Rebar binding detection method and system based on instance segmentation and dynamic topology modeling

Abstract

The invention discloses a reinforcing steel bar binding detection method and system based on instance segmentation and dynamic topology modeling. The method comprises the steps that reinforcing steel bar mesh three-dimensional point cloud data are collected through a multi-sensor fusion system; identifying a reinforcing steel bar main body and a binding point by adopting a mask optimization instance segmentation network; constructing a dynamic topology model based on skeleton extraction and curvature weighting strategies, and expressing a steel bar space connection relationship; and generating a quality evaluation matrix containing a tightening degree quantitative index according to the construction specification. By adopting the multi-sensor fusion system, the problem of perspective distortion of traditional monocular vision can be solved; through an improved mask optimization instance segmentation network, accurate recognition of the reinforcing steel bars and binding points is realized, and recognition of a reinforcing steel bar intersection-to-union ratio is improved; a dynamic topology model is constructed, and the modeling error of the spiral stirrup is reduced; and finally, a multi-dimensional quality evaluation matrix is generated, rapid and efficient steel bar binding full-automatic acceptance check is achieved, the limitation on a plane orthogonal grid is broken through, and complex working conditions such as a curved surface steel bar mesh are comprehensively supported.

Description

Technical Field

[0001] This invention relates to the field of building engineering quality inspection technology, and more specifically to a method and system for detecting rebar tying based on instance segmentation and dynamic topology modeling. Background Technology

[0002] Currently, in the field of automated inspection of rebar binding quality, existing technical solutions suffer from the following three significant technical defects, which limit their application effectiveness and promotion value in real-world complex engineering scenarios: Target detection is highly sensitive to occlusion: Existing methods heavily rely on the clear visibility of the rebar segments or nodes themselves when identifying rebar nodes and tying points. For example, the technical solution disclosed in patent document CN112627538A relies on the identification of rebar segments between nodes to locate the node positions. However, in real construction sites, rebar meshes are often densely arranged, components are interwoven, or construction materials are temporarily piled up, which can easily lead to partial or complete occlusion of target rebar segments or nodes. This occlusion directly causes the target detection algorithm to fail, resulting in inaccurate node positioning or even missed detections, seriously affecting the reliability of subsequent tying status judgments.

[0003] The orthogonal geometric assumption limits applicability: Current mainstream rebar tying detection algorithms are typically based on the strong geometric assumption that the rebar mesh is arranged in orthogonal straight lines. This assumption greatly simplifies the complexity of image processing and geometric reasoning. However, in actual engineering (such as irregular structures, curved surfaces, or complex node areas), the rebar mesh often presents curved, oblique, or non-orthogonal arrangements. Existing technologies cannot effectively handle these non-orthogonal geometric relationships, leading to complete failure or numerous misjudgments in node localization, rebar tracking, and tying relationship determination in these scenarios, greatly limiting the universality and application scope of the technology.

[0004] Lack of multi-dimensional quantitative assessment capabilities: Existing technologies for assessing binding quality are too simplistic and crude, typically only providing a binary qualitative judgment of "bound" or "not bound." This simplistic judgment cannot meet the needs of refined management. It completely lacks the ability to quantitatively assess key parameters of binding quality, such as: whether the number of turns of the binding wire meets specifications, the tightness of the binding points, or whether there are intermediate states such as partial detachment or loosening of the binding points. This lack of assessment capabilities makes it difficult for automated inspection results to accurately reflect the actual construction quality level, failing to provide a deeper and more valuable basis for quality control.

[0005] In summary, existing automatic detection technologies for rebar binding quality have significant shortcomings in terms of robustness against obstruction, geometric adaptability (especially for non-orthogonal and curved structures), and quantitative precision of evaluation dimensions, making it difficult to meet the complex, variable, and high-standard quality control needs of modern building construction.

