A three-dimensional point cloud reconstruction method based on geometric consistency probability fusion

Abstract

The application provides a three-dimensional point cloud reconstruction method based on geometric consistency probability fusion; three-view data are acquired and normalized and spatially aligned; three-dimensional candidate points are generated based on the projection correspondence relationship between the views; an adaptive probability fusion model is constructed to evaluate the confidence of the matching error, local sampling density and neighborhood support relationship of the three-dimensional candidate points; the three-dimensional candidate points are screened according to the confidence and topological consistency to generate initial point clouds; curve structures are recovered based on geometric priors and view topological constraints; and point cloud data are output; the geometric consistency probability fusion mechanism is used to improve the accuracy and integrity of the three-dimensional point cloud reconstruction, and the method is suitable for automatic reconstruction of three-dimensional models from engineering three-view data.

Description

Technical Field

[0001] This invention belongs to the field of computer vision, and specifically relates to a method for reconstructing three-dimensional point clouds based on geometric consistency probabilistic fusion. Background Technology

[0002] With the development of computer-aided design, intelligent manufacturing, digital twins and 3D vision technology, the reconstruction of 3D models based on 2D engineering drawings has become an important research direction in mechanical design, industrial inspection, reverse engineering and digital modeling. In practical engineering applications, a large number of historical parts, mechanical structures and industrial drawings are still saved in the form of 2D three-view drawings. Therefore, how to quickly and accurately recover the 3D spatial structure from the 2D views has important theoretical significance and engineering value.

[0003] Existing 3D reconstruction methods typically establish spatial point matching by establishing the projection correspondence between the front view, top view, and side view, thereby recovering the 3D point cloud or wireframe structure. However, most traditional methods use fixed threshold matching or simple geometric constraint mechanisms, lacking a comprehensive evaluation of the reliability of matching between different views. When there are uneven sampling density, missing local views, complex curve structures, or projection ambiguities, problems such as incorrect matching, increased false points, and structural breaks are easily generated, resulting in poor stability of the reconstruction results.

[0004] Therefore, there is still a lack of a 3D point cloud reconstruction method that can integrate multi-view geometric consistency, adaptive probability evaluation, and topological constraints to improve the accuracy of 3D candidate point selection, enhance curve structure recovery capability, and improve the integrity and reliability of the overall 3D reconstruction results. Summary of the Invention

[0005] This invention proposes a 3D point cloud reconstruction method based on geometric consistency probabilistic fusion. The method involves acquiring three-view data and performing normalization and spatial alignment; generating 3D candidate points based on the projection correspondence between views; constructing an adaptive probabilistic fusion model to evaluate the confidence of the 3D candidate points based on matching error, local sampling density, and neighborhood support relationships; filtering 3D candidate points based on confidence and topological consistency to generate an initial point cloud; recovering the curve structure based on geometric priors and view topological constraints; and outputting the point cloud data. This invention improves the accuracy and completeness of 3D point cloud reconstruction through a geometric consistency probabilistic fusion mechanism, and is suitable for the automatic reconstruction of engineering three-view drawings into 3D models.

[0006] This invention proposes a 3D point cloud reconstruction method based on geometric consistency probabilistic fusion; comprising the following steps:

[0007] S1. Acquisition and parsing of three-view data; acquire the two-dimensional view data of the object to be reconstructed, and obtain the front view, side view and top view data by layer name;

[0008] S2. View Normalization and Spatial Alignment; To align the three views under a unified reference coordinate system, the obtained three views are... Figure 2 The coordinates of the points in each dimension are normalized.

[0009] S3. Generate 3D candidate points based on the geometric consistency of the three views; utilize the spatial projection correspondence between the three views to perform cross-view matching on the 2D sampling points to generate candidate 3D points;

[0010] S4. Construct an adaptive probabilistic fusion model to evaluate the confidence of 3D candidate points, including matching error, local sampling density, neighborhood support relationship and dynamic weight allocation mechanism;

[0011] S5. Based on confidence level and topological neighborhood consistency, three-dimensional candidate points are filtered to generate an initial point cloud;

[0012] S6. Based on geometric priors, view topological constraints, and structural continuity, the curve structure is restored;

[0013] S7. Point cloud output; Output the obtained point cloud data and visualize it to obtain a ply file.

[0014] Preferably, in step S1, the three-view data is acquired and parsed, characterized in that: Read the DXF file, iterate through all views in the model space from the DXF file, then adaptively sample the obtained views, obtain sample points, and label them; front view. Top view Side view Where x represents the horizontal coordinate, y represents the vertical coordinate, and z represents the depth coordinate; In addition, type identification and classification are required: LINE, LWPOLYLINE, CIRCLE, and ARC; for LWPOLYLINE entities, determine whether they are closed shapes based on their closed attribute; for ARC entities, perform angle normalization processing for abnormal start and end angles. The adaptive sampling method is as follows: For a straight line segment, the sampling interval is proportional to the length of the line segment, and the number of sampling points... ,in The length of the line segment. The preset sampling interval; For circles, curvature adaptive sampling is used, and the number of sampling points is... ,in This is the curvature gain coefficient. It is the radius of the circle, n is clipped to interval; For circular arcs, a dual adaptive sampling method is used, considering both arc length and curvature, with a sampling point count of [number missing]. ,in Let the arc length be , Arc length weighting coefficient, This is the curvature gain coefficient; Meanwhile, the original connection order between sampling points of the same type is maintained during the sampling process, serving as local topological edges; sampling points with the same geometric features between different views establish cross-view topological associations through type and parameter indexes, forming a geometrically consistent topological graph structure oriented towards 3D reconstruction.

