Method, device and equipment for detecting unsafe behaviors of capital construction personnel

Abstract

The invention discloses an unsafe behavior detection method, device and equipment for capital construction personnel, and relates to the field of image detection. The method comprises the following steps: forming a constructor state data set; the protective equipment of the constructor comprises a safety helmet and a reflective vest; a YOLOv8-E model is constructed; the YOLOv8-E model adopts a YOLOv8 model structure, uses an HGStem module and an HGBlock module as feature extraction modules, uses depth separable convolution to perform up-sampling, uses a VoVGSCSP module as an attention unit, and has four layers of feature interaction structures; training a YOLOv8-E model by utilizing the state data set of the construction personnel to obtain an unsafe state recognition model; and inputting the monitoring image of the construction personnel on the infrastructure construction site, which is acquired in real time, into the unsafe state identification model, and outputting an identification result of the wearing state of the protective equipment of the construction personnel. According to the invention, the detection precision of the unsafe states of multiple types of protection equipment of constructors can be improved.

Description

Technical Field

[0001] The present application relates to the field of image detection, and in particular to a method, device and equipment for detecting unsafe behaviors of infrastructure construction workers. Background Art

[0002] During construction and building, personnel safety protection inspection is a key link to ensure the safety of construction workers and reduce the occurrence of production safety accidents. With the development of modern construction industry, the complexity and scale of construction projects are increasing, and the safety risks at construction sites are becoming increasingly prominent. Although it is stipulated that all personnel involved in construction must wear protective equipment such as reflective clothing, safety helmets, and safety shoes during construction operations, in actual construction sites, there are still some construction workers who have weak safety awareness and fail to wear protective equipment as required. These phenomena not only increase the safety hazards at the construction site, but may also lead to serious production safety accidents, causing casualties and property losses.

[0003] Traditional construction worker safety status monitoring usually relies on supervisors to review real-time monitoring images for a long time to determine whether the construction workers are wearing protective equipment in accordance with regulations. However, the efficiency of manual detection is limited by work experience and subjective factors, resulting in inconsistent detection standards and difficult to ensure detection accuracy; at the same time, long-term work can cause visual fatigue in supervisors, leading to risks such as missed detection and false detection. In recent years, with the rapid development of artificial intelligence and computer vision technology, target detection methods based on deep learning theory have been widely used in industrial manufacturing, face recognition, and smart transportation due to their advantages of fast detection speed and high recognition accuracy. Target detection methods use a large number of widely used data sets to pre-train a recognition model. In the application, the pre-processed image to be detected is input into the recognition model with a set threshold to complete the recognition and classification of the target. In order to effectively improve the efficiency and intelligence of personnel safety status detection at infrastructure sites, researchers have explored the use of automatic target detection methods to replace manual detection based on the massive image data accumulated from infrastructure sites to improve the safety of construction sites. By designing and deploying a construction worker unsafe state detection model and using on-site construction monitoring images to complete the detection model training, it is possible to detect in real time whether the construction workers are wearing protective equipment correctly, such as whether they are wearing safety helmets and reflective clothing. When the model identifies that a construction worker is in an unsafe state, it will issue a warning in a timely manner to remind relevant managers to take necessary safety measures.

[0004] At present, target detection methods are mainly divided into three categories. The first category is the region-based convolutional neural network (R-CNN) series, the second category is the YOLO (You Only Look Once) series, and the third category is the Transformer-based target detection model DETR (Detection Transformer). Although the existing detection methods can identify the unsafe state of construction workers to a certain extent, considering that the size of construction workers in the monitoring images is small, the redundant background in the image greatly affects the recognition quality of the detection method; at the same time, most of the existing detection methods are only for the safety state of wearing a helmet, and there are problems such as low detection accuracy in multiple types of unsafe states. Summary of the invention

[0005] The purpose of this application is to provide a method, device and equipment for detecting unsafe behaviors of infrastructure construction workers, which can improve the detection accuracy of unsafe conditions of various types of protective equipment of construction workers.

[0006] To achieve the above objectives, this application provides the following solutions:

[0007] In the first aspect, the present application provides a method for detecting unsafe behaviors of infrastructure construction workers, including: obtaining monitoring images of construction workers at infrastructure construction sites collected from multiple perspectives; marking the wearing status of protective equipment of construction workers in each monitoring image, and forming a construction worker status data set together with the monitoring image; the protective equipment includes a safety helmet and a reflective vest; constructing a YOLOv8-E model; the YOLOv8-E model adopts the YOLOv8 model structure, uses the HGStem module and the HGBlock module as feature extraction modules, uses deep separable convolution for upsampling, uses the VoVGSCSP module as an attention unit, and has a four-layer feature interaction structure; uses the construction worker status data set to train the YOLOv8-E model to obtain an unsafe state recognition model; inputs the real-time monitoring images of construction workers at the infrastructure construction site into the unsafe state recognition model, and outputs the recognition results of the wearing status of the construction workers' protective equipment.

