A tunnel drilling position positioning method based on auxiliary drilling target detection

Abstract

The present invention provides a method for locating tunnel drilling positions based on auxiliary drilling target detection, and relates to the technical field of target detection and tunnel drilling positioning. The method first obtains a dataset of shield plate auxiliary drilling target images, labels the auxiliary drilling targets in the dataset, performs image preprocessing on the labeled training set, and produces a training set in VOC format; then, extracts the priori box size information of the auxiliary drilling targets on the tunnel shield plates in the training set; then, builds a target detection model based on an improved YOLOV5 network; and trains the target detection model using the training set data; finally, uses the trained target detection model to detect the shield plate auxiliary drilling targets; and uses the subway tunnel shield plate auxiliary drilling targets detected by the target detection model to determine the appropriate drilling position. The target detection model designed by this method improves the accuracy of target detection in low light conditions, thereby improving the precision of target detection in tunnel environments.

Description

Technical Field

[0001] The present invention relates to the technical field of target detection and tunnel drilling positioning, and in particular to a tunnel drilling hole positioning method based on auxiliary drilling target detection. Background Art

[0002] With the development of technologies such as network communications, artificial intelligence, and video analysis, the traditional subway tunnel construction industry has been closely following the development opportunities of artificial intelligence. Combining the business characteristics of subway tunnel construction, it has actively explored the practical application of artificial intelligence. How to apply this technology to the safe operation of subway tunnel construction and the implementation and management of drilling to achieve intelligence, Internet of Things, and practicality has always been the focus of subway tunnel construction companies.

[0003] Automated tunnel drilling technology is a new type of technology. Its development requires the combination of different technical disciplines and the strengthening of research on construction methods, construction environment, and working conditions to achieve a more efficient construction process, improve construction quality and efficiency, and reduce construction costs.

[0004] During subway tunnel construction, holes need to be drilled in the tunnel sidewalls for cable and pipe installation. Tunnel drilling is often done manually, with workers using ladder trucks to support electric drills. To address this, a high-definition industrial camera is installed on the robotic arm of a tunnel drilling robot. When the automatic drilling device stops, it captures the moment. The industrial computer's built-in target detection algorithm then detects auxiliary drilling targets, such as shield plate grooves (in four directions: up, down, left, and right) and anchor bolts. This auxiliary drilling target information is then used to plan appropriate drilling target locations. This location information is then sent back to the industrial computer, which controls the tunnel drilling robot to automatically perform the drilling operation. This results in a more efficient drilling process, improved construction quality and efficiency, and reduced costs.

[0005] The object detection process is plagued by numerous uncertainties, such as the uncertain number of objects in an image, the varying appearances, shapes, and poses of objects, and the interference of factors such as illumination, occlusion, and low detection accuracy for small objects during imaging. This makes detection algorithms challenging. Since the advent of the deep learning era, the development of object detection algorithms has primarily focused on two directions: two-stage object detectors and single-stage object detectors. Two-stage object detectors typically use a region proposal network (RPN) to generate region proposals, followed by classification and bounding box regression. Examples of single-stage detectors include RCNN, Fast RCNN, and Faster RCNN. Single-stage object detectors, such as YOLO, SSD, and RetinaNet, combine region proposal and detection steps into a single network. Compared to other detection algorithms, the YOLOV5 algorithm offers the best balance between accuracy and speed. However, its detection performance in subway tunnels is subpar, and it is unable to quickly and effectively identify objects in shield-assisted drilling in subway tunnels. At present, in addition to the traditional method of planning drilling positions with the help of workers' experience, there are also some methods such as using robots, sensors and artificial intelligence algorithms to plan drilling positions to improve the safety and efficiency of drilling operations. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to address the deficiencies of the above-mentioned prior art and provide a tunnel drilling hole positioning method based on auxiliary drilling target detection, so that the auxiliary drilling robot can perform automated drilling operations more simply, efficiently and orderly.

