A steel structure pre-buried anchor bolt group batch intelligent detection method and system

Abstract

The application provides a steel structure pre-buried anchor bolt group batch intelligent detection method and system, the method comprising: determining a to-be-detected construction area; target layout is carried out on the to-be-detected construction area, the absolute space coordinates of each of a plurality of peripheral targets and the group target code of each of a plurality of anchor bolt groups are determined; based on the absolute space coordinates of each of the plurality of peripheral targets, image acquisition is carried out on the to-be-detected construction area, and a four-scale image sequence is obtained; based on the group target code of each of the plurality of anchor bolt groups, image preprocessing and anchor bolt group identification marking are carried out on the four-scale image sequence, a mask image sequence and a scale type corresponding to the mask image sequence are obtained; based on the mask image sequence, the scale type corresponding to the mask image sequence and the absolute space coordinates of each of the plurality of peripheral targets, anchor bolt detection is carried out through a lightweight space position mapping network, and a global top view is obtained. The above scheme can improve the efficiency, accuracy and traceability of large-scale steel structure pre-buried anchor bolt group detection.

Description

Technical Field

[0001] This application relates to the field of civil engineering technology, specifically to a method and system for batch intelligent detection of pre-embedded anchor bolt groups in steel structures. Background Technology

[0002] In steel structure construction, the positioning accuracy of pre-embedded anchor bolt groups directly affects the installation quality and structural safety of upper steel columns, beams, and other components. In actual construction, anchor bolt groups are typically deployed in batches over a large construction area, with each group containing dozens to hundreds of anchor bolts, and the number of anchor bolt groups reaching dozens. Their spatial distribution is characterized by its large quantity, wide coverage, and scattered distribution. The final position of the anchor bolts is affected by multiple factors during concrete pouring, such as mold offset, concrete shrinkage, vibration disturbance, and changes in ambient temperature. Deviations exist between the actual spatial position and the theoretical design position. Therefore, it is necessary to conduct position checks and acceptance tests on all anchor bolts before and after the initial setting of the concrete.

[0003] Traditional anchor bolt inspection methods primarily rely on surveying equipment such as total stations, theodolites, or RTK instruments to measure the spatial coordinates of each anchor bolt point by point, and then compare the deviation with the design values. While this method is reliable in accuracy, it involves a large workload, long operation cycle, and high labor costs when dealing with hundreds or thousands of anchor bolts. Furthermore, the measurement data is mostly recorded in the form of discrete points, making it difficult to form a continuous location distribution map covering the entire area, which is not conducive to the overall macro-assessment of installation quality and the tracing of problems.

[0004] In recent years, drone aerial photography and computer vision technologies have been increasingly applied to engineering surveying, providing a new technical approach for the rapid detection of anchor bolt groups. However, existing vision-based detection schemes are mostly designed for single anchor bolts or small anchor bolt groups. For large construction areas containing multiple anchor bolt groups, each aerial image only covers a local area, lacking a unified physical spatial reference. How to efficiently and accurately map the anchor bolt information detected in multiple local images to a global physical coordinate system while maintaining the identification of each anchor bolt group is a pressing technical challenge.

[0005] Meanwhile, the images collected by UAVs in a single inspection span a wide range of scales—from global images containing all peripheral targets to detailed images covering only a single anchor bolt group. The spatial coverage, resolution, and number of reference targets included in each frame vary significantly. Existing image stitching or feature matching methods often suffer from cumulative errors or mapping failures when processing such multi-scale, non-uniformly covered image sequences due to large scale differences and inconsistent overlapping areas, making it difficult to maintain sub-anchor bolt level positioning accuracy over a large area. Summary of the Invention

[0006] This application aims to provide a method and system for batch intelligent inspection of pre-embedded anchor bolt groups in steel structures, which can improve the efficiency, accuracy and traceability of inspection of pre-embedded anchor bolt groups in large-scale steel structures.

[0007] The technical solution of this application is implemented as follows: In a first aspect, embodiments of this application provide a method for batch intelligent detection of pre-embedded anchor bolt groups in steel structures, the method comprising: The construction area to be inspected is determined; and targets are set up in the construction area to be inspected, and the absolute spatial coordinates of each of the multiple peripheral targets and the group target codes of each of the multiple anchor bolt groups are determined. Based on the absolute spatial coordinates of the multiple peripheral targets, images are acquired from the construction area to be detected to obtain a four-scale image sequence; wherein, the four-scale image sequence contains images of four different scale types; Based on the group target encoding of each of the multiple anchor bolt groups, the four-scale image sequence is preprocessed and the anchor bolt group is identified and marked to obtain the mask image sequence and the scale type corresponding to the mask image sequence; Based on the mask image sequence, the scale type corresponding to the mask image sequence, and the absolute spatial coordinates of the multiple peripheral targets, anchor bolt detection is performed through a lightweight spatial location mapping network to obtain a global top view; wherein, the global top view is a pixel-level distribution map of the actual spatial positions of all anchor bolts in the construction area to be detected.

[0008] In the above scheme, the step of setting up targets in the construction area to be inspected, and determining the absolute spatial coordinates of each of the multiple peripheral targets and the group target codes of each of the multiple anchor bolt groups, includes: Targets are deployed around the construction area to be inspected, and the global absolute reference frame of the target reference frame is determined. Measure the spatial distance between multiple peripheral targets in the peripheral target reference frame; The absolute spatial coordinates of each of the multiple peripheral targets are determined based on the spatial distance between them and the global absolute reference system. Multiple target groups are deployed in the construction area to be inspected. The multiple target groups are encoded to obtain the target group codes for each of the multiple anchor bolt groups; wherein, there is a correspondence between the multiple target groups and the multiple anchor bolt groups.

[0009] In the above scheme, the step of acquiring images of the construction area to be detected based on the absolute spatial coordinates of the multiple peripheral targets to obtain a four-scale image sequence includes: Based on the absolute spatial coordinates of the multiple peripheral targets, flight path planning is performed to determine the UAV's cruising route. The patrol route is used to collect images of the construction area to be inspected, and a four-scale image sequence is obtained.

[0010] In the above scheme, the step of acquiring images of the construction area to be inspected and obtaining a four-scale image sequence by means of the cruise route includes: The first-scale image is acquired by using the cruise route to obtain a global target image; wherein, the global target image represents an image covering all peripheral targets; The second-scale image acquisition is performed on the construction area to be detected via the cruise route to obtain target and group association images; wherein, the target and group association images represent images covering at least one peripheral target and at least one group target; The patrol route is used to acquire images of the construction area to be inspected at a third scale, resulting in multi-group area images; wherein, the multi-group area images represent images covering at least two groups of targets; The patrol route is used to acquire images of the construction area to be inspected at a fourth scale, resulting in a single-group detailed survey image; wherein, the single-group detailed survey image represents an image covering only one group of targets; Based on the global target image, the target and swarm association image, the multi-swarm region image, and the single-swarm detailed investigation image, the four-scale image sequence is determined.