[0008] In the second aspect, the present application provides a device for detecting unsafe behaviors of infrastructure construction workers, including: a multi-view video monitoring device and a processing device. The multi-view video monitoring device is set at the infrastructure construction site; the multi-view video monitoring device is connected to the processing device; the multi-view video monitoring device is used to collect monitoring images of construction workers during the infrastructure construction process; the processing device is used to obtain the recognition result of the wearing status of the protective equipment of the construction workers based on the monitoring images and using the above-mentioned unsafe behaviors detection method for infrastructure construction workers.

[0009] In a third aspect, the present application provides a computer device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement any of the above-described methods for detecting unsafe behaviors of infrastructure construction workers.

[0010] According to the specific embodiments provided in this application, this application has the following technical effects:

[0011] The present application provides a method, device and equipment for detecting unsafe behaviors of infrastructure construction workers. The YOLOv8-E model uses the HGStem module and the HGBlock module as feature extraction modules. The HGStem module can retain and fuse more local feature information while reducing the size of the feature map through a combination of multi-layer convolution and pooling operations. The HGBlock adopts compression-excitation operations and more matrix merging operations to enhance the ability to capture important features. The deep separable convolution retains the effectiveness of feature extraction by decomposing the standard convolution into two steps: deep convolution and point-by-point convolution, and can realize multi-scale feature extraction. VoVGSCSP can effectively reduce redundant gradient flows and improve learning ability in the feature interaction process compared to the C2F unit by combining GSConv and CSPNet strategies. By introducing more feature interaction layers, more complex feature transfer and optimization are achieved, ensuring that the network can effectively fuse feature information of different scales, thereby improving the detection accuracy of unsafe states of multiple types of protective equipment for construction workers and reducing the risk of production safety accidents at infrastructure construction sites. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0013] Figure 1 This is a flow chart of a method for detecting unsafe behaviors of infrastructure construction workers in one embodiment of the present application;

[0014] Figure 2 A schematic diagram of an interface for labeling data using the Make Sense software provided in another embodiment of the present application;

[0015] Figure 3 A schematic diagram of the network structure of a YOLOv8-E model provided in another embodiment of the present application;

[0016] Figure 4A schematic diagram of the network structure of an HGStem module provided in another embodiment of the present application;

[0017] Figure 5 A schematic diagram of the network structure of the HGBlock module provided in another embodiment of the present application;

[0018] Figure 6 A schematic diagram of the GSConv network structure provided in another embodiment of the present application;

[0019] Figure 7 A schematic diagram of the GSBottleneck network structure provided for another embodiment of the present application;

[0020] Figure 8 A schematic diagram of the VoVGSCSP network structure provided in another embodiment of the present application;

[0021] Fig. 9 A schematic diagram of a working framework of a method for detecting unsafe behaviors of infrastructure construction workers provided in another embodiment of the present application;

[0022] Fig.10 A schematic diagram of a confusion matrix of a YOLOv8 model provided in another embodiment of the present application;

[0023] Fig.11 A schematic diagram of a confusion matrix of a YOLOv8-E model provided in another embodiment of the present application;

[0024] Fig.12 A schematic diagram of a PR curve of a YOLOv8 model provided in another embodiment of the present application;

[0025] Fig.13 A schematic diagram of a PR curve of a YOLOv8-E model provided in another embodiment of the present application;

[0026] Fig.14 A schematic diagram of an F-1 score curve of a YOLOv8 model provided in another embodiment of the present application;

[0027] Fig.15 A schematic diagram of an F-1 score curve of a YOLOv8-E model provided in another embodiment of the present application;

[0028] Fig.16 A schematic diagram of actual detection effect of a target recognition model provided by another embodiment of the present application;

[0029] Fig.17 A schematic diagram of the structure of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0030] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0031] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0032] At present, there are three main methods for detecting unsafe conditions of construction workers on construction sites. The first one is the Region-based Convolutional Neural Network (R-CNN) series, which realizes target detection by "extracting candidate regions first and then classifying". R-CNN is the earliest proposed target detection method based on deep learning. This method uses Selective Search to generate candidate regions, and then performs convolutional feature extraction and classification on each region. However, since each candidate region must be individually feature extracted, the computational efficiency is low. In order to improve computational efficiency, researchers designed Fast R-CNN based on R-CNN. Fast R-CNN reduces the time consumed by the feature extraction process by mapping candidate regions to shared convolutional feature maps; and by introducing the ROI pooling layer, feature extraction and classification are integrated in the same network, which reduces the model size while ensuring detection accuracy. Faster R-CNN is based on Fast R-CNN and adds Region Proposal Networks (RPN) to further optimize the detection process. RPN enables the generation of candidate regions to be completed through convolutional neural networks, reducing the complexity of target detection methods and becoming one of the mainstream methods in practical applications. The second type of detection method attempts to directly predict the bounding box and category label of the target from the image, skipping the step of generating candidate regions. The main model is the YOLO (You Only Look Once) series. The YOLO series model transforms the target detection problem into a regression problem and directly predicts the coordinates of the bounding box and the category probability. The YOLOv5 and YOLOv8 models currently used in the industry have made many improvements in network architecture and loss function design to improve the accuracy and speed of detection. In recent years, the success of the Transformer structure in the field of natural language processing has also been introduced into target detection, opening up a new technical route for target detection methods. The Transformer-based target detection model DETR (Detection Transformer) transforms target detection into a set prediction problem, trains the network in an end-to-end manner, and does not require candidate region generation or non-maximum suppression (NMS) steps. It uses the self-attention mechanism of the transformer and performs outstandingly in modeling long-distance dependencies. However, DETR takes a long time to train, and its performance in small target detection needs to be improved, so it has few practical applications.

