A deep learning-based template engineering frame body intelligent recognition system

Abstract

The application discloses a kind of based on deep learning's template engineering frame body intelligent identification system, comprising: image acquisition module, for obtaining the image information of template engineering frame body And corresponding sparse point cloud data are synchronized;Component identification module, for the image information and corresponding sparse point cloud data are synchronized Input pre-trained identification model, and output the identification result of the frame body component in image is obtained;Three-dimensional reconstruction module, for based on sparse point cloud data and the three-dimensional reconstruction of template engineering is carried out to frame body component identification result, and reconstruction model is obtained;Engineering detection module, for judging whether template engineering frame body satisfies construction standard according to reconstruction model.The present application can realize the automatic quality inspection work of template engineering frame body, reduce the input of quality inspection personnel and improve the acceptance efficiency of template engineering frame body.

Description

Technical Field

[0001] This invention belongs to the field of intelligent recognition technology in building engineering, specifically relating to an intelligent recognition system for template engineering structures based on deep learning. Background Technology

[0002] In the construction industry, formwork engineering, especially high-rise formwork projects, suffers from problems such as low acceptance efficiency and repetitive issues. Large-area formwork acceptance often results in incomplete problem identification, posing risks of not being able to effectively and comprehensively rectify the scaffolding. When multiple projects are being inspected, management costs are undoubtedly increased. In formwork engineering scaffolding acceptance, especially for high-rise projects, the erection of components such as uprights, diagonal braces, safety nets, main beams, and scissor braces must all meet the requirements of the design plan. How to intelligently identify and effectively rectify potential hazards in formwork engineering scaffolding is a pressing problem that needs to be solved. Summary of the Invention

[0003] In view of this, the purpose of this invention is to propose a deep learning-based intelligent identification system for template engineering frames, so as to realize the automatic quality inspection of template engineering frames, reduce the input of quality inspection personnel, and improve the acceptance efficiency of template engineering frames.

[0004] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a deep learning-based intelligent recognition system for template engineering structures, comprising: The image acquisition module is used to acquire image information of the template engineering frame and synchronize the corresponding sparse point cloud data; The component recognition module is used to input image information and synchronously corresponding sparse point cloud data into a pre-trained recognition model, and output the recognition results of the frame components in the image. The 3D reconstruction module is used to perform 3D reconstruction of the template project based on sparse point cloud data and the identification results of the frame components, so as to obtain the reconstruction model. The engineering inspection module is used to determine whether the formwork engineering frame meets the construction standards based on the reconstructed model.

[0005] Preferably, the component recognition module includes: The image recognition network unit is used to perform regression and classification processing on image information using a trained target detection model based on the Faster-RCNN algorithm to obtain the first type of target detection result; The point cloud data segmentation unit is used to determine the 2D target bounding box of the detected target in the visual image based on the first type of target detection results, and to map its four corner points into four rays in the three-dimensional world through coordinate system transformation, thereby determining the range of the view frustum where the detected target is located, and to segment the sparse point cloud data based on the range of the view frustum to obtain the region point cloud data. The point cloud feature extraction unit is used to process the regional point cloud data by using PointNet with an attention mechanism as the point cloud feature extraction network, and outputs the category of each point cloud as the second type of target detection result. The feature structure matching unit is used to determine the type of frame component corresponding to the detected target based on the first type of target detection result, determine the surface structure features of the frame component based on the second type of target detection result, and determine the position and orientation of the frame component in three-dimensional space based on the frame component type and its corresponding surface structure features as the recognition result of the frame component in the image.

[0006] Preferably, the Faster-RCNN algorithm comprises four parts: a multi-scale feature extraction network, a region feature extraction network, a ROI pooling layer, and classification and regression layers. The algorithm flow includes: Image information is used as input data, and multi-scale convolutional feature maps are obtained by fusing the ResNet101 residual network and the feature pyramid network. A region feature extraction network is used to extract candidate bounding boxes of regions that may contain the target from the convolutional feature map. The ROI pooling layer performs pooling operations on the candidate bounding boxes of the region, converting inputs of different sizes into outputs of a fixed size. Regression and classification processes are performed using classification and regression layers to obtain target bounding boxes and scores; During the training process of the Faster-RCNN algorithm: The weights of the 2D object detection model were pre-trained on the ImageNet classification and COCO object detection datasets; Fine-tuning of the weights of the 2D object detection model on the KITTI 2D object detection dataset to classify and predict the bounding boxes of 2D objects.