[0015] Preferably, in step S1, an adaptive sampling method is used to dynamically adjust the sampling density according to the geometric features of different primitive types. A sampling strategy proportional to the length is adopted for straight line segments, and a curvature adaptive mechanism is introduced for circles and arcs to automatically densify sampling points where the curvature is large, effectively preserving geometric details. At the same time, the closure judgment of polylines and the normalization of arc angles ensure the accuracy of data parsing. The ignoring of unrecognizable entities improves the robustness of the method.

[0016] Preferably, in step S2, view normalization and spatial alignment are performed, characterized in that: Perform coordinate normalization on the front view, top view, and side view respectively, so that the coordinate origin of the three views is unified to the same reference system; The normalization process specifically involves calculating the minimum X-coordinate of all sampled points in each view. minimum value of Y coordinate minimum Z-coordinate Perform coordinate transformation on each sampling point in each view: For the front view, perform the transformation: , This ensures that the origin of the front view coordinate system is located at the lower left corner of the view, with the positive Y-axis pointing vertically upwards. For the side view, perform the transformation: , This ensures that the origin of the side view coordinate system is located at the lower left corner of the view, with the positive Y-axis pointing vertically upwards. For the top view, perform the transformation: , This ensures that the origin of the top view coordinate system is located at the upper left corner of the view, with the positive X-axis pointing horizontally to the right. Spatial alignment further includes: after normalization, the front view's These correspond to (X,Y) coordinates in three-dimensional space, respectively, and are shown in the top view. These correspond to (X, Z) coordinates in three-dimensional space, respectively, and are the side view coordinates. These correspond to the (Z, Y) coordinates in three-dimensional space, enabling the coordinate alignment of the three views within the same three-dimensional space reference system.

[0017] Preferably, in step S2, by performing coordinate normalization on the front view, top view, and side view respectively, the origin of each view coordinate is unified to the same reference system, eliminating the coordinate inconsistency problem caused by the offset of the view position in the original data; the correspondence between the coordinate axes of each view and the three-dimensional space coordinates is clarified, so that the three views can achieve precise alignment under a unified three-dimensional reference system, laying the coordinate foundation for subsequent cross-view matching and three-dimensional candidate point generation based on geometric consistency.

[0018] Preferably, in step S3, three-dimensional candidate points are generated based on the geometric consistency of the three views, characterized in that: Using the set of sample points in the front view as a reference, iterate through each sample point in the front view. Geometric consistency matching is performed with the top view sampling point set and the side view sampling point set respectively to generate three-dimensional candidate points; Geometric consistency matching specifically includes the following sub-steps: S31. Top view matching: For the front view In the top view sampling point set Searching for the content that satisfies All candidate points, of which A preset tolerance threshold is set for the X-direction; depth coordinates are provided for each successfully matched top viewpoint. Forming two-dimensional candidate pairs ; S32, Side view verification: For each generated candidate pair In the side view sampling point set Searching for the content that satisfies and ,in A preset tolerance threshold is set; if a side view point that meets the conditions exists, the candidate pair is determined to pass the side view verification. S33, Generation of 3D Candidate Points: For combinations that simultaneously satisfy top view matching and side view verification, generate 3D candidate points. Record the three-view matching information of the 3D candidate point, including: the coordinates of the front view point, the coordinates of the matching top view point, and the coordinates of the matching side view point; S34. Missing view fault handling: When the top view is successfully matched but the side view has no matching point, the 3D candidate point is marked as "side view missing" and its confidence weight is reduced, but it is still retained as a 3D candidate point. When the top view has no matching point, the generation of candidate points for the front view point is skipped, and no 3D candidate point is generated. Through the above steps, a set of 3D candidate points with matching error information is generated to provide input for subsequent confidence calculation.

[0019] Preferably, in step S3, cross-view matching is performed based on the front view, using the top view to provide depth information and the side view for dual verification, forming a "matching-verification" mechanism that effectively ensures the geometric consistency of the 3D candidate points; by setting a tolerance threshold... and It allows for a certain matching error, enhancing the method's tolerance to data noise and sampling bias; it introduces missing view fault tolerance processing, which retains 3D candidate points and reduces their weights when the side view matching is missing, thus avoiding reconstruction loopholes caused by incomplete information from a single view.

[0020] Preferably, in step S4, an adaptive probabilistic fusion model is constructed to evaluate the confidence level of the three-dimensional candidate points, characterized in that: For each generated 3D candidate point Calculate the confidence score, which measures the degree to which the 3D candidate point is supported by all three views. The confidence assessment specifically includes the following steps: S41. Single-view matching error calculation: Calculate the top view matching error: ,in The X coordinates of the matched top-down viewpoint; Calculate the X-direction matching error of the side view: ,in The Z-coordinate of the matched side viewpoint; Calculate the Y-direction matching error of the side view: ,in The Y coordinate of the matched side viewpoint; S42. Matching error-confidence component conversion: The matching error is converted into a confidence component using an exponential decay function: ,in The X-axis scale parameter; ,in The Z-axis scale parameter; ,in The Y-axis scale parameter; S43, Adaptive Weight Allocation: The weights of each confidence component are dynamically adjusted based on the sampling density and entity type of the three views: Weight The weight is determined based on the local sampling density of the front view; the denser the sampling, the higher the weight. ;in This refers to the local sampling density of the front view. The strength of the support in the neighborhood of the top view; The strength of the support in the neighborhood of the side view; , The number of sampling points in the neighborhood. This represents the area of ​​the corresponding neighborhood. ,in The neighborhood support strength of the corresponding view. The spatial distance between neighboring points and the target point. The neighborhood radius, For indicator functions, when the condition Output 1 if true, otherwise output 0; Weight Determined based on the number of neighboring support points of the matching point in the top view; ; Weight Determined based on the number of neighboring support points of the matching point in the side view; ; S44. Confidence Calculation: The confidence level is calculated using the following formula: ,in The local confidence level of the front view sampling points; S45, Adaptive adjustment for missing views: When the side view is missing, the weights are redistributed. , The confidence level calculation formula has been adjusted to and multiplied by the attenuation coefficient .