[0011] In the above scheme, the step of performing image preprocessing and anchor group identification marking on the four-scale image sequence based on the group target encoding of each of the multiple anchor bolt groups to obtain the mask image sequence and the scale type corresponding to the mask image sequence includes: By using the pre-acquired camera distortion coefficients and camera pose angles, distortion correction and perspective correction are performed on each frame of the four-scale image sequence to obtain the corrected four-scale image sequence. Target coding recognition is performed on the corrected four-scale image sequence to determine the group target code and peripheral target of each of the multiple anchor bolt groups in the corrected four-scale image sequence; Anchor bolt detection is performed on the corrected four-scale image sequence to determine the pixel coordinates of each anchor bolt. Based on the group target encoding, the pixel coordinates of the outer target and the multiple anchor bolts, values ​​are assigned and a mask is output to obtain the mask image sequence and the scale type corresponding to the mask image sequence.

[0012] In the above scheme, the step of detecting anchor bolts and obtaining a global top view based on the mask image sequence, the scale type corresponding to the mask image sequence, and the absolute spatial coordinates of the multiple peripheral targets, using a lightweight spatial location mapping network, includes: Get the preset physical resolution; The mask image sequence, the scale type corresponding to the mask image sequence, and the absolute spatial coordinates of each of the multiple peripheral targets are preprocessed to obtain the mask stacking vector, the scale encoding vector, and the normalized coordinate vector. Based on the mask stacking vector, the scale encoding vector, the normalized coordinate vector, and the preset physical resolution, anchor bolt detection is performed through the lightweight spatial location mapping network to obtain a global top view.

[0013] In the above scheme, the step of detecting anchor bolts and obtaining a global top view based on the mask stacking vector, the scale encoding vector, the normalized coordinate vector, and the preset physical resolution through the lightweight spatial location mapping network includes: Based on the mask stacking vector and the scale encoding vector, convolutional pooling is performed through the encoding layer of the lightweight spatial location mapping network to obtain a high-level semantic feature map. The bottleneck layer of the lightweight spatial location mapping network is used to perform convolution processing on the high-level semantic feature map to obtain the bottleneck feature tensor. The normalized coordinate vector and the preset physical resolution are encoded by the spatial prior injection layer of the lightweight spatial location mapping network to obtain coordinate embedding feature map and resolution embedding feature map respectively. Based on the bottleneck feature tensor, the coordinate embedding feature map, and the resolution embedding feature map, the prior enhancement features are obtained by concatenation and convolution through the spatial prior injection layer of the lightweight spatial location mapping network. Based on the aforementioned prior enhancement features, convolution and concatenation processing are performed through the decoding layer of the lightweight spatial location mapping network to obtain the final feature map; Based on the final feature map, anchor bolt detection is performed on the region to be detected through the global image reconstruction layer of the lightweight spatial location mapping network to obtain the global top view.

[0014] Secondly, embodiments of this application provide a batch intelligent detection system for pre-embedded anchor bolt groups in steel structures, comprising: a determination module, a data acquisition module, a preprocessing and marking module, and a detection module, wherein... The determining module is used to determine the construction area to be inspected; and to set up targets in the construction area to be inspected, and to determine the absolute spatial coordinates of each of the multiple peripheral targets and the group target codes of each of the multiple anchor bolt groups. The acquisition module is used to acquire images of the construction area to be detected based on the absolute spatial coordinates of the multiple peripheral targets, and obtain a four-scale image sequence; wherein, the four-scale image sequence contains images of four different scale types; The preprocessing and labeling module is used to perform image preprocessing and anchor group identification labeling on the four-scale image sequence based on the group target encoding of each of the multiple anchor groups, to obtain a mask image sequence and the scale type corresponding to the mask image sequence; The detection module is used to perform anchor bolt detection through a lightweight spatial location mapping network based on the mask image sequence, the scale type corresponding to the mask image sequence, and the absolute spatial coordinates of the multiple peripheral targets, to obtain a global top view; wherein, the global top view is a pixel-level distribution map of the actual spatial positions of all anchor bolts in the construction area to be detected.

[0015] Thirdly, embodiments of this application provide a batch intelligent detection device for pre-embedded anchor bolt groups in steel structures, comprising: a processor and a memory; wherein, Memory, used to store computer programs; A processor is used to retrieve and run computer programs from memory to perform methods such as those described in the first aspect.

[0016] Fourthly, embodiments of this application provide a computer-readable storage medium storing executable instructions for causing a processor to perform the method as described in the first aspect.

[0017] This application provides a method and system for batch intelligent detection of pre-embedded anchor bolt groups in steel structures. The method determines the construction area to be detected; targets are deployed in the construction area to be detected, and the absolute spatial coordinates of multiple peripheral targets and the group target codes of multiple anchor bolt groups are determined; based on the absolute spatial coordinates of the multiple peripheral targets, images of the construction area to be detected are acquired to obtain a four-scale image sequence; wherein, the four-scale image sequence contains images of four different scale types; based on the group target codes of the multiple anchor bolt groups, the four-scale image sequence is preprocessed and anchor bolt group identifiers are added to obtain a mask image sequence and the scale type corresponding to the mask image sequence; based on the mask image sequence, the scale type corresponding to the mask image sequence, and the absolute spatial coordinates of the multiple peripheral targets, anchor bolt detection is performed through a lightweight spatial location mapping network to obtain a global top view; wherein, the global top view is a pixel-level distribution map of the actual spatial positions of all anchor bolts in the construction area to be detected. In the above scheme, the method of using the absolute spatial coordinates of the outer targets as a reference and directly regressing the global top view using a lightweight neural network avoids the cumulative errors of traditional image stitching and achieves a complete single-frame representation of the anchor bolt positions across the entire region. By deploying coded group targets at the geometric center of each anchor bolt group, each anchor bolt is assigned a unique code to its group during the detection stage, ensuring that each anchor bolt pixel in the output global top view carries clear group identity information. By encoding the absolute spatial coordinates of the outer targets, the preset physical resolution, the mask image sequence, and the scale type of each frame image as auxiliary information and inputting it into the lightweight neural network, the model can adaptively adjust the contribution weight of each frame feature in the global reconstruction according to the image scale. When restoring the global top view, it is directly constrained by the physical reference, ensuring that the output pixel positions have a clear physical spatial correspondence, avoiding the cumulative errors of traditional stitching methods. At the same time, it achieves unified batch output of any number of anchor bolt groups and any number of anchor bolts, thereby improving the efficiency, accuracy, and traceability of large-scale steel structure pre-embedded anchor bolt group detection. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the specification, serve to explain the technical solutions of this application. Obviously, the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0019] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0020] Figure 1 This is an optional flowchart illustrating a batch intelligent detection method for pre-embedded anchor bolt groups in steel structures, provided as an embodiment of this application. Figure 2 A schematic diagram of a batch intelligent detection system for pre-embedded anchor bolt groups in steel structures provided in this application embodiment; Figure 3 This is a structural schematic diagram of a batch intelligent detection device for pre-embedded anchor bolt groups in steel structures, provided as an embodiment of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the specific technical solutions of this application will be further described in detail below with reference to the accompanying drawings of the embodiments of this application. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application.

[0022] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this application is for the purpose of describing embodiments of this application only and is not intended to be limiting of this application.

[0023] In the following description, references to "some embodiments," "this embodiment," "this application embodiment," and examples, etc., describe a subset of all possible embodiments. However, it is understood that "some embodiments" may be the same subset or different subset of all possible embodiments and may be combined with each other without conflict.