[0007] Preferably, the algorithm flow for the PointNet network incorporating an attention mechanism includes: regional point cloud data The data is output in the form of n, where n is the number of point clouds and 4 is the dimension of the point clouds; The point cloud data of the region is used to extract point cloud features through a two-layer perceptron model with shared parameters, and the point cloud features are aligned by a feature space transformation matrix prediction network. The self-attention mechanism layer is used to encode the point cloud features to obtain feature vectors that can be used for point cloud classification; The feature vector is concatenated with the global features obtained by max pooling to achieve instance segmentation of point cloud data. Finally, the category to which each point cloud belongs is output as the second type of object detection result.

[0008] Preferably, the self-attention mechanism layer encodes point cloud features based on the Multi-Head Attention mechanism. During the encoding process: The Multi-Head Attention mechanism linearly maps Q (Query), K (Key), and V (Value) through matrix parameters on the basis of the original Attention mechanism. The aligned point cloud features are then subjected to a preset number of non-shared parameter Attention operations. The point cloud features are fused through scaled dot product and Softmax normalization. Finally, the point cloud features are connected to the corresponding V (Value) values ​​through residuals to output a feature vector that can be used for point cloud classification. The computational model for the self-attention mechanism layer includes: Based on the self-attention mechanism, the three inputs Q (Query), K (Key), and V (Value) are set as the same matrix X; The encoding result for the point cloud features is: The traditional Multi-Head Attention mechanism utilizes three input components: Q (Query), K (Key), and V (Value), which can be represented as follows: in, For the point cloud feature set of regional point cloud data, The encoding result of point cloud features, For a learnable parameter matrix, For feature splicing work, ; For any point cloud feature The encoding process is as follows: in, Represented as the normalization factor, It is a matrix containing s key values. yes The dimensions of each query condition.

[0009] Preferably, the 3D reconstruction module performs the following: Based on the recognition results, the type of each of the multiple frame components is determined, as well as the position and orientation of the frame components in three-dimensional space; The component model corresponding to each type of frame component is determined by searching the pre-set component model library; Based on the position and orientation of the frame components in three-dimensional space, all component models are rotated, translated, and scaled to obtain a simple reconstructed three-dimensional model of the template engineering frame. The sparse point cloud data is used to cover the 3D model of the template engineering frame that is simply reconstructed. The model of each component is adjusted in a feedback manner according to the contact rate between the point cloud and the model surface. The reconstructed model is then output after the adjustment is completed.

[0010] Preferably, the engineering testing module performs the following: Obtain the reconstruction model and determine the coordinate positions of the key feature nodes corresponding to each frame component in the reconstruction model. Based on the coordinate positions of key feature nodes in the reconstruction model, the second type of relative positional relationship between frame components is determined; Obtain the standard construction model of the template engineering frame obtained in advance through modeling before construction, and extract the first type of relative positional relationship between the frame components of the standard construction model; The matching degree is obtained by matching the first type of relative positional relationship and the second type of relative positional relationship. When the matching degree is greater than the preset matching degree threshold, it is determined that the formwork engineering frame meets the construction standards.

[0011] The present invention has achieved at least the following beneficial effects: 1. It can realize automatic quality inspection of formwork engineering frames, reduce the input of quality inspection personnel and improve the acceptance efficiency of formwork engineering frames; 2. The system prioritizes identification based on image information, then segments the locked target data to obtain point cloud data belonging to the target for further identification and judgment. While ensuring identification accuracy, it reduces the volume of point cloud data and the amount of computation required for point cloud data, thereby significantly improving the identification efficiency of frame components. 3. To achieve the scanning and modeling of the three-dimensional model of the formwork structure after construction is completed.

[0012] Other advantages, objectives, and features of the invention will be set forth in the following description and will be apparent to those skilled in the art in some respects, or may be learned by practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0013] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration: Figure 1 This is a schematic diagram of the structure of a template engineering frame intelligent recognition system based on deep learning in an embodiment of the present invention; Figure 2This is a flowchart of the Faster-RCNN algorithm in an embodiment of the present invention; Figure 3 This is a flowchart of the PointNet network algorithm that incorporates an attention mechanism in an embodiment of the present invention. Figure 4 This is a flowchart of the self-attention mechanism layer in an embodiment of the present invention. Detailed Implementation

[0014] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0015] The present invention provides a deep learning-based intelligent recognition system for template engineering structures, comprising: The image acquisition module is used to acquire image information of the template engineering frame and synchronize the corresponding sparse point cloud data; The component recognition module is used to input image information and synchronously corresponding sparse point cloud data into a pre-trained recognition model, and output the recognition results of the frame components in the image. The 3D reconstruction module is used to perform 3D reconstruction of the template project based on sparse point cloud data and the identification results of the frame components, so as to obtain the reconstruction model. The engineering inspection module is used to determine whether the formwork engineering frame meets the construction standards based on the reconstructed model.

