Safety identification method for high-altitude operation protection in hole

Abstract

The invention relates to the technical field of tunnel construction safety protection, and discloses an in-hole high-altitude operation protection safety identification method, which comprises the following steps: acquiring tunnel face construction image data; constructing an in-hole high-altitude operation target identification model and performing pre-training; the in-hole high-altitude operation target comprises a trolley, a worker and a protective fence; obtaining the recognition type and position information of the in-hole aerial work target according to the tunnel face construction image data by using the trained in-hole aerial work target recognition model; determining a target space relation according to the identification type and the position information of the in-hole aerial work target; determining an in-hole aerial work scene according to the spatial relationship between the trolley and the worker; and based on the in-hole aerial work scene, the aerial work protection safety state is determined according to the space relation between the workers and the protective fences. According to the invention, accurate identification of the working procedure of working aloft on the tunnel face and the protection safety state of working aloft of workers is realized.

Description

Technical Field

[0001] The invention relates to the technical field of tunnel construction safety protection, and in particular to a method for identifying safety of high-altitude work protection in a tunnel. Background Art

[0002] As an important part of the infrastructure field, tunnel (road) engineering plays a pivotal role in transportation, hydropower development, water resource utilization, etc. During the construction process, tunnels are often restricted by complex and changeable geological conditions, narrow tunnel space, harsh construction environment and other factors. Construction workers are more likely to suffer from safety accidents such as falling from heights, being hit by objects, tunnel collapse, and mechanical collision, resulting in huge losses of life and property.

[0003] Tunnel construction involves the coordinated construction of many processes, including manual drilling, charging, erecting frames, and secondary lining, all of which need to be performed at high altitudes. Unsafe behaviors of workers and the lack of corresponding safety protection equipment are the main causes of high-altitude fall accidents. The status of high-altitude protective equipment at the construction site is usually determined by manual inspections and manual analysis of video surveillance images, which is often time-consuming and labor-intensive and difficult to ensure the timeliness of the inspection. In addition, tunnel projects are mostly long linear projects, and as the excavation of the face continues to deepen, the construction conditions in the tunnel become more complex. Manual inspections have low visibility and often have blind spots, making it impossible to conduct a global inspection of construction safety.

[0004] With the development of sensor technology and computer vision technology, automatic real-time detection of the safety status of workers working at heights has become a reality. However, the sensor method requires workers to wear additional sensor equipment, which is expensive and affects the normal operation of workers; the computer vision method is low-cost and has little impact on workers, but faces problems such as poor lighting conditions in tunnels, complex image backgrounds, and changes in the size of the detection target. Therefore, it is necessary to invent a high-altitude work protection safety identification method suitable for complex environments in tunnels to achieve rapid identification and feedback of high-altitude work protection safety. Summary of the invention

[0005] In view of the above-mentioned deficiencies in the prior art, the present invention provides a method for safety identification of high-altitude work protection in a cave.

[0006] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: A method for identifying safety during high-altitude work protection in a cave comprises the following steps: Obtain tunnel face construction image data; Constructing a high-altitude working target recognition model in a tunnel and performing pre-training; the high-altitude working targets in the tunnel include trolleys, workers and guardrails; The trained in-tunnel aerial work target recognition model is used to obtain the in-tunnel aerial work target recognition type and location information based on the tunnel face construction image data; Determine the target spatial relationship based on the target identification type and position information of the high-altitude operation in the cave; Determine the high-altitude operation scene in the tunnel based on the spatial relationship between the trolley and the workers; Based on the high-altitude work scene in the tunnel, the high-altitude work protection safety status is determined according to the spatial relationship between the workers and the guardrails.

[0007] Preferably, the pre-training of the in-hole aerial work target recognition model includes: The CycleGAN network is used to enhance the tunnel face construction images; Annotate the enhanced tunnel face construction image with examples.

