High-precision real scene three-dimensional reconstruction method for architectural heritage

Abstract

The application provides a high-precision real scene three-dimensional reconstruction method for building heritage, and relates to the technical field of data processing.The method comprises the following steps: acquiring image data and dividing a region to be reconstructed into multiple sub-regions according to the overall structure of the building heritage; collecting point clouds by using a collecting device for each sub-region, adopting a first sampling interval in a conventional region, and dynamically shortening the first sampling interval to a second sampling interval in a carved recess or narrow space region; performing coarse registration processing, and executing iterative correction on corresponding points of features in an overlapping region to generate a first sub-region point cloud model; performing quality detection to determine a missing region with a point cloud density that does not meet a density judgment standard; combining image data to extract local edge contour and gray gradient information, and projecting the information to the missing region to generate compensation points; performing weighted fusion processing and executing curved surface interpolation in a fusion region; performing boundary feature alignment and global splicing to obtain a three-dimensional reconstruction model; and the application improves the accuracy of three-dimensional reconstruction.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for high-precision real-scene 3D reconstruction of architectural heritage. Background Technology

[0002] Existing methods for 3D reconstruction of architectural heritage typically employ drone aerial imagery combined with terrestrial photogrammetry to generate point cloud data, which is then stitched together using conventional 3D reconstruction algorithms. While these methods can obtain relatively complete external outlines for large-scale architectural reconstruction, they often rely on fixed-interval sampling strategies and single-path scanning methods to construct point cloud data for complex structures or interior spaces.

[0003] However, in practical applications, such as reconstructing historical buildings with numerous sculpted details or narrow passages, existing technologies often suffer from insufficient point cloud density in recessed areas due to fixed sampling intervals. This results in localized missing points or blurred details. For instance, insufficient sampling points at arches or stone carvings can make it difficult to fully reproduce edge features, leading to voids or surface discontinuities in the final 3D model. This not only affects the accuracy and realism of the overall 3D model but also causes gaps and misleading information when used for subsequent digital preservation or virtual display of cultural heritage. Summary of the Invention

[0004] The purpose of this invention is to provide a high-precision real-scene 3D reconstruction method for architectural heritage, aiming to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: A method for high-precision real-scene 3D reconstruction of architectural heritage, the method comprising: Acquire multi-angle image data of the exterior and interior of the architectural heritage, and divide the area to be reconstructed into multiple sub-areas according to the overall structure of the architectural heritage to obtain image data; Based on the image data, point cloud data of each sub-region is acquired using acquisition equipment. The first sampling interval is used in the regular area, and the second sampling interval is dynamically shortened in the carved depression or narrow space area. Sub-region point cloud data is generated through multi-path cross acquisition. The point cloud data of the sub-region is coarsely registered, and iterative correction is performed using the feature corresponding points of the overlapping area to generate the point cloud model of the first sub-region. The point cloud model of the first sub-region is subjected to quality inspection to identify the point cloud density distribution and integrity, and to determine the missing regions where the point cloud density does not meet the density judgment criteria. In the missing region, local edge contours and gray-level gradient information are extracted by combining image data, and the information is projected onto the missing region to generate compensation points, thereby generating a second sub-region point cloud model. The point cloud model of the second sub-region is subjected to weighted fusion processing, and surface interpolation is performed within the fusion region to form a 3D model of the sub-region. The 3D models of the sub-regions are aligned by boundary features and stitched together globally. A smooth fusion is achieved through multi-scale transition surfaces, resulting in a high-precision real-scene 3D reconstruction model that covers the entire architectural heritage.

[0006] The above-described solution of the present invention has at least the following beneficial effects: First, by acquiring multi-angle image data of the exterior and interior of the architectural heritage, and dividing the area to be reconstructed into multiple sub-regions based on the overall structure, differentiated processing of the reconstruction object can be performed before data acquisition. Compared with existing technologies that rely on fixed-interval sampling, this image data-based segmentation strategy can ensure the integrity of the overall framework in large-scale areas, and achieve refined modeling in complex areas or areas rich in detail, thereby avoiding the coexistence of data redundancy and local missing data. For example, the flat walls of the palace facade can be divided into large-scale sub-regions to improve acquisition efficiency, while complex details such as arches and stone carvings can be divided into small-scale sub-regions to enhance point cloud density.

[0007] Next, point cloud data is acquired using acquisition equipment within sub-regions. A first sampling interval is used in regular areas, while a second sampling interval is dynamically shortened in carved recesses or narrow spaces. Combined with a multi-path cross-acquisition method, the uniformity and integrity of the point cloud data can be effectively improved. This dynamic sampling combined with multi-path acquisition overcomes the limitations of traditional fixed intervals and single paths, ensuring high-density point clouds can still be acquired on complex curved surfaces or in occluded spaces. For example, inside the corridors of ancient temples, dynamically shortening the sampling interval and using cross-path acquisition can avoid data loss at corners due to occlusion.

[0008] Secondly, by performing coarse registration on the sub-region point cloud data and iterative correction based on the feature corresponding points of overlapping areas, the stitching error between data collected from different paths can be significantly reduced. Compared with existing technologies that rely solely on conventional registration algorithms, the multi-stage registration strategy of this invention can ensure higher accuracy of the sub-region point cloud model during global stitching, thereby achieving a balance between detail representation and overall structural preservation. For example, in the multi-angle scanning of the stone tower, iterative correction can gradually eliminate misalignment, making the angular structure of the tower body complete and clear.

[0009] Furthermore, by performing quality checks on the point cloud model of the sub-regions, identifying the density distribution and integrity of the point cloud, and determining any missing areas that do not meet the density judgment criteria, potential defects can be detected and marked in advance. This detection method avoids exposing voids only after the final model is generated, thus enabling dynamic repair during the modeling process. For example, in the arched section of a grotto mural, if occlusion causes sparse point clouds, this can be detected during the detection phase and trigger subsequent compensation operations.

[0010] Finally, by extracting local edge contours and grayscale gradient information from image data in the missing areas, and projecting this information to generate compensation points, weighted fusion and surface interpolation are performed on the compensated point cloud model. Ultimately, boundary feature alignment and smooth stitching are completed globally, resulting in a high-precision, real-world 3D reconstruction model covering the entire architectural heritage. This modeling mechanism, based on the complementarity of images and point clouds, can simultaneously ensure the continuity of large-scale structures and the realism of small-scale details. For example, in virtual displays, viewers can perceive the complete outline of the palace roof and observe the exquisite textures of the eaves carvings, avoiding information loss and misleading information caused by missing point clouds in existing technologies. Attached Figure Description

[0011] Figure 1 This is a flowchart of a high-precision real-scene 3D reconstruction method for architectural heritage provided in an embodiment of the present invention. Detailed Implementation

[0012] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0013] like Figure 1 As shown, embodiments of the present invention propose a method for high-precision real-scene 3D reconstruction of architectural heritage, the method comprising: Acquire multi-angle image data of the exterior and interior of the architectural heritage, and divide the area to be reconstructed into multiple sub-areas according to the overall structure of the architectural heritage to obtain image data; Based on the image data, point cloud data of each sub-region is acquired using acquisition equipment. The first sampling interval is used in the regular area, and the second sampling interval is dynamically shortened in the carved depression or narrow space area. Sub-region point cloud data is generated through multi-path cross acquisition. The point cloud data of the sub-region is coarsely registered, and iterative correction is performed using the feature corresponding points of the overlapping area to generate the point cloud model of the first sub-region. The point cloud model of the first sub-region is subjected to quality inspection to identify the point cloud density distribution and integrity, and to determine the missing regions where the point cloud density does not meet the density judgment criteria. In the missing region, local edge contours and gray-level gradient information are extracted by combining image data, and the information is projected onto the missing region to generate compensation points, thereby generating a second sub-region point cloud model. The point cloud model of the second sub-region is subjected to weighted fusion processing, and surface interpolation is performed within the fusion region to form a 3D model of the sub-region. The 3D models of the sub-regions are aligned by boundary features and stitched together globally. A smooth fusion is achieved through multi-scale transition surfaces, resulting in a high-precision real-scene 3D reconstruction model that covers the entire architectural heritage.

[0014] In this embodiment of the invention, by acquiring multi-angle image data of the exterior and interior of the architectural heritage, and dividing the area to be reconstructed into sub-regions based on the overall structure, a comprehensive and hierarchical spatial data foundation can be established in the initial stage, avoiding data redundancy or insufficient local acquisition caused by overall scanning. Large-scale division is used on the exterior facade, while small-scale division is used for sculpted details or recessed areas, thereby achieving differentiated processing.

[0015] Next, point cloud acquisition is performed using acquisition equipment based on sub-region information. The first sampling interval is used in regular areas, and the second sampling interval is dynamically shortened in complex or narrow areas. Combined with multi-path cross acquisition, the integrity and density distribution of the point cloud are improved, preserving the overall structure while ensuring clear details.

[0016] Secondly, by performing coarse registration on the sub-region point cloud data and iterative correction by combining the feature corresponding points of the overlapping areas, the error caused by path differences is gradually reduced, and a reliable sub-region point cloud model is generated.

[0017] Furthermore, by identifying areas with insufficient or missing point cloud density through quality inspection, holes can be avoided in the final model, and a basis for subsequent compensation can be provided.

[0018] Subsequently, edge contours and grayscale gradient information are extracted from the missing areas by combining image data, and these are projected to generate compensation points, thus constructing a complete point cloud model and achieving complementarity between two-dimensional textures and three-dimensional data.

[0019] Next, the compensated point cloud model is weighted and fused, and surface interpolation is performed in sparse regions to ensure surface continuity and smoothness and reduce redundant interference.