[0016] The working principle and beneficial effects of the above technical solution are as follows: An image acquisition module acquires image information of the formwork frame and corresponding sparse point cloud data. A component recognition module inputs the image information and corresponding sparse point cloud data into a pre-trained recognition model and outputs the recognition results of the frame components in the image. These recognition results include the identification of the frame component type, as well as the identification of the frame component's position and pose, thus facilitating the reconstruction of the formwork frame model based on the position and pose of each frame component. The frame structure includes, but is not limited to, uprights, tie rods, safety nets, main beams, scissor braces, top supports, and main beam square steel. Then, a 3D reconstruction module performs 3D reconstruction of the formwork project based on the sparse point cloud data and the recognition results of the frame components, obtaining a reconstructed model. Finally, engineering inspection determines whether the formwork frame meets construction standards based on the reconstructed model, thereby achieving the recognition of the formwork frame. This invention automatically collects and processes image and point cloud data, and uses deep learning neural networks to intelligently identify various scaffold components. Finally, based on the identification results, it reconstructs the template project in three-dimensional space. This allows for easy comparison between the reconstructed 3D model and a preset standard 3D model to determine whether the template project meets construction standards. No manual on-site acceptance is required; simply importing the template frame image and scanned sparse point cloud data into the system enables automatic identification. This effectively improves the efficiency of template project acceptance and reduces management costs by eliminating the need for on-site personnel.

[0017] In a preferred embodiment, the image acquisition module acquires image information of the template engineering frame and simultaneously collects corresponding sparse point cloud data in the same time and space. After acquiring the image information and sparse point cloud data, the image information and sparse point cloud data are filtered to reduce noise interference to subsequent algorithms. Because subsequent algorithms involve data fusion from multiple sensors or cameras, it is necessary to perform data matching and alignment on image information and sparse point cloud data to ensure that image information and sparse point cloud data are in the same coordinate system on the 2D data plane. To reduce the amount of data to be computed, feature extraction can be performed on image information and sparse point cloud data. For example, the edges, textures and colors corresponding to the image information can be used for target recognition, and the features such as ground planes and object boundaries extracted from sparse point clouds can be used for modeling and segmentation. In some cases, image information and sparse point cloud data may be very large and difficult to process directly. Dimensionality reduction techniques such as image compression, sampling, or point cloud downsampling can be used to reduce the data volume and simplify the processing. In some cases, to improve the generalization ability and robustness of the model, data augmentation operations such as rotation, translation and mirroring can be performed on image information and sparse point cloud data to expand the training dataset used for model training.

[0018] In a preferred embodiment, to achieve the fusion of image information and sparse point cloud data, a Bursa model is used to translate the origin of the original coordinate system, rotate the coordinate axes to align them with the target coordinate system, and scale the coordinate axes to ultimately achieve coordinate system transformation. Due to the synchronous acquisition of image information and sparse point cloud data, the scaling ratio of the coordinate system in each direction is approximately equal, represented by a single scaling parameter. The rotation angles of the three degrees of freedom are also relatively small, and their sine values ​​are replaced by angle values. The coordinate transformation model is as follows: In the formula, , , Initial coordinate system coordinates in , , For the target coordinate system coordinates in , , For translation parameters, , , M is the rotation parameter, and M is the scaling parameter; The data transformed by the Bursa model satisfies the solution posture of the iterative nearest neighbor algorithm, that is, the nearest point is the matching point, and the rotation relationship R and the translation relationship t are minimized: in, This indicates the accuracy of the point cloud after spatial reference transformation; The rotation relation R and the translation relation t can be obtained separately and expressed as: in, , Let the centroids of the two point clouds be... , Let n be the coordinates of the two sets of point clouds after centroid removal, and n be the total number of point clouds. The above method achieves coordinate unification between image information and sparse point cloud data.