[0021] Preferably, in step S4, an exponential decay function is used to convert the matching error into a confidence component, which can smoothly reflect the impact of the error on the confidence and avoid the abrupt change problem caused by hard thresholds; by integrating local sampling density and neighborhood support strength for adaptive weight allocation, the confidence assessment is more in line with the distribution characteristics of the data itself; a missing view adaptive adjustment mechanism is introduced, which automatically reallocates weights and multiplies them by a decay coefficient when information of a certain view is missing, which not only preserves effective information but also reduces the impact of unreliable 3D candidate points.

[0022] Preferably, in step S5, the three-dimensional candidate points are screened based on confidence level and topological neighborhood consistency, characterized in that: Based on the confidence score calculated in step S4, the set of 3D candidate points generated in step S3 is filtered and fused to generate an initial 3D point cloud. The confidence score filtering and initial point cloud generation specifically include the following steps: S51, Dual Threshold Filtering: Set a high confidence threshold and low confidence threshold ,in Regarding confidence level The three-dimensional candidate points are directly marked as "high confidence points" and retained; for the confidence level... The 3D candidate points are marked as "medium confidence points" and must pass the local consistency test before being retained; for confidence... The three-dimensional candidate points are directly eliminated; S52, Neighborhood Consistency Check: For points with medium confidence, construct their spatial neighborhood. ,in The neighborhood radius, For other points in the surrounding area, For the current target point, Given the Euclidean distance, count the number of high-confidence points in the neighborhood. ;like If a confidence point passes the neighborhood consistency test, it is retained; otherwise, it is removed. The minimum neighbor threshold; S53, Point Cloud Fusion and Optimization: All high-confidence points and medium-confidence points that passed the test are merged into an initial point cloud set; spatial deduplication is performed on the initial point cloud: duplicate points are merged and deleted; outlier filtering is performed on the deduplicated point cloud: the average neighborhood distance of each point is calculated and outliers with average neighborhood distances that deviate significantly from the overall distribution are removed. S54. Topological connectivity constraints: Using the topology graph structure constructed in step S1, each retained point in the point cloud must have at least one neighboring point also retained in the topology graph of its corresponding view; for isolated points, a secondary check is performed: if its confidence level... If so, keep it; otherwise, discard it. Through the above steps, an initial 3D point cloud is generated after confidence filtering, neighborhood consistency constraints, point cloud fusion and optimization, and topological connectivity constraints.

[0023] Preferably, in step S5, a dual threshold screening strategy is adopted to divide the three-dimensional candidate points into three categories: high, medium, and low. High-confidence points are directly retained, low-confidence points are directly eliminated, and medium-confidence points are further identified through neighborhood consistency test, thereby achieving refined screening of three-dimensional candidate points. The neighborhood consistency test uses the spatial distribution of high-confidence points to verify the reliability of medium-confidence points, enhancing the robustness of the screening. Combined with spatial deduplication, outlier filtering, and topological connectivity constraints, redundant points and isolated noise points are effectively eliminated.

[0024] Preferably, in step S6, the curve structure is restored based on geometric priors, view topological constraints, and structural continuity, characterized in that: The completion of curve and arc structures specifically includes the following steps: S61. Circular Arc Entity Recognition and Classification: Iterate through all CIRCLE and ARC type entities identified in step S1; classify them into three categories based on their view labels: front view arcs, top view arcs, and side view arcs; for each arc entity, extract its geometric parameters: center coordinates. Radius R, start and end angles , ; S62, View-adaptive 3D enhancement: Front view arc elevation: Curvature adaptive sampling is used to perform two-dimensional sampling of the arc, generating a set of sampling points. For each sampling point Searching for the X coordinate in the initial point cloud generated by S5. For each point, obtain the corresponding Z-coordinate candidate set; take the mode of the Z-coordinate candidate set as the depth value of that point. Generate 3D points ; Top view arc elevation: Perform two-dimensional sampling on the arc to generate a set of sampling points. ,in Corresponding to the Z coordinate in three-dimensional space; for each sampling point Searching for the X coordinate in the initial point cloud. And the Z coordinate is For each point, obtain the corresponding Y-coordinate candidate set; take the mode of the Y-coordinate candidate set as the height value of that point. Generate 3D points ; Side view arc lifting: Perform two-dimensional sampling on the arc to generate a set of sampling points. ,in Corresponding to the Z coordinate in three-dimensional space, Corresponding to the Y coordinate; for each sampling point Searching for the Z coordinate in the initial point cloud is And the Y coordinate is For each point, obtain the corresponding candidate set of X-coordinates; take the mode of the candidate set of X-coordinates as the X-coordinate of that point. Generate 3D points ; S63. Confidence assessment of circular arc completion: For each generated 3D arc point, calculate its arc completion confidence. : ,in To successfully match the number of sampling points supported by the initial point cloud, This represents the total number of sampling points. This is the view integrity coefficient; S64, Arc-shaped point cloud fusion: Set the confidence threshold for arc completion. ;for The arc point is merged with the initial point cloud generated by S5; for For the arc points, only the arc points with high matching reliability among the sampled points are retained; the merged complete point cloud is then deduplicated again. S65, Optimization of Circular Arc Continuity: For the completed circular arc structure, the spatial distance between adjacent sampling points is detected; if the distance between adjacent points exceeds a preset threshold... Linear interpolation is performed between two points to complete the arc completion; the interpolation does not add new arc completion confidence labels, but inherits the average arc completion confidence of the two endpoints; The above steps complete the curve and arc structure.