[0007] To solve the above technical problems, the present invention adopts a technical solution: a tunnel drilling hole location method based on auxiliary drilling target detection, comprising the following steps:

[0008] Step 1: Obtain a dataset of shield plate assisted drilling target images, adjust the resolution of the dataset to a uniform size, and divide it into a training set and a test set according to the ratio;

[0009] Step 2: Label the auxiliary drilling targets in the shield plate auxiliary drilling target image dataset and check whether there are images with incorrect or missing category names; the auxiliary drilling targets include shield plate grooves and anchor bolts;

[0010] The auxiliary drilling targets in the shield plate auxiliary drilling target image dataset are annotated, including five target categories: shield plate upward groove, shield plate downward groove, shield plate left groove, shield plate right groove, and anchor bolt. After annotation, the size, position, and category information of each target in the image are saved at the same time.

[0011] Step 3: Perform image preprocessing on the labeled training set to produce a training set in VOC format;

[0012] The process of making a VOC format dataset is specifically as follows: the image data and annotation information are combined to make a training dataset in the PASCAL VOC format. The training dataset contains three folders: Annotations, ImageSets, and images. The Annotations folder stores XML files with image annotation information, and the ImageSets folder stores txt files. Each line in the txt file contains the name of an image. These txt files divide the images in the dataset into training and test sets. The images folder contains all training and test verification images.

[0013] Step 4: Extract the prior frame size information of the auxiliary drilling target on the tunnel shield plate in the training set;

[0014] In the training set, the labeled box dimensions of the tunnel shield groove and anchor frame area were obtained through labeling information. K-means clustering was used to obtain the prior box dimensions of the shield groove and anchor bolt. The three scale prior boxes were used to cluster the prior boxes into nine sizes.

[0015] Step 5: Build a target detection model based on the improved YOLOV5 network;

[0016] Based on the YOLOv5 detection model, in its backbone network, all ordinary convolutions except the CSP structure are replaced with GhostConv lightweight convolutions, and a CA attention module is added before the SPPF structure; in the feature extraction network part, all ordinary convolutions are replaced with GhostConv lightweight convolutions, and a PSA attention module is added after the CSP structure. The Softmax classifier is used for training. In the training parameters, the learning rate is set to 0.0001 and the batch size is 12. Batch normalization is used for iterative training of weight parameters and bias parameters. The weight decay rate weight_decay is set to 0.0005 by default, the power is 0.937, the optimizer is AdamW, the pooling layer adopts the average pooling method, and the padding mode of the maximum pooling is set to SAME. The training is iterated 100 times.

[0017] Step 6: Use the training set data to train the target detection model;

[0018] Modify the target category number and corresponding target category names of the target detection model, and replace the preset 9 prior box sizes in the model configuration file yaml with the prior box sizes clustered in step 4. Then, add the relevant codes of the GhostConv lightweight convolution, PSA attention module, and CA attention module to the corresponding positions of the YOLOv5 network structure. Finally, run the train.py file of the YOLOv5 network to train the improved YOLOV5 target detection model on the training set.

[0019] Step 7: Use the trained target detection model to detect the shield plate auxiliary drilling target;

[0020] Images of subway tunnel shield plates captured in real time by industrial cameras are fed into a trained target detection model for detection. Preliminary detection results for shield plate-assisted drilling targets are obtained under a given accuracy confidence threshold. These preliminary detection results include target categories and target bounding box coordinates. A non-maximum suppression algorithm is then used to remove redundant target bounding boxes, thereby identifying targets for shield plate-assisted drilling in subway tunnels.

[0021] Step 8: Use the target detection model to detect the subway tunnel shield plate auxiliary drilling targets to determine the appropriate drilling hole locations;

[0022] 1) Find all matching pairs of shield plate grooves in subway tunnels. The matching pairs must meet the following two conditions:

[0023] a. Horizontal direction: The left groove in the image is on the left, the right groove is on the right, and the distance between the left and right grooves is greater than the minimum distance and less than the maximum distance;

[0024] b. The left groove and the right groove intersect in the vertical direction;

[0025] 2) Group all groove matching pairs according to the shield plate and select the matching pair group closest to the center position in the image. The specific grouping steps are as follows:

[0026] All matching pairs are grouped, and the matching pairs on the same shield plate are all in the same group;

[0027] Select the matching pairs closest to the center of the image;

[0028] 3) Analyze all groove matching pairs within the selected groove matching pair group and select the appropriate punch hole position; the appropriate punch hole position is 1 / 3 of the distance between the rightmost side of the right shield plate groove and the leftmost side of the left shield plate groove on the y-axis; select the height position of the specified punch hole on the x-axis, and the intersection of the selected x-axis and y-axis is used as the coordinate of the appropriate punch hole position.