[0019] In a preferred embodiment, the component identification module includes: The image recognition network unit is used to perform regression and classification processing on image information using a trained target detection model based on the Faster-RCNN algorithm to obtain the first type of target detection result; The point cloud data segmentation unit is used to determine the 2D target bounding box of the detected target in the visual image based on the first type of target detection results, and to map its four corner points into four rays in the three-dimensional world through coordinate system transformation, thereby determining the range of the view frustum where the detected target is located, and to segment the sparse point cloud data based on the range of the view frustum to obtain the region point cloud data. The point cloud feature extraction unit is used to process the regional point cloud data by using PointNet with an attention mechanism as the point cloud feature extraction network, and outputs the category of each point cloud as the second type of target detection result. The feature structure matching unit is used to determine the type of frame component corresponding to the detected target based on the first type of target detection result, determine the surface structure features of the frame component based on the second type of target detection result, and determine the position and orientation of the frame component in three-dimensional space based on the frame component type and its corresponding surface structure features as the recognition result of the frame component in the image.

[0020] The working principle and beneficial effects of the above technical solution are as follows: The image recognition network unit uses a trained target detection model based on the Faster-RCNN algorithm to perform regression and classification processing on image information, obtaining the first type of target detection result. The point cloud data segmentation unit determines the 2D target bounding box in the visual image based on the first type of target detection result, and maps its four corner points to four rays in the three-dimensional world through coordinate system transformation, thereby determining the view frustum range where the target is located. Based on the view frustum range, the sparse point cloud data is segmented to obtain regional point cloud data, thus realizing the segmentation of point cloud data belonging to the same structural framework. This method significantly reduces the computational cost of point cloud feature extraction. The point cloud feature extraction unit employs PointNet with an attention mechanism to process regional point cloud data and outputs the category of each point cloud as the second-class target detection result, thus achieving point cloud data detection. The feature structure matching unit determines the type of the frame component corresponding to the detected target based on the first-class target detection result, determines the surface structure features of the frame component based on the second-class target detection result, and determines the position and orientation of the frame component in three-dimensional space based on the frame component type and its corresponding surface structure features, serving as the recognition result of the frame component in the image. This allows for priority recognition using image information, followed by data segmentation of the locked detection target to obtain point cloud data belonging to the detection target for further recognition and judgment. While ensuring recognition accuracy, this method reduces the volume of point cloud data and the computational cost, thereby significantly improving the recognition efficiency of frame components.

[0021] In a preferred embodiment, the Faster-RCNN algorithm comprises four parts: a multi-scale feature extraction network, a region feature extraction network, an ROI pooling layer, and classification and regression layers. The algorithm flow includes: Image information is used as input data, and multi-scale convolutional feature maps are obtained by fusing the ResNet101 residual network and the feature pyramid network. A region feature extraction network is used to extract candidate bounding boxes of regions that may contain the target from the convolutional feature map. The ROI pooling layer performs pooling operations on the candidate bounding boxes of the region, converting inputs of different sizes into outputs of a fixed size. Regression and classification processes are performed using classification and regression layers to obtain target bounding boxes and scores; During the training process of the Faster-RCNN algorithm: The weights of the 2D object detection model were pre-trained on the ImageNet classification and COCO object detection datasets; Fine-tuning of the weights of the 2D object detection model on the KITTI 2D object detection dataset to classify and predict the bounding boxes of 2D objects.

[0022] The working principle and beneficial effects of the above technical solution are as follows: A feature pyramid network is integrated into the original Faster-RCNN algorithm to combine high-resolution low-level features with high semantic information of high-level features, thus improving detection accuracy. The improved Faster-RCNN algorithm consists of four parts: a multi-scale feature extraction network, a region feature extraction network, an ROI pooling layer, and classification and regression layers. Figure 2 As shown, the algorithm first takes image information as input, which is then fused with a ResNet101 residual network and a feature pyramid network to obtain multi-scale convolutional feature maps. A region feature extraction network is then used to extract candidate bounding boxes (ROIs) that may contain the target from these ROIs. ROI pooling layers are then used to pool these candidate bounding boxes, converting inputs of different sizes into fixed-size outputs. Finally, a detection network is used for regression and classification to obtain the target bounding boxes and scores. The training process involves pre-training the 2D target detection model weights on the ImageNet classification and COCO target detection datasets. The weights are then fine-tuned on the KITTI 2D target detection dataset to classify and predict the bounding boxes of 2D targets. These techniques are used to train the image recognition network.