[0025] Preferably, in step S6, the two-dimensional arc sampling points are matched with the initial point cloud by arc entity recognition and view adaptive 3D enhancement, and the missing dimension coordinates are obtained by using the mode, effectively restoring the curve structure; arc completion confidence evaluation is introduced, and the completion reliability is quantified by combining the matching success rate and the view integrity coefficient, and a threshold is set for fusion screening to ensure the completion quality; the continuity of the completed arc is optimized, and adjacent points with excessive spacing are filled by linear interpolation, making the curve smoother and more complete.

[0026] Compared with the prior art, the present invention has the following technical effects:

[0027] By constructing an adaptive probabilistic fusion model, confidence is evaluated by comprehensively considering matching error, local sampling density, and neighborhood support relationships, significantly improving the accuracy of 3D point cloud reconstruction and reducing the number of mismatched points. A missing view tolerance and dynamic weight allocation mechanism are introduced to enhance the algorithm's robustness to incomplete view information or uneven sampling. Combining geometric priors and view topological constraints to restore curves and arc structures improves upon the problems of broken curve structures and incomplete point clouds found in traditional methods. The entire process does not rely on large-scale training data, exhibits a high degree of automation, and is suitable for the automatic reconstruction of engineering three-view diagrams into 3D point clouds. Attached Figure Description

[0028] Figure 1 This is a flowchart of a three-dimensional point cloud reconstruction method based on geometric consistency probabilistic fusion proposed in this invention.

[0029] Figure 2 This is a diagram of the dual-path adaptive sensing network structure proposed in this invention.

[0030] Figure 3 This is a prototype guidance operation structure diagram proposed in this invention.

[0031] Figure 4 This is the dynamic sampling structure proposed in this invention. Detailed Implementation

[0032] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0033] This invention acquires the front, side, and top views of the object to be reconstructed, and performs normalization and spatial alignment. Based on the spatial projection correspondence between the three views, it generates 3D candidate points. An adaptive probabilistic fusion model is constructed to evaluate the confidence of the 3D candidate points based on matching error, local sampling density, neighborhood support relationship, and dynamic weight allocation. 3D candidate points are then filtered based on confidence and topological neighborhood consistency to generate an initial point cloud. The structure is restored based on geometric priors, view topological constraints, and structural continuity. Finally, the point cloud data is output. This invention improves the accuracy and completeness of 3D point cloud reconstruction through a geometric consistency probabilistic fusion mechanism.

[0034] Please see Figures 1 to 4 The present application presents a three-dimensional point cloud reconstruction method based on geometric consistency probabilistic fusion, which aims to improve the accuracy and completeness of three-dimensional point cloud reconstruction through the geometric consistency probabilistic fusion mechanism.

[0035] S1. Acquisition and parsing of three-view data; acquire the two-dimensional view data of the object to be reconstructed, and obtain the front view, side view and top view data by layer name;

[0036] Furthermore, in step S1, the three-view data is acquired and parsed, characterized by: Read the DXF file, iterate through all views in the model space, and determine which view is being retrieved based on the layer name: layer names containing "front," "top," or "top" indicate a top view, and "side," "left," or "right" indicate a side view. Layer name determination is case-insensitive and supports both Chinese and English layer names. Then, adaptive sampling is performed on the obtained views to obtain and label sampling points. (Front view) Top view Side view Where x represents the horizontal coordinate, y represents the vertical coordinate, and z represents the depth coordinate; In addition, type identification and classification are required: LINE, LWPOLYLINE, CIRCLE, and ARC; for LWPOLYLINE entities, determine whether they are closed shapes based on their closed attribute; for ARC entities, perform angle normalization processing for abnormal start and end angles. The adaptive sampling method is as follows: For a straight line segment, the sampling interval is proportional to the length of the line segment, and the number of sampling points... ,in The length of the line segment. The preset sampling interval; ; For circles, curvature adaptive sampling is used, and the number of sampling points is... ,in This is the curvature gain coefficient. It is the radius of the circle, n is clipped to interval; ; For circular arcs, a dual adaptive sampling method is used, considering both arc length and curvature, with a sampling point count of [number missing]. ,in Let the arc length be , Arc length weighting coefficient, This is the curvature gain coefficient. , ; Meanwhile, the original connection order between sampling points of the same type is maintained during the sampling process, serving as local topological edges; sampling points with the same geometric features between different views establish cross-view topological associations through type and parameter indexes, forming a geometrically consistent topological graph structure oriented towards 3D reconstruction.

[0037] S2. View Normalization and Spatial Alignment; To align the three views under a unified reference coordinate system, the obtained three views are... Figure 2 The coordinates of the points in each dimension are normalized.