[0006] Therefore, there is an urgent need for a new method or system for detecting the quality of rebar tying that can overcome the above-mentioned technical defects and has stronger environmental adaptability, a wider range of geometric structure processing capabilities, and more refined quality assessment dimensions. Summary of the Invention

[0007] In view of this, the present invention provides a method and system for detecting rebar tying based on instance segmentation and dynamic topology modeling, which at least partially solves the problems of existing automatic rebar tying quality detection technologies in terms of robustness against occlusion, geometric adaptability, and quantitative precision of evaluation dimensions.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, embodiments of the present invention provide a method for detecting rebar tying based on instance segmentation and dynamic topology modeling, comprising the following steps: S1. Acquire 3D point cloud data of steel mesh through a multi-sensor fusion system; S2. For the three-dimensional point cloud data of the steel mesh, a mask-optimized instance segmentation network is used to identify the main body of the steel bars and the binding points; S3. Construct a dynamic topology model based on skeleton extraction and curvature weighting strategy to express the spatial connection relationship of steel bars; S4. Generate a quality assessment matrix containing quantitative indicators of tightening according to the construction specifications.

[0009] Further, step S1 includes: S11. Acquire millimeter-precision point cloud data using a laser scanner. L Color depth point cloud P is captured synchronously using an RGBD camera. R And based on IMU positioning data, initial registration of multi-view point clouds is performed; S12, Extract laser point cloud P L Medium curvature extreme feature points and depth point cloud P R Cross-modal matching using SIFT descriptors in the dataset: S13. The RANSAC algorithm is used to solve for the optimal spatial transformation matrix, outliers are eliminated through iterative optimization, and the final point cloud data is obtained through weighted fusion.

[0010] Further, the mask-optimized instance segmentation network in step S2 includes: The feature pyramid network infrastructure is used to fuse deep semantic features of the overall structure with shallow detail features that preserve edge information through a top-down path. The reinforcement feature attention module is used to dynamically focus on key areas of the reinforcement through an attention mechanism, suppressing background interference and compensating for missing information. An edge-enhanced loss function is used to jointly optimize mask accuracy and edge localization.

[0011] Furthermore, in step S3, constructing the dynamic topology model includes: S31. Skeleton Extraction: Morphological thinning algorithm is used to process the rebar mask image to generate a single-pixel wide skeleton.

[0012] M represents the binary mask of the main steel reinforcement structure output by the instance segmentation network; B represents the 3×3 structural kernel; k represents the number of iterations; This represents the erosion operation; S32, Node Generation: Create topological nodes at the extreme points of skeleton curvature: , This indicates the rate of change and direction of change of the tangent direction of the steel reinforcement cage curve at a certain point s; This represents the threshold for determining the extreme point of curvature; S33, Adaptive Neighborhood Search: Dynamically determining connected neighborhoods using curvature-driven radius.

[0013] S34, Curvature-weighted edge connection: Calculate the similarity weights between nodes and establish connections: γ=2.0 represents the control edge connection weight; when When the value is greater than 0.35, an edge e is established. ij ;k i and k j Represents the curvature values ​​of adjacent nodes i and j; S35. Topology Verification: Correcting Connection Relationships Based on Construction Specification Constraints. like Delete edge e ij ;like Add virtual tethering points; P i P j t represents the coordinate vector of adjacent nodes i and j. i t j This indicates the tangent direction between adjacent nodes i and j.

[0014] Further, in step S4, the quantification index of tightness is defined as:

[0015] Where k is the number of sampling directions; θ i For the angle between the tie wire and the reinforcing bar, σ d d is the distance decay coefficient, which controls the rate of exponential decay; i Where n is the offset distance of the binding point, n is the actual number of wrapping turns, and n0 is the minimum number of turns specified.

[0016] In a second aspect, embodiments of the present invention also provide a rebar tying detection system based on instance segmentation and dynamic topology modeling, which applies the rebar tying detection method based on instance segmentation and dynamic topology modeling as described in any one of the first aspects. The system includes: The point cloud data processing module is used to collect three-dimensional point cloud data of steel mesh through a multi-sensor fusion system; The identification module is used to identify the main body of the steel reinforcement and the binding points by using a mask-optimized instance segmentation network on the three-dimensional point cloud data of the steel reinforcement mesh. A model building module is constructed based on skeleton extraction and curvature weighting strategies to build a dynamic topology model that expresses the spatial connection relationship of steel bars; The quality assessment module is used to generate a quality assessment matrix containing quantitative indicators of tightness based on construction specifications.