[0023] In a preferred embodiment, refer to Figure 3 The algorithm flow of the PointNet network with an attention mechanism includes: regional point cloud data The data is output in the form of n, where n is the number of point clouds and 4 is the dimension of the point clouds; The point cloud data of the region is used to extract point cloud features through a two-layer perceptron model with shared parameters, and the point cloud features are aligned by a feature space transformation matrix prediction network. The self-attention mechanism layer is used to encode the point cloud features to obtain feature vectors that can be used for point cloud classification; The feature vector is concatenated with the global features obtained by max pooling to achieve instance segmentation of point cloud data. Finally, the category to which each point cloud belongs is output as the second type of object detection result.

[0024] The working principle and beneficial effects of the above technical solution are as follows: It achieves the recognition of sparse point cloud data by introducing an attention mechanism into the PointNet network. The algorithm flow of the PointNet network with the attention mechanism includes: processing the regional point cloud data as follows: Data is output in the form of n, where n is the number of point clouds and 4 is the dimension of the point clouds. The regional point cloud data is processed using a two-layer perceptron model with shared parameters to extract point cloud features, and the features are aligned using a feature space transformation matrix prediction network. A self-attention mechanism layer is used to encode the point cloud features to obtain feature vectors that can be used for point cloud classification. The feature vectors are concatenated with the global features obtained through max pooling to achieve instance segmentation of the point cloud data. Finally, the category of each point cloud is output as the second-class object detection result. The overall framework is illustrated in the diagram below. Figure 3 The above technical solutions enable the construction of a point cloud feature extraction network.

[0025] In a preferred embodiment, refer to Figure 4 The self-attention mechanism layer encodes point cloud features based on the Multi-Head Attention mechanism. During the encoding process: The Multi-Head Attention mechanism linearly maps Q (Query), K (Key), and V (Value) through matrix parameters on the basis of the original Attention mechanism. The aligned point cloud features are then subjected to a preset number of non-shared parameter Attention operations. The point cloud features are fused through scaled dot product and Softmax normalization. Finally, the point cloud features are connected to the corresponding V (Value) values ​​through residuals to output a feature vector that can be used for point cloud classification. The computational model for the self-attention mechanism layer includes: Based on the self-attention mechanism, the three inputs Q (Query), K (Key), and V (Value) are set as the same matrix X; The encoding result for the point cloud features is: The traditional Multi-Head Attention mechanism utilizes three input components: Q (Query), K (Key), and V (Value), which can be represented as follows: in, For the point cloud feature set of regional point cloud data, The encoding result of point cloud features, For a learnable parameter matrix, For feature splicing work, ; For any point cloud feature The encoding process is as follows: in, Represented as the normalization factor, It is a matrix containing s key values. yes The dimensions of each query condition.

[0026] The working principle and beneficial effects of the above technical solution are as follows: by introducing a self-attention mechanism to realize the model, the cloud feature extraction network can fully explore the local feature information of the point cloud, thereby improving the recognition accuracy of the point cloud feature extraction network.

[0027] In a preferred embodiment, the 3D reconstruction module performs the following: Based on the recognition results, the type of each of the multiple frame components is determined, as well as the position and orientation of the frame components in three-dimensional space; The component model corresponding to each type of frame component is determined by searching the pre-set component model library; Based on the position and orientation of the frame components in three-dimensional space, all component models are rotated, translated, and scaled to obtain a simple reconstructed three-dimensional model of the template engineering frame. The sparse point cloud data is used to cover the 3D model of the template engineering frame that is simply reconstructed. The model of each component is adjusted in a feedback manner according to the contact rate between the point cloud and the model surface. The reconstructed model is then output after the adjustment is completed.