[0008] Preferably, using the CycleGAN network to enhance the tunnel face construction image includes: The first discriminator judges the input low-illumination tunnel face construction image to determine whether it meets the characteristics of the image in the low-illumination tunnel dataset, and outputs a probability value indicating that the input image is a true low-illumination image; The input low-illuminance tunnel face construction image is input into the first generator to generate a normal-illuminance tunnel face construction image, and then the generated image is judged by the second discriminator to determine whether it meets the characteristics of the image in the normal-illuminance tunnel data set, and the probability value representing that the input image is a true normal-illuminance image is output; The tunnel face construction image with normal illumination generated by the first generator is input into the second generator, and the tunnel face construction image with low illumination is regenerated, so that the generated tunnel face construction image with low illumination is consistent with the input tunnel face construction image with low illumination, thereby forming an image conversion cycle; The adversarial loss, cycle consistency loss, and identity mapping loss in the image conversion process are calculated, and the optimizer is used to calculate and update the gradient value of the loss function and the parameter values ​​of the generator and discriminator, and the gradient descent method is used to iterate continuously until the overall loss function is minimized.

[0009] Preferably, the in-hole aerial work target recognition model includes: Backbone network, neck network and prediction network; The backbone network includes a first CBS module, a first C3 module, a second CBS module, a second C3 module and an SPPF module which are arranged in sequence, as well as a first attention module arranged at the output end of the second CBS module, a second attention module arranged at the output end of the second C3 module, and a third attention module arranged at the output end of the SPPF module.

[0010] Preferably, the first attention module, the second attention module and the third attention module each include: A height global average pooling unit, a width global average pooling unit, a stacking unit, a first convolution unit, a normalization unit, a first activation unit, a split unit, a transposition unit, a second convolution unit, a third convolution unit, a second activation unit, a third activation unit, and a multiplication unit; The height global average pooling unit and the width global average pooling unit respectively perform global average pooling operations in the height direction and global average pooling operations in the width direction on the input feature map to obtain a first global average pooling feature map and a second global average pooling feature map; After the stacking unit performs a stacking operation on the first global average pooling feature map and the second global average pooling feature map, the stacking unit sequentially passes through a first convolution unit, a normalization unit, and a first activation unit to obtain a first spatial feature map; After the segmentation unit separates the width direction feature map and the height direction feature map from the first spatial feature map, the transposition unit performs transposition operations on the width direction feature map and the height direction feature map respectively, and then passes the two feature maps through the second convolution unit and the third convolution unit, and the second activation unit and the third activation unit respectively, to obtain the width direction attention and the height direction attention; The multiplication unit multiplies the input feature map with the width direction attention and the height direction attention to obtain an attention feature map.

[0011] Preferably, determining the target spatial relationship according to the identification type and position information of the aerial work target in the cave includes: Determine the spatial inclusion, intersection or separation relationship between the worker and trolley target category boundary boxes according to the coordinate information of the worker detection box and the coordinate information of the trolley detection box; The spatial inclusion, intersection or separation relationship between the worker and the guardrail target category boundary box is determined according to the coordinate information of the worker detection box and the coordinate information of the guardrail detection box.

[0012] Preferably, determining the spatial inclusion, intersection or separation relationship between the worker and trolley target category boundary boxes according to the coordinate information of the worker detection box and the coordinate information of the trolley detection box includes: When the coordinate information of the worker detection frame A meets the coordinate information of the trolley detection frame B , then there is a spatial inclusion relationship between the worker and the trolley target category boundary box; When the coordinate information of the worker detection frame A meets the coordinate information of the trolley detection frame B , then there is a spatial intersection relationship between the worker and the trolley target category boundary box; When the coordinate information of the worker detection frame A meets the coordinate information of the trolley detection frame B , then the worker and the trolley target category boundary box are spatially separated.

[0013] Preferably, determining the spatial inclusion, intersection or separation relationship between the worker and the guardrail target category boundary box according to the coordinate information of the worker detection box and the coordinate information of the guardrail detection box includes: When the coordinate information of the worker detection frame A meets the coordinate information of the guardrail detection frame C , then there is a spatial inclusion relationship between the worker and the guardrail target category boundary box; When the coordinate information of the worker detection frame A meets the coordinate information of the guardrail detection frame C , then there is a spatial intersection relationship between the worker and the guardrail target category boundary box; When the coordinate information of the worker detection frame A meets the coordinate information of the guardrail detection frame C , then the worker and the guardrail target category boundary box are spatially separated.