[0017] As can be seen from the above technical solution, compared with the prior art, the present invention has the following technical advantages: First, a multi-sensor fusion system is used to acquire millimeter-level precision 3D point clouds, solving the perspective distortion problem of traditional monocular vision. Second, an improved mask-optimized instance segmentation network is used to achieve accurate identification of rebars and tying points, improving the recognition of rebar intersection-to-union ratio. Then, a curvature-driven dynamic topology model is constructed to reduce the modeling error of helical stirrups. Finally, a multi-dimensional quality assessment matrix is ​​generated based on construction specifications to achieve rapid and efficient fully automated acceptance of rebar tying, with efficiency far exceeding that of manual inspection and a low false negative rate. This invention breaks through the limitations of traditional methods on planar orthogonal meshes and fully supports complex conditions such as curved rebar meshes; it is especially suitable for automated acceptance of complex conditions such as irregular structures and high-density rebar meshes. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 The flowchart of the rebar tying detection method based on instance segmentation and dynamic topology modeling provided by the present invention is shown.

[0020] Figure 2 The flowchart illustrates the multi-sensor fusion system for processing 3D point cloud data provided by this invention.

[0021] Figure 3 The processing flowchart of the rebar feature attention module provided by the present invention is shown.

[0022] Figure 4 This invention provides a flowchart for constructing a dynamic topology model.

[0023] Figure 5 The block diagram of the rebar tying detection system based on instance segmentation and dynamic topology modeling provided by the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Reference Figure 1 As shown, this embodiment of the invention discloses a method for detecting rebar tying based on instance segmentation and dynamic topology modeling, including the following steps S1~S4: S1. Acquire 3D point cloud data of steel mesh through a multi-sensor fusion system; The multi-sensor fusion system includes: a laser scanner (0.1mm accuracy): to acquire high-precision spatial coordinates; RGBD cameras (such as Azure Kinect): Simultaneously capture color and depth information; IMU positioning module: Enables automatic registration of multi-view data; In this step, the registration and fusion of the three-dimensional power source can solve the problem of data loss in strong light / dust environment of a single sensor. By using the above three types of sensor devices, the point cloud registration error can also be reduced from the traditional ±5mm to ±0.8mm.

[0026] Step S1 specifically includes: S11. Acquire millimeter-precision point cloud data using a laser scanner. L Color depth point cloud P is captured synchronously using an RGBD camera. R And based on IMU positioning data, initial registration of multi-view point clouds is performed; S12, Extract laser point cloud P L Medium curvature extreme feature points and depth point cloud P RCross-modal matching using SIFT descriptors in the dataset: S13. The RANSAC algorithm is used to solve for the optimal spatial transformation matrix, outliers are eliminated through iterative optimization, and the final point cloud data is obtained through weighted fusion.

[0027] For the specific integration process, refer to Figure 2 As shown: (1) Feature point extraction enter: Laser point cloud: , This represents the 3D coordinates of the i-th point in the laser coordinate system, N represents the number of point clouds acquired in a single scan, and i represents the index.

[0028] RGBD point cloud: , This represents the three-dimensional coordinates of the j-th point in the camera coordinate system. This represents a three-channel color vector; M represents the number of point clouds acquired in a single frame (usually 10). 5 ~10 6 ); operate: Extracting curvature extrema points from laser point clouds as key points: ; Extracting SIFT descriptors from RGBD point clouds: Key point coordinates

[0029] (2) Cross-modal feature matching Extract K L The curvature extrema and K R Using SIFT descriptors, establish a feature descriptor distance model:

[0030] Filter matching pairs:

[0031] (3) Robust Registration (RANSAC) Solve for the transformation matrix:

[0032] Iterative optimization to remove outliers.

[0033] (4) Weighted fusion calculation Final point cloud generation formula: The weight w is 0.7.

[0034] Density optimization: Voxel filtering of overlapping regions .