[0028] The working principle and beneficial effects of the above technical solution are as follows: The identification results determine the type of each of the multiple frame components, as well as their position and orientation in three-dimensional space; a pre-set component model library is used to find the corresponding component model for each type of frame component; based on the position and orientation of the frame components in three-dimensional space, all component models are rotated, translated, and scaled to obtain a simple reconstructed three-dimensional model of the template engineering frame; finally, sparse point cloud data is used to cover the simple reconstructed three-dimensional model of the template engineering frame, and each component model is adjusted in a feedback manner according to the contact rate between the point cloud and the model surface. During the adjustment process, the model is continuously adjusted by scaling, rotating, and translating to maximize the contact between the point cloud and the model surface without including the point cloud, thereby achieving fast and accurate alignment of the model. After adjustment, the reconstructed model is output. The above technical solution enables the scanning and modeling of a 3D model of the completed formwork structure. Compared to the traditional method of directly using point clouds for compensation modeling, the above technical solution uses the splicing and adjustment of prefabricated components (pre-established 3D models of the structure components) to construct the overall model of the formwork structure. This allows the constructed 3D model to fully preserve the corner details of the formwork structure (such as the corners, protrusions, and depressions). Under this technology, the density of point cloud data acquisition can be effectively reduced, and the amount of data processing and calculation can be decreased, thereby indirectly improving the recognition speed of the formwork structure.

[0029] In a preferred embodiment, the engineering inspection module performs the following: Obtain the reconstruction model and determine the coordinate positions of the key feature nodes corresponding to each frame component in the reconstruction model. Based on the coordinate positions of key feature nodes in the reconstruction model, the second type of relative positional relationship between frame components is determined; Obtain the standard construction model of the template engineering frame obtained in advance through modeling before construction, and extract the first type of relative positional relationship between the frame components of the standard construction model; The matching degree is obtained by matching the first type of relative positional relationship and the second type of relative positional relationship. When the matching degree is greater than the preset matching degree threshold, it is determined that the formwork engineering frame meets the construction standards.

[0030] The working principle and beneficial effects of the above technical solution are as follows: The engineering inspection module acquires the reconstructed model and determines the coordinate positions of the key feature nodes corresponding to each frame component in the reconstructed model; based on the coordinate positions of the key feature nodes in the reconstructed model, it determines the second type of relative positional relationship between the frame components; it acquires the standard construction model of the formwork engineering frame obtained in advance through modeling before construction, and extracts the first type of relative positional relationship between the frame components of the standard construction model; it matches and calculates the matching degree between the first type of relative positional relationship and the second type of relative positional relationship, and when the matching degree is greater than the preset matching degree threshold, it is determined that the formwork engineering frame meets the construction standards. Thus, the automatic quality inspection of the formwork engineering frame is ultimately achieved. As an alternative technical solution, the quality inspection rules between the frame components can be pre-set (e.g., the distance between two uprights under the beam of a tall formwork project, the spacing of the diagonal bracing in each span, the arrangement of scissor bracing on the first floor, etc.) to determine whether the second type of relative positional relationship meets the quality inspection rules. Items that do not meet the quality inspection rules can be displayed and marked, thereby realizing the quality inspection of the formwork frame. This alternative technical solution can be applied to the quality inspection of different formwork frames.

[0031] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made to it in form and detail without departing from the scope defined by the claims of the present invention.

Claims

1. A deep learning-based intelligent recognition system for template engineering structures, characterized in that, include: The image acquisition module is used to acquire image information of the template engineering frame and synchronize the corresponding sparse point cloud data; The component recognition module is used to input image information and synchronously corresponding sparse point cloud data into a pre-trained recognition model, and output the recognition results of the frame components in the image. The 3D reconstruction module is used to perform 3D reconstruction of the template project based on sparse point cloud data and the identification results of the frame components, so as to obtain the reconstruction model. The engineering inspection module is used to determine whether the formwork structure meets the construction standards based on the reconstructed model. The engineering testing module performs the following: Obtain the reconstruction model and determine the coordinate positions of the key feature nodes corresponding to each frame component in the reconstruction model. Based on the coordinate positions of key feature nodes in the reconstruction model, the second type of relative positional relationship between frame components is determined; Obtain the standard construction model of the template engineering frame obtained in advance through modeling before construction, and extract the first type of relative positional relationship between the frame components of the standard construction model; The matching degree is obtained by matching the first type of relative positional relationship and the second type of relative positional relationship. When the matching degree is greater than the preset matching degree threshold, it is determined that the formwork engineering frame meets the construction standards.