[0038] Furthermore, in step S2, view normalization and spatial alignment are performed, characterized by: Perform coordinate normalization on the front view, top view, and side view respectively, so that the coordinate origin of the three views is unified to the same reference system; The normalization process specifically involves calculating the minimum X-coordinate of all sampled points in each view. minimum value of Y coordinate minimum Z-coordinate Perform coordinate transformation on each sampling point in each view: For the front view, perform the transformation: , This ensures that the origin of the front view coordinate system is located at the lower left corner of the view, with the positive Y-axis pointing vertically upwards. For the side view, perform the transformation: , This ensures that the origin of the side view coordinate system is located at the lower left corner of the view, with the positive Y-axis pointing vertically upwards. For the top view, perform the transformation: , This ensures that the origin of the top view coordinate system is located at the upper left corner of the view, with the positive X-axis pointing horizontally to the right. Spatial alignment further includes: after normalization, the front view's These correspond to (X,Y) coordinates in three-dimensional space, respectively, and are shown in the top view. These correspond to (X, Z) coordinates in three-dimensional space, respectively, and are the side view coordinates. These correspond to the (Z, Y) coordinates in three-dimensional space, enabling the coordinate alignment of the three views within the same three-dimensional space reference system.

[0039] S3. Generate 3D candidate points based on the geometric consistency of the three views; utilize the spatial projection correspondence between the three views to perform cross-view matching on the 2D sampling points to generate candidate 3D points;

[0040] Furthermore, in step S3, three-dimensional candidate points are generated based on the geometric consistency of the three views, characterized in that: Using the set of sample points in the front view as a reference, iterate through each sample point in the front view. Geometric consistency matching is performed with the top view sampling point set and the side view sampling point set respectively to generate three-dimensional candidate points; Geometric consistency matching specifically includes the following sub-steps: S31. Top view matching: For the front view In the top view sampling point set Searching for the content that satisfies All candidate points, of which To preset the tolerance threshold in the X direction, Each successfully matched top viewpoint provides depth coordinates. Forming two-dimensional candidate pairs ; S32, Side view verification: For each generated candidate pair In the side view sampling point set Searching for the content that satisfies and ,in To preset the tolerance threshold, If a side viewpoint that meets the conditions exists, then the candidate pair is determined to pass the side view verification. S33, Generation of 3D Candidate Points: For combinations that simultaneously satisfy top view matching and side view verification, generate 3D candidate points. Record the three-view matching information of the candidate point, including: the coordinates of the front view point, the coordinates of the matching top view point, and the coordinates of the matching side view point; S34. Missing view fault handling: When the top view is successfully matched but the side view has no matching point, the 3D candidate point is marked as "side view missing" and its confidence weight is reduced, but it is still retained as a 3D candidate point; when the top view has no matching point, the generation of 3D candidate points for the front view point is skipped and no 3D candidate point is generated. The above steps generate a set of three-dimensional candidate points with matching error information, providing input for subsequent confidence calculation.

[0041] S4. Construct an adaptive probabilistic fusion model to evaluate the confidence of 3D candidate points, including matching error, local sampling density, neighborhood support relationship and dynamic weight allocation mechanism;

[0042] Furthermore, in step S4, an adaptive probabilistic fusion model is constructed to evaluate the confidence level of the three-dimensional candidate points, characterized by: For each generated 3D candidate point Calculate the confidence score, which measures the degree to which the 3D candidate point is supported by all three views. The confidence assessment specifically includes the following steps: S41. Single-view matching error calculation: Calculate the top view matching error: ,in The X coordinates of the matched top-down viewpoint; Calculate the X-direction matching error of the side view: ,in The Z-coordinate of the matched side viewpoint; Calculate the Y-direction matching error of the side view: ,in The Y coordinate of the matched side viewpoint; S42. Matching error-confidence component conversion: The matching error is converted into a confidence component using an exponential decay function: ,in The X-axis scale parameter; ,in The mean of the matching error in the X direction. The standard deviation of the matching error in the X direction. This is the error adjustment coefficient. ; ,in The Z-axis scale parameter; ,in The mean of the matching error in the Z direction. The standard deviation of the matching error in the Z direction. This is the error adjustment coefficient. ; ,in The Y-axis scale parameter; ,in The mean of the matching error in the Y direction. The standard deviation of the matching error in the Y direction. This is the error adjustment coefficient. ; S43, Adaptive Weight Allocation: The weights of each confidence component are dynamically adjusted based on the sampling density and entity type of the three views: weights The weight is determined based on the local sampling density of the front view; the denser the sampling, the higher the weight. ;in This refers to the local sampling density of the front view. The strength of the support in the neighborhood of the top view; The strength of the support in the neighborhood of the side view; , The number of sampling points in the neighborhood. This represents the area of ​​the corresponding neighborhood. ,in The neighborhood support strength of the corresponding view. The spatial distance between neighboring points and the target point. The neighborhood radius, , For indicator functions, when the condition Output 1 if true, otherwise output 0; weight Determined based on the number of neighboring support points of the matching point in the top view; Weight Determined based on the number of neighboring support points of the matching point in the side view; The weights satisfy the normalization constraint: ; S44. Confidence Calculation: The confidence level is calculated using the following formula: ,in The local confidence level of the front view sampling points. ; S45, Adaptive adjustment for missing views: When the side view is missing, the weights are redistributed. , The confidence level calculation formula has been adjusted to and multiplied by the attenuation coefficient , .