[0035] The relative error of the fused point cloud is ≤ 0.8mm (laser single-source error: 1.5mm), while preserving RGBD color attributes. This supports the extraction of texture features for subsequent instance segmentation.

[0036] S2. For the three-dimensional point cloud data of the steel mesh, a mask-optimized instance segmentation network is used to identify the main body of the steel bars and the binding points; Further, the mask-optimized instance segmentation network in step S2 includes: 1) The Feature Pyramid Network (FPN) infrastructure serves as a foundational architecture for multi-scale feature fusion, addressing the detection bottleneck caused by scale differences in rebar mesh images. It can be used to fuse deep semantic features of the overall structure with shallow detail features that preserve edge information via a top-down path; adapting to varying rebar diameters (e.g., Φ12mm~Φ32mm rebars) and dense node scenarios. By utilizing upsampling and lateral connections, semantic information from high-level feature maps is injected into low-level features, enhancing the recognition capability of small-sized binding points (e.g., occluded or distant nodes). Furthermore, by sharing backbone network weights (e.g., ResNet), redundant computation is reduced, improving inference speed.

[0037] In rebar tying inspection, FPN addresses the scale sensitivity of traditional methods (such as the reliance on fixed-size detection boxes in related technologies), enabling the model to simultaneously handle high-resolution details in the foreground and global structure in the background, significantly improving the intersection-over-union (IoU) ratio.

[0038] 2) The Reinforcing Steel Feature Attention Module (RAM) is used to dynamically focus on key areas of the reinforcing steel through an attention mechanism, suppressing background interference and compensating for missing information; its calculation process satisfies:

[0039] σ is the sigmoid activation function, and GAP represents global average pooling. Occlusion robustness is improved; in partially occluded scenarios (such as concrete pouring formwork occlusion), RAM compensates for missing information through attention weights, resulting in a measured recall improvement of 18.7%.

[0040] Furthermore, the computation is lightweight, employing 1×1 convolutional compression channels, reducing the number of parameters by 50%, making it suitable for mobile deployment. Compared to simple geometric fitting methods, RAM introduces semantic-level feature optimization to address false detection issues in complex scenes (such as shadows and rust), improving the attention rate of rebar features.

[0041] 3) Edge-enhancing loss function, used to jointly optimize mask accuracy and edge localization; the formula is as follows:

[0042] Represents the mask cross-entropy. Indicates edge reinforcement terms; Edge reinforcement coefficient: σ=1.5. Image gradient. As loss weights, the learning of edge features is enhanced, and its components represent the rate of change of the image in the horizontal and vertical directions: In the edge region ( When the value of λ approaches 1, the edge loss constraint is strengthened; when the value of λ approaches 0 in the smooth region, the interference is reduced; through gradient calculation, the algorithm can accurately capture the physical boundary features of the steel bars, providing key geometric information for subsequent instance segmentation and topology modeling.

[0043] A mask-optimized instance segmentation network was used to identify the main body of the rebar and the binding points. Compared with the standard Mask R-CNN, the rebar feature response was improved by 2.3 times; the IoU of the binding wire recognition reached 89.2%.

[0044] This loss function overcomes the insensitivity of traditional binary cross-entropy loss to boundaries, reduces edge localization error to 1.2 pixels, and achieves a binding point offset detection accuracy of 96%.

[0045] Among them, reference Figure 3 The image shows the processing procedure of the Reinforcing Bar Feature Attention Module (RAM): (1) Feature extraction (3×3 convolution) The input feature map is processed by a 3×3 convolution kernel to extract spatial features of the reinforcing bars, enhancing local contextual information. The formula is as follows:

[0046] Capture local patterns of rebar edges and tying points.

[0047] (2) Global Information Compression (GAP) Perform global average pooling on the feature maps:

[0048] Spatial features are compressed into channel descriptors while preserving the global context.

[0049] (3) Channel adjustment (1×1 convolution dimensionality reduction → dimensionality increase) Use a bottleneck structure to reduce computational load: Dimension reduction: Compress the channels to 1 / 4 of their original size (e.g., 256 → 64); Upgrade: Restore to the original number of channels (64→256); Learn cross-channel nonlinear relationships to improve feature representation capabilities.

