and the weight ratios of each modality's input data during the completion process in that region. This helps determine if the deviation is due to missing input data, noise, or unreasonable weight allocation. Simultaneously, by combining the uncertainty distribution and deviation source markers recorded in the traceable generation chain, the tracing scope is narrowed, quickly locating the logical breakpoint that causes a significant deviation between the completion result and the real observations. The corresponding processing steps, deviation type, and impact range of this logical breakpoint are recorded, providing precise location data for subsequent incremental fine-tuning.

[0115] Preferably, in step 52, when performing incremental fine-tuning of the spatial intelligent object representation for traceable occluded regions based on the real observations, the processing objects are the logical breakpoints that have been located and the real observations in the new frame spatial intelligent perception feature data. A local incremental correction method is used instead of a global reconstruction method to perform optimization processing on the spatial intelligent object representation for traceable occluded regions. First, for the processing step corresponding to the logical breakpoint, an incremental correction loss function is constructed using the real observations as the baseline data. Fine-tuning is performed on the model parameters, feature weights, computational logic, or input data allocation ratio corresponding to this step. Only the core steps causing the deviation are corrected, while the parameters and states of other non-deviation steps remain unchanged, controlling the consumption of computational resources while correcting the deviation. Subsequently, based on the fine-tuned processing steps, the geometric structure information and semantics corresponding to the deviation region are regenerated. The attribute information is used to locally update the content of the corresponding region in the traceable spatial intelligent object representation for occluded areas. At the same time, the confidence score and uncertainty distribution of the region are recalculated based on the real observations, replacing the original confidence information. Then, the corresponding records in the traceable generation chain are updated synchronously, including the parameter status of logical breakpoints, intermediate feature mapping status, and output result information, so as to preserve the complete traceability link for the updated spatial object representation. At the same time, the corresponding time sequence mark and real observation source identifier are added to this correction operation to ensure the integrity and continuity of the traceability generation chain. Finally, the real observations used in this correction, the deviation localization results, and the incremental optimization strategy are archived into the experience database corresponding to self-supervised learning for subsequent cross-scene knowledge graph updates and model self-evolution optimization, so that the accuracy and reliability of the spatial object representation can be continuously improved during operation.

[0116] Optionally, the semantic transfer filling process is implemented by retrieving a pre-established spatial semantic knowledge graph, specifically including:

[0117] Retrieve prior models of object categories in the spatial semantic knowledge graph that match the semantic context and spatial attributes of the target reconstruction region, and construct global semantic topological constraints for semantic transfer filling based on the prior models of object categories;

[0118] The semantic relevance is numerically corrected based on the global semantic topology constraints, and the context semantic weight is proportionally corrected based on the global semantic topology constraints; the known region semantic features adapted to the target reconstruction region are determined according to the corrected semantic relevance and the corrected context semantic weight; the known region semantic features are directionally migrated to the target reconstruction region to complete the semantic migration filling process.

[0119] Preferably, retrieving prior models of object categories in the spatial semantic knowledge graph that match the surrounding context semantics and spatial attributes of the target reconstruction region includes: the spatial semantic knowledge graph is a structured knowledge storage system built on massive spatial perception data, object geometric structure data, and semantic attribute association data in different scenarios; its core includes key elements such as object category nodes, geometric structure feature edges, semantic association feature edges, and scene adaptation attributes; wherein the object category nodes cover common scene object types such as pedestrians, walls, furniture, vehicles, and ground; each object category node is associated with and stores the corresponding standard geometric structure template, typical semantic attribute set, common adjacent object types, and morphological variation rules in different scenarios; geometric structure feature edges are used to record the geometric topological association rules between different object categories, semantic association feature edges are used to record the semantic interaction logic between different object categories, and scene adaptation attributes are used to mark the adaptation parameters and morphological preferences of object categories in different application scenarios such as indoor navigation, augmented reality, and digital twins. During retrieval and matching, the core features of the target reconstruction region are first extracted, including semantic attribute markers from the occlusion mask image, geometric structural features of the visible areas surrounding the target reconstruction region, and scene type information recorded in the spatial intelligent perception feature data. Then, using these core features as search keywords, a similarity matching query is performed in the spatial semantic knowledge graph. Priority is given to matching object category nodes with consistent semantic attributes, similar geometric structural features, and suitable scene types. Simultaneously, the matching similarity between the target reconstruction region features and candidate object category nodes is calculated. The matching similarity comprehensively considers three dimensions of quantitative indicators: semantic consistency, geometric structural compatibility, and scene adaptability. Finally, the standard geometric structural template, semantic attribute set, and adjacency association rules corresponding to the object category node with the highest matching similarity are selected and integrated to form a prior model of the object category matching the target reconstruction region. This prior model of the object category provides prior knowledge support based on massive scene experience for semantic transfer filling processing, avoiding semantic filling biases that may result from relying solely on local context features, and improving the accuracy and rationality of semantic transfer filling processing.