[0043] S5. Based on confidence level and topological neighborhood consistency, three-dimensional candidate points are filtered to generate an initial point cloud;

[0044] Furthermore, in step S5, the three-dimensional candidate points are screened based on confidence level and topological neighborhood consistency, characterized by: Based on the confidence score calculated in step S4, the set of 3D candidate points generated in step S3 is filtered and fused to generate an initial 3D point cloud. The confidence score filtering and initial point cloud generation specifically include the following steps: S51, Dual Threshold Filtering: Set a high confidence threshold and low confidence threshold ,in Regarding confidence level The three-dimensional candidate points are directly marked as "high confidence points" and retained; for the confidence level... The 3D candidate points are marked as "medium confidence points" and must pass the local consistency test before being retained; for confidence... The three-dimensional candidate points are directly eliminated; , ; S52, Neighborhood Consistency Check: For points with medium confidence, construct their spatial neighborhood. ,in The neighborhood radius, ; For other points in the surrounding area, For the current target point, Given the Euclidean distance, count the number of high-confidence points in the neighborhood. ;like If a confidence point passes the neighborhood consistency test, it is retained; otherwise, it is removed. The minimum neighbor threshold, ; S53, Point Cloud Fusion and Optimization: All high-confidence points and medium-confidence points that passed the test are merged into an initial point cloud set; spatial deduplication is performed on the initial point cloud: duplicate points are merged, and the coordinates are rounded to three decimal places before deleting duplicate points; outlier filtering is performed on the deduplicated point cloud: the average neighborhood distance of each point is calculated, and outliers whose average neighborhood distance deviates significantly from the overall distribution are removed. S54. Topological connectivity constraints: Using the topology graph structure constructed in step S1, each retained point in the point cloud must have at least one neighboring point also retained in the topology graph of its corresponding view; for isolated points, a secondary check is performed: if its confidence level... If so, keep it; otherwise, discard it. Through the above steps, an initial 3D point cloud is generated after confidence filtering, neighborhood consistency constraints, point cloud fusion and optimization, and topological connectivity constraints.

[0045] S6. Based on geometric priors, view topological constraints, and structural continuity, the curve structure is restored;

[0046] Furthermore, in step S6, the curve structure is restored based on geometric priors, view topological constraints, and structural continuity, characterized by: The completion of curve and arc structures specifically includes the following steps: S61. Circular Arc Entity Recognition and Classification: Iterate through all CIRCLE and ARC type entities identified in step S1; classify them into three categories based on their view labels: front view arcs, top view arcs, and side view arcs; for each arc entity, extract its geometric parameters: center coordinates. Radius R, start and end angles , ; S62, View-adaptive 3D enhancement: The front view arc elevation uses curvature adaptive sampling to perform two-dimensional sampling of the arc, generating a set of sampling points. For each sampling point Searching for the X coordinate in the initial point cloud generated by S5. For each point, obtain the corresponding Z-coordinate candidate set; take the mode of the Z-coordinate candidate set as the depth value of that point. Generate 3D points ; Top view arc elevation: Perform two-dimensional sampling on the arc to generate a set of sampling points. ,in Corresponding to the Z coordinate in three-dimensional space; for each sampling point Searching for the X coordinate in the initial point cloud. And the Z coordinate is For each point, obtain the corresponding Y-coordinate candidate set; take the mode of the Y-coordinate candidate set as the height value of that point. Generate 3D points ; Side view arc lifting: Perform two-dimensional sampling on the arc to generate a set of sampling points. ,in Corresponding to the Z coordinate in three-dimensional space, Corresponding to the Y coordinate; for each sampling point Searching for the Z coordinate in the initial point cloud is And the Y coordinate is For each point, obtain the corresponding candidate set of X-coordinates; take the mode of the candidate set of X-coordinates as the X-coordinate of that point. Generate 3D points ; S63. Confidence assessment of circular arc completion: For each generated 3D arc point, calculate its arc completion confidence. : ,in To successfully match the number of sampling points supported by the initial point cloud, This represents the total number of sampling points. For view integrity coefficient, front view Top view Side view ; S64, Arc-shaped point cloud fusion: Set the confidence threshold for arc completion. ;for The arc point is merged with the initial point cloud generated by S5; for For the arc points, only the arc points with high matching reliability among the sampled points are retained; the merged complete point cloud is deduplicated again, retaining three decimal places of precision; S65, Optimization of Circular Arc Continuity: For the completed circular arc structure, the spatial distance between adjacent sampling points is detected; if the distance between adjacent points exceeds a preset threshold... Set to twice the arc sampling interval, perform linear interpolation to fill the points between the two points; the interpolation does not add new arc completion confidence labels, but inherits the average arc completion confidence of the two endpoints; The above steps complete the curve and arc structure.

[0047] S7. Point cloud output; Output the obtained point cloud data and visualize it to obtain a ply file.

[0048] Furthermore, in step S7, complete point cloud data is obtained, visualized, and saved as a ply file.

[0049] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. A 3D point cloud reconstruction method based on geometric consistency probabilistic fusion, characterized in that, Includes the following steps: S1. Acquisition and parsing of three-view data; acquire the two-dimensional view data of the object to be reconstructed, and obtain the front view, side view and top view data by layer name; S2, View Normalization and Spatial Alignment; To align the three views under a unified reference coordinate system, the coordinates of the two-dimensional points in the obtained three views are normalized. S3. Generate 3D candidate points based on the geometric consistency of the three views; By utilizing the spatial projection correspondence between the three views, cross-view matching is performed on the two-dimensional sampling points to generate candidate three-dimensional points; S4. Construct an adaptive probabilistic fusion model to evaluate the confidence of 3D candidate points, including matching error, local sampling density, neighborhood support relationship and dynamic weight allocation mechanism; S5. Based on confidence level and topological neighborhood consistency, three-dimensional candidate points are filtered to generate an initial point cloud; S6. Based on geometric priors, view topological constraints, and structural continuity, the curve structure is restored; S7, point cloud output; The obtained point cloud data is output and visualized to obtain a ply file.