[0050] (4) Attention generation (ReLU→Sigmoid) Spatial weights are generated using an activation function:

[0051] Generate an attention map in the [0,1] interval, with high-value areas focusing on key parts of the reinforcing bars.

[0052] (5) Feature enhancement (point-by-point multiplication) Weighted sum of original feature map and attention map:

[0053] Enhance the characteristic response of reinforcing bars and suppress irrelevant backgrounds (such as formwork shadows).

[0054] (6) Information retention (residual join) The final output incorporates the original information:

[0055] Avoid gradient vanishing and ensure feature integrity.

[0056] This design achieves precise focusing of steel reinforcement features through a three-stage architecture of "convolution enhancement → channel compression → spatial activation," providing high signal-to-noise ratio input features for subsequent topology modeling.

[0057] S3. Construct a dynamic topology model based on skeleton extraction and curvature weighting strategy to express the spatial connection relationship of steel bars; Among them, reference Figure 4 As shown, constructing a dynamic topology model includes: S31. Skeleton Extraction: Morphological thinning algorithm is used to process the rebar mask image to generate a single-pixel wide skeleton.

[0058] M represents the binary mask of the main steel reinforcement structure output by the instance segmentation network; B represents the 3×3 structural kernel; k represents the number of iterations; This indicates the erosion operation; during preprocessing, a morphological closing operation is performed on M to eliminate small holes and burrs generated during segmentation.

[0059] S32, Node Generation: Create topological nodes at the extreme points of skeleton curvature: , This indicates the rate of change and direction of change of the tangent direction of the steel reinforcement cage curve at a certain point s; This represents the threshold for determining the curvature extremum; where d(dr / ds) is the minute change (also a vector) in the tangent direction vector (dr / ds) as the arc length increases by ds. d²r / ds² is the change in tangent direction d(dr / ds) divided by the change in arc length ds. This vector describes how quickly the tangent direction changes and in which direction. It directly measures the degree and direction of curvature of the curve. It points to the center of curvature (center of curvature).

[0060] The symbol ‖...‖ represents the magnitude of the orientation quantity. ‖d²r / ds²‖ is the length of the vector d²r / ds². It only cares about the degree of curvature, not the direction of curvature. This is the curvature κ(s): it quantifies the degree of curvature of the curve at point s. The larger the value, the more curved the curve is at that point (the smaller the radius of curvature); the smaller the value (even 0), the straighter the curve is.

[0061] S33, Adaptive Neighborhood Search: Dynamically determining connected neighborhoods using curvature-driven radius.

[0062] Where k is the local curvature, for example, when the curvature k>0.25, the radius increases by 68%, which can solve the problem of broken curved steel bar connections; α is the reference radius, which is 20mm; β is the curvature adjustment coefficient.

[0063] S34, Curvature-weighted edge connection: Calculate the similarity weights between nodes and establish connections: γ=2.0 represents the control edge connection weight; when When the value is greater than 0.35, an edge e is established. ij ;k i and k j Represents the curvature values ​​of adjacent nodes i and j; S35. Topology Verification: Correcting Connection Relationships Based on Construction Specification Constraints. like Delete edge e ij ;like Add virtual tethering points; P i P j t represents the coordinate vector of adjacent nodes i and j. i t j This indicates the tangent direction between adjacent nodes i and j.

[0064] This step can improve the integrity rate of curved rebar connections.

[0065] For example: Taking the modeling of steel mesh for a curved roof as an example, the problem scenario is: a two-way curved steel mesh (rate of curvature change). ) Technical Implementation: 1) Node generation density: 0.8 nodes / mm for every millimeter change in curvature. 2) Adaptive radius effect:

[0066] 3) Topology correction: Automatically deletes 12 out-of-range connections and completes 8 binding points.

[0067] Comparison of experimental data:

[0068] This step constructs an inseparable process chain by refining morphology, detecting curvature extrema, adaptive connection, and standard verification. Through the dual mechanism of curvature-driven and standard verification, it breaks through the dependence of traditional methods on orthogonal structures, and is especially suitable for complex scenarios such as nuclear power plant containment structures and curved bridge towers.