[0024] If the application documents contain similar descriptions such as "first / second", the following explanation shall be added: In the following description, the terms "first / second / third" are used only to distinguish similar objects and do not represent a specific order of objects. It is understood that "first / second / third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0025] This application provides a method for batch intelligent detection of pre-embedded anchor bolt groups in steel structures. Figure 1 This is an optional flowchart illustrating a batch intelligent detection method for pre-embedded anchor bolt groups in steel structures, provided in an embodiment of this application. Figure 1 The steps shown are explained.

[0026] S101. Determine the construction area to be inspected; and set up targets in the construction area to be inspected, and determine the absolute spatial coordinates of each of the multiple peripheral targets and the group target codes of each of the multiple anchor bolt groups.

[0027] In this embodiment, two levels of spatial reference targets are set up in the construction area to be inspected, and the design parameters of the anchor bolt group are recorded to provide an absolute spatial reference, anchor bolt group identification and theoretical comparison basis for subsequent image processing.

[0028] In some embodiments of this application, peripheral targets are deployed in the construction area to be inspected, and a global absolute reference frame for the peripheral target reference frame is determined; the spatial distance between multiple peripheral targets in the peripheral target reference frame is measured; the absolute spatial coordinates of each of the multiple peripheral targets are determined based on the spatial distance between the multiple peripheral targets and the global absolute reference frame; group targets are deployed in the construction area to be inspected to obtain multiple group targets; the multiple group targets are encoded to obtain the group target codes for each of the multiple anchor bolt groups; wherein, there is a correspondence between the multiple group targets and the multiple anchor bolt groups.

[0029] For example, the deployment of peripheral targets is illustrated using N_tar peripheral targets. N_tar peripheral targets (N_tar≥4) are deployed on the outermost perimeter of the construction area to be inspected, forming a peripheral target reference frame and establishing a global absolute reference system. The spatial distance between each peripheral target is determined using measuring equipment, and the absolute spatial coordinates of the j-th peripheral target are calibrated as P_tar_abs[j]=(X_tar_abs[j],Y_tar_abs[j]), where j=1,2,...,N_tar. The peripheral targets can be either coded targets or checkerboard targets. The spatial coordinates of the peripheral targets are calibrated using a total station or RTK measuring equipment, with a calibration accuracy better than 1cm. P_tar_abs[j] is stored as a two-dimensional coordinate vector in cm units.

[0030] The deployment of group targets is illustrated using N_groups of anchor bolts. One group target is deployed at the geometric center or within the coverage area of ​​each anchor bolt group. Assume there are N_groups of anchor bolt groups in the construction area to be inspected. The group target code ID of the k-th anchor bolt group is denoted as ID_group[k] (k=1,2,...,N_group). The group target codes ID of different anchor bolt groups are unique and used to distinguish the anchor bolt groups in subsequent operations. The group targets are AprilTag or ArUco coded targets, and their codes ID_group[k] are pre-associated with and entered into the system with the corresponding anchor bolt group code k.

[0031] It should be noted that the outer target coding and the group target coding adopt a unified coding system. Let the maximum value of all group target codes be N_group, that is, the code ID_tar[j] of the j-th outer target = N_group + j (j=1,2,...,N_tar), to ensure that the numerical ranges of the two types of target codes do not overlap.

[0032] Understandably, coded targets or checkerboard targets are deployed on the outermost perimeter of the construction area to be inspected to establish a global absolute spatial reference frame, and its absolute spatial coordinates are calibrated using a total station or RTK. Simultaneously, AprilTag or ArUco coded target groups are deployed at the geometric center of each anchor bolt group as unique identifiers for the anchor bolt group. Both types of targets use a unified coding system to ensure that the coded value ranges do not overlap, providing an absolute spatial reference and a basis for anchor bolt group identification in subsequent image processing.

[0033] S102. Based on the absolute spatial coordinates of multiple peripheral targets, images are acquired from the construction area to be inspected to obtain a four-scale image sequence; wherein, the four-scale image sequence contains images of four different scale types.

[0034] In some embodiments of this application, flight path planning is performed based on the absolute spatial coordinates of multiple peripheral targets to determine the cruise route of the UAV; through the cruise route, images of the construction area to be inspected are acquired to obtain a four-scale image sequence.

[0035] In some embodiments of this application, S102 can be implemented by S201-S202, as follows: S201. Based on the absolute spatial coordinates of multiple peripheral targets, flight path planning is performed to determine the UAV's cruise route.

[0036] S202. By following the cruise route, images are collected from the construction area to be inspected, and a four-scale image sequence is obtained.

[0037] In some embodiments of this application, a first-scale image acquisition is performed on the construction area to be inspected via a cruise route to obtain a global target image; wherein the global target image represents an image covering all peripheral targets; a second-scale image acquisition is performed on the construction area to be inspected via the cruise route to obtain a target-group association image; wherein the target-group association image represents an image covering at least one peripheral target and at least one group target; a third-scale image acquisition is performed on the construction area to be inspected via the cruise route to obtain a multi-group region image; wherein the multi-group region image represents an image covering at least two group targets; a fourth-scale image acquisition is performed on the construction area to be inspected via the cruise route to obtain a single-group detailed investigation image; wherein the single-group detailed investigation image represents an image covering only one group target; a four-scale image sequence is determined based on the global target image, the target-group association image, the multi-group region image, and the single-group detailed investigation image.

[0038] For example, based on the absolute spatial coordinates P_tar_abs[j] of the j-th outer target mentioned above, an S-shaped reciprocating cruise route for the UAV is planned. The cruise route satisfies the following constraints: The cruise route starts at the location of P_tar_abs[1] and ends at the location of P_tar_abs[N_tar]. The cruise route passes sequentially over or near all N_tar peripheral targets. The cruise route covers the entire area of ​​all N_group of anchor bolt groups to be inspected.

[0039] During flight, the camera gimbal maintains a fixed pitch angle α (45° ≤ α ≤ 90°) and shoots downwards.

[0040] The drone flies along the planned cruise route and collects image sequences at four scales during the flight according to a preset triggering strategy.

[0041] Suppose that a total of M frames of images are acquired, forming a four-scale image sequence I[m] from the M frames; where the scale type of the m-th frame is denoted as Scale[m]∈{1,2,3,4} (m=1,2,...,M), and each scale is defined as follows: First scale (Scale[m]=1): Global target image, covering all N_tar peripheral targets, with M1≥1 frames acquired; Second scale (Scale[m]=2): Target-group association image, covering at least one peripheral target and at least one anchor bolt group of target groups, with M2 frames acquired; Third scale (Scale[m]=3): Multi-group region image, covering at least 2 anchor bolt groups of target groups, with M3 frames acquired; Fourth scale (Scale[m]=4): Single-group detailed image, covering only the target of one anchor bolt group, with M4 frames acquired.

[0042] The above M1, M2, M3, and M4 satisfy M = M1 + M2 + M3 + M4. Each shot must ensure that the target is clearly and identifiable in the image. Images at four scales are achieved through variable altitude flight; or through zoom, adjusting the focal length at the same flight altitude to obtain different fields of view.