[0029] The beneficial effects of adopting the above technical solution are as follows: the present invention provides a tunnel drilling hole location method based on auxiliary drilling target detection. Starting from the actual environment of subway tunnels, the present invention studies the problems of uneven lighting, local overexposure, severe occlusion in small tunnel spaces, and inaccurate small target detection. A target detection model based on improved YOLOV5 is designed. In the backbone network and feature extraction network, ordinary convolution is replaced with GhostConv lightweight convolution, and a CA attention module is added before the SPPF structure, and a PSA attention module is added after the CSP structure. The designed target detection model uses the attention module to help improve the performance of the model while reducing parameters, allowing the model to focus on the most relevant parts of the data and improve the accuracy of target detection in low light, thereby improving the accuracy of target detection in tunnel environments. The actual detection time is 250ms under 1280 resolution using a 1060 graphics card, with high accuracy. It has high application value in the field of tunnel drilling detection. At the same time, the K-Means clustering algorithm is used to statistically analyze the ground truth annotation boxes to obtain the scale of the anchor bounding box, thereby generating higher-quality candidate windows. This improves the accuracy of the predicted box regression, thereby enhancing the detection accuracy of the auxiliary drilling targets in subway tunnel shield plates. Finally, the structure of the auxiliary drilling targets is used to better and adaptively locate the coordinate information of the appropriate drilling positions. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 A flowchart of a tunnel drilling position positioning method based on auxiliary drilling target detection provided by an embodiment of the present invention;

[0031] Figure 2 A schematic diagram of the network structure of an object detection model based on the improved YOLOV5 network provided in an embodiment of the present invention;

[0032] Figure 3 Schematic diagram of the network structure of GhostConv lightweight convolution provided by an embodiment of the present invention;

[0033] Figure 4 A schematic diagram of the network structure of the CA attention module provided in an embodiment of the present invention;

[0034] Figure 5 A schematic diagram of the network structure of the PAS attention module provided by an embodiment of the present invention;

[0035] FIG6 is a diagram showing the actual effect of detecting targets and locating suitable hole positions for shield plate-assisted drilling in a subway tunnel provided by an embodiment of the present invention, wherein (a) is the detection result of the target detection model, (b) is the suitable y-axis position of the drilling position, and (c) is the coordinates of the suitable drilling position. Implementation Method

[0036] The following embodiments of the present invention are described in further detail with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention.

[0037] A tunnel drilling hole location method based on auxiliary drilling target detection, such as Figure 1 As shown, the following steps are included:

[0038] Step 1: Obtain a dataset of shield plate assisted drilling target images, adjust the resolution of the dataset to a uniform size, and divide it into a training set and a test set according to the ratio;

[0039] In this example, we simulated the motion of a robotic arm and used a similar handheld industrial camera to capture a large number of video images of the shield plate under different lighting scenarios. We saved an image every two seconds, deleted images with poor quality, and divided the captured images into training and test sets in a 4:1 ratio.

[0040] Step 2: Label the auxiliary drilling targets (shield plate grooves and anchor bolts) in the shield plate auxiliary drilling target image dataset and check whether there are images with incorrect or missing category names;

[0041] The auxiliary drilling targets in the dataset are labeled, including five target categories: shield plate groove facing up (up), shield plate groove facing down (down), shield plate groove facing left (left), shield plate groove facing right (right), and anchor bolt (circle). After labeling, the size, position, and category information of each target in the image are also saved.

[0042] As a possible implementation of this embodiment, a training set image enhancement step is also included before training the object detection model. This training set image enhancement includes adjusting the image's exposure, HSV color gamut, and saturation, generating more training data using data jittering and horizontal flipping strategies, and employing the Mosaic data enhancement method. Its main idea is to randomly crop and scale four images, then randomly arrange and splice them to form a single image. This enriches the dataset while increasing the number of small sample targets and speeding up network training.