[0020] Finally, by aligning the boundaries of the three-dimensional models of multiple sub-regions and stitching them together globally, and introducing multi-scale transition surfaces at the stitching points, we can achieve both overall continuity and local detail, thus obtaining a high-precision three-dimensional model that covers the entire architectural heritage.

[0021] The density determination criteria can be set based on the average characteristics of the point cloud distribution. For example, when the number of points in a certain grid cell is less than half of the global average number of points, it can be identified as a low-density candidate region; furthermore, when the average distance between local points exceeds twice the global average distance, the region can be identified as a potential missing region; if obvious breaks appear in the image texture feature comparison at the same time, it is finally identified as a missing region. Through the above multi-level standard setting, misjudgments and omissions can be effectively avoided.

[0022] This involves acquiring multi-angle image data of the exterior and interior of the architectural heritage, and dividing the area to be reconstructed into multiple sub-regions based on the overall structure of the architectural heritage to obtain image data, specifically including: First, multiple image acquisition methods were used to obtain multi-angle data on the exterior and interior of the architectural heritage. Exterior images were captured by aerial photography using high-resolution cameras mounted on acquisition equipment, obtaining complete exterior data from multiple directions such as the top, sides, and angles of the building. Interior images were captured by ground cameras, panoramic cameras, or mobile scanning equipment, obtaining multi-angle images of interior spaces such as corridors, halls, and vaults. This combined acquisition method ensured that both the external outline and internal details were covered.

[0023] Next, the acquired image data undergoes preprocessing, including distortion correction, brightness equalization, and image stitching. Distortion correction is mainly used to eliminate distortion caused by the camera lens, brightness equalization is used to improve the difference in brightness between images caused by uneven lighting, and image stitching is used to create a wider-angle image scene from the same perspective.

[0024] Secondly, based on the overall structural characteristics of the architectural heritage, the area to be reconstructed is divided into multiple sub-regions. This division process relies on the extraction of structural features from the images, such as detecting the edges of the building's exterior walls, the ridgelines of the roof, and the partition walls of the interior spaces, to form zoning results with boundary integrity. During the division process, differentiated processing can be applied based on the geometric complexity of the regions: areas with small curvature changes and flat geometric shapes are divided into large-scale sub-regions, while areas with dense sculptural details and complex spatial curvature are divided into small-scale sub-regions.

[0025] Finally, the segmentation results are correlated with the corresponding image data, generating a dedicated image dataset for each sub-region. This partitioned image data provides targeted foundational data for subsequent point cloud acquisition and 3D reconstruction, ensuring both the efficiency of overall modeling and the accuracy of acquiring complex and detailed areas.

[0026] In a preferred embodiment of the present invention, point cloud acquisition is performed on each sub-region using an acquisition device based on image data, including: Based on the acquired image data, the curvature features of the building surface are extracted, and the sub-regions are divided into regions with gentle curvature and regions with abrupt curvature changes, generating curvature division results; Based on the curvature division results, in the region with gentle curvature, the acquisition device is controlled to acquire images at a first motion speed and a first sampling interval to obtain the first sampled point cloud data. Based on the curvature division results, in the curvature abrupt change region, by real-time monitoring of the image grayscale change rate, the acquisition device is triggered to reduce its movement speed, increase the sampling frequency, and shorten it to the second sampling interval to obtain the second sampling point cloud data; Based on the obstacle avoidance sensor feedback of the acquisition device in a narrow space, the attitude and acquisition angle of the acquisition device are adjusted, and compensation acquisition is performed on the occluded area to generate compensated point cloud data. The first sampled point cloud data, the second sampled point cloud data, and the compensated point cloud data are merged, and redundant points are weighted using the overlapping data generated by multi-path cross-scanning to obtain the original point cloud data of the sub-region.

[0027] In this embodiment of the invention, by extracting the curvature features of the building surface based on image data, the region is divided into a region with gentle curvature and a region with abrupt curvature changes. This allows for differentiated control before data acquisition, avoiding insufficient or redundant sampling. Using a first motion speed and sampling interval in the gentle curvature region ensures efficiency; in the abrupt curvature region, by monitoring image grayscale changes in real time, the acquisition parameters are dynamically adjusted and shortened to a second sampling interval, which improves point cloud density and ensures that complex details are fully captured. Simultaneously, adjusting the device posture and scanning angle based on obstacle avoidance sensor feedback effectively covers occluded areas. Combined with data fusion from different paths and redundancy weighting, complete and consistent original point cloud data for the sub-regions is ultimately obtained.

[0028] The conventional area refers to the area of ​​the building surface with little curvature and relatively flat geometric shape, such as a flat wall or column surface; the carved and recessed area refers to the part with obvious curvature and complex geometric details, such as stone carvings, arch edges, etc.; the narrow space area is the space-restricted location determined by the feedback of obstacle avoidance sensors, such as the inside of the passage, the gap between beams and columns, or behind the window frame.

[0029] Specifically, based on the acquired image data, the curvature features of the building surface are extracted, and the sub-regions are divided into regions with gentle curvature and regions with abrupt curvature changes, generating curvature division results, including: First, surface features are extracted from the image data corresponding to the sub-regions. By analyzing the grayscale changes in the images, the continuity and abrupt changes in surface morphology are identified. Areas with smooth grayscale distribution usually correspond to the gentle parts of the building surface, i.e., the regular areas; while areas with drastic grayscale changes are often associated with sculpted edges, broken lines, or complex textures.

[0030] Next, based on the extracted image features, curvature estimation is performed using preliminary point cloud data. Curvature estimation analyzes the geometric relationships between neighboring points, that is, by statistically analyzing the height and orientation differences between adjacent pixels or point clouds, parameters reflecting the degree of surface undulation are formed. The smaller the height difference between adjacent points, the smoother the surface; the greater the height and orientation differences between adjacent points, the greater the surface curvature.

[0031] Secondly, based on the calculated curvature parameters, the sub-regions are divided into regions with gentle curvature and regions with abrupt curvature changes. Gentle curvature regions correspond to large areas of flat structures on the building surface, such as the smooth surface of exterior walls or stone columns; regions with abrupt curvature changes correspond to locally complex parts, such as the edges of arches, the concave and convex areas of carvings, or the fine grilles of window frames.

[0032] Finally, curvature segmentation results are generated and bound to sub-region image data, providing a basis for dynamic adjustment of subsequent sampling intervals and scanning strategies. This process allows for focused optimization of complex areas while maintaining efficiency in large-area acquisition.

[0033] Specifically, based on the curvature division results, in regions with gentle curvature, the acquisition device is controlled to acquire images at a first motion speed and a first sampling interval to obtain first sampled point cloud data, which includes: First, guided by the curvature division results, regions with gentle curvature are designated as key areas for motion path planning. For these regions, the control system generates motion commands based on a pre-set initial motion speed, ensuring that the data acquisition device operates at a uniform speed above or around the region.

[0034] Next, during the movement, the camera and acquisition module are triggered to acquire data according to the preset first sampling interval. The first sampling interval is a relatively large interval set based on the characteristics of flat areas, which can ensure the integrity of point cloud coverage and avoid unnecessary redundant sampling, thereby improving the overall acquisition efficiency.

[0035] Secondly, during the data acquisition process, the motion control module adjusts the attitude of the acquisition device in real time to ensure that the camera's optical axis maintains a suitable angle with the surface of the area, avoiding blind spots caused by viewing angle deviations. For example, in vertical plane areas such as the outer walls of palaces, the control system automatically adjusts the pitch angle of the acquisition device to keep the camera at a vertical viewing angle.

[0036] Finally, the point cloud data and image data obtained in this process are initially paired to generate the first sampled point cloud data. This point cloud data covers the flat areas of the architectural heritage, forming the basic framework for 3D reconstruction and laying the data foundation for subsequent detailed acquisition of areas with abrupt curvature changes.

[0037] Among these, based on feedback from obstacle avoidance sensors of the acquisition device in a narrow space, the attitude and acquisition angle of the acquisition device are adjusted, and compensatory acquisition is performed on the occluded area to generate compensated point cloud data, specifically including: First, when the data acquisition device enters a narrow space or an area with obstructions, its onboard obstacle avoidance sensors begin collecting environmental information in real time. These sensors typically include ultrasonic sensors, infrared sensors, or lidar, used to detect the relative distance and orientation between the data acquisition device and surrounding obstacles. The feedback information from the sensors forms a spatial obstacle distribution map, providing a reference for adjusting the data acquisition device's path.

[0038] Next, the control system calculates the attitude adjustment parameters of the data acquisition device based on the obstacle distribution map. These parameters are used to control the roll angle, pitch angle, and yaw angle of the data acquisition device, enabling it to continue data acquisition without colliding with obstacles. For example, when the data acquisition device moves inside the arched passage of an ancient building, if the sensor detects that the top is too close, the control system will adjust the pitch angle of the data acquisition device to avoid collisions and ensure the coverage area of ​​the camera's field of view.

[0039] Secondly, by incorporating obstacle information, the acquisition angle is dynamically adjusted to ensure that obstructed areas can be effectively scanned. This adjustment is achieved by changing the horizontal or vertical emission angle of the laser, compensating for areas that were originally not directly covered. For example, in a narrow space behind a carved window frame, the laser can be adjusted to cover the previously blind spot, thus completing the point cloud data.

[0040] Finally, the data acquired during this process is separately labeled as compensated point cloud data. Compensated point clouds not only compensate for gaps caused by occlusion or narrowness, but also enhance the overall spatial integrity when merged with normal point cloud data. For example, in the area beneath the beam frame structure inside ancient temples, compensated acquisition can effectively reconstruct the three-dimensional information of the connection between the beams and the roof.