2. The three-dimensional point cloud reconstruction method based on geometric consistency probabilistic fusion according to claim 1, wherein in step S1, three-view data acquisition and parsing are performed, characterized in that: Read the DXF file, iterate through all views in the model space from the DXF file, then adaptively sample the obtained views, obtain sample points, and label them; front view. Top view Side view Where x represents the horizontal coordinate, y represents the vertical coordinate, and z represents the depth coordinate; In addition, type identification and classification are required: LINE, LWPOLYLINE, CIRCLE, and ARC; for LWPOLYLINE entities, determine whether they are closed shapes based on their closed attribute; for ARC entities, perform angle normalization processing for abnormal start and end angles. The adaptive sampling method is as follows: For a straight line segment, the sampling interval is proportional to the length of the line segment, and the number of sampling points... ,in The length of the line segment. The preset sampling interval; For circles, curvature adaptive sampling is used, and the number of sampling points is... ,in This is the curvature gain coefficient. It is the radius of the circle, n is clipped to interval; For circular arcs, a dual adaptive sampling method is used, considering both arc length and curvature, with a sampling point count of [number missing]. ,in Let the arc length be , Arc length weighting coefficient, This is the curvature gain coefficient; Meanwhile, the original connection order between sampling points of the same type is maintained during the sampling process, serving as local topological edges; sampling points with the same geometric features between different views establish cross-view topological associations through type and parameter indexes, forming a geometrically consistent topological graph structure oriented towards 3D reconstruction.

3. The 3D point cloud reconstruction method based on geometric consistency probabilistic fusion according to claim 2, wherein in step S2, view normalization and spatial alignment are performed, characterized in that: Perform coordinate normalization on the front view, top view, and side view respectively, so that the coordinate origin of the three views is unified to the same reference system; The normalization process specifically involves calculating the minimum X-coordinate of all sampled points in each view. minimum value of Y coordinate minimum Z-coordinate Perform coordinate transformation on each sampling point in each view: For the front view, perform the transformation: , This ensures that the origin of the front view coordinate system is located at the lower left corner of the view, with the positive Y-axis pointing vertically upwards. For the side view, perform the transformation: , This ensures that the origin of the side view coordinate system is located at the lower left corner of the view, with the positive Y-axis pointing vertically upwards. For the top view, perform the transformation: , This ensures that the origin of the top view coordinate system is located at the upper left corner of the view, with the positive X-axis pointing horizontally to the right. Spatial alignment further includes: after normalization, the front view's These correspond to (X, Y) coordinates in three-dimensional space, respectively, and the top view... These correspond to (X, Z) coordinates in three-dimensional space, respectively, and are the side view coordinates. These correspond to the (Z, Y) coordinates in three-dimensional space, enabling the coordinate alignment of the three views within the same three-dimensional space reference system.

4. The three-dimensional point cloud reconstruction method based on geometric consistency probabilistic fusion according to claim 3, wherein in step S3, three-dimensional candidate points are generated based on the geometric consistency of the three views, characterized in that: Using the set of sample points in the front view as a reference, iterate through each sample point in the front view. Geometric consistency matching is performed with the top view sampling point set and the side view sampling point set respectively to generate three-dimensional candidate points; Geometric consistency matching specifically includes the following sub-steps: S31. Top view matching: For the front view In the top view sampling point set Searching for the content that satisfies All candidate points, of which A preset tolerance threshold is set for the X-direction; depth coordinates are provided for each successfully matched top viewpoint. Forming two-dimensional candidate pairs ; S32, Side view verification: For each generated candidate pair In the side view sampling point set Searching for the content that satisfies and ,in A preset tolerance threshold is set; if a side view point that meets the conditions exists, the candidate pair is determined to pass the side view verification. S33, Generation of 3D Candidate Points: For combinations that simultaneously satisfy top view matching and side view verification, generate 3D candidate points. ; Record the three-view matching information of the 3D candidate point, including: the coordinates of the front view point, the coordinates of the matching top view point, and the coordinates of the matching side view point; S34. Missing view fault handling: When the top view is successfully matched but the side view has no matching point, the 3D candidate point is marked as "side view missing" and its confidence weight is reduced, but it is still retained as a candidate point; when the top view has no matching point, the generation of 3D candidate points for the front view point is skipped and no 3D candidate points are generated. The above steps generate a set of three-dimensional candidate points with matching error information, providing input for subsequent confidence calculation.