[0120] Preferably, the global semantic topological constraints for semantic transfer filling based on the object category prior model include: constructing global semantic topological constraints covering the target reconstruction region and its surrounding environment based on the matched object category prior model and the distribution patterns of contextual features around the target reconstruction region; firstly, based on the standard geometric structure templates and typical semantic attribute sets recorded in the object category prior model, determining the core semantic attributes, semantic boundary ranges, and geometric distribution patterns of the semantic attributes that the object corresponding to the target reconstruction region should possess, forming semantic attribute constraint rules. These rules clarify the semantic categories that different spatial locations within the target reconstruction region should possess, as well as the boundary delineation standards between semantic categories; subsequently, based on the common adjacent object types and geometric topological association rules recorded in the object category prior model, and combined with the object distribution in the visible area surrounding the target reconstruction region, determining the relationship between the object corresponding to the target reconstruction region and its surrounding environment. Knowing the semantic relationships that objects should possess, semantic association constraint rules are formed. These rules clarify the compatibility, adjacency, and exclusion relationships between the semantic attributes of the target reconstruction region and the semantic attributes of surrounding objects. Then, based on the scene adaptation attributes recorded in the object category prior model, and combined with the current scene type and application requirements, scene adaptation adjustments are made to the semantic attribute constraint rules and semantic association constraint rules. For example, in indoor navigation scenarios, semantic constraints related to navigation path planning are strengthened; in augmented reality interaction scenarios, semantic constraints related to virtual-real fusion display are strengthened. Finally, the adjusted semantic attribute constraint rules and semantic association constraint rules are structurally integrated to form global semantic topology constraints. These global semantic topology constraints are presented in the form of a semantic relationship graph, where nodes correspond to the semantic categories of the target reconstruction region and surrounding objects, and edges correspond to the constraint relationships and constraint strengths between different semantic categories. Global semantic topology constraints provide global-level semantic logic guidance for semantic transfer filling, avoiding the problem of local semantics being reasonable but global semantics contradicting each other during semantic transfer filling. This ensures that the result of semantic transfer filling is not only consistent with the local context but also maintains consistency with the semantic topology structure of the global scene.

[0121] Preferably, the specific process of performing semantic relevance numerical correction, context semantic weight ratio correction, and semantic migration filling based on global semantic topology constraints is as follows: First, perform numerical correction on the semantic relevance based on the global semantic topology constraints, eliminating false relevance scores that violate the constraint rules, calibrating the numerical value of effective relevance, and eliminating invalid relevance data with low matching effectiveness; simultaneously, perform weight ratio correction on the context semantic weight based on the global semantic topology constraints, adjusting the weight ratio of semantics in each surrounding known region, strengthening the weight of compliant semantic features, weakening or even eliminating the weight of non-compliant semantic features, and calibrating the rationality of weight allocation; then, based on the corrected semantic relevance and the corrected context semantic weight, select known region semantic features that are suitable for the target reconstruction region and conform to the global semantic logic, and eliminate redundant semantic features with low relevance and non-compliant weights; finally, migrate the selected known region semantic features to the target reconstruction region to complete the semantic migration filling process.