[0041] Specifically, the first sampled point cloud data, the second sampled point cloud data, and the compensated point cloud data are merged, and redundant points are weighted using the overlapping data generated by multi-path cross-scanning to obtain the original point cloud data of the sub-region, which includes: First, the first sampled point cloud data obtained in regions with gentle curvature, the second sampled point cloud data obtained in regions with abrupt curvature changes, and the compensated point cloud data collected in occluded regions are aggregated. The aggregation process requires unifying the coordinate system and performing preliminary spatial alignment to ensure that data from different paths and methods can be superimposed within the same three-dimensional reference frame.

[0042] Next, for data from different sources, redundant points in multi-path acquisition are identified through spatial overlap detection. Specifically, the spatial differences between points in the same area are compared, and points with differences less than a set distance threshold are identified as redundant.

[0043] Secondly, when processing redundant points, they are not directly deleted, but optimized through a weighted approach. The weighting strategy is based on two factors: first, the sampling density of the point, i.e., the local point cloud density of the area where the point is located; points with higher density have lower weights. Second, path confidence, i.e., the stability and angular superiority of the acquisition path; points with higher confidence have higher weights. This ensures that data redundancy is compressed while prioritizing the retention of higher-quality points.

[0044] Furthermore, a set of fused points is formed through weighted processing, and the fused result is smoothed to avoid surface roughness caused by minor differences between point clouds from different sources. For example, in multipath acquisition of a stone tower wall, point clouds acquired through different paths may appear slightly misaligned on the surface; weighted fusion and smoothing can eliminate these inconsistencies.

[0045] Finally, the fused point set is used as the raw data for the sub-region point cloud for subsequent registration and model generation. This raw data ensures coverage integrity while also possessing high accuracy and high consistency, thus providing a solid data foundation for subsequent 3D reconstruction.

[0046] In a preferred embodiment of the present invention, coarse registration processing is performed on the sub-region point cloud data, and iterative correction is performed using the feature corresponding points of the overlapping regions to generate a first sub-region point cloud model, including: Based on the original point cloud data of the sub-region, calculate the geometric center and normal vector distribution to form the initial pose parameters; Based on the initial attitude parameters, coarse registration is performed on the point clouds collected from different paths to obtain preliminarily aligned sub-region point cloud data; Based on the preliminarily aligned sub-region point cloud data, feature points of overlapping regions are extracted. These feature points are then filtered based on normal vector consistency and local curvature extrema to obtain corresponding point pairs. Based on the corresponding point pairs, calculate the Euclidean distance and angle difference between the points to form the error distribution results; Based on the error distribution results, iterative correction processing is performed to gradually update the attitude parameters until the overall registration error is lower than the preset error threshold, thus obtaining the point cloud model of the first sub-region.

[0047] In this embodiment of the invention, by performing coarse registration on the sub-region point cloud data and forming initial pose parameters using the geometric center and normal vector, a unified coordinate reference can be quickly established, reducing misalignment between data from different paths. Furthermore, by extracting feature points from overlapping areas and filtering corresponding point pairs, the stability of the matching points is ensured. Subsequently, an error distribution is formed by combining Euclidean distance and orientation differences, and the pose parameters are gradually optimized during the iteration process until the overall error is below a threshold. This method can achieve high-precision alignment between multi-path data, resulting in a reliable first sub-region point cloud model.

[0048] Specifically, based on the original point cloud data of the sub-region, the geometric center and normal vector distribution are calculated to form the initial pose parameters, including: First, the raw point cloud data of the sub-region is imported into the point cloud processing system, and the 3D coordinates of the points are read as a whole. To obtain the overall positional distribution of the point cloud, the geometric center of the point cloud in this region is obtained by averaging the coordinates of all points. This geometric center reflects the overall positional orientation of the point cloud in this region and serves as the initial reference point for subsequent registration.

[0049] Next, normal vector estimation is performed on local regions within the point cloud. Specifically, several neighboring points are selected within the neighborhood of each point, and the spatial distribution of these points is used to determine the local orientation of the point cloud surface. The overall distribution of the normal vectors can be used to characterize the spatial orientation features of the sub-region point cloud.

[0050] Finally, the geometric center and normal vector distribution are used as inputs to form the initial pose parameters, which serve as the reference coordinate system for coarse registration. This ensures a reasonable initial pose for subsequent point cloud alignment, reducing the computational load and error range of subsequent iterations.

[0051] Specifically, based on the initial pose parameters, coarse registration is performed on point clouds acquired through different paths to obtain preliminarily aligned sub-region point cloud data, including: First, initial attitude parameters are applied to point cloud data acquired along different paths to unify their spatial reference frame. Specifically, the geometric center of the point cloud along each path is translated to the same position, and the rotation angle of the point cloud is adjusted according to the normal vector distribution to keep them roughly aligned in the same direction.

[0052] Next, the overall position of the initially aligned point cloud is corrected. This is done by comparing the overlapping areas between point clouds to ensure that the overall shape is basically consistent at a macroscopic level. At this point, although there may still be some discrepancies in the details, the overall geometric contours have overlapped.

[0053] Finally, preliminary aligned sub-region point cloud data are obtained. This data provides the foundation for subsequent feature point extraction and fine-grained iterative correction.

[0054] Specifically, based on the pre-aligned sub-region point cloud data, feature points of overlapping regions are extracted. These feature points are then filtered based on normal vector consistency and local curvature extrema to obtain corresponding point pairs, including: First, in the initially aligned sub-region point cloud data, overlapping spatial regions between different paths were identified. These regions are the focus of subsequent fine-grained registration.

[0055] Next, feature points are extracted within the overlapping region. The selection of feature points is based on two conditions: first, consistency of normal vectors, that is, when adjacent points have similar orientations, it is considered that the point can reflect the continuity characteristics of the surface; second, local curvature extrema, that is, points where the surface undergoes obvious bending or abrupt changes, these points can reflect geometric differences.

[0056] Secondly, the feature points extracted from the two paths are compared, and a correspondence is established based on the similarity of spatial location and geometric features to obtain corresponding point pairs.

[0057] The resulting corresponding point pairs provide benchmark data for subsequent error calculation and iterative correction.

[0058] Specifically, based on corresponding point pairs, the Euclidean distance and angular differences between the points are calculated to form the error distribution results, which include: First, for each pair of corresponding points, calculate their spatial distance in the three-dimensional coordinate system. The smaller the straight-line distance between corresponding points, the better the registration effect; the larger the distance, the more likely there is a deviation.

[0059] Next, in addition to spatial distance, the angular difference between normal vectors must also be calculated. The magnitude of the angular difference reflects the degree of consistency between the two points in the surface direction. The smaller the difference, the closer the surfaces are to a match; the larger the difference, the more likely there is still a local registration error.

[0060] Secondly, the two indicators mentioned above are combined to form the error distribution result. This result reflects the matching of the entire point cloud in terms of spatial position and direction, and can be visually represented in a two-dimensional heatmap or numerical matrix.

[0061] By analyzing the error distribution, we can identify which regions have larger errors, providing a reference direction for subsequent iterative corrections.

[0062] Specifically, based on the error distribution results, iterative correction processing is performed to gradually update the pose parameters until the overall registration error is lower than a preset error threshold, thus obtaining the point cloud model of the first sub-region, which includes: First, the error distribution results are used as input to adjust the translation and rotation parameters of the point cloud. Each adjustment aims to reduce the distance and angular differences between corresponding point pairs, gradually making the point cloud more uniform.

[0063] Next, after an adjustment, the new error distribution is recalculated, and the parameters are updated again. This process iterates continuously, gradually approaching the optimal solution.

[0064] Secondly, a preset error threshold is set. When the overall registration error is lower than this threshold, the point cloud registration is considered to have reached satisfactory accuracy, and the iteration process stops. If the error is still higher than the threshold, a new iteration continues.

[0065] Finally, the point cloud data obtained through multiple rounds of iterative updates has achieved high-precision alignment in spatial location and orientation, thus forming the first sub-region point cloud model. This model not only retains complete data collected from different paths but also ensures consistency in geometry and surface orientation, laying a reliable foundation for subsequent sub-region stitching and global reconstruction.

[0066] In a preferred embodiment of the present invention, the point cloud model of the first sub-region is subjected to quality inspection to identify the point cloud density distribution and integrity, and to determine the existence of missing regions where the point cloud density does not meet the density judgment criteria, including: Based on the point cloud model of the first sub-region, a spatial partitioning structure based on a three-dimensional mesh is established to obtain the point cloud density matrix; Based on the point cloud density matrix, the number of points in each grid cell is compared with the global average to form low-density candidate regions; Based on the low-density candidate regions, the average distance between points is calculated and compared with the global average distance to generate potential missing regions. Based on the potential missing regions, and combining the normal consistency of the boundary points with the texture changes of the corresponding images, boundary continuity detection is performed to obtain the missing regions.

[0067] In this embodiment of the invention, by performing quality inspection on the point cloud model of the first sub-region, forming a point cloud density matrix using a 3D mesh, and comparing the local and global average values, low-density candidate regions can be quickly identified. Further calculation of the average distance between points, combined with distribution uniformity, allows for the screening of potential missing regions. Finally, a joint determination is made by combining boundary point normal consistency and image texture changes to obtain the actual missing regions. This multi-level detection method avoids misjudgment and accurately locates regions requiring compensation.