[0014] As a preferred method, the high-altitude working scene in the cave is determined according to the spatial relationship between the trolley and the worker, including: If the worker and the trolley target category bounding boxes are spatially contained or spatially intersected, the scene is judged to be an aerial work process in a cave; If only the trolley target is detected, only the worker target is detected, or both are detected and the trolley and all the worker target category boundary boxes are spatially separated, it is judged as a non-high-altitude operation process.

[0015] As a preferred embodiment, based on the high-altitude operation scene in the cave, determining the high-altitude operation protection safety status according to the spatial relationship between the worker and the guardrail includes: According to the shooting range of the surveillance camera and the size of the trolley, the area above the trolley detection frame is delineated as the high-altitude working area, and the spatial relationship between the worker and the guardrail is judged in this area. If the high-altitude working worker detection frame is spatially separated from a certain guardrail detection frame, it is judged that the worker is in a dangerous working state where the guardrail is missing; if the high-altitude working worker detection frame is spatially intersected or spatially contained with all guardrail detection frames, it is judged that the worker is in a safe working state.

[0016] The present invention has the following beneficial effects: 1. The present invention performs image enhancement on images in tunnels. The processed images have obvious improvements in four indicators: brightness, average gradient, energy gradient and information entropy, which improves the visual effect of the images and improves the detection accuracy of subsequent image recognition methods.

[0017] 2. The present invention realizes the automatic processing and recognition of tunnel construction image data, and based on the spatial relationship of identified targets such as workers, trolleys, guardrails, etc., it realizes the accurate recognition of the high-altitude operation process at the heading face and the safety status of workers' high-altitude operation protection, providing technical support for all-weather, all-round and efficient safety management of high-altitude edge areas in the tunnel. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A schematic diagram of a safety identification method for high-altitude work protection in a cave; Figure 2 Schematic diagram of the CycleGAN network structure for preprocessing images inside the cave; Figure 3 This is a schematic diagram of the structure of the target recognition model for high-altitude operations in the cave; Figure 4 This is a schematic diagram showing an example of the spatial relationship between the trolley and the worker; Figure 5 Schematic diagram of the confusion matrix for high-altitude scene recognition; Figure 6 Schematic diagram of confusion matrix for protection status identification; Figure 7 This is a schematic diagram showing an example of the spatial relationship between workers and guardrails; Figure 8 Schematic diagram of the safety identification results for workers working at heights. DETAILED DESCRIPTION

[0019] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0020] like Figure 1 As shown, a method for identifying safety of aerial work protection in a cave provided by an embodiment of the present invention includes the following steps S1 to S6: S1. Acquire tunnel face construction image data; In an optional embodiment of the present invention, step S1 uses the video data obtained by video monitoring of the tunnel face to capture frames to obtain the construction image inside the tunnel. In order to adapt to the characteristics of tunnel engineering such as changeable lighting conditions and complex construction environment inside the tunnel, a camera with high definition, wide angle, dustproof and explosion-proof functions is selected.

[0021] S2. Constructing a model for identifying targets of aerial work in a tunnel and pre-training the model; the targets of aerial work in a tunnel include trolleys, workers and guardrails; In an optional embodiment of the present invention, when step S2 pre-trains the target recognition model for high-altitude operations in the tunnel, real tunnel construction scenes from different distances, angles, locations, and processes are screened out from the images acquired from the camera, including construction workers, trolleys, guardrails in high-altitude operation scenes such as drilling, charging and blasting, and construction workers in non-high-altitude operation scenes. A total of 690 images are collected to construct a data set for feature analysis of construction images in the tunnel.

[0022] In this embodiment, the pre-training of the indoor high-altitude operation target recognition model includes: The CycleGAN network is used to enhance the tunnel face construction images; Annotate the enhanced tunnel face construction image with examples.