[0122] Preferably, in a specific implementation of the dynamic adjustment and application of global semantic topology constraints in a scenario, the global semantic topology constraints are integrated into the entire process of semantic transfer and filling. When calculating semantic relevance, the semantic relevance constraint rules in the global semantic topology constraints are used as the basis for weight adjustment to correct the semantic relevance between the target reconstruction region and the surrounding known semantic categories, so that the calculation result of semantic relevance is more in line with the global semantic logic. When performing contextual semantic weight transfer, the semantic attribute constraint rules in the global semantic topology constraints are used as the filling boundary limit. First, the weight ratio calibration is completed, and then the adapted semantic features are selected based on the calibrated parameters to ensure that the semantic attributes of the transferred filling do not exceed the semantic boundary range defined by the constraint rules, and that the distribution of semantic attributes meets the requirements of the constraint rules. At the same time, after the semantic transfer and filling process is completed, the global semantic topology constraints are used as the verification standard to perform global semantic verification on the generated semantic consistency features, identify and correct the regions where the semantic attributes do not match the global semantic topology constraints, and ensure that the semantic consistency features fully meet the requirements of the global semantic topology constraints. Furthermore, when the contextual features surrounding the target reconstruction region change dynamically, such as when surrounding objects move and alter adjacency relationships, the system updates the global semantic topology constraints in real time. It readjusts the semantic attribute constraint rules and semantic association constraint rules, and simultaneously re-adjusts the semantic association degree values ​​and contextual semantic weight ratios based on the updated constraints. Then, it performs a new round of semantic feature filtering and targeted migration, ensuring that the global semantic topology constraints always remain consistent with the actual situation of the current scene. This provides continuous and effective global semantic guidance for semantic migration and filling processing. This dynamically adaptable global semantic topology constraint mechanism enables semantic migration and filling processing to adapt to dynamically changing scene environments, further improving the stability and accuracy of semantically consistent features.

[0123] like Figure 2As shown, this is a spatial intelligent object representation device for traceable generation of occluded areas, comprising:

[0124] The occlusion recognition module 601 is used to generate spatial intelligent perception feature data that records the geometric structure and dynamic evolution trend of the scene based on the spatial environment sensing data stream for the scene, and extract the depth gradient component and optical flow displacement vector from it to determine the occlusion area in the scene and generate the corresponding occlusion mask map.

[0125] The region reconstruction module 602 is used to define the target reconstruction region in the spatial intelligent perception feature data using the occlusion mask map, and extract the context features around the target reconstruction region to perform multi-scale geometric completion and semantic transfer filling processing to determine the geometric features and semantic consistency features of the object.

[0126] The attribute correction module 603 is used to perform cross-modal alignment processing and physical rationality correction processing on the geometric features and semantic consistency features of the object to generate corrected object attribute features;

[0127] The object representation module 604 is used to perform scene uncertainty modeling and evaluation based on the modified object attribute features to generate a traceable spatial intelligent object representation for occlusion areas.

[0128] like Figure 3 As shown, an electronic device includes a processor 701 and a memory 702, which communicate with each other via a bus.

[0129] The memory 702 is used to store computer programs;

[0130] When the processor 701 executes the program stored in the memory 702, it implements the functions of each module of the spatial intelligent object representation device for traceable generation of occluded areas as described in this application, or implements the steps of the spatial intelligent object representation method for traceable generation of occluded areas as described in this application.