[0069] S4. Generate a quality assessment matrix containing quantitative indicators of tightening according to the construction specifications.

[0070] In step S4, the quantification index of tightness is defined as follows:

[0071] Where k is the number of sampling directions; θ i For the angle between the tie wire and the reinforcing bar, σ d d is the distance decay coefficient, which controls the rate of exponential decay; i The offset distance of the binding point, n is the actual number of wrapping turns, and n0 is the minimum number of wrapping turns specified (taken as 3). Tightness. The standard threshold is 0.75; a value greater than 0.75 is considered acceptable; otherwise, it is considered unacceptable. In addition, spacing compliance was also taken into consideration:

[0072] δ: The average deviation of the binding point position, representing the overall deviation between the actual binding point position and the position required by the specification. δ≤5mm is considered acceptable; otherwise, it is considered unacceptable. K: The total number of binding points (sampling points) within the detection area.

[0073] P k The actual position coordinates of the kth binding point (usually three-dimensional coordinates, such as (x, y, z)).

[0074] P spec According to construction specifications, the theoretical coordinates of the kth binding point (and P) k (Corresponding expected position).

[0075] : Sum the absolute values ​​of the differences between the actual and theoretical positions of all binding points (usually the absolute value of the Euclidean distance, i.e., |P|). k -P spec | indicates the distance difference.

[0076] In addition, the wire winding count identification model:

[0077] n count : The number of wire wrapping turns output by the model (integer or floating-point predicted value).

[0078] ConvLSTM: Convolutional Long Short-Term Memory, a deep learning model that combines convolutional neural networks (CNN) and recurrent neural networks (RNN). Compared to traditional image processing (such as Hough transform), ConvLSTM is robust to occlusion and lighting changes, and can handle multi-frame image sequences of dynamic winding processes (such as videos of binding operations).

[0079] V ROI The input region of interest is a sequence of dynamic images of the tying points, which is the image region near the tying points (usually a local image block cropped from the segmented rebar mask).

[0080] The principle is as follows: Input is a sequence of ROI images near the binding point (multiple consecutive images or stacked images from different angles). ConvLSTM extracts spatial features through convolution operations and captures temporal or angular continuity information through an LSTM module. Output is the predicted number of wrapping loops, n. count (e.g., 2.0, 2.5, 3.0, etc., can be rounded to the nearest whole number).

[0081] The quality assessment matrix is ​​as follows:

[0082] This allows us to expand binary judgments into a four-dimensional continuous evaluation.

[0083] The rebar tying detection method based on instance segmentation and dynamic topology modeling provided by this invention overcomes the three major limitations of "occlusion sensitivity, orthogonal assumption, and binary judgment" through the technical chain of instance segmentation, dynamic topology, and quantitative evaluation.

[0084] Based on the same inventive concept, this invention also provides a rebar tying detection system based on instance segmentation and dynamic topology modeling. Since the principle of the problem solved by this system is similar to the aforementioned rebar tying detection method based on instance segmentation and dynamic topology modeling, the implementation of this system can refer to the implementation of the aforementioned method, and the repeated parts will not be described again.

[0085] Reference Figure 5 As shown, this invention also provides a rebar tying detection system based on instance segmentation and dynamic topology modeling, applying the rebar tying detection method based on instance segmentation and dynamic topology modeling as described in any of the above embodiments. The system includes: The point cloud data processing module is used to collect three-dimensional point cloud data of steel mesh through a multi-sensor fusion system; The identification module is used to identify the main body of the steel reinforcement and the binding points by using a mask-optimized instance segmentation network on the three-dimensional point cloud data of the steel reinforcement mesh. A model building module is constructed based on skeleton extraction and curvature weighting strategies to build a dynamic topology model that expresses the spatial connection relationship of steel bars; The quality assessment module is used to generate a quality assessment matrix containing quantitative indicators of tightness based on construction specifications.