2. The intelligent recognition system for template engineering frames based on deep learning according to claim 1, characterized in that, The component identification module includes: The image recognition network unit is used to perform regression and classification processing on image information using a trained target detection model based on the Faster-RCNN algorithm to obtain the first type of target detection result; The point cloud data segmentation unit is used to determine the 2D target bounding box of the detected target in the visual image based on the first type of target detection results, and to map its four corner points into four rays in the three-dimensional world through coordinate system transformation, thereby determining the range of the view frustum where the detected target is located, and to segment the sparse point cloud data based on the range of the view frustum to obtain the region point cloud data. The point cloud feature extraction unit is used to process the regional point cloud data by using PointNet with an attention mechanism as the point cloud feature extraction network, and outputs the category of each point cloud as the second type of target detection result. The feature structure matching unit is used to determine the type of frame component corresponding to the detected target based on the first type of target detection result, determine the surface structure features of the frame component based on the second type of target detection result, and determine the position and orientation of the frame component in three-dimensional space based on the frame component type and its corresponding surface structure features as the recognition result of the frame component in the image.

3. The intelligent recognition system for template engineering frames based on deep learning according to claim 2, characterized in that, The Faster-RCNN algorithm comprises four parts: a multi-scale feature extraction network, a region feature extraction network, a ROI pooling layer, and classification and regression layers. The algorithm flow includes: Image information is used as input data, and multi-scale convolutional feature maps are obtained by fusing the ResNet101 residual network and the feature pyramid network. A region feature extraction network is used to extract candidate bounding boxes of regions that may contain the target from the convolutional feature map. The ROI pooling layer performs pooling operations on the candidate bounding boxes of the region, converting inputs of different sizes into outputs of a fixed size. Regression and classification processes are performed using classification and regression layers to obtain target bounding boxes and scores; During the training process of the Faster-RCNN algorithm: The weights of the 2D object detection model were pre-trained on the ImageNet classification and COCO object detection datasets; Fine-tuning of the weights of the 2D object detection model on the KITTI 2D object detection dataset to classify and predict the bounding boxes of 2D objects.

4. The intelligent recognition system for template engineering frames based on deep learning according to claim 2, characterized in that, The algorithm flow of the PointNet network with the introduced attention mechanism includes: regional point cloud data The data is output in the form of n, where n is the number of point clouds and 4 is the dimension of the point clouds; The point cloud data of the region is used to extract point cloud features through a two-layer perceptron model with shared parameters, and the point cloud features are aligned by a feature space transformation matrix prediction network. The self-attention mechanism layer is used to encode the point cloud features to obtain feature vectors that can be used for point cloud classification; The feature vector is concatenated with the global features obtained by max pooling to achieve instance segmentation of point cloud data. Finally, the category to which each point cloud belongs is output as the second type of object detection result.

5. The intelligent recognition system for template engineering frames based on deep learning according to claim 4, characterized in that, The self-attention mechanism layer encodes point cloud features based on the Multi-Head Attention mechanism. During the encoding process: The Multi-Head Attention mechanism linearly maps Q (Query), K (Key), and V (Value) through matrix parameters on the basis of the original Attention mechanism. The aligned point cloud features are then subjected to a preset number of non-shared parameter Attention operations. The point cloud features are fused through scaled dot product and Softmax normalization. Finally, the point cloud features are connected to the corresponding V (Value) values ​​through residuals to output a feature vector that can be used for point cloud classification. The computational model for the self-attention mechanism layer includes: Based on the self-attention mechanism, the three inputs Q (Query), K (Key), and V (Value) are set as the same matrix X; The encoding result for the point cloud features is: The traditional Multi-Head Attention mechanism utilizes three input components: Q (Query), K (Key), and V (Value), which can be represented as follows: in, For the point cloud feature set of regional point cloud data, The encoding result of point cloud features, For a learnable parameter matrix, For feature splicing work, ; For any point cloud feature The encoding process is as follows: in, Represented as the normalization factor, It is a matrix containing s key values. yes The dimensions of each query condition.

6. The intelligent recognition system for template engineering frames based on deep learning according to claim 1, characterized in that, The three-dimensional reconstruction module performs the following: Based on the recognition results, the type of each of the multiple frame components is determined, as well as the position and orientation of the frame components in three-dimensional space; The component model corresponding to each type of frame component is determined by searching the pre-set component model library; Based on the position and orientation of the frame components in three-dimensional space, all component models are rotated, translated, and scaled to obtain a simple reconstructed three-dimensional model of the template engineering frame. The sparse point cloud data is used to cover the 3D model of the template engineering frame that is simply reconstructed. The model of each component is adjusted in a feedback manner according to the contact rate between the point cloud and the model surface. The reconstructed model is then output after the adjustment is completed.

Images