[0023] Due to the poor lighting conditions in the construction tunnel, the video surveillance images are disturbed by smoke and dust in the tunnel, and have the characteristics of low illumination and low definition. In this embodiment, CycleGAN is used to enhance the tunnel surveillance images to improve the subsequent detection accuracy. CycleGAN uses adversarial loss function, cycle consistency loss function and identity mapping loss function to constrain and optimize the target image data, so that the original low illumination image domain features are converted to normal illumination image domain features to the greatest extent. The model network structure is as follows: Figure 2 As shown, the total loss function and the overall optimization objective function formula are as follows:

[0024]

[0025] in, For the generator , For the generator , For the discriminator , For the discriminator , is the total loss function, To combat the loss function, is the cycle consistency loss function, is the identity mapping loss function, is the weight of cycle consistency loss, is the weight of the identity mapping loss, is the overall optimization objective function.

[0026] CycleGAN includes forward conversion and reverse conversion processes. The two conversion processes are carried out simultaneously. Taking the forward conversion process as an example, the specific steps are as follows: (1) Through the first discriminator The input low-light tunnel image is judged to determine whether it meets the characteristics of the images in the low-light tunnel dataset, and a probability value is output to represent the probability that the input image is a true low-light image.

[0027] (2) Input the low-light tunnel image into the first generator The tunnel image with normal illumination is generated by the second discriminator The generated image is judged to determine whether it meets the characteristics of the images in the normal illumination tunnel dataset, and a probability value is output to represent the probability that the input image is a true normal illumination image.

[0028] (3) The first generator The normal illumination tunnel image generated in the second generator is input into the second generator In the process, the low illumination tunnel image is regenerated so that the generated low illumination tunnel image is as consistent as possible with the input low illumination tunnel image, forming a cycle.

[0029] (4) Calculate the adversarial loss, cycle consistency loss, and identity mapping loss in the image conversion process, use the optimizer to calculate and update the gradient value of the loss function and the parameter values ​​of the generator and discriminator, and iterate continuously through the gradient descent method until the overall loss function is minimized.

[0030] After preprocessing, the LabelImg annotation tool is used to annotate the workers, trolleys, and guardrail instances in each image in the dataset, thus forming image samples for identifying the safety status of aerial work protection in tunnels.

[0031] Using the previous tunnel face monitoring video images, we obtained image samples of the safety status of aerial work protection in the tunnel through preprocessing, and divided them into training image dataset and test image dataset at a ratio of 4:1 (each image has been labeled with the type labels of trolley, worker, guardrail, etc.). Among them, the image dataset is used as the input of the model, and the categories and location information of the three types of targets detected are used as the output of the model.

[0032] In an optional embodiment of the present invention, the in-hole aerial work target recognition model constructed in step S2 includes: Backbone network, neck network and prediction network; The backbone network includes a first CBS module, a first C3 module, a second CBS module, a second C3 module and an SPPF module which are arranged in sequence, as well as a first attention module arranged at the output end of the second CBS module, a second attention module arranged at the output end of the second C3 module, and a third attention module arranged at the output end of the SPPF module.

[0033] Among them, the first attention module, the second attention module and the third attention module all include: A height global average pooling unit, a width global average pooling unit, a stacking unit, a first convolution unit, a normalization unit, a first activation unit, a split unit, a transposition unit, a second convolution unit, a third convolution unit, a second activation unit, a third activation unit, and a multiplication unit; The height global average pooling unit and the width global average pooling unit respectively perform global average pooling operations in the height direction and global average pooling operations in the width direction on the input feature map to obtain a first global average pooling feature map and a second global average pooling feature map; After the stacking unit performs a stacking operation on the first global average pooling feature map and the second global average pooling feature map, the stacking unit sequentially passes through a first convolution unit, a normalization unit, and a first activation unit to obtain a first spatial feature map; After the segmentation unit separates the width direction feature map and the height direction feature map from the first spatial feature map, the transposition unit performs transposition operations on the width direction feature map and the height direction feature map respectively, and then passes the two feature maps through the second convolution unit and the third convolution unit, and the second activation unit and the third activation unit respectively, to obtain the width direction attention and the height direction attention; The multiplication unit multiplies the input feature map with the width direction attention and the height direction attention to obtain an attention feature map.