[0043] Step 3: Perform image preprocessing on the labeled training set to produce a training set in VOC format;

[0044] The process of making a VOC format dataset is specifically as follows: the image data and annotation information are combined to make a training dataset in the PASCAL VOC format. The training dataset contains three folders: Annotations, ImageSets, and images. The Annotations folder stores XML files with image annotation information, and the ImageSets folder stores txt files. Each line in the txt file contains the name of an image. These txt files divide the images in the dataset into training and test sets. The images folder contains all training and test verification images.

[0045] Step 4: Extract the prior frame size information of the auxiliary drilling target on the tunnel shield plate in the training set;

[0046] In the training set, the labeled information was used to obtain the annotated box size information for the tunnel shield groove and anchor box area. K-means clustering was used to determine the prior box size of the shield groove and anchor bolt. The three scale prior boxes were then clustered to produce nine prior box sizes. The nine prior box sizes are: (10x13), (16x30), (33x23), (30x61), (62x45), (59x119), (116x90), (156x198), and (373x326). In this embodiment, larger prior boxes (116x90), (156x198), and (373x326) were applied to the smallest 4*4 and 2*2 feature maps (with the largest receptive field), making them suitable for detecting larger objects. Medium-sized prior boxes (30x61), (62x45), and (59x119) are applied to medium-sized 16*16 and 8*8 feature maps (medium receptive field), which are suitable for detecting medium-sized objects. Smaller prior boxes (10x13), (16x30), and (33x23) are applied to larger 64*64 and 32*32 feature maps (smaller receptive field), which are suitable for detecting small objects.

[0047] Step 5: Build a target detection model based on the improved YOLOV5 network;

[0048] Build a target detection model in the Pytorch deep learning framework, such as Figure 2As shown in the figure, based on the YOLOv5 detection model, all ordinary convolutions except the CSP structure in its backbone network are replaced by GhostConv lightweight convolutions, and a CA attention module is added before the SPPF structure; the above strategy is also adopted in the feature extraction network part, and all ordinary convolutions are replaced by GhostConv lightweight convolutions, and a PSA attention module is added after the CSP structure. The Softmax classifier is used for training, and the learning rate is set to 0.0001 and the batch size is 12 in the training parameters. Batch normalization is used to iteratively train the weight parameters and bias parameters, and the weight decay rate weight_decay is set to 0.0005 by default, the power is 0.937, the optimizer is AdamW, the pooling layer adopts the average pooling method, and the padding mode of the maximum pooling is set to SAME. The training is iterated 100 times;

[0049] In this embodiment, the basic network structure of GhostConv lightweight convolution is as follows Figure 3 As shown in the figure, the original one-step convolution is converted to a two-step convolution. The first step performs a regular convolution, but with a reduced number of output channels. The second step performs a depthwise separable convolution based on the first step. In addition, the second step has a parallel connection branch, which is the output of the first step. The number of output channels of GhostConv lightweight convolution is equal to the number of channels c after the first convolution step plus the number of channels n*c after the second convolution step, so the final number of channels is (n+1)*c.

[0050] The basic network structure of the CA attention module is as follows Figure 4 As shown in Figure 2, two 1D global pooling operations are used to aggregate the input features in the vertical and horizontal directions into two separate directional feature maps, respectively. These two feature maps with embedded direction-specific information are then encoded into two attention maps, each of which captures the long-range correlation of the input feature map along a spatial direction. Therefore, the position information can be preserved and generated into two attention maps. The two attention maps are then applied to the input feature map by multiplication to emphasize the representation of interest.

[0051] The basic network structure of the PAS attention module is as follows Figure 5 As shown in the figure. First, this module divides the input tensor into S groups based on the channel. Each group is convolved with a different convolution kernel size, and the convolution kernel size gradually increases, such as K = 3, 5, 7, and 9, to obtain the receptive field of feature maps of different scales. In this way, PSA integrates contextual information of different scales and has better pixel-level attention. Then, the SE module is used to extract the weight value of each group channel, and finally the weight value of the S group is weighted. Finally, the weight value of each group channel attention is spliced, softmax normalized, and the output feature map of the S group is weighted.