[0043] Understandably, instead of using a single-scale acquisition method with a fixed flight altitude or fixed focal length, an S-shaped reciprocating inspection route is planned. In a single flight, image sequences of four scales are acquired sequentially: global target image, target-group correlation image, multi-group regional image, and single-group detailed investigation image. This forms a complete multi-scale information coverage from global overview to local detailed investigation, taking into account both the requirements of large-scale regional coverage and local accuracy.

[0044] S103. Based on the group target encoding of each of the multiple anchor bolt groups, perform image preprocessing and anchor bolt group identification marking on the four-scale image sequence to obtain the mask image sequence and the scale type corresponding to the mask image sequence.

[0045] In some embodiments of this application, distortion correction and perspective correction are performed on each frame of the four-scale image sequence using pre-acquired camera distortion coefficients and camera attitude angles to obtain a corrected four-scale image sequence; target encoding recognition is performed on the corrected four-scale image sequence to determine the group target encoding and peripheral targets of each of the multiple anchor bolt groups in the corrected four-scale image sequence; anchor bolt detection is performed on the corrected four-scale image sequence to determine the pixel coordinates of each of the multiple anchor bolts; based on the group target encoding, peripheral targets, and pixel coordinates of each of the multiple anchor bolts, values ​​are assigned and a mask is output to obtain a mask image sequence and the scale type corresponding to the mask image sequence.

[0046] For example, obtain the camera distortion coefficients and camera pose angles, and use the camera distortion coefficients Kdist = (k1, k2, p1, p2, k3) and camera pose angles θ[m] = (θpitch[m], θroll[m]) to sequentially process each frame of the four-scale image sequence I[m]: Lens distortion correction eliminates radial and tangential distortion. Perspective transformation correction is performed by constructing a perspective transformation matrix based on the gimbal pitch angle θpitch[m] and roll angle θroll[m] to correct the tilted image into a top-down plan view.

[0047] Output the corrected four-scale image sequence I_rect[m].

[0048] In the corrected four-scale image sequence I_rect[m], identify and decode all visible targets: Group targets (AprilTag or ArUco): Decoding the group targets yields the encoded set ID_group_detected ={idg[1],idg[2],...idg[Km]}, where Km is the number of group targets detected in the m-th frame image; Peripheral targets (coded targets or checkerboard): Decode the peripheral targets to obtain the coded set ID_tar_detected={idt[1],idt[2],...idt[Tm]}, where Tm is the number of peripheral targets detected in the m-th frame image.

[0049] In the corrected planar image I_rect[m], the pixel positions of all anchor bolts are detected by the image processing algorithm, and the pixel coordinates pbolt[i]=(ui,vi) of the i-th anchor bolt are recorded, where i=1,2,...,Nbolt[m], and Nbolt[m] is the total number of anchor bolts detected in the m-th frame image.

[0050] Based on the identified target group code ID_group_detected and its position in the image, and combined with the geometric prior relationship between each target group and the corresponding anchor bolt group (such as the Euclidean distance principle from the anchor bolt to the center of the target group being the shortest, or the classification based on the preset anchor bolt group coverage area), the pixel coordinates pbolt[i] of all detected anchor bolts are assigned to the corresponding anchor bolt groups.

[0051] For the k-th detected target group, it is encoded as idg[k], and all anchor bolt pixels within its corresponding anchor bolt group area are assigned the value idg[k]. If an anchor bolt cannot be attributed to any identified target group (e.g., it is located at the edge of the image or the target is occluded), its pixel value is assigned to 0, to be filled in later frames or subsequent processing steps; and for non-anchor bolt positions, its pixel value is also assigned to 0.

[0052] Similarly, for the t-th detected peripheral target, it is encoded as idt[t], and the pixel of its corresponding pixel is assigned the value idt[t].

[0053] For each anchor bolt, only one pixel at its geometric center or centroid is retained as a representative point, and the remaining anchor bolt pixels are not marked. The output mask image sequence Mask[m] is generated, with the same size as I_rect[m] (i.e., W×H pixels). Finally, the mask image sequence Mask[m] and the scale type Scale[m] are associated and stored.

[0054] Understandably, by using perspective correction and target coding recognition technology, the pixel positions of the group targets, peripheral targets, and anchor bolts are detected simultaneously in each frame of the corrected planar image. Based on the group target coding and its position in the image, combined with the prior relationship that the group targets are placed at the geometric center of the anchor bolt group, each detected anchor bolt pixel is assigned a unique coding ID of its group, generating a mask image that retains only the representative points of the anchor bolts, thus realizing the integrated marking of anchor bolt position and group identity.

[0055] S104. Based on the mask image sequence, the scale type corresponding to the mask image sequence, and the absolute spatial coordinates of multiple peripheral targets, anchor bolt detection is performed through a lightweight spatial location mapping network to obtain a global top view; wherein, the global top view is a pixel-level distribution map of the actual spatial positions of all anchor bolts in the construction area to be detected.

[0056] This embodiment first describes the lightweight spatial location mapping network. The lightweight spatial location mapping network is designed to accurately learn the mapping relationship between the mask image sequence and the global top view. The lightweight spatial location mapping network adopts a lightweight encoder-decoder structure. The lightweight spatial location mapping network includes an input layer, an encoding layer, a bottleneck layer, a spatial prior injection layer, a decoding layer, and a global image reconstruction layer. The encoding layer includes encoder block 1, encoder block 2, and encoder block 3; the decoding layer includes decoder block 1 and decoder block 2.

[0057] The training method includes three parts: training dataset construction, network architecture and training objective design, and model training and optimization strategies.

[0058] The construction of the training dataset includes data sources, sample construction, and dataset partitioning. Data sources include: collecting multiple sets of mask image data and corresponding peripheral target coordinates within the construction area to be inspected, following the method described above. Each valid sample contains: Input: a set of M mask images, Mask[m] (m=1,2,...,M), and the corresponding scale type, Scale[m]; Label: a global top-view I_global_gt, with dimensions H_global×W_global, where each pixel value is the group target code ID_group[k] of the anchor bolt at that location (0 for non-anchor bolt locations). The true label I_global_gt is generated by manually calibrating or measuring the actual spatial position of the anchor bolts point-by-point with a high-precision total station, projecting it according to the preset physical resolution R_physical, and simultaneously generating corresponding category labels based on the target code of the anchor bolt's group.

[0059] Sample construction involves taking one pixel at the geometric center or centroid of each frame of the mask image as the anchor bolt representative point, and setting the remaining pixels to 0. All frame mask images and their scale types collected within the same construction area to be inspected are grouped into a single sample. Each sample contains: Input: a multi-channel input concatenated from the mask image tensor X∈R(M×H×W×1) and the scale encoding tensor S∈R(M×4); Label: Global top-down view .

[0060] The dataset is divided according to the construction areas to be inspected, ensuring that the same area does not appear in both the training and validation sets simultaneously. For example, the dataset partition ratio could be: training set: 70%, validation set: 15%, test set: 15%. The training set is used for model parameter learning, the validation set is used for model tuning and early stopping detection, and the test set is used for final model performance evaluation.