[0086] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0087] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting rebar tying based on instance segmentation and dynamic topology modeling, characterized in that, Includes the following steps: S1. Acquire 3D point cloud data of steel mesh through a multi-sensor fusion system; S2. For the three-dimensional point cloud data of the steel mesh, a mask-optimized instance segmentation network is used to identify the main body of the steel bars and the binding points; S3. Construct a dynamic topology model based on skeleton extraction and curvature weighting strategy to express the spatial connection relationship of steel bars; S4. Generate a quality assessment matrix containing quantitative indicators of tightening according to the construction specifications.

2. The rebar tying detection method based on instance segmentation and dynamic topology modeling according to claim 1, characterized in that, Step S1 includes: S11. Acquire millimeter-precision point cloud data using a laser scanner. L Color depth point cloud P is captured synchronously using an RGBD camera. R And based on IMU positioning data, initial registration of multi-view point clouds is performed; S12, Extract laser point cloud P L Mid-curvature extreme feature points and depth point cloud P R Cross-modal matching using SIFT descriptors in the dataset: S13. The RANSAC algorithm is used to solve for the optimal spatial transformation matrix, outliers are eliminated through iterative optimization, and the final point cloud data is obtained through weighted fusion.

3. The rebar tying detection method based on instance segmentation and dynamic topology modeling according to claim 1, characterized in that, The mask-optimized instance segmentation network in step S2 includes: The feature pyramid network infrastructure is used to fuse deep semantic features of the overall structure with shallow detail features that preserve edge information through a top-down path. The reinforcement feature attention module is used to dynamically focus on key areas of the reinforcement through an attention mechanism, suppressing background interference and compensating for missing information. An edge-enhanced loss function is used to jointly optimize mask accuracy and edge localization.

4. The rebar tying detection method based on instance segmentation and dynamic topology modeling according to claim 1, characterized in that, In step S3, constructing the dynamic topology model includes: S31. Skeleton Extraction: Morphological thinning algorithm is used to process the rebar mask image to generate a single-pixel wide skeleton. M represents the binary mask of the main steel reinforcement structure output by the instance segmentation network; B represents the 3×3 structural kernel; k represents the number of iterations; This represents the erosion operation; S32, Node Generation: Create topological nodes at the extreme points of skeleton curvature: , This indicates the rate of change and direction of change of the tangent direction of the steel reinforcement cage curve at a certain point s; This represents the threshold for determining the extreme point of curvature; S33, Adaptive Neighborhood Search: Dynamically determining connected neighborhoods using curvature-driven radius. S34, Curvature-weighted edge connection: Calculate the similarity weights between nodes and establish connections: γ=2.0 represents the control edge connection weight; when When the value is greater than 0.35, an edge e is established. ij ;k i and k j Represents the curvature values ​​of adjacent nodes i and j; S35. Topology Verification: Correcting Connection Relationships Based on Construction Specification Constraints. like Delete edge e ij ;like Add virtual tethering points; P i P j t represents the coordinate vector of adjacent nodes i and j. i t j This indicates the tangent direction between adjacent nodes i and j.

5. The rebar tying detection method based on instance segmentation and dynamic topology modeling according to claim 1, characterized in that, In step S4, the tightness measurement index is defined as: Where k is the number of sampling directions; θ i For the angle between the tie wire and the reinforcing bar, σ d d is the distance decay coefficient, which controls the rate of exponential decay; i Where n is the offset distance of the binding point, n is the actual number of wrapping turns, and n0 is the minimum number of turns specified.

6. A rebar tying detection system based on instance segmentation and dynamic topology modeling, characterized in that, The system, employing the rebar tying detection method based on instance segmentation and dynamic topology modeling as described in any one of claims 1-5, comprises: The point cloud data processing module is used to collect three-dimensional point cloud data of steel mesh through a multi-sensor fusion system; The identification module is used to identify the main body of the steel reinforcement and the binding points by using a mask-optimized instance segmentation network on the three-dimensional point cloud data of the steel reinforcement mesh. A model building module is constructed based on skeleton extraction and curvature weighting strategies to build a dynamic topology model that expresses the spatial connection relationship of steel bars; The quality assessment module is used to generate a quality assessment matrix containing quantitative indicators of tightness based on construction specifications.

Images