[0052] Step 6: Use the training set data to train the target detection model;

[0053] The COCO dataset is a public image dataset, and the pre-trained weight file of the YOLOv5 model is trained on this dataset. During migration training, according to the target category requirements of the target detection algorithm in this instance, the original target category 80 is changed to 5, and the target category names are modified to up, down, left, right, and circle. Modify the number of target categories and the corresponding target category names of the target detection model, and replace the preset 9 prior box sizes in the model configuration file yaml with the prior box sizes clustered in step 4, then add the relevant codes of the GhostConv lightweight convolution, PSA attention module, and CA attention module to the corresponding positions of the YOLOv5 network structure, and finally run the train.py file of the YOLOv5 network to train the improved YOLOV5 target detection model on the training set;

[0054] Step 7: Use the trained target detection model to detect the shield plate auxiliary drilling target;

[0055] Images of subway tunnel shield plates captured in real time by an industrial camera are fed into a trained target detection model for detection. Preliminary detection results for shield plate-assisted drilling targets are obtained under a given accuracy confidence threshold. These preliminary detection results include target categories and target bounding box coordinates. The non-maximum suppression (NMS) algorithm is then used to remove redundant target bounding boxes, thereby identifying targets for subway tunnel shield plate-assisted drilling. The target detection model detection results are shown in Figure 6(a).

[0056] The process of removing redundant target bounding boxes using the non-maximum suppression algorithm is specifically as follows:

[0057] (1) Obtain all target detection windows and their scores through target detection model detection;

[0058] (2) Sort the scores of all windows, and select the window with the highest score and its corresponding window as the suppression window;

[0059] (3) The remaining windows with lower scores than the suppression window are regarded as suppressed windows, and the overlap area (IOU) between the suppressed window and the suppression window is calculated;

[0060] (4) If the overlapping area is greater than the set threshold T, the window is suppressed, that is, the window is removed;

[0061] (5) The process ends when only one detection window is left. Otherwise, the next window with the highest score is selected as the suppression window and the process goes to step (4). After non-maximum suppression, the final detection result is obtained.

[0062] Step 8: Use the target detection model to detect the subway tunnel shield plate auxiliary drilling targets to determine the appropriate drilling hole locations;

[0063] 1) Find all matching pairs of shield plate grooves in subway tunnels. The matching pairs must meet the following two conditions:

[0064] a. Horizontal direction: In the image, the left groove is on the left, and the right groove is on the right, and the distance between the left and right grooves is greater than the minimum distance and less than the maximum distance (the minimum distance is the closest distance between the right edge of the left groove and the left edge of the right groove. The maximum distance is determined based on actual conditions. In this embodiment, the minimum distance is set to 100 pixels and the maximum distance is set to 1500 pixels).

[0065] b. The left groove and the right groove intersect in the vertical direction;

[0066] 2) Group all groove matching pairs according to the shield plate and select the matching pair group closest to the center position in the image. The specific grouping steps are as follows:

[0067] All matching pairs are grouped, and the matching pairs on the same shield plate are all in the same group;

[0068] Select the matching pairs closest to the center of the image;

[0069] 3) Analyze all groove matching pairs within the selected groove matching pair group and select the appropriate punch hole position. The appropriate punch hole position is 1 / 3 of the distance between the rightmost groove of the right shield plate and the leftmost groove of the left shield plate on the y-axis. The appropriate y-axis position of the punch hole position is shown in Figure 6 (b). Select the height position of the specified punch hole on the x-axis, and the intersection of the selected x-axis and y-axis is used as the coordinate of the appropriate punch hole position, as shown in Figure 6 (c).