[0061] Second, the training objective is to minimize the pixel-wise cross-entropy loss, ensuring that the category (anchor group ID or background) of each pixel in the network's output global top-down view I_global is as consistent as possible with the true distribution. The loss function is pixel-wise cross-entropy loss (since each pixel is a category label, the number of categories is N_group + 1, where N_group is the total number of anchor groups, and 1 represents the background class):

[0062] Here, y_gt(u,v,c) is an indicator function, which takes the value 1 when the true class of pixel (u,v) is c, and 0 otherwise; y_pred(u,v,c) is the class probability predicted by the network.

[0063] Third, model training and optimization strategies. The optimizer used was AdamW; the initial learning rate was 0.001; learning rate scheduling: the ReduceLROnPlateau scheduler was used, monitoring the validation set loss, with patience=10, factor=0.5, and min_lr=1×10. -6 The weight decay is 0.0001; the batch size is 4; and the maximum number of training epochs is 200. A batch is randomly sampled from the training set (input tensor shape (B, M, H, W, 5), label shape (B, H_global, W_global)); forward propagation yields the predicted global top view I_global_pred; pixel-wise cross-entropy loss is calculated; backpropagation is performed, gradient clipping (norm 1.0) is applied, and parameters are updated; the training loss is output every 50 mini-batches; after each epoch, the loss and pixel-wise classification accuracy are evaluated on the validation set; if the learning rate decreases or early stopping conditions are met, the corresponding operations are performed. Early stopping occurs if the validation loss does not decrease for 20 consecutive epochs; otherwise, training continues to 200 epochs, and the optimal weights are restored.

[0064] In some embodiments of this application, a preset physical resolution is obtained; the mask image sequence, the scale type corresponding to the mask image sequence, and the absolute spatial coordinates of each of the multiple peripheral targets are preprocessed to obtain a mask stacking vector, a scale encoding vector, and a normalized coordinate vector; based on the mask stacking vector, the scale encoding vector, the normalized coordinate vector, and the preset physical resolution, anchor bolt detection is performed through a magnitude spatial location mapping network to obtain a global top view.

[0065] For example, a lightweight spatial location mapping network is constructed based on the output mask image sequence Mask[m] and its scale type Scale[m], combined with the absolute spatial coordinates P_tar_abs[j]=(X_tar_abs[j],Y_tar_abs[j]) of the multiple peripheral targets calibrated in S101. Since there are differences in viewpoint, occlusion and coverage between images of different scales, it is difficult to recover the complete global anchor distribution by relying solely on traditional geometric stitching. Therefore, a lightweight spatial location mapping network is used to learn the spatial correspondence between multi-scale images and output a unified global top view.

[0066] Understandably, the method of using the absolute spatial coordinates of the outer targets as a reference and directly regressing the global top view using a lightweight neural network avoids the cumulative errors of traditional image stitching and achieves a complete single-frame representation of the anchor bolt positions across the entire region. By deploying coded group targets at the geometric center of each anchor bolt group, each anchor bolt is assigned a unique code to its group during the detection phase, ensuring that each anchor bolt pixel in the output global top view carries clear group identity information. By encoding the absolute spatial coordinates of the outer targets, the preset physical resolution, the mask image sequence, and the scale type of each frame as auxiliary information and inputting it into the lightweight neural network, the model can adaptively adjust the contribution weight of each frame feature in the global reconstruction according to the image scale. When restoring the global top view, it is directly constrained by the physical reference, ensuring that the output pixel positions have a clear physical spatial correspondence, avoiding the cumulative errors of traditional stitching methods. At the same time, it achieves unified batch output of any number of anchor bolt groups and any number of anchor bolts, thereby improving the efficiency, accuracy, and traceability of large-scale steel structure pre-embedded anchor bolt group detection.

[0067] In some embodiments of this application, S104 can be implemented using S401-S403, as follows: S401, Obtain the preset physical resolution.

[0068] S402. Preprocess the mask image sequence, the scale type corresponding to the mask image sequence, and the absolute spatial coordinates of each of the multiple peripheral targets to obtain the mask stacking vector, the scale encoding vector, and the normalized coordinate vector.

[0069] For example, the input data for a lightweight spatial location mapping network includes the following types: Mask[m] is a sequence of mask images, m=1,2,...,M, and is a single-channel label image of shape (H,W). All mask images have the same resolution. The scale type Scale[m] ∈ {1, 2, 3, 4} corresponds to the mask image sequence; The absolute spatial coordinates of multiple peripheral targets are P_tar_abs[j]=(X_tar_abs[j],Y_tar_abs[j]), j=1,2,...,N_tar; The preset physical resolution R_physical (unit: cm / pixel) is the actual physical size represented by each pixel in the global top view. The preset physical resolution value can be calculated by the ratio of the pixel distance of the outer target in the image to its known actual physical distance, or by setting an empirical value in advance based on the drone's aerial photography altitude and camera parameters, for example, R_physical = 0.5 cm / pixel.

[0070] The preprocessing method is as follows: 1. Stack each frame of the mask image sequence Mask[m] directly into a multi-channel mask stack vector X∈R(M×H×W×1).

[0071] 2. The scale type Scale[m] encodes a one-hot vector S_vec[m]∈R4, which is stacked into a scale-encoded vector S∈R(M×4), and then concatenated with the mask image features before being input into the network. The values ​​of scale type Scale[m] 1, 2, 3, and 4 correspond to "global target image", "target and group association image", "multi-group region image", and "single-group detailed investigation image", respectively.

[0072] 3. The absolute spatial coordinates P_tar_abs[j] of multiple peripheral targets are processed by min-max normalization: X_tar_norm[j]= (X_tar_abs[j]-X_tar_min) / (X_tar_max-X_tar_min); Y_tar_norm[j]= (Y_tar_abs[j]-Y_tar_min) / (Y_tar_max-Y_tar_min); Where X_tar_min, X_tar_max, Y_tar_min, and Y_tar_max are the minimum and maximum values ​​of the X and Y coordinates of all peripheral targets, respectively. The normalized coordinate tensor P_tar_norm[j]∈R(N_tar×2) is used as the spatial prior input.

[0073] 4. The preset physical resolution R_physical is used as a scalar parameter R input to a lightweight spatial location mapping network to help determine the global size of the output image.

[0074] S403. Based on mask stacking vectors, scale encoding vectors, normalized coordinate vectors, and preset physical resolution, anchor bolt detection is performed through a lightweight spatial location mapping network to obtain a global top view.

[0075] In some embodiments of this application, based on the mask stacking vector and the scale encoding vector, convolutional pooling is performed through the encoding layer of a lightweight spatial location mapping network to obtain a high-level semantic feature map; the bottleneck layer of the lightweight spatial location mapping network is used to convolve the high-level semantic feature map to obtain a bottleneck feature tensor; the spatial prior injection layer of the lightweight spatial location mapping network encodes the normalized coordinate vector and the preset physical resolution respectively to obtain a coordinate embedding feature map and a resolution embedding feature map; based on the bottleneck feature tensor, the coordinate embedding feature map, and the resolution embedding feature map, the spatial prior injection layer of the lightweight spatial location mapping network performs concatenation and convolution processing to obtain a priori enhancement features; based on the prior enhancement features, the decoding layer of the lightweight spatial location mapping network performs convolution and concatenation processing to obtain a final feature map; based on the final feature map, the global image reconstruction layer of the lightweight spatial location mapping network performs anchor bolt detection on the region to be detected to obtain a global top view.