[0034] The network structure of the in-tunnel aerial work target recognition model constructed in this embodiment is to add an attention module after the CBS module, C3 module and SPPF module of the original YOLOv5 model backbone network. The network structure is as follows: Figure 3 As shown. First, according to the detection task requirements and computing resource constraints, YOLOv5s is selected as the detection model, and the CA attention mechanism is introduced into the detection model to improve the detection accuracy; then, the divided training set is used to train the YOLOv5-CA model, and the hyperparameters are set as follows: the input image size is 640×640, the batch training data size is set to 32, the initial learning rate is set to 0.001, the weight decay coefficient is set to 0.0005, the training momentum is set to 0.937, the training round is set to 400, and the optimizer is Adam. After multiple iterative trainings, the model network structure, loss function and optimizer parameters are continuously adjusted to obtain the YOLOv5-CA target detection model of construction workers, trolleys and guardrails in the tunnel construction scene with optimal parameters; finally, the model detection performance and generalization ability are evaluated based on the divided test set, and the accuracy, recall and average precision mAP are selected as evaluation indicators to evaluate the accuracy of the model in detecting workers, trolleys and guardrails. The calculation formula is as follows:

[0035]

[0036]

[0037] In the formula, , They are precision and recall respectively; True Positive (TP) is the number of correctly detected targets whose Intersection over Union (IoU) is greater than a set threshold (usually 0.5); False Positive (FP) is the number of actual non-targets detected; False Negative (FN) is the number of undetected targets; The number of types of PPE tested.

[0038] At this time, the model's precision, recall and average precision (mAP) reached 97.8%, 93.3% and 96.0% respectively, which is 1.69% higher than the mAP of 94.4% of the model before the improvement. The model can identify the trolleys, workers and guardrails in the tunnel construction images with high accuracy, so it is believed that the model can be used to identify high-altitude working targets in the tunnel.

[0039] S3, using the trained in-tunnel aerial work target recognition model to obtain the in-tunnel aerial work target recognition type and location information according to the tunnel face construction image data; In an optional embodiment of the present invention, step S3 uses the in-hole aerial work target recognition model established in step S2 to identify the coordinates of the worker detection frame A. , coordinates of the trolley detection frame B And the coordinates of the guardrail detection frame C .

[0040] S4, determining the target spatial relationship according to the target identification type and position information of the high-altitude operation in the cave; In an optional embodiment of the present invention, step S4 determines the target spatial relationship according to the identification type and position information of the aerial work target in the cave, including: Determine the spatial inclusion, intersection or separation relationship between the worker and trolley target category boundary boxes according to the coordinate information of the worker detection box and the coordinate information of the trolley detection box; The spatial inclusion, intersection or separation relationship between the worker and the guardrail target category boundary box is determined according to the coordinate information of the worker detection box and the coordinate information of the guardrail detection box.

[0041] This embodiment is based on the coordinates of the worker detection frame A , coordinates of the trolley detection frame B And the coordinates of the guardrail detection frame C , determine the spatial inclusion, intersection or separation relationship between the bounding boxes of the worker and the trolley, and between the worker and the guardrail.

[0042] The spatial inclusion, intersection or separation relationship between the worker and trolley target category boundary boxes is determined based on the coordinate information of the worker detection box and the coordinate information of the trolley detection box, including: When the coordinate information of the worker detection frame A meets the coordinate information of the trolley detection frame B ,Right now:

[0043] in Indicates the coordinates of the upper left corner of the worker detection box, Indicates the coordinates of the lower right corner of the worker detection box, Indicates the coordinates of the upper left corner of the trolley detection frame, Indicates the coordinates of the lower right corner of the trolley detection frame.

[0044] There is a spatial inclusion relationship between the worker and the trolley target category boundary box.

[0045] When the coordinate information of the worker detection frame A meets the coordinate information of the trolley detection frame B ,Right now:

[0046] Then there is a spatial intersection relationship between the worker and the trolley target category boundary boxes.