[0068] Specifically, based on the point cloud model of the first sub-region, a spatial partitioning structure based on a 3D mesh is established to obtain the point cloud density matrix, which includes: First, the point cloud model of the first sub-region is imported into the point cloud processing platform, and a mesh structure is established based on its three-dimensional spatial extent. This structure divides the overall space into multiple regular cubic units, enabling the point cloud distribution to be represented in a more refined spatial manner.

[0069] Next, the number of points contained in each small cell is counted, and the result is used as the point cloud density value of the cell. In this way, the point cloud of the entire sub-region can be represented by a three-dimensional matrix, with the value of each cell corresponding to its point cloud density.

[0070] Secondly, to avoid blurry density information due to low resolution, the size of the cells is dynamically adjusted based on the amount of point cloud data when dividing the grid. For example, smaller grid cells are used in areas with a large number of point clouds to ensure accuracy, while larger grid cells are used in areas with a small number of point clouds to improve efficiency.

[0071] The resulting point cloud density matrix can intuitively reflect the point distribution at each spatial location, providing data support for subsequent low-density area detection.

[0072] Specifically, based on the point cloud density matrix, the number of points in each grid cell is compared with the global average to form low-density candidate regions, including: First, a comprehensive statistical analysis of the point cloud density matrix is ​​performed, calculating the average number of points across all units. This average value represents the overall point cloud acquisition density level for the current sub-region.

[0073] Next, the number of points in each grid cell is compared with this average value. If the number of points in a particular cell is significantly lower than the global average, it is identified as a low-density candidate region. This comparison avoids misjudgments caused by differences in overall density.

[0074] Secondly, to improve the accuracy of the judgment, a relative proportion judgment method can be introduced during the comparison process. For example, a region is only considered a candidate region when the number of points in a cell is less than half of the average. This can filter out normal slight density fluctuations and retain only regions where the point cloud is truly sparse.

[0075] Finally, the identified low-density candidate regions can provide a preliminary range for subsequent missing region detection, avoiding large-scale redundant analysis and thus improving overall computational efficiency.

[0076] Specifically, based on low-density candidate regions, the average distance between points is calculated and compared with the global average distance to generate potential missing regions, including: First, within the low-density candidate region, point cloud data is selected, and the average distance between these points is calculated. The average distance reflects the sparsity of the point cloud in that region; a larger distance indicates that the points are more dispersed.

[0077] Next, the local average distance is compared with the global average distance of the entire sub-region point cloud. If the local average distance is significantly greater than the global average distance, it indicates that the candidate region does indeed have a sparse point cloud problem, and it can be further confirmed as a potential missing region.

[0078] Secondly, in the comparison process, in addition to the distance between points, the uniformity of point distribution is also considered. For example, when the point cloud is sparse but evenly distributed, it can be judged as a normal situation; while when the point cloud is both sparse and unevenly distributed, it is more likely to be identified as a missing region.

[0079] Finally, this determination step provides basic data for the preliminary detection of potential missing regions, ensuring that the identification of missing regions has both global reference value and preserves local detail differences.

[0080] In a preferred embodiment of the present invention, in the missing region, local edge contours and gray-level gradient information are extracted by combining image data, and the information is projected onto the missing region to generate compensation points, thereby generating a second sub-region point cloud model, including: Based on the missing area, extract the corresponding image data segment to obtain the missing area image input; Based on the image input of the missing region, the local edge contour is extracted using an edge detection algorithm, and the texture gradient information is extracted based on the gray-level gradient change rate to form the edge gradient corresponding feature data; Based on the feature data corresponding to the edge gradient, the contour direction and gray-level distribution trend of the missing region are calculated to obtain the projection compensation parameters. Based on the projection compensation parameters, combined with the imaging model of the image acquisition device and its spatial position and attitude parameters during the acquisition process, the edge contour and gray-level gradient information are projected onto the three-dimensional spatial position of the missing area to generate a compensation point dataset. Based on the compensation point dataset, it is merged with the first sub-region point cloud model to obtain the second sub-region point cloud model.

[0081] In this embodiment of the invention, by extracting the edge contours and grayscale gradient information of the image in the missing area and projecting this information to generate compensation points, the missing point cloud caused by insufficient sampling or occlusion can be effectively compensated. This method fully utilizes the texture and boundary features of the two-dimensional image to complement the three-dimensional point cloud data, thus enabling the restoration of architectural surface details even when the point cloud is incomplete. For example, in stone pillar carvings or arch edges, local gaps often appear because lasers cannot completely cover them; however, by projecting compensation using the image's grayscale gradient, the continuity of the pattern or edge can be restored. Next, the generated compensation points are fused with the first sub-region point cloud model, ensuring a natural transition of the compensation points without compromising the realism of the original point cloud. The resulting second sub-region point cloud model significantly improves both overall integrity and local continuity, making the textures and contours of complex buildings clearer and providing high-quality data support for subsequent three-dimensional model stitching and overall display.

[0082] Specifically, based on the image input of the missing region, an edge detection algorithm is used to extract local edge contours, and texture gradient information is extracted based on the gray-level gradient change rate to form edge gradient corresponding feature data, which includes: First, the image data segment corresponding to the missing region is input into the image processing module. By scanning the image grayscale values ​​pixel by pixel, the boundary locations with significant brightness changes are identified. Then, common edge detection methods, such as algorithms based on pixel difference or directional gradient, are used to extract the contour of the local region. This yields the contour shape information of the missing region.

[0083] Next, the rate of change of pixel grayscale is calculated near the extracted edges. The greater the rate of change of grayscale, the more obvious the texture change in that area. In this way, texture details in a 2D image can be converted into quantifiable gradient information.

[0084] Secondly, the edge contours are combined with the grayscale gradients to form edge gradient corresponding feature data. This data not only records the edge morphology of the region but also includes the directionality and intensity of texture changes. This information provides a complete spatial feature description for the subsequent projection of compensation points.

[0085] Specifically, based on the feature data corresponding to the edge gradient, the contour direction and gray-level distribution trend of the missing region are calculated to obtain the projection compensation parameters, which include: First, the feature data corresponding to the edge gradients are structured. By sorting the edge points according to their spatial continuity, the overall direction of the contour can be determined. The contour direction reflects the extension direction of the surface structure of the missing region, such as straight line extension, arc transition, or polygonal transformation.

[0086] Next, based on the contour direction, the grayscale gradient is statistically analyzed. By analyzing the magnitude and direction of the gradient values, the distribution trend of the texture can be extracted. For example, when the gradient direction remains consistent in a certain area, it indicates that the texture in that area has extensibility; while when the gradient direction changes abruptly, it indicates the presence of boundary or detailed features.

[0087] Secondly, the contour direction is combined with the grayscale distribution trend to generate projection compensation parameters. These parameters include three key elements: the contour extension direction, the texture intensity variation pattern, and the spatial correspondence. They will guide the generation of compensation points in subsequent steps, ensuring that the point cloud distribution is consistent with the real building surface.

[0088] Specifically, the compensation point dataset is merged with the first sub-region point cloud model to obtain the second sub-region point cloud model, which includes: First, the compensation point dataset generated by projection is loaded into the point cloud processing system and spatially aligned with the point cloud model of the first sub-region. During the alignment process, the boundaries of the missing regions are used as references to ensure that the compensation points can accurately fill the empty regions without deviating from the actual structure.

[0089] Next, a consistency check is performed on the data during the merging process. Specifically, the spatial spacing and normal direction between the compensation points and the original point cloud are compared. When the difference between the compensation points and the original point cloud surface is too large, the weights are automatically adjusted or the points are removed to avoid surface abruptness caused by unreasonable compensation.

[0090] Secondly, the compensation points are gradually integrated into the original point cloud through a weighted fusion method. The weights are set based on two aspects: first, the reliability of the source of the compensation points, for example, the more obvious the edge gradient information, the greater the weight; second, the consistency between the compensation points and the original points' neighborhoods, the closer the compensation points are to the original points, the higher the weight.

[0091] Finally, after merging, the resulting second sub-region point cloud model repairs the missing areas, maintaining the authenticity of the original point cloud while introducing points generated by image feature compensation, thereby improving the overall continuity and integrity.

[0092] In a preferred embodiment of the present invention, a weighted fusion process is performed on the point cloud model of the second sub-region, and surface interpolation is performed within the fusion region to form a three-dimensional model of the sub-region, including: Based on the point cloud model of the second sub-region, the overlapping regions between adjacent sub-regions are extracted to form a set of overlapping region point clouds. Based on the point cloud set of the overlapping region, the density index and acquisition path confidence of each point are calculated to obtain the weighted parameter matrix; Based on the weighted parameter matrix, the point clouds of the overlapping regions are weighted and fused to generate fused point cloud data; Based on the fused point cloud data, surface interpolation is performed in the sparse region of the point cloud, and an interpolation strategy based on the consistency of the normal of neighboring points is adopted to generate interpolated surface data. Based on the fused point cloud data and interpolated surface data, the mesh structure of the sub-region 3D model is constructed, and surface continuity optimization is performed to obtain the sub-region 3D model.

[0093] In this embodiment of the invention, by performing weighted fusion processing on the point cloud model of the second sub-region, the problems of excessive redundant points and uneven quality in multi-path acquisition can be effectively solved. The weighting strategy considers both the local density of points and the confidence of the acquisition path, thereby ensuring that high-reliability points have a greater weight in the fusion, while the influence of low-confidence points is weakened. This mechanism can improve the overall accuracy of the fused data and reduce surface roughness caused by redundant points. Subsequently, surface interpolation is performed in the overlapping area to compensate for sparse or small-scale missing points in the point cloud. In particular, the interpolation strategy based on the consistency of the normal of neighboring points can ensure a natural surface transition without geometric abruptness. This not only restores the continuity of the point cloud but also ensures the smoothness of complex curved surface structures. For example, in structures such as vaults, stone carvings, or domes, the interpolated surface can balance detail and overall smoothness. The final generated 3D model of the sub-region is more stable at local joints, and its overall performance is more consistent with the real structure of architectural heritage.