[0076] For example, this network employs a lightweight encoder and decoder architecture, outputting a global top-down view I_global with a resolution of [resolution missing]. H_global×W_global (determined by the actual range of the construction area to be inspected [X_min,X_max]×[Y_min,Y_max] and the preset physical resolution R_physical), each pixel value is the target code of the anchor bolt group or 0; the network structure configuration of the lightweight spatial location mapping network is shown in Table 1.

[0077] Table 1

[0078] Encoder block 1 consists of two convolutional layers (3×3 kernels, 32 filters), batch normalization, ReLU activation, and max pooling (2×2). The input is a mask stacking vector and a scale encoding vector. These are concatenated to obtain a concatenated tensor (shape (B, M, 512, 512, 5)). The first convolutional layer compresses and fuses the mask features of M frames along the channel dimension into a 32-channel feature map, while simultaneously incorporating scale type information into the mask features of each frame, outputting a first-channel feature map with shape (B, 256, 256, 32). Encoder block 1 utilizes the scale type Scale[m] as auxiliary information, enabling the network to distinguish the weights and spatial coverage of mask images of different scales in global stitching. Max pooling compresses spatial resolution to expand the receptive field.

[0079] Encoder block 2 contains two convolutional layers (3×3 kernels, 64 filters), batch normalization, ReLU activation, and max pooling (2×2). The input is the first channel feature map output from encoder block 1, and the output is a mid-level spatial layout feature map with shape (B, 128, 128, 64). Encoder block 2 further extracts the distribution pattern of anchor bolt groups and the relative positional relationships between groups, providing spatial context features for global reconstruction.

[0080] Encoder block 3 contains two convolutional layers (3×3 kernels, 128 filters), batch normalization, ReLU activation, and max pooling (2×2). Its input is the mid-level spatial layout feature map output from encoder block 2, and its output is a high-level semantic feature map with shape (B, 64, 64, 128). Encoder block 3 captures the overall spatial arrangement of the anchor bolt group, the topological relationships between groups, and their relative orientation to the surrounding targets.

[0081] The bottleneck layer comprises two convolutional layers (3×3 kernels, 128 filters), batch normalization, and ReLU activation. The input is the high-level semantic feature map output from encoder block 3, and the output is a bottleneck feature tensor F_bottleneck with shape (B, 64, 64, 128). At the lowest spatial resolution, this layer adaptively recalibrates and aggregates the multi-scale features extracted by the encoder from each frame's mask image using channel-dimension optimization. Through channel-wise response adjustment of the two convolutional layers, it strengthens the feature channels that play a crucial role in discriminating the spatial layout of the anchor bolt group, while suppressing low-confidence responses caused by changes in shooting angle, lighting, or local occlusion. This results in an output with a more powerful abstract feature of the global spatial layout. This bottleneck feature tensor serves as the core feature source for the decoder to recover the global top-down view. During subsequent upsampling, it is gradually fused with the spatial detail features retained by each layer of the encoder through skip connections, guiding the decoder to recover an anchor bolt distribution map with precise location and clear group boundaries.

[0082] The spatial prior injection layer is responsible for injecting the normalized coordinate vector and the user-defined physical resolution as prior knowledge into the network, providing physical spatial reference constraints for global top-view reconstruction.

[0083] The specific steps are as follows: The normalized coordinate tensor P_tar_norm ∈ R(N_tar×2) is encoded into a coordinate embedding feature map with dimensions (B,64,64,C) that matches the spatial dimension of the bottleneck feature tensor, using a multilayer perceptron (MLP). This allows the network to perceive the global spatial position of each anchor bolt group relative to the outer target during the reconstruction process.

[0084] The preset physical resolution R is sinusoidally position-encoded and then mapped to a resolution embedding vector through a fully connected layer. This vector is then expanded into a resolution embedding feature map with the same spatial size as the bottleneck feature, enabling the network to clearly output the actual physical scale corresponding to each pixel in the image.

[0085] The coordinate embedding feature map, resolution embedding feature map, and bottleneck feature tensor F_bottleneck are concatenated along the channel dimension, and then subjected to a 1×1 convolution for feature fusion and channel compression. The output is a prior enhancement feature with shape (B,64,64,128) for further processing by decoder block 1.

[0086] This injection mechanism introduces absolute spatial coordinates and physical scale information into the decoding path early on, enabling the subsequent upsampling process to directly recover the anchor bolt position under a unified physical spatial reference. This avoids the cumulative errors caused by unclear scale or missing coordinate references during the conversion from image pixel space to physical space, ensuring that the final output global top view I_global has an accurate physical spatial correspondence.

[0087] Decoder block 1 contains upsampling convolutions (UpConv), skip connections, two convolutional layers (3×3 kernels, 64 filters), batch normalization, and ReLU activation. First, the prior enhancement features output from the spatial prior injection layer are restored to (B, 128, 128, 64) through upsampling convolutions. Then, the mid-layer spatial layout feature map (shape (B, 256, 256, 64)) saved after the first convolutional layer and before the max pooling operation is extracted from encoder block 2. This feature needs to be downsampled (e.g., by average pooling or convolution with a stride of 2) to (B, 128, 128, 64) before being concatenated and fused with the upsampling feature along the channel dimension. Finally, the features are further fused through two 3×3 convolutional layers (64 filters), outputting a restored feature map with shape (B, 128, 128, 64). This skip connection mechanism introduces the high-resolution spatial detail information retained during the encoding stage into the decoding path, effectively compensating for the positional accuracy lost during upsampling, and making the final reconstructed anchor group boundary clearer and more accurate.

[0088] Decoder block 2 includes upsampling convolution (UpConv), skip connections, two convolutional layers (3×3 kernels, 32 filters), batch normalization, and ReLU activation. The recovered feature map output from decoder block 1 is upsampled to (B, 256, 256, 32) and concatenated with the first channel feature map output from encoder block 1 to produce a final feature map with shape (B, 256, 256, 32).

[0089] The global image reconstruction layer upsamples the final feature map (B, 256, 256, 32) output from decoder block 2 to a global size H_global × W_global using bilinear interpolation, where: H_global =(Y_max -Y_min) / R_physical; W_global =(X_max -X_min) / R_physical; Among them, X_min, X_max, Y_min, and Y_max are determined by the absolute coordinates P_tar_abs of the outer target, i.e., X_min = min (X_tar_abs[j]), X_max = max (X_tar_abs[j]), and Y_min and Y_max are determined in the same way.

[0090] The upsampled feature map is compressed to 1 channel by a 1×1 convolution layer, and the output is a global top view I_global∈R(H_global ×W_global ×1). The value of each pixel in the image is the target encoding ID_group[k] of the corresponding anchor bolt (a non-zero value in the mask) or 0 (background).

[0091] The final output is the global top view I_global, which has the following properties: The image size is H_global × W_global (covering the entire construction area to be inspected); The physical resolution is R_physical (cm / pixel), which is consistent with the pre-calibrated accuracy (better than 1 cm); The pixel value represents the group target code ID_group[k] (≥1) of the anchor bolt's group; background position: 0; The coverage includes all anchors in all anchor groups across the entire area, with locations marked with pixel-level precision. I_global, which is a pixel-level distribution map of the actual spatial location of all anchors in the entire construction area to be inspected, can be used as the basic data input for subsequent project acceptance, deviation analysis, quality assessment and other stages.