[0047] When the coordinate information of the worker detection frame A meets the coordinate information of the trolley detection frame B , then the worker and the trolley target category boundary box are spatially separated.

[0048] Determining the spatial inclusion, intersection or separation relationship between the worker and the guardrail target category boundary box based on the coordinate information of the worker detection box and the coordinate information of the guardrail detection box includes: When the coordinate information of the worker detection frame A meets the coordinate information of the guardrail detection frame C ,Right now:

[0049] Then there is a spatial inclusion relationship between the worker and the guardrail target category bounding box; When the coordinate information of the worker detection frame A meets the coordinate information of the guardrail detection frame C ,Right now:

[0050] Then there is a spatial intersection relationship between the worker and the guardrail target category boundary box; When the coordinate information of the worker detection frame A meets the coordinate information of the guardrail detection frame C , then the worker and the guardrail target category boundary box are spatially separated.

[0051] S5. Determine the high-altitude operation scene in the tunnel based on the spatial relationship between the trolley and the workers; In an optional embodiment of the present invention, step S5 determines whether it belongs to an aerial work scene according to the spatial relationship between the trolley detection frame and the worker detection frame, specifically: If the worker and the trolley target category bounding boxes are spatially contained or spatially intersected, the scene is judged to be an aerial work process in a cave; If only the trolley target is detected, only the worker target is detected, or both are detected and the trolley and all the worker target category boundary boxes are spatially separated, it is judged as a non-high-altitude operation process.

[0052] by Figure 4 Taking the above as an example, the spatial relationship between the trolley detection frame and the worker detection frame is specifically analyzed to determine whether it belongs to a high-altitude operation scene. Figure 4 The middle trolley and workers 2, 3, 4, 5, 6, 7, 8, and 9 are all in a spatial inclusion relationship. In order to achieve the overall construction of the tunnel face, the size of the trolley is often similar to the cross-sectional size of the tunnel. The spatial position of the workers working at the edge of the high altitude must be included in the trolley. This situation is regarded as an operation process in the high-altitude scene; Figure 4 The trolley and worker 1 are in a spatial intersection relationship, that is, construction on the top of the trolley or on the ground may result in spatial intersection, which can also be regarded as an operation process in an aerial scene. When only the trolley is detected, only the worker is detected, or the worker and the trolley are far apart (because the image viewing angle is limited, the situation where the person and the trolley coexist but the distance is too far generally does not occur), it is a spatial separation relationship and is regarded as a non-aerial operation process.

[0053] 87 tunnel images were randomly selected as the data set for model verification, of which 36 images belong to high-altitude work processes and 51 images belong to non-high-altitude work processes. The data set was input into the model, and the confusion matrix obtained was as follows: Figure 5 As shown, it can be seen that there are 33 high-altitude work processes and 54 non-high-altitude work processes, among which TP is 32, FP is 1, and FN is 4, that is, the Precision and Recall of the model are 96.97% and 88.89% respectively, with high recognition accuracy.

[0054] S6. Based on the high-altitude work scene in the tunnel, determine the high-altitude work protection safety status according to the spatial relationship between the workers and the guardrails.

[0055] In an optional embodiment of the present invention, step S6 continues to identify the safety protection status of the aerial workers based on the aerial work scene identified in step S5. According to the shooting range of the surveillance camera and the size of the trolley, a certain area above the trolley detection frame is defined as the aerial work area (this value depends on the size of the tunnel), and the spatial relationship between the workers and the guardrails is analyzed in this area to identify the safety protection status of the aerial workers. If the aerial worker detection frame is spatially separated from a certain guardrail detection frame, it is judged that the worker is in a dangerous working state with a missing guardrail; if the aerial worker detection frame is spatially intersecting or spatially containing all guardrail detection frames, it is judged that the worker is in a safe working state.