[0094] Specifically, based on the point cloud model of the second sub-region, overlapping regions between adjacent sub-regions are extracted to form a set of overlapping region point clouds, including: First, the point cloud model of the second sub-region is input into the point cloud management system, and spatial boundary detection is performed on the overall point cloud. By identifying the boundary range of each sub-region point cloud, it is determined which regions have spatial overlap.

[0095] Next, points in the overlapping areas are extracted. Specifically, the point cloud boundary coordinates of different sub-regions are compared. When the boundaries of two regions coincide or are close to each other within a certain spatial threshold range, the points in these regions are extracted as candidate points for overlap.

[0096] Secondly, the candidate points are further filtered to remove isolated or noisy points, retaining only the data that truly represents the overlapping features of the regions. The resulting point cloud set of overlapping regions reflects the common parts between adjacent sub-regions and provides data support for subsequent fusion processing.

[0097] Specifically, based on the point cloud set of the overlapping region, the density index and acquisition path confidence of each point are calculated to obtain a weighted parameter matrix, which includes: First, density analysis is performed on the points within the overlapping region. The density index is calculated by counting the number of neighboring points within a certain radius of each point. The more points there are, the denser the point cloud in that region.

[0098] Next, the path confidence score is calculated by combining the acquired path information. Path confidence score is a measure of the reliability of point cloud data during the acquisition process, such as whether the movement posture of the acquisition device is stable and whether the acquisition angle is ideal. Generally, points acquired directly from the target surface have higher confidence scores than points acquired by oblique scanning.

[0099] Secondly, the point density index is combined with the path confidence to generate a weighted parameter matrix. This matrix assigns a comprehensive weight value to each point, taking into account both the local density of the point cloud and the reliability of the data source.

[0100] Specifically, based on the weighted parameter matrix, the point clouds of overlapping regions are weighted and fused to generate fused point cloud data, including: First, spatial matching is performed on point cloud data from different acquisition paths in the overlapping area to ensure that points at the same location can correspond to the same spatial range.

[0101] Next, these corresponding points are fused according to weighted parameters. The weighting strategy is as follows: in the overlapping region, if a point has a high confidence and density value, its location and attributes will have a greater weight in the fusion process; while points with low confidence or density will have a relatively weaker impact in the fusion process.

[0102] Secondly, after weighted fusion, the point cloud data of the overlapping areas will be optimized into a unified set of points. This set avoids the problem of too many points caused by repeated collection and also eliminates local inconsistencies caused by errors from different paths.

[0103] Specifically, based on the fused point cloud data, surface interpolation is performed in the sparse region of the point cloud, using an interpolation strategy based on the consistency of the normal direction of neighboring points to generate interpolated surface data, including: First, density detection is performed on the fused point cloud data to identify areas where the point cloud distribution is sparse or contains small holes.

[0104] Next, surface interpolation is performed in these sparse regions. Specifically, the normal directions of neighboring points are found using points in the sparse region as centers, and the consistency of these normal directions is used to infer the surface trend of the missing regions.

[0105] Secondly, supplementary points are generated according to this trend, gradually forming a continuous surface. In this way, the interpolated surface can maintain the natural extensibility of the surface while avoiding abrupt geometric changes.

[0106] The resulting interpolated surface data can effectively compensate for the lack of point cloud data, making the three-dimensional structure of the entire sub-region more complete and smooth.

[0107] Specifically, based on the fused point cloud data and interpolated surface data, the mesh structure of the sub-region 3D model is constructed, and surface continuity optimization is performed to obtain the sub-region 3D model, which includes: First, the fused point cloud data and interpolated surface data are imported into the 3D modeling module and integrated. Through point-to-surface matching, the discrete point cloud is transformed into a mesh structure.

[0108] Next, after the mesh is built, the continuity of the model surface is checked to find whether there are breaks, overlaps or irregular protrusions on the surface.

[0109] Secondly, a smoothing algorithm is used to optimize the surface, ensuring a natural transition between adjacent meshes and avoiding subtle discontinuities caused by interpolation and blending. During optimization, the original geometric features are preserved first, avoiding excessive smoothing that could lead to loss of detail.

[0110] Finally, the output sub-region 3D model has a complete mesh structure and continuous surface properties, which can be directly used for subsequent sub-region stitching and global modeling.

[0111] In a preferred embodiment of the present invention, the 3D model of the sub-region is aligned by boundary features and globally stitched together, and smooth fusion is achieved through multi-scale transition surfaces to obtain a high-precision real-scene 3D reconstruction model covering the entire architectural heritage, including: Based on the 3D model of each sub-region, extract the set of boundary feature points to form a global boundary index; Based on the global boundary index, perform boundary alignment processing, calculate the difference vector between boundary points of adjacent sub-regions, and obtain alignment correction parameters; Based on the alignment correction parameters, adjust the boundary positions of the 3D model of the sub-region to generate a preliminary stitched model; Based on the preliminary splicing model, a multi-scale transition surface is constructed, which consists of a large-scale overall smooth surface and a small-scale local detail surface. Based on the multi-scale transition surface, a smooth fusion process is performed on the splicing area of ​​the preliminary splicing model to eliminate boundary discontinuities and obtain a high-precision real-scene 3D reconstruction model.

[0112] In this embodiment of the invention, by aligning the boundary features of the 3D models of sub-regions and globally stitching them together, the problem of discontinuous boundaries between different sub-regions in traditional stitching methods can be solved. This method first extracts a set of boundary feature points and establishes a global index to ensure a unified reference system during stitching. Then, it calculates the difference vectors of the boundary points of adjacent sub-regions and generates correction parameters to adjust the boundaries, thereby reducing misalignment and gaps. For example, when stitching together temple corridors and courtyard models, boundary deviations may occur, and correction processing can make the stitching more natural. Next, based on the initial stitched model, multi-scale transition surfaces are constructed. Large-scale curved surfaces ensure overall smoothness, while small-scale curved surfaces retain details such as carvings, thus achieving natural integration at both the macro and micro levels. Finally, smoothing processing eliminates boundary abruptness, making the overall model continuous and complete. This not only restores the general structure of architectural heritage but also maintains the authenticity of details, providing high-precision 3D results for cultural preservation and virtual display.

[0113] Specifically, based on the 3D model of each sub-region, a set of boundary feature points is extracted to form a global boundary index, which includes: First, multiple sub-region 3D models are imported into a unified processing platform, and the boundary regions of each model are automatically detected. Boundary detection is performed by scanning the neighborhood distribution of points in the point cloud or mesh model. When a point lacks sufficient neighboring points in its neighborhood, it is marked as a boundary point.

[0114] Next, the boundary points of each sub-region are extracted, numbered and stored according to their coordinate positions to form a boundary point set.

[0115] Secondly, the boundary points of all sub-regions are aggregated to establish a global index structure. This index is equivalent to a global boundary map, providing a unified reference framework for subsequent alignment processing.

[0116] Specifically, based on the global boundary index, boundary alignment processing is performed, calculating the difference vector between boundary points of adjacent sub-regions to obtain alignment correction parameters, including: First, the boundary points of adjacent sub-regions in the global boundary index are matched. The matching method is to find points that are close to each other based on spatial proximity.

[0117] Next, the difference vector between these matching points is calculated. The difference vector reflects the spatial deviation between the two sub-regions at the boundary, including positional offset and directional difference.

[0118] Secondly, all the difference vectors are statistically analyzed and summarized to obtain a set of correction parameters. These parameters represent the translation and rotation amounts that need to be adjusted between sub-regions, and are the key basis for subsequent model stitching.

[0119] Specifically, based on alignment correction parameters, the boundary positions of the 3D models of the sub-regions are adjusted to generate a preliminary stitched model, including: First, the correction parameters obtained in the previous step are applied to the corresponding sub-region model to adjust its position and orientation, either as a whole or locally.

[0120] Next, the boundaries of adjacent sub-regions are corrected one by one, so that they gradually approach and overlap with the corresponding regions.

[0121] Secondly, the adjusted sub-regions were merged together to form a preliminary splicing model. Although there may still be slight discontinuities at the boundaries, the overall architectural form has been basically restored on a global scale.

[0122] Based on the preliminary splicing model, a multi-scale transition surface is constructed. This multi-scale transition surface consists of a large-scale overall smooth surface and a small-scale local detail surface, specifically including: First, the boundary transition area of ​​the preliminary splicing model is detected, and areas with obvious discontinuities are marked.

[0123] Next, multi-scale transition surfaces are generated in these regions. Large-scale surfaces form smooth transitions by fitting a large number of boundary points to eliminate macroscopic geometric abrupt changes; small-scale surfaces focus on detailed regions and maintain the original fine structure by fitting local point cloud features.

[0124] Secondly, large-scale and small-scale curved surfaces are combined to form a layered structure. This ensures the overall smoothness and continuity of the model while preserving local sculpted details or angular features.

[0125] Specifically, based on the multi-scale transition surface, a smoothing and blending process is performed on the stitching region of the initial stitching model to eliminate boundary discontinuities, resulting in a high-precision real-scene 3D reconstruction model, including: First, multi-scale transition surfaces are applied to the splicing area. Through the fusion calculation of curved surfaces and point clouds, the boundaries of adjacent sub-regions are naturally connected.