Claims

1. A method for representing spatial intelligent objects that can be traced and generated for occlusion regions, characterized in that, include: Step 1: Generate spatial intelligent perception feature data that records the scene's geometric structure and dynamic evolution trend based on the spatial environment sensing data stream for the scene. Specifically, this includes: acquiring the color image component, depth information component, and point cloud data component contained in the spatial environment sensing data stream; performing timestamp alignment processing on the color image component, the depth information component, and the point cloud data component using a unified time reference to generate a time-synchronized sensing stream; and performing reprojection and spatial coordinate system mapping processing on the time-synchronized sensing stream using a spatial calibration matrix to generate the spatial intelligent perception feature data. Then, the depth gradient component and optical flow displacement vector are extracted from them to determine the occlusion area in the scene and generate the corresponding occlusion mask map. Step 2: Use the occlusion mask image to define the target reconstruction region in the spatial intelligent perception feature data, and extract the context features around the target reconstruction region to perform multi-scale geometric completion and semantic transfer filling processing to determine the object's geometric features and semantic consistency features. The multi-scale geometric completion process includes: based on the spatial boundary to be reconstructed of the target reconstruction region determined in the spatial intelligent perception feature data by the occlusion mask map, performing geometric extrapolation using a multi-scale voxel generation algorithm to obtain a coarse geometric shell; and performing fine-grained feature refinement on the coarse geometric shell using surface curvature constraints to determine the object geometric features that reflect the complete structure of the occlusion region. Perform semantic transfer filling processing, including: analyzing the feature distribution pattern of known semantic categories in the spatial intelligent perception feature data around the target reconstruction area, calculating the semantic correlation between the feature distribution pattern and the boundary of the space to be reconstructed, and transferring the corresponding contextual semantic weight to the geometric feature space according to the semantic correlation to determine the semantic consistency feature that is consistent with the surrounding environment attributes; And construct a traceable generation chain, specifically including: real-time recording of the input sensor source identifier, model version operator, and intermediate feature mapping state when performing the multi-scale geometric completion and semantic transfer filling processes; encapsulating the input sensor source identifier, the model version operator, and the intermediate feature mapping state into an audit payload with logical dependencies, so as to construct the traceable generation chain with a unique traceability identifier that runs through the perception layer to the representation layer; Step 3: Perform cross-modal alignment and physical rationality correction on the object's geometric features and semantic consistency features to generate corrected object attribute features, including: retrieving the traceable generation chain, extracting coordinate information from the object's geometric features and performing cross-modal feature alignment with the category attributes extracted from the semantic consistency features to eliminate the matching deviation between spatial location and semantic attributes; The physical rationality correction process is performed on the aligned cross-modal features using preset gravity support constraints and object contact logic to generate the corrected object attribute features; Step 4: Perform scene uncertainty modeling and evaluation based on the corrected object attribute features to generate a traceable spatial intelligent object representation for occluded areas.

2. The spatial intelligent object representation method for traceable generation of occluded regions according to claim 1, characterized in that, The step 1 of generating the occlusion mask image specifically includes: Local depth gradient calculation is performed on the spatial intelligent sensing feature data to obtain the depth gradient components that characterize abrupt changes in spatial structure; Perform pixel trajectory tracking processing between consecutive frames on the spatial intelligent perception feature data to obtain the optical flow displacement vector characterizing the dynamic evolution trend; Based on the depth gradient components and the optical flow displacement vector, feature-deficient regions are identified, and semantic attribute labeling is performed on the feature-deficient regions to generate the occlusion mask map.

3. The spatial intelligent object representation method for traceable generation of occluded regions according to claim 1, characterized in that, Step 4, which involves performing scenario uncertainty modeling and assessment, specifically includes: The modified object attribute features are subjected to multi-sample probability estimation using the Monte Carlo sampling algorithm to generate a confidence map representing the credibility of the prediction results. Based on the confidence map, the uncertainty distribution of the occluded area is determined, and then the corrected object attribute features are associated and encapsulated with the uncertainty distribution to generate the traceable spatial intelligent object representation for the occluded area.

4. The spatial intelligent object representation method for traceable generation of occluded regions according to claim 1, characterized in that, The method also includes dynamic update actions: Real-time monitoring of the residual values ​​between the completed content in the traceable spatial intelligent object representation facing the occluded area and the actual observation values ​​in the subsequently received spatial environment sensing data stream; In response to the residual value exceeding a preset deviation threshold, the logic breakpoint where the deviation occurred in the traceable generation chain is retrieved, and incremental fine-tuning of the spatial intelligent object representation for the traceable occluded area is performed based on the actual observation value.

5. The spatial intelligent object representation method for traceable generation of occluded regions according to claim 1, characterized in that, The semantic transfer filling process is implemented by retrieving a pre-established spatial semantic knowledge graph, specifically including: Retrieve prior models of object categories in the spatial semantic knowledge graph that match the semantic context and spatial attributes of the target reconstruction region, and construct global semantic topological constraints for semantic transfer filling based on the prior models of object categories; Numerical correction is performed on the semantic relevance based on the global semantic topology constraints, and weight ratio correction is performed on the context semantic weight based on the global semantic topology constraints; The known region semantic features that adapt to the target reconstruction region are determined based on the corrected semantic relevance and the corrected context semantic weight. The semantic features of the known region are directionally transferred to the target reconstructed region to complete the semantic transfer filling process.

Images