[0056] by Figure 7 Taking the example of the above, the spatial relationship between the worker detection frame and the guardrail detection frame is specifically analyzed to determine whether it belongs to a safe working state. Figure 7 In the figure, workers 1, 2, 3, 4 and the guardrails are all in a spatial inclusion relationship. When workers are working at heights, the presence of guardrails near their edges can effectively prevent them from falling. This spatial relationship means that workers are protected by guardrails when they squat or bend over to work, which means they are in a safe working state. Figure 7 The middle guardrail and worker 6 are in a spatial intersection relationship, which means that the worker is protected by the guardrail when working upright, which is a safe working state; when there is no guardrail in front of the worker in the high-altitude working area, it is a spatial separation relationship, and the worker is in a dangerous working state.

[0057] After judging the image as a high-altitude work process, the worker's safety status is further judged. Among the 32 TP samples, there are a total of 267 worker instances, 144 of which are in a safe state and 123 are in a dangerous state. Figure 6 From the output confusion matrix, we can see that the model output results are 146 people in a safe state and 121 people in a dangerous state, among which TP is 140, FP is 6, and FN is 4. That is, the Precision and Recall of the model are 95.89% and 97.22% respectively, and the accuracy meets the engineering requirements. Figure 8 It is part of the output result of the model on the safety identification of workers working at heights. The green detection box indicates that the worker is in a safe state, and the red detection box indicates that the worker is in a dangerous edge state.

[0058] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0059] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0060] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0061] The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

[0062] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A method for identifying safety during high-altitude operations in a cave, characterized in that: The following steps are involved: Obtain tunnel face construction image data; Constructing a high-altitude working target recognition model in a tunnel and performing pre-training; the high-altitude working targets in the tunnel include trolleys, workers and guardrails; The trained in-tunnel aerial work target recognition model is used to obtain the in-tunnel aerial work target recognition type and location information based on the tunnel face construction image data; Determine the target spatial relationship based on the target identification type and position information of the high-altitude operation in the cave; Determine the high-altitude operation scene in the tunnel based on the spatial relationship between the trolley and the workers; Based on the high-altitude work scene in the tunnel, the high-altitude work protection safety status is determined according to the spatial relationship between the workers and the guardrails.

2. A method for identifying safety during aerial work protection in a cave according to claim 1, characterized in that: Pre-training of the indoor aerial work target recognition model includes: The CycleGAN network is used to enhance the tunnel face construction images; Annotate the enhanced tunnel face construction image with examples.

3. A method for identifying safety during aerial work protection in a cave according to claim 1, characterized in that: Using the CycleGAN network to enhance tunnel face construction images includes: The first discriminator judges the input low-illumination tunnel face construction image to determine whether it meets the characteristics of the image in the low-illumination tunnel dataset, and outputs a probability value indicating that the input image is a true low-illumination image; The input low-illuminance tunnel face construction image is input into the first generator to generate a normal-illuminance tunnel face construction image, and then the generated image is judged by the second discriminator to determine whether it meets the characteristics of the image in the normal-illuminance tunnel data set, and the probability value representing that the input image is a true normal-illuminance image is output; The tunnel face construction image with normal illumination generated by the first generator is input into the second generator, and the tunnel face construction image with low illumination is regenerated, so that the generated tunnel face construction image with low illumination is consistent with the input tunnel face construction image with low illumination, thereby forming an image conversion cycle; The adversarial loss, cycle consistency loss, and identity mapping loss in the image conversion process are calculated, and the optimizer is used to calculate and update the gradient value of the loss function and the parameter values ​​of the generator and discriminator, and the gradient descent method is used to iterate continuously until the overall loss function is minimized.

4. A method for identifying safety during aerial work in a cave according to claim 1, characterized in that: The target recognition model for high-altitude operations in tunnels includes: Backbone network, neck network and prediction network; The backbone network includes a first CBS module, a first C3 module, a second CBS module, a second C3 module and an SPPF module which are arranged in sequence, as well as a first attention module arranged at the output end of the second CBS module, a second attention module arranged at the output end of the second C3 module, and a third attention module arranged at the output end of the SPPF module.