[0126] Next, the continuity of the merged splicing area is checked to ensure that there are no cracks, overlaps or abrupt transitions.

[0127] Secondly, the boundary areas are smoothed when necessary to ensure that the surface of the model presents a consistent appearance both visually and geometrically.

[0128] Ultimately, the resulting 3D model not only covers the overall structure of the architectural heritage but also possesses continuous and complete boundary features, providing reliable support for subsequent digital preservation, virtual display, and 3D analysis.

[0129] In a preferred embodiment of the present invention, based on the curvature division result, in the curvature abrupt change region, by real-time monitoring of the image grayscale change rate, the acquisition device is triggered to reduce its movement speed, increase its sampling frequency, and shorten it to a second sampling interval to obtain second sampled point cloud data, including: Based on the image data of the curvature abrupt change region, the pixel grayscale difference values ​​between consecutive image frames are extracted to generate a local grayscale difference matrix. The local gray-level difference matrix is ​​calculated over time to obtain the gray-level change rate per unit time, thus forming local texture change parameters; Based on the local texture change parameters, a high-detail region marking result is generated by comparing them with a preset change rate threshold. Based on the high detail region marking results, the motion control module of the acquisition device is controlled to perform adjustment processing to reduce the motion speed and increase the sampling frequency, and to generate updated motion and sampling parameters; Based on the updated motion and sampling parameters, the sampling interval is dynamically shortened to the second sampling interval, generating a dense point cloud acquisition command. According to the dense point cloud acquisition command, perform a local scan and generate a second sampled point cloud data.

[0130] In this embodiment of the invention, by monitoring the real-time grayscale changes of image data in areas of abrupt curvature changes, complex and detailed areas can be dynamically identified, and the acquisition device can be automatically triggered to adjust its acquisition strategy. Specifically, this involves reducing the movement speed, increasing the sampling frequency, and shortening the laser acquisition interval, thereby forming high-density point clouds in these areas. This dynamic response mechanism avoids the shortcomings of fixed-interval acquisition in complex structures, ensuring that more data points are acquired in areas such as carvings, arches, and patterns. For example, at the edge of a stone arch, due to drastic grayscale changes, the system automatically densifies the sampling, thus ensuring that the curved edge is continuous and complete in the point cloud model. Furthermore, this mechanism can significantly improve the data quality of high-detail areas without significantly increasing the overall acquisition time. The resulting second-sampled point cloud data is superior to traditional methods in terms of local detail reproduction, ensuring that the overall model is both efficient and detailed, making it particularly suitable for the digital reconstruction of complex architectural heritage.

[0131] Specifically, based on the image data of the curvature abrupt change region, pixel gray-level difference values ​​between consecutive image frames are extracted to generate a local gray-level difference matrix, which includes: First, the multi-frame images corresponding to the curvature abrupt change region are input into the image processing module. The degree of difference is recorded by comparing the pixel brightness values ​​in adjacent images frame by frame.

[0132] Next, the difference values ​​of each pixel are stored according to their spatial location to form a two-dimensional matrix. The values ​​in the matrix represent the grayscale variation of that point between different image frames.

[0133] The resulting local grayscale difference matrix can intuitively reflect which areas in the image have rapid brightness or texture changes, providing basic data for subsequent texture parameter calculations.

[0134] Specifically, time-series calculations are performed on the local gray-level difference matrix to obtain the gray-level change rate per unit time, forming local texture change parameters, including: First, the acquisition time of consecutive image frames is combined with the data in the gray-level difference matrix to calculate the gray-level change of each pixel per unit time.

[0135] Next, these changes are converted into a rate index, which measures how quickly the brightness value of a pixel changes over a continuous time period. The higher the rate of change, the more significant the texture or detail at that location.

[0136] The resulting local texture variation parameters can quantitatively represent the detail complexity of regions with abrupt curvature changes.

[0137] Specifically, based on local texture change parameters, a comparison is made with a preset change rate threshold to generate high-detail region marking results, including: First, the texture change parameters of each pixel are compared with a pre-set change rate threshold. This threshold, derived from experimental statistics, is used to distinguish between ordinary areas and high-detail areas.

[0138] Next, when the rate of change of a pixel exceeds a threshold, the pixel is marked as a high detail pixel; when the rate of change is below the threshold, it is marked as a normal pixel.

[0139] Finally, by summarizing the labeling results of all pixels, a spatial distribution map of the high-detail area is generated, providing a basis for adjusting the motion strategy of the subsequent acquisition device.

[0140] Specifically, based on the high-detail region marking results, the motion control module of the acquisition device is controlled to perform adjustments to reduce the motion speed and increase the sampling frequency, generating updated motion and sampling parameters, including: First, the high-detail area marking results are input into the motion control system of the acquisition device. The control system dynamically adjusts the motion path parameters based on the location of these areas.

[0141] Next, above the high detail area, the movement speed of the acquisition device is automatically reduced so that it can acquire more image frames per unit time.

[0142] At the same time, the camera's sampling frequency was increased, meaning more images were captured in the same amount of time to enhance data coverage.

[0143] The updated motion speed and sampling parameters are generated, enabling the acquisition device to adopt differentiated acquisition strategies in different areas.

[0144] Specifically, based on the updated motion and sampling parameters, the sampling interval is dynamically shortened to the second sampling interval, generating dense point cloud acquisition instructions, including: First, in the high detail region, the acquisition frequency of the acquisition device is controlled to shorten the time interval between adjacent laser pulses.

[0145] Next, by shortening the interval, more laser sampling points are generated per unit area, forming a higher density point cloud.

[0146] Finally, the control system generates a dense point cloud acquisition command, which includes coordinated parameters for motion speed, camera sampling frequency, and acquisition frequency to ensure the quality of data acquisition in high detail areas.

[0147] Specifically, according to the dense point cloud acquisition command, a local scan is performed to generate second sampled point cloud data, including: First, the acquisition device enters the high-detail area for scanning according to the dense point cloud acquisition instructions. At this time, its movement speed decreases, the camera sampling frequency increases, and the acquisition interval shortens.

[0148] Subsequently, the number of data points generated by the acquisition device in high-detail areas increased significantly, covering detailed areas that were originally easy to miss.

[0149] Finally, these newly collected data are organized into second sampling point cloud data, which complements the first sampling point cloud data of ordinary areas, ensuring that the overall modeling is both efficient and accurate.

[0150] In a preferred embodiment of the present invention, based on the potential missing region, and combining the normal consistency of the boundary points with the texture changes of the corresponding image, boundary continuity detection is performed to obtain the missing region, including: Based on the point cloud data of potential missing regions, extract the normal vectors of the boundary points to form a normal distribution dataset; Based on the normal distribution dataset, calculate the angle difference between adjacent normal vectors and generate a direction consistency index matrix; Based on the directional consistency index matrix, boundary points with included angles greater than a threshold are selected to obtain a set of abnormal boundary points; Based on the set of abnormal boundary points, map them to the corresponding image regions, extract the image texture gradient directions, and form a texture direction distribution dataset. Based on the texture orientation distribution dataset, a joint comparison is performed with the normal consistency index to generate texture normal matching results; Based on the texture normal matching degree, when the matching degree is lower than the preset matching threshold, the corresponding region is determined to be a missing region, and the boundary contour of the missing region is generated.

[0151] In this embodiment of the invention, by combining point cloud geometric information with image texture features for joint detection, the accuracy of missing region identification can be effectively improved. First, by calculating the normal consistency of the boundary points of potential missing regions, the locations of abrupt normal changes can be preliminarily screened, indicating possible data breaks. Then, by comparing with the texture gradient direction in the image, it is further confirmed whether it is a true missing region. This dual-determination method can significantly reduce the probability of misjudgment. For example, when the point cloud on the surface of a stone tablet is sparse due to material reflection, but the image texture remains continuous, the system will not misjudge it as a missing region. Conversely, in areas obscured by vaults or window frames, if the normal is discontinuous and breaks appear in the image texture, the missing region can be accurately identified. The resulting boundary contour of the missing region is not only accurate but also provides a clear range for the generation of compensation points, making subsequent repair more efficient and reliable. Overall, this method demonstrates high robustness and accuracy in the complex environment of architectural heritage.

[0152] Specifically, based on the point cloud data of potentially missing regions, the normal vectors of boundary points are extracted to form a normal vector distribution dataset, which includes: First, the point cloud data of the potentially missing regions is imported into the point cloud processing system. By selecting several neighboring points within the neighborhood of each point, the surface orientation direction of that point, i.e., the normal vector, is estimated using the spatial distribution trend of these points.

[0153] Next, perform the above calculations on all boundary points within the region, and record the resulting normal vectors into a unified dataset.

[0154] This dataset can reflect the overall directional characteristics of boundary points in space, providing basic data for subsequent consistency detection.

[0155] Specifically, based on the normal distribution dataset, the angle difference between adjacent normal vectors is calculated to generate a direction consistency index matrix, which includes: First, in the normal distribution dataset, the normal vector directions of adjacent points are compared pairwise. If the surface directions of two points are similar, the difference in their included angle is small; if the surface directions differ greatly, the difference in their included angle is significant.

[0156] Next, the results of these angular differences are converted into directional consistency indices and stored in matrix form according to spatial location.

[0157] This index matrix can intuitively show the degree of consistency of boundary points in a certain area in spatial direction, providing a quantitative basis for subsequent identification of abnormal boundary points.