[0092] This application has the following beneficial effects: Firstly, a two-level spatial reference target deployment strategy is adopted. Encoded targets are deployed around the construction area to be inspected and their absolute spatial coordinates are marked. At the same time, a unique coded group target is deployed at the geometric center of each anchor bolt group. This provides a global physical benchmark and anchor bolt group identification for subsequent image processing, fundamentally solving the problem of lack of unified spatial reference between multiple frames of images and the inability to assign anchor bolt detection results to specific groups.

[0093] Secondly, a four-scale image acquisition strategy is adopted, allowing the UAV to acquire multi-scale image sequences, from global target images to single-group detailed investigation images, in a single flight along an S-shaped route, balancing the dual requirements of large-area coverage and local positioning accuracy. Combining perspective correction and target coding recognition technology, target decoding and anchor bolt detection are completed simultaneously in each frame of the corrected image, and the detected anchor bolt pixels are assigned real-time ID values ​​according to the group target coding. The local detection results scattered in multiple frames are marked as groups, providing structured input data for subsequent global reconstruction.

[0094] Furthermore, the constructed lightweight spatial location mapping network takes the mask image and its scale encoding as input, and the absolute spatial coordinates and physical resolution of the outer target as prior constraints. It directly regresses to generate a global top-down view covering the entire region in an end-to-end manner. The scale encoding mechanism enables the network to adaptively distinguish the spatial coverage and information weights of images at different scales. The spatial prior injection layer introduces the physical reference into the decoding path early, ensuring that each pixel in the output image has a clear physical coordinate correspondence. This avoids the problems of accumulated errors or mapping failures caused by large scale differences and inconsistent overlapping areas in traditional image stitching methods.

[0095] Finally, the entire process from drone inspection to global top-down view output was automated, upgrading the traditional inefficient inspection mode that relied on total station point-by-point measurement and manual data aggregation to a one-time panoramic inspection mode based on visual perception and neural network intelligent mapping. A single flight can output a complete distribution map containing all anchor bolt groups in the entire area, with each anchor bolt having its group code. This data can be directly used as the basis for subsequent anchor bolt position deviation analysis, group installation quality assessment, and project acceptance, significantly improving the efficiency, accuracy, and traceability of large-scale steel structure pre-embedded anchor bolt group inspection.

[0096] Based on the above embodiments of the intelligent batch detection method for pre-embedded anchor bolt groups in steel structures, this application also provides an intelligent batch detection system for pre-embedded anchor bolt groups in steel structures, such as... Figure 2 As shown, Figure 2 This is a schematic diagram of a batch intelligent detection system for pre-embedded anchor bolt groups in steel structures, provided in an embodiment of this application. The system includes: a determination module 201, a data acquisition module 202, a preprocessing and marking module 203, and a detection module 204. The determining module 201 is used to determine the construction area to be inspected; and to set up targets in the construction area to be inspected, and to determine the absolute spatial coordinates of each of the multiple peripheral targets and the group target codes of each of the multiple anchor bolt groups. The acquisition module 202 is used to acquire images of the construction area to be detected based on the absolute spatial coordinates of the multiple peripheral targets, and obtain a four-scale image sequence; wherein, the four-scale image sequence contains images of four different scale types; The preprocessing and labeling module 203 is used to perform image preprocessing and anchor group identification labeling on the four-scale image sequence based on the group target encoding of each of the multiple anchor groups, to obtain a mask image sequence and the scale type corresponding to the mask image sequence; The detection module 204 is used to perform anchor bolt detection through a lightweight spatial location mapping network based on the mask image sequence, the scale type corresponding to the mask image sequence, and the absolute spatial coordinates of the multiple peripheral targets, to obtain a global top view; wherein, the global top view is a pixel-level distribution map of the actual spatial positions of all anchor bolts in the construction area to be detected.

[0097] Based on the above embodiments of the method for batch intelligent detection of pre-embedded anchor bolt groups in steel structures, this application also provides a batch intelligent detection device for pre-embedded anchor bolt groups in steel structures, such as... Figure 3 As shown, Figure 3 This is a schematic diagram of a batch intelligent inspection device for pre-embedded anchor bolts in steel structures, provided in an embodiment of this application. The device includes a processor 301 and a memory 302. The memory 302 stores a computer program; the processor 301 retrieves and runs the computer program from the memory to execute a batch intelligent inspection method for pre-embedded anchor bolts in steel structures as described in the above embodiment.

[0098] In the embodiments of this application, the processor 301 described above can be at least one of the following: Application-Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), Central Processing Unit (CPU), Controller, Microcontroller, and Microprocessor. It is understood that for different devices, the electronic device used to implement the above processor function can also be other types, and the embodiments of this application do not specifically limit it.

[0099] This application provides a computer-readable storage medium storing a computer program for implementing, when executed by a processor, a method for batch intelligent detection of pre-embedded anchor bolt groups in steel structures as described in any of the above embodiments.

[0100] For example, the program instructions corresponding to the batch intelligent detection method for pre-embedded anchor bolts in steel structures in this embodiment can be stored on storage media such as optical discs, hard disks, and USB flash drives. When the program instructions corresponding to the batch intelligent detection method for pre-embedded anchor bolts in steel structures in the storage media are read or executed by an electronic device, the batch intelligent detection method for pre-embedded anchor bolts in steel structures as described in any of the above embodiments can be realized.

[0101] Furthermore, in the embodiments of this application, the functional modules can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional module.

[0102] If the integrated unit is implemented as a software functional module and is not sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this embodiment, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the method of this embodiment. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0103] It should be understood that the phrases "one embodiment," "an embodiment," or "some embodiments" mentioned throughout the specification mean that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment," "in one embodiment," or "in some embodiments" appearing throughout the specification do not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the embodiments in this application are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. The descriptions of the various embodiments above tend to emphasize the differences between the various embodiments; their similarities or commonalities can be referred to mutually, and for the sake of brevity, these will not be repeated here.

[0104] The modules described above as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules. They may be located in one place or distributed across multiple network units. Some or all of the modules may be selected to achieve the purpose of this embodiment according to actual needs.

[0105] In addition, each functional module in the various embodiments of this application can be integrated into one processing unit, or each module can be a separate unit, or two or more modules can be integrated into one unit; the integrated modules can be implemented in hardware or in the form of hardware plus software functional units.

[0106] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as mobile storage devices, read-only memory (ROM), magnetic disks, or optical disks.

[0107] The methods disclosed in the several method embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments.

[0108] The features disclosed in the several product embodiments provided in this application can be arbitrarily combined without conflict to obtain new product embodiments.

[0109] The features disclosed in the several method or device embodiments provided in this application can be arbitrarily combined without conflict to obtain new method or device embodiments.