5. The 3D point cloud reconstruction method based on geometric consistency probabilistic fusion according to claim 4, wherein in step S4, an adaptive probabilistic fusion model is constructed to evaluate the confidence of 3D candidate points, characterized in that: For each generated 3D candidate point Calculate the confidence score, which measures the degree to which the 3D candidate point is supported by all three views. The confidence assessment specifically includes the following steps: S41. Single-view matching error calculation: Calculate the top view matching error: ,in For the matched top view point X coordinates; calculate the X-direction matching error of the side view: ,in Given the Z-coordinate of the matched side viewpoint; calculate the Y-direction matching error of the side view: ,in The Y coordinate of the matched side viewpoint; S42. Matching error-confidence component conversion: The matching error is converted into a confidence component using an exponential decay function: ,in The X-axis scale parameter; ,in The Z-axis scale parameter; ,in The Y-axis scale parameter; S43, Adaptive Weight Allocation: The weights of each confidence component are dynamically adjusted based on the sampling density and entity type of the three views: Weight The weight is determined based on the local sampling density of the front view; the denser the sampling, the higher the weight. ;in This refers to the local sampling density of the front view. The strength of the support in the neighborhood of the top view; The strength of the support in the neighborhood of the side view; , The number of sampling points in the neighborhood. This represents the area of ​​the corresponding neighborhood. ,in The neighborhood support strength of the corresponding view. The spatial distance between neighboring points and the target point. The neighborhood radius, For indicator functions, when the condition Output 1 if true, otherwise output 0; Weight Determined based on the number of neighboring support points of the matching point in the top view; ; Weight Determined based on the number of neighboring support points of the matching point in the side view; ; S44. Confidence Calculation: The confidence level is calculated using the following formula: ,in The local confidence level of the front view sampling points; S45, Adaptive adjustment for missing views: When the side view is missing, the weights are redistributed. , The confidence level calculation formula has been adjusted to and multiplied by the attenuation coefficient .

6. The 3D point cloud reconstruction method based on geometric consistency probabilistic fusion according to claim 5, wherein in step S5, 3D candidate points are screened based on confidence level and topological neighborhood consistency, characterized in that: Based on the confidence level calculated in step S4, the set of 3D candidate points generated in step S3 is filtered and fused to generate an initial 3D point cloud. The confidence screening and initial point cloud generation process includes the following steps: S51, Dual Threshold Filtering: Set a high confidence threshold and low confidence threshold ,in Regarding confidence level The three-dimensional candidate points are directly marked as "high confidence points" and retained; for the confidence level... The 3D candidate points are marked as "medium confidence points" and must pass the local consistency test before being retained; for confidence... The three-dimensional candidate points are directly eliminated; S52, Neighborhood Consistency Check: For points with medium confidence, construct their spatial neighborhood. ,in The neighborhood radius, For other points in the surrounding area, For the current target point, Given the Euclidean distance, count the number of high-confidence points in the neighborhood. ;like If a confidence point passes the neighborhood consistency test, it is retained; otherwise, it is removed. The minimum neighbor threshold; S53, Point Cloud Fusion and Optimization: All high-confidence points and medium-confidence points that passed the test are merged into an initial point cloud set; spatial deduplication is performed on the initial point cloud: duplicate points are merged and deleted; outlier filtering is performed on the deduplicated point cloud: the average neighborhood distance of each point is calculated and outliers with average neighborhood distances that deviate significantly from the overall distribution are removed. S54. Topological connectivity constraints: Using the topology graph structure constructed in step S1, each retained point in the point cloud must have at least one neighboring point also retained in the topology graph of its corresponding view; for isolated points, a secondary check is performed: if its confidence level... If so, keep it; otherwise, discard it. Through the above steps, an initial 3D point cloud is generated after confidence filtering, neighborhood consistency constraints, point cloud fusion and optimization, and topological connectivity constraints.

7. A three-dimensional point cloud reconstruction method based on geometric consistency probabilistic fusion according to claim 6, wherein in step S6, the curve structure is restored based on geometric priors, view topological constraints, and structural continuity, characterized in that: The completion of curve and arc structures specifically includes the following steps: S61. Circular Arc Entity Recognition and Classification: Iterate through all CIRCLE and ARC type entities identified in step S1; classify them into three categories based on their view labels: front view arcs, top view arcs, and side view arcs; for each arc entity, extract its geometric parameters: center coordinates. Radius r, start and end angles , ; S62, View-adaptive 3D enhancement: Front view arc elevation: Two-dimensional sampling of the arc is performed using curvature adaptive sampling to generate a set of sampling points. For each sampling point Searching for the X coordinate in the initial point cloud generated by S5 is... For each point, obtain the corresponding Z-coordinate candidate set; take the mode of the Z-coordinate candidate set as the depth value of that point. Generate 3D points ; Top view arc elevation: Perform two-dimensional sampling on the arc to generate a set of sampling points. ,in Corresponding to the Z coordinate in three-dimensional space; for each sampling point Searching for the X coordinate in the initial point cloud. And the Z coordinate is For each point, obtain the corresponding Y-coordinate candidate set; take the mode of the Y-coordinate candidate set as the height value of that point. Generate 3D points ; Side view arc lifting: Perform two-dimensional sampling on the arc to generate a set of sampling points. ,in Corresponding to the Z coordinate in three-dimensional space, Corresponding to the Y coordinate; for each sampling point Searching for the Z coordinate in the initial point cloud is And the Y coordinate is For each point, obtain the corresponding candidate set of X-coordinates; take the mode of the candidate set of X-coordinates as the X-coordinate of that point. Generate 3D points ; S63. Confidence assessment of circular arc completion: For each generated 3D arc point, calculate its arc completion confidence. : ,in To successfully match the number of sampling points supported by the initial point cloud, This represents the total number of sampling points. This is the view integrity coefficient; S64, Arc-shaped point cloud fusion: Set the confidence threshold for arc completion. ;for The arc point is merged with the initial point cloud generated by S5; for For the arc points, only the arc points with high matching reliability among the sampled points are retained; the merged complete point cloud is then deduplicated again. S65, Optimization of Circular Arc Continuity: For the completed circular arc structure, the spatial distance between adjacent sampling points is detected; if the distance between adjacent points exceeds a preset threshold... Linear interpolation is performed between two points to complete the arc completion; the interpolation does not add new arc completion confidence labels, but inherits the average arc completion confidence of the two endpoints; The above steps complete the curve and arc structure.

Images