[0158] Specifically, based on the direction consistency index matrix, boundary points with included angles greater than a threshold are selected to obtain a set of abnormal boundary points, which includes: First, a threshold is preset to distinguish between normal continuous boundaries and abnormal mutation boundaries.

[0159] Next, the values ​​in the direction consistency index matrix are compared with the threshold point by point. If the angle difference exceeds the threshold, it is considered that there is an abnormal direction deviation at that point.

[0160] Finally, these outliers are extracted to form a set of outlier boundary points, providing candidate regions for subsequent image data comparison.

[0161] Specifically, based on the set of abnormal boundary points, the corresponding image regions are mapped, and the image texture gradient directions are extracted to form a texture direction distribution dataset, which includes: First, the positions of the abnormal boundary points in the 3D point cloud are projected onto the corresponding 2D image to determine the image region in which they are located.

[0162] Next, the directionality of the image grayscale values ​​is detected within this region. By analyzing the direction of grayscale changes, the direction of the texture gradient is extracted.

[0163] Secondly, the extracted texture orientation data are organized into a texture orientation distribution dataset. This dataset can be correlated with normal information to help determine the authenticity of point cloud missing data.

[0164] Specifically, based on the texture orientation distribution dataset, a joint comparison is performed with the normal consistency index to generate texture normal matching results, including: First, the texture direction is compared point by point with the direction of the normal vector. If the two directions are consistent or close, it means that the point cloud in that area matches the image features; if the difference is significant, it means that there is a potential missing point cloud in that area.

[0165] Next, the comparison results of all points are summarized to generate an overall texture normal matching result, which is used to quantify the degree of consistency between the image and the point cloud.

[0166] This result provides an objective basis for subsequent determination of missing regions, avoiding misjudgments caused by relying on a single data source.

[0167] Specifically, based on the texture normal matching degree result, when the matching degree is lower than a preset matching threshold, the corresponding region is determined to be a missing region, and the boundary contour of the missing region is generated, including: First, set a matching threshold to distinguish between normal regions and missing regions.

[0168] Next, the texture normal matching results are compared, and when the matching value at a certain location is lower than the threshold, it is marked as a missing region.

[0169] Secondly, within the missing region, its boundary points are further traced, and a continuous missing boundary contour is generated.

[0170] The resulting boundary of the missing region not only accurately locates the missing area of ​​the point cloud, but also provides a clear spatial range for the subsequent generation of compensation points.

[0171] In a preferred embodiment of the present invention, based on the projection compensation parameters and combined with the imaging model of the image acquisition device and its spatial position and posture parameters during the acquisition process, the edge contour and grayscale gradient information are projected onto the three-dimensional spatial position of the missing area to generate a compensation point dataset, including: Based on the image fragments of the missing region, perform edge detection, extract edge pixels, and form an edge contour dataset; Based on the edge contour dataset, calculate the grayscale gradient value of each edge pixel and generate a texture gradient vector field; Based on the texture gradient vector field, combined with the imaging model of the image acquisition device and its spatial position and attitude parameters during the acquisition process, the projection path parameters of the pixel in three-dimensional space are generated. Based on the projection path parameters, the edge contour and texture gradient are mapped point by point to the three-dimensional space to generate a preliminary set of compensation points; Based on the initial set of compensation points, weights are assigned to each compensation point according to the gray-level gradient magnitude to generate a weighted compensation point dataset. Based on the weighted compensation point dataset, it is fused with the first sub-region point cloud model to form the second sub-region point cloud model.

[0172] In this embodiment of the invention, by utilizing the edge contours and grayscale gradient information of the image, and combining them with the imaging model and pose parameters for 3D projection, reasonably positioned compensation points can be generated in the missing areas. This process ensures that the features of the 2D image can be accurately mapped to 3D space, avoiding compensation errors caused by coordinate system mismatch. For example, in the area at the top of a towering stone tower, even with insufficient laser data, correct compensation points can be generated through image projection. Next, by assigning weights to the compensation points based on the grayscale gradient magnitude, boundary and texture features can be highlighted, making the compensation points more closely resemble the real surface. Finally, fusing the weighted compensation point data with the first sub-region point cloud model can repair the missing areas while maintaining a natural transition, ensuring overall realism and avoiding surface abruptness caused by over-compensation. The final generated second sub-region point cloud model possesses continuity and integrity, laying a solid foundation for subsequent global stitching and 3D reconstruction.

[0173] Specifically, based on the image fragments of the missing region, edge detection is performed to extract edge pixels and form an edge contour dataset, which includes: First, the original image segments corresponding to the missing regions are imported into the image processing module for preprocessing, including noise filtering and brightness equalization, to ensure the accuracy of subsequent feature extraction.

[0174] Next, edge detection algorithms are used to identify pixels in the image that show significant changes in brightness or texture. These pixels are usually located at the contour edges or texture boundaries of the building surface.

[0175] Finally, the detected edge pixels are connected and grouped according to their positions in the image to form an edge contour dataset, which provides a basis for subsequent grayscale gradient calculation.

[0176] Specifically, based on the edge contour dataset, the grayscale gradient values ​​of each edge pixel are calculated, and a texture gradient vector field is generated, including: First, in the edge contour dataset, the neighborhood region of each edge pixel is scanned to measure the magnitude and direction of its grayscale value change.

[0177] Next, the magnitude of the grayscale change is used as the gradient magnitude, and the direction of change is used as the gradient direction to generate corresponding gradient features for each edge point.

[0178] Ultimately, the gradient magnitude and direction of all edge points together constitute the texture gradient vector field, which can reflect the texture direction and detail distribution of the edge region in the image.

[0179] Specifically, based on the texture gradient vector field, and combined with the imaging model of the image acquisition device and its spatial position and orientation parameters during the acquisition process, the projection path parameters of the pixels in three-dimensional space are generated, including: First, the imaging model parameters of the acquisition device are called, including focal length, imaging scale factor and lens distortion correction parameters, to establish the mapping relationship between the two-dimensional image and the three-dimensional space.

[0180] Next, the spatial position and attitude of the equipment during the image acquisition process (such as position coordinates, pitch angle, yaw angle and roll angle) are input into the mapping model to ensure that the image features correspond to the actual spatial scene.

[0181] Secondly, for each edge point in the texture gradient vector field, combined with the imaging model and pose parameters, its projection path parameters in 3D space are generated. These parameters describe the possible 3D spatial trajectory of each image point, providing a directional basis for subsequent point cloud compensation.

[0182] Specifically, based on the projection path parameters, the edge contours and texture gradients are mapped point-by-point to three-dimensional space to generate a preliminary set of compensation points, including: First, extend each image edge point along its projection path to find a position that matches the boundary of the missing area, and generate a compensation point at that position.

[0183] Next, during the mapping process, the direction of the gray-level gradient and the spatial location of the compensation points are considered simultaneously, so that the distribution of the compensation points not only conforms to the spatial geometric structure, but also remains consistent with the texture information of the image.

[0184] Ultimately, the set of compensation points generated by all mappings constitutes the preliminary compensation point data, filling in the three-dimensional spatial coordinates of the missing regions.

[0185] Specifically, based on the initial set of compensation points, weights are assigned to each compensation point according to the gray-level gradient magnitude to generate a weighted compensation point dataset, which includes: First, for each compensation point in the initial compensation point set, read the magnitude of its corresponding grayscale gradient in the image. The larger the gradient magnitude, the more prominent the texture or boundary features of that point in the image.

[0186] Next, the grayscale gradient magnitude is converted into a weight value, assigning different importance to the compensation points. For example, points with larger gradient magnitudes have higher weights and are more representative in subsequent fusion; while points with smaller gradient magnitudes have lower weights and only play an auxiliary role.

[0187] Ultimately, the resulting weighted compensation point dataset retains key texture and edge features from the image while avoiding model distortion caused by an excessive number of low-quality points.

[0188] Specifically, the weighted compensation point dataset is fused with the first sub-region point cloud model to form the second sub-region point cloud model, which includes: First, the weighted compensation point dataset is imported into the point cloud processing system and its coordinate system is unified with that of the first sub-region point cloud model to ensure that the two types of point cloud data are superimposed under the same spatial reference.

[0189] Next, during the fusion process, the system prioritizes retaining the original point cloud and supplements the missing regions with compensation points according to their weight allocation principles. When there is a conflict between compensation points and original points, the points with higher weights are retained to ensure the final data quality.

[0190] Secondly, smoothing and consistency processing are performed on the merged region to ensure a natural transition between the compensation points and the original points, avoiding geometric abruptness.

[0191] Ultimately, the resulting second sub-region point cloud model repaired the missing areas, significantly improving both overall continuity and local detail integrity, providing a solid data foundation for subsequent sub-region stitching and global modeling.

[0192] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for high-precision real-scene 3D reconstruction of architectural heritage, characterized in that, The method includes: Acquire multi-angle image data of the exterior and interior of the architectural heritage, and divide the area to be reconstructed into multiple sub-areas according to the overall structure of the architectural heritage to obtain image data; Based on the image data, point cloud data of each sub-region is acquired using acquisition equipment. The first sampling interval is used in the regular area, and the second sampling interval is dynamically shortened in the carved depression or narrow space area. Sub-region point cloud data is generated through multi-path cross acquisition. The point cloud data of the sub-region is coarsely registered, and iterative correction is performed using the feature corresponding points of the overlapping area to generate the point cloud model of the first sub-region. The point cloud model of the first sub-region is subjected to quality inspection to identify the point cloud density distribution and integrity, and to determine the missing regions where the point cloud density does not meet the density judgment criteria. In the missing region, local edge contours and gray-level gradient information are extracted by combining image data, and the information is projected onto the missing region to generate compensation points, thereby generating a second sub-region point cloud model. The point cloud model of the second sub-region is subjected to weighted fusion processing, and surface interpolation is performed within the fusion region to form a 3D model of the sub-region. The 3D models of the sub-regions are aligned by boundary features and stitched together globally. A smooth fusion is achieved through multi-scale transition surfaces, resulting in a high-precision real-scene 3D reconstruction model that covers the entire architectural heritage.