[0110] The above description is merely an embodiment of this application, but the protection scope of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for batch intelligent inspection of pre-embedded anchor bolt groups in steel structures, characterized in that, The method includes: The construction area to be inspected is determined; and targets are set up in the construction area to be inspected, and the absolute spatial coordinates of each of the multiple peripheral targets and the group target codes of each of the multiple anchor bolt groups are determined. Based on the absolute spatial coordinates of the multiple peripheral targets, images are acquired from the construction area to be detected to obtain a four-scale image sequence; wherein, the four-scale image sequence contains images of four different scale types; Based on the group target encoding of each of the multiple anchor bolt groups, the four-scale image sequence is preprocessed and the anchor bolt group is identified and marked to obtain the mask image sequence and the scale type corresponding to the mask image sequence; Based on the mask image sequence, the scale type corresponding to the mask image sequence, and the absolute spatial coordinates of the multiple peripheral targets, anchor bolt detection is performed through a lightweight spatial location mapping network to obtain a global top view; wherein, the global top view is a pixel-level distribution map of the actual spatial positions of all anchor bolts in the construction area to be detected.

2. The method according to claim 1, characterized in that, The step of setting up targets in the construction area to be inspected, and determining the absolute spatial coordinates of multiple peripheral targets and the group target codes of multiple anchor bolt groups, includes: Targets are deployed around the construction area to be inspected, and the global absolute reference frame of the target reference frame is determined. Measure the spatial distance between multiple peripheral targets in the peripheral target reference frame; The absolute spatial coordinates of each of the multiple peripheral targets are determined based on the spatial distance between them and the global absolute reference system. Multiple target groups are deployed in the construction area to be inspected. The multiple target groups are encoded to obtain the target group codes for each of the multiple anchor bolt groups; wherein, there is a correspondence between the multiple target groups and the multiple anchor bolt groups.

3. The method according to claim 1, characterized in that, The process involves acquiring images of the construction area to be detected based on the absolute spatial coordinates of the multiple peripheral targets, resulting in a four-scale image sequence, including: Based on the absolute spatial coordinates of the multiple peripheral targets, flight path planning is performed to determine the UAV's cruising route. The patrol route is used to collect images of the construction area to be inspected, and a four-scale image sequence is obtained.

4. The method according to claim 3, characterized in that, The process of acquiring images of the construction area to be inspected along the cruise route to obtain a four-scale image sequence includes: The first-scale image is acquired by using the cruise route to obtain a global target image; wherein, the global target image represents an image covering all peripheral targets; The second-scale image acquisition is performed on the construction area to be detected via the cruise route to obtain target and group association images; wherein, the target and group association images represent images covering at least one peripheral target and at least one group target; The patrol route is used to acquire images of the construction area to be inspected at a third scale, resulting in multi-group area images; wherein, the multi-group area images represent images covering at least two groups of targets; The patrol route is used to acquire images of the construction area to be inspected at a fourth scale, resulting in a single-group detailed survey image; wherein, the single-group detailed survey image represents an image covering only one group of targets; Based on the global target image, the target and swarm association image, the multi-swarm region image, and the single-swarm detailed investigation image, the four-scale image sequence is determined.

5. The method according to claim 1, characterized in that, The step of preprocessing and marking the four-scale image sequence based on the group target encoding of each of the multiple anchor bolt groups to obtain a mask image sequence and the scale type corresponding to the mask image sequence includes: By using the pre-acquired camera distortion coefficients and camera pose angles, distortion correction and perspective correction are performed on each frame of the four-scale image sequence to obtain the corrected four-scale image sequence. Target coding recognition is performed on the corrected four-scale image sequence to determine the group target code and peripheral target of each of the multiple anchor bolt groups in the corrected four-scale image sequence; Anchor bolt detection is performed on the corrected four-scale image sequence to determine the pixel coordinates of each anchor bolt. Based on the group target encoding, the pixel coordinates of the outer target and the multiple anchor bolts, values ​​are assigned and a mask is output to obtain the mask image sequence and the scale type corresponding to the mask image sequence.

6. The method according to claim 1, characterized in that, The step of detecting anchor bolts using a lightweight spatial location mapping network based on the mask image sequence, the scale type corresponding to the mask image sequence, and the absolute spatial coordinates of the multiple peripheral targets, to obtain a global top view, includes: Get the preset physical resolution; The mask image sequence, the scale type corresponding to the mask image sequence, and the absolute spatial coordinates of each of the multiple peripheral targets are preprocessed to obtain the mask stacking vector, the scale encoding vector, and the normalized coordinate vector. Based on the mask stacking vector, the scale encoding vector, the normalized coordinate vector, and the preset physical resolution, anchor bolt detection is performed through the lightweight spatial location mapping network to obtain a global top view.

7. The method according to claim 6, characterized in that, The process of detecting anchor bolts using the lightweight spatial location mapping network based on the mask stacking vector, the scale encoding vector, the normalized coordinate vector, and the preset physical resolution to obtain a global top view includes: Based on the mask stacking vector and the scale encoding vector, convolutional pooling is performed through the encoding layer of the lightweight spatial location mapping network to obtain a high-level semantic feature map. The bottleneck layer of the lightweight spatial location mapping network is used to perform convolution processing on the high-level semantic feature map to obtain the bottleneck feature tensor. The normalized coordinate vector and the preset physical resolution are encoded by the spatial prior injection layer of the lightweight spatial location mapping network to obtain coordinate embedding feature map and resolution embedding feature map respectively. Based on the bottleneck feature tensor, the coordinate embedding feature map, and the resolution embedding feature map, the prior enhancement features are obtained by concatenation and convolution through the spatial prior injection layer of the lightweight spatial location mapping network. Based on the aforementioned prior enhancement features, convolution and concatenation processing are performed through the decoding layer of the lightweight spatial location mapping network to obtain the final feature map; Based on the final feature map, anchor bolt detection is performed on the region to be detected through the global image reconstruction layer of the lightweight spatial location mapping network to obtain the global top view.

8. A batch intelligent inspection system for pre-embedded anchor bolt groups in steel structures, characterized in that, The system includes: a determination module, a data acquisition module, a preprocessing and labeling module, and a detection module, wherein, The determining module is used to determine the construction area to be inspected; and to set up targets in the construction area to be inspected, and to determine the absolute spatial coordinates of each of the multiple peripheral targets and the group target codes of each of the multiple anchor bolt groups. The acquisition module is used to acquire images of the construction area to be detected based on the absolute spatial coordinates of the multiple peripheral targets, and obtain a four-scale image sequence; wherein, the four-scale image sequence contains images of four different scale types; The preprocessing and labeling module is used to perform image preprocessing and anchor group identification labeling on the four-scale image sequence based on the group target encoding of each of the multiple anchor groups, to obtain a mask image sequence and the scale type corresponding to the mask image sequence; The detection module is used to perform anchor bolt detection through a lightweight spatial location mapping network based on the mask image sequence, the scale type corresponding to the mask image sequence, and the absolute spatial coordinates of the multiple peripheral targets, to obtain a global top view; wherein, the global top view is a pixel-level distribution map of the actual spatial positions of all anchor bolts in the construction area to be detected.

9. A batch intelligent inspection device for pre-embedded anchor bolt groups in steel structures, characterized in that, include: Processor and memory, of which, The memory is used to store computer programs; The processor is configured to call and run the computer program from the memory to perform the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores executable instructions for causing a processor to execute, thereby implementing the method of any one of claims 1 to 7.

Images