5. The method for identifying safety during aerial work protection in a cave according to claim 1, characterized in that: The first attention module, the second attention module and the third attention module each include: A height global average pooling unit, a width global average pooling unit, a stacking unit, a first convolution unit, a normalization unit, a first activation unit, a split unit, a transposition unit, a second convolution unit, a third convolution unit, a second activation unit, a third activation unit, and a multiplication unit; The height global average pooling unit and the width global average pooling unit respectively perform global average pooling operations in the height direction and global average pooling operations in the width direction on the input feature map to obtain a first global average pooling feature map and a second global average pooling feature map; After the stacking unit performs a stacking operation on the first global average pooling feature map and the second global average pooling feature map, the stacking unit sequentially passes through a first convolution unit, a normalization unit, and a first activation unit to obtain a first spatial feature map; After the segmentation unit separates the width direction feature map and the height direction feature map from the first spatial feature map, the transposition unit performs transposition operations on the width direction feature map and the height direction feature map respectively, and then passes the two feature maps through the second convolution unit and the third convolution unit, and the second activation unit and the third activation unit respectively, to obtain the width direction attention and the height direction attention; The multiplication unit multiplies the input feature map with the width direction attention and the height direction attention to obtain an attention feature map.

6. A method for identifying safety during aerial work protection in a cave according to claim 1, characterized in that: Determining the target spatial relationship based on the target identification type and location information of the high-altitude operation in the cave includes: Determine the spatial inclusion, intersection or separation relationship between the worker and trolley target category boundary boxes according to the coordinate information of the worker detection box and the coordinate information of the trolley detection box; The spatial inclusion, intersection or separation relationship between the worker and the guardrail target category boundary box is determined according to the coordinate information of the worker detection box and the coordinate information of the guardrail detection box.

7. A method for identifying safety during aerial work protection in a cave according to claim 6, characterized in that: The spatial inclusion, intersection or separation relationship between the worker and trolley target category boundary boxes is determined based on the coordinate information of the worker detection box and the coordinate information of the trolley detection box, including: When the coordinate information of the worker detection frame A meets the coordinate information of the trolley detection frame B , then there is a spatial inclusion relationship between the worker and the trolley target category boundary box; When the coordinate information of the worker detection frame A meets the coordinate information of the trolley detection frame B , then there is a spatial intersection relationship between the worker and the trolley target category boundary box; When the coordinate information of the worker detection frame A meets the coordinate information of the trolley detection frame B , then the worker and the trolley target category boundary box are spatially separated.

8. A method for identifying safety during aerial work protection in a cave according to claim 6, characterized in that: Determining the spatial inclusion, intersection or separation relationship between the worker and the guardrail target category boundary box based on the coordinate information of the worker detection box and the coordinate information of the guardrail detection box includes: When the coordinate information of the worker detection frame A meets the coordinate information of the guardrail detection frame C , then there is a spatial inclusion relationship between the worker and the guardrail target category boundary box; When the coordinate information of the worker detection frame A meets the coordinate information of the guardrail detection frame C , then there is a spatial intersection relationship between the worker and the guardrail target category boundary box; When the coordinate information of the worker detection frame A meets the coordinate information of the guardrail detection frame C , then the worker and the guardrail target category boundary box are spatially separated.

9. A method for identifying safety during aerial work protection in a cave according to claim 1, characterized in that: According to the spatial relationship between the trolley and the workers, the aerial work scenes in the tunnel include: If the worker and the trolley target category bounding boxes are spatially contained or spatially intersected, the scene is judged to be an aerial work process in a cave; If only the trolley target is detected, only the worker target is detected, or both are detected and the trolley and all the worker target category boundary boxes are spatially separated, it is judged as a non-high-altitude operation process.

10. A method for identifying safety during aerial work in a cave according to claim 1, characterized in that: Based on the aerial work scene in the cave, the aerial work protection safety status is determined according to the spatial relationship between the worker and the guardrail, including: According to the shooting range of the surveillance camera and the size of the trolley, the area above the trolley detection frame is delineated as the high-altitude working area, and the spatial relationship between the worker and the guardrail is judged in this area. If the high-altitude working worker detection frame is spatially separated from a certain guardrail detection frame, it is judged that the worker is in a dangerous working state where the guardrail is missing; if the high-altitude working worker detection frame is spatially intersected or spatially contained with all guardrail detection frames, it is judged that the worker is in a safe working state.

Images