2. The method for high-precision real-scene 3D reconstruction of architectural heritage according to claim 1, characterized in that, Based on the image data, point cloud acquisition was performed on each sub-region using acquisition equipment, including: Based on the acquired image data, the curvature features of the building surface are extracted, and the sub-regions are divided into regions with gentle curvature and regions with abrupt curvature changes, generating curvature division results; Based on the curvature division results, in the region with gentle curvature, the acquisition device is controlled to acquire images at a first motion speed and a first sampling interval to obtain the first sampled point cloud data. Based on the curvature division results, in the curvature abrupt change region, by real-time monitoring of the image grayscale change rate, the acquisition device is triggered to reduce its movement speed, increase the sampling frequency, and shorten it to the second sampling interval to obtain the second sampling point cloud data; Based on the obstacle avoidance sensor feedback of the acquisition device in a narrow space, the attitude and acquisition angle of the acquisition device are adjusted, and compensation acquisition is performed on the occluded area to generate compensated point cloud data. The first sampled point cloud data, the second sampled point cloud data, and the compensated point cloud data are merged, and redundant points are weighted using the overlapping data generated by multi-path cross-scanning to obtain the original point cloud data of the sub-region.

3. The method for high-precision real-scene 3D reconstruction of architectural heritage according to claim 1, characterized in that, Coarse registration is performed on the sub-region point cloud data, and iterative correction is performed using the feature corresponding points of the overlapping areas to generate the first sub-region point cloud model, including: Based on the original point cloud data of the sub-region, calculate the geometric center and normal vector distribution to form the initial pose parameters; Based on the initial attitude parameters, coarse registration is performed on the point clouds collected from different paths to obtain preliminarily aligned sub-region point cloud data; Based on the preliminarily aligned sub-region point cloud data, feature points of overlapping regions are extracted. These feature points are then filtered based on normal vector consistency and local curvature extrema to obtain corresponding point pairs. Based on the corresponding point pairs, calculate the Euclidean distance and angular differences between the points to form the error distribution results; Based on the error distribution results, iterative correction processing is performed to gradually update the attitude parameters until the overall registration error is lower than the preset error threshold, thus obtaining the point cloud model of the first sub-region.

4. The method for high-precision real-scene 3D reconstruction of architectural heritage according to claim 1, characterized in that, The point cloud model of the first sub-region is subjected to quality inspection to identify the point cloud density distribution and integrity, and to determine the missing regions where the point cloud density does not meet the density judgment criteria, including: Based on the point cloud model of the first sub-region, a spatial partitioning structure based on a three-dimensional mesh is established to obtain the point cloud density matrix; Based on the point cloud density matrix, the number of points in each grid cell is compared with the global average to form low-density candidate regions; Based on the low-density candidate regions, the average distance between points is calculated and compared with the global average distance to generate potential missing regions. Based on the potential missing regions, and combining the normal consistency of the boundary points with the texture changes of the corresponding images, boundary continuity detection is performed to obtain the missing regions.

5. The method for high-precision real-scene 3D reconstruction of architectural heritage according to claim 1, characterized in that, In the missing region, local edge contours and grayscale gradient information are extracted by combining image data, and the information is projected onto the missing region to generate compensation points, thus generating a second sub-region point cloud model, including: Based on the missing area, extract the corresponding image data segment to obtain the missing area image input; Based on the image input of the missing region, the local edge contour is extracted using an edge detection algorithm, and the texture gradient information is extracted based on the gray-level gradient change rate to form the edge gradient corresponding feature data; Based on the feature data corresponding to the edge gradient, the contour direction and gray-level distribution trend of the missing region are calculated to obtain the projection compensation parameters. Based on the projection compensation parameters, combined with the imaging model of the image acquisition device and its spatial position and attitude parameters during the acquisition process, the edge contour and gray-level gradient information are projected onto the three-dimensional spatial position of the missing area to generate a compensation point dataset. Based on the compensation point dataset, it is merged with the first sub-region point cloud model to obtain the second sub-region point cloud model.

6. The method for high-precision real-scene 3D reconstruction of architectural heritage according to claim 1, characterized in that, The point cloud model of the second sub-region is subjected to weighted fusion processing, and surface interpolation is performed within the fusion region to form a 3D model of the sub-region, including: Based on the point cloud model of the second sub-region, the overlapping regions between adjacent sub-regions are extracted to form a set of overlapping region point clouds. Based on the point cloud set of the overlapping region, the density index and acquisition path confidence of each point are calculated to obtain the weighted parameter matrix; Based on the weighted parameter matrix, the point clouds of the overlapping regions are weighted and fused to generate fused point cloud data; Based on the fused point cloud data, surface interpolation is performed in the sparse region of the point cloud, and an interpolation strategy based on the consistency of the normal of neighboring points is adopted to generate interpolated surface data. Based on the fused point cloud data and interpolated surface data, the mesh structure of the sub-region 3D model is constructed, and surface continuity optimization is performed to obtain the sub-region 3D model.

7. The method for high-precision real-scene 3D reconstruction of architectural heritage according to claim 1, characterized in that, The 3D models of the sub-regions are aligned by boundary features and globally stitched together. Smooth fusion is achieved through multi-scale transition surfaces, resulting in a high-precision, real-world 3D reconstruction model covering the entire architectural heritage site. This includes: Based on the 3D model of each sub-region, extract the set of boundary feature points to form a global boundary index; Based on the global boundary index, perform boundary alignment processing, calculate the difference vector between boundary points of adjacent sub-regions, and obtain alignment correction parameters; Based on the alignment correction parameters, adjust the boundary positions of the 3D model of the sub-region to generate a preliminary stitched model; Based on the preliminary splicing model, a multi-scale transition surface is constructed, which consists of a large-scale overall smooth surface and a small-scale local detail surface. Based on the multi-scale transition surface, a smooth fusion process is performed on the splicing area of ​​the preliminary splicing model to eliminate boundary discontinuities and obtain a high-precision real-scene 3D reconstruction model.

8. The method for high-precision real-scene 3D reconstruction of architectural heritage according to claim 2, characterized in that, Based on the curvature division results, in regions of abrupt curvature change, by real-time monitoring of the image grayscale change rate, the acquisition device is triggered to reduce its movement speed, increase its sampling frequency, and shorten the interval to a second sampling interval, thus obtaining the second sampled point cloud data, including: Based on the image data of the curvature abrupt change region, the pixel grayscale difference values ​​between consecutive image frames are extracted to generate a local grayscale difference matrix. The local gray-level difference matrix is ​​calculated over time to obtain the gray-level change rate per unit time, thus forming local texture change parameters; Based on the local texture change parameters, a high-detail region marking result is generated by comparing them with a preset change rate threshold. Based on the high detail region marking results, the motion control module of the acquisition device is controlled to perform adjustment processing to reduce the motion speed and increase the sampling frequency, and to generate updated motion and sampling parameters; Based on the updated motion and sampling parameters, the sampling interval is dynamically shortened to the second sampling interval, generating a dense point cloud acquisition command. According to the dense point cloud acquisition command, perform a local scan and generate a second sampled point cloud data.

9. The method for high-precision real-scene 3D reconstruction of architectural heritage according to claim 4, characterized in that, Based on the potential missing regions, and combining the normal consistency of boundary points with the texture changes of the corresponding images, boundary continuity detection is performed to obtain the missing regions, including: Based on the point cloud data of potential missing regions, extract the normal vectors of the boundary points to form a normal distribution dataset; Based on the normal distribution dataset, calculate the angle difference between adjacent normal vectors and generate a direction consistency index matrix; Based on the directional consistency index matrix, boundary points with included angles greater than a threshold are selected to obtain a set of abnormal boundary points; Based on the set of abnormal boundary points, map them to the corresponding image regions, extract the image texture gradient directions, and form a texture direction distribution dataset. Based on the texture orientation distribution dataset, a joint comparison is performed with the normal consistency index to generate texture normal matching results; Based on the texture normal matching degree, when the matching degree is lower than the preset matching threshold, the corresponding region is determined to be a missing region, and the boundary contour of the missing region is generated.

10. The method for high-precision real-scene 3D reconstruction of architectural heritage according to claim 5, characterized in that, Based on the projection compensation parameters, combined with the imaging model of the image acquisition device and its spatial position and orientation parameters during the acquisition process, the edge contours and grayscale gradient information are projected onto the three-dimensional spatial position of the missing area, generating a compensation point dataset, including: Based on the image fragments of the missing region, perform edge detection, extract edge pixels, and form an edge contour dataset; Based on the edge contour dataset, calculate the grayscale gradient value of each edge pixel and generate a texture gradient vector field; Based on the texture gradient vector field, combined with the imaging model of the image acquisition device and its spatial position and attitude parameters during the acquisition process, the projection path parameters of the pixel in three-dimensional space are generated. Based on the projection path parameters, the edge contour and texture gradient are mapped point by point to the three-dimensional space to generate a preliminary set of compensation points; Based on the initial set of compensation points, weights are assigned to each compensation point according to the gray-level gradient magnitude to generate a weighted compensation point dataset. Based on the weighted compensation point dataset, it is fused with the first sub-region point cloud model to form the second sub-region point cloud model.

Images