[0070] By sending the appropriate drilling hole locations to the drilling robot, drilling operations can be carried out in the subway tunnel.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A tunnel drilling hole location method based on auxiliary drilling target detection, characterized by: The following steps are involved: Step 1: Obtain a dataset of shield plate assisted drilling target images, adjust the resolution of the dataset to a uniform size, and divide it into a training set and a test set according to the ratio; Step 2: Label the auxiliary drilling targets in the shield plate auxiliary drilling target image dataset and check whether there are images with incorrect or missing category names; the auxiliary drilling targets include shield plate grooves and anchor bolts; Step 3: Perform image preprocessing on the labeled training set to produce a training set in VOC format; Step 4: Extract the prior frame size information of the auxiliary drilling target on the tunnel shield plate in the training set; Step 5: Build a target detection model based on the improved YOLOV5 network; Based on the YOLOv5 detection model, in its backbone network, all ordinary convolutions except the CSP structure are replaced with GhostConv lightweight convolutions, and a CA attention module is added before the SPPF structure; in the feature extraction network part, all ordinary convolutions are replaced with GhostConv lightweight convolutions, and a PSA attention module is added after the CSP structure. The Softmax classifier is used for training. In the training parameters, the learning rate is set to 0.0001 and the batch size is 12. Batch normalization is used for iterative training of weight parameters and bias parameters. The weight decay rate weight_decay is set to 0.0005 by default, the power is 0.937, the optimizer is AdamW, the pooling layer adopts the average pooling method, and the padding mode of the maximum pooling is set to SAME. The training is iterated 100 times. Step 6: Use the training set data to train the target detection model; Step 7: Use the trained target detection model to detect the shield plate auxiliary drilling target; Step 8: Use the target detection model to detect the subway tunnel shield plate auxiliary drilling targets to determine the appropriate drilling hole locations; 1) Find all matching pairs of shield plate grooves in subway tunnels. The matching pairs must meet the following two conditions: a. Horizontal direction: The left groove in the image is on the left, the right groove is on the right, and the distance between the left and right grooves is greater than the minimum distance and less than the maximum distance; b. The left groove and the right groove intersect in the vertical direction; 2) Group all groove matching pairs according to the shield plate and select the matching pair group closest to the center position in the image. The specific grouping steps are as follows: All matching pairs are grouped, and the matching pairs on the same shield plate are all in the same group; Select the matching pairs closest to the center of the image; 3) Analyze all groove matching pairs within the selected groove matching pair group and select the appropriate punch hole position; the appropriate punch hole position is 1 / 3 of the distance between the rightmost side of the right shield plate groove and the leftmost side of the left shield plate groove on the y-axis; select the height position of the specified punch hole on the x-axis, and the intersection of the selected x-axis and y-axis is used as the coordinate of the appropriate punch hole position.

2. The method for locating tunnel drilling positions based on auxiliary drilling target detection according to claim 1, characterized in that: The step 2 labels the auxiliary drilling targets in the shield plate auxiliary drilling target image data set, including five target categories: shield plate upward groove, shield plate downward groove, shield plate left groove, shield plate right groove and anchor bolt. After labeling, the size, position and category information of each target in the image are saved at the same time.

3. The tunnel drilling hole location method based on auxiliary drilling target detection according to claim 1 is characterized by: The process of making a VOC format dataset is specifically as follows: the image data and annotation information are combined to make a training dataset in the PASCAL VOC format. The training dataset contains three folders: Annotations, ImageSets, and images. The Annotations folder stores XML files of image annotation information, and the ImageSets folder stores txt files. Each line in the txt file contains the name of an image. These txt files divide the images in the dataset into training set and test set sets. The images folder contains all training and test verification images.

4. The method for locating tunnel drilling positions based on auxiliary drilling target detection according to claim 1, characterized in that: The specific method of step 4 is: In the training set, the labeled box size information of the tunnel shield groove and anchor frame area is obtained through labeling information. K-means clustering is used to obtain the prior box size of the shield groove and anchor bolt, and the three set scale prior boxes are clustered into nine sizes of prior boxes.

5. The method for locating tunnel drilling positions based on auxiliary drilling target detection according to claim 1, characterized in that: The specific method of step 6 is: modify the number of target categories and the corresponding target category names of the target detection model, and replace the preset 9 prior box sizes in the model configuration file yaml with the prior box sizes clustered in step 4, then add the relevant codes of GhostConv lightweight convolution, PSA attention module and CA attention module to the corresponding positions of the YOLOv5 network structure, and finally run the train.py file of the YOLOv5 network to train the improved YOLOV5 target detection model on the training set.

6. The method for locating tunnel drilling positions based on auxiliary drilling target detection according to claim 2, characterized in that: The specific method of step 7 is as follows: inputting the subway tunnel shield plate image captured in real time by the industrial camera into the trained target detection model for detection, and obtaining a preliminary detection result of the shield plate assisted drilling target under a given accuracy confidence threshold, the preliminary detection result including the target category and the target bounding box coordinates; Then, the non-maximum suppression algorithm is used to remove redundant target bounding boxes, thereby identifying the subway tunnel shield plate assisted drilling targets.

Images