[0033] The detection performance of the above existing models needs to be improved. In view of this, in order to accurately detect and identify the unsafe state of personnel during infrastructure construction, this application is based on the YOLOv8 model. By improving the feature extraction technology and feature interaction fusion algorithm, an automatic detection model YOLOv8-Enhanced (YOLOv8-E) for identifying the unsafe state of personnel during infrastructure construction is proposed. YOLOv8-E can more accurately identify and detect the unsafe state of construction personnel, thereby reducing the risk of production safety accidents at infrastructure construction sites.

[0034] In an exemplary embodiment, Figure 1 As shown, a method for detecting unsafe behaviors of infrastructure construction workers is provided, including the following steps 101 to 105. Among them:

[0035] Step 101: Acquire monitoring images of construction workers at a capital construction site collected from multiple perspectives.

[0036] Step 102: Annotate the wearing status of the protective equipment of the construction workers in each monitoring image, and together with the monitoring image, form a construction worker status dataset; the protective equipment includes a safety helmet and a reflective vest.

[0037] Step 103: construct a YOLOv8-E model; the YOLOv8-E model adopts the YOLOv8 model structure, uses the HGStem module and the HGBlock module as feature extraction modules, uses the depthwise separable convolution for upsampling, uses the VoVGSCSP module as an attention unit, and has a four-layer feature interaction structure.

[0038] Step 104: Use the construction worker status dataset to train the YOLOv8-E model to obtain an unsafe state recognition model.

[0039] Step 105: input the monitoring images of the construction workers at the infrastructure construction site collected in real time into the unsafe state recognition model, and output the recognition results of the wearing status of the construction workers' protective equipment.

[0040] In implementing the above steps 101 to 105, the YOLOv8-E model uses the HGStem module and the HGBlock module as feature extraction modules. The HGStem module can retain and fuse more local feature information while reducing the size of the feature map through a combination of multi-layer convolution and pooling operations. HGBlock uses compression-excitation operations and more matrix merging operations to enhance the ability to capture important features. The deep separable convolution retains the effectiveness of feature extraction by decomposing the standard convolution into two operations: deep convolution and point-by-point convolution, and can achieve multi-scale feature extraction. VoVGSCSP can effectively reduce redundant gradient flows and improve learning capabilities in the feature interaction process compared to the C2F unit by combining GSConv and CSPNet strategies. By introducing more feature interaction layers, more complex feature transfer and optimization are achieved, ensuring that the network can effectively fuse feature information of different scales, thereby improving the detection accuracy of unsafe states of multiple types of protective equipment for construction workers and reducing the risk of production safety accidents at infrastructure construction sites.

[0041] In another exemplary embodiment of the present application, multi-view video surveillance equipment installed at the infrastructure construction site is used to collect image data of the status of personnel during the infrastructure construction process. In order to improve the applicability and accuracy of the target detection model in the location scenario, it is necessary to expand the collected image data set. The expansion method includes random flipping, random translation, partial cutting and brightness adjustment of the image. In order to facilitate model processing, the processed image data is resized so that the size of all images is 512*512. Then the above step 101 can be replaced by the following steps 201~steps:

[0042] Step 201: Collect monitoring images of construction workers during the infrastructure construction process through multi-view video monitoring equipment set up at the infrastructure construction site.

[0043] Step 202: Expand the collected surveillance image; the expansion includes random flipping, random translation, partial cutting and brightness adjustment.

[0044] Step 203: Adjust the size of the expanded surveillance image to a uniform size.

[0045] In another exemplary embodiment of the present application, considering that hard hats and reflective vests are the main protective equipment to protect the safety of personnel during construction, the present application marks the status of the construction personnel's hard hats and reflective vests. The present application designs four protective equipment wearing states, including hard hat wearing, hard hat missing, reflective vest wearing, and reflective vest missing. In the method of the present application, the data uses Make Sense annotation software to complete the target anchoring and state calibration, such as Figure 2 shown.

[0046] In another exemplary embodiment of the present application, in order to improve the detection accuracy of unsafe construction conditions from a high-altitude perspective, the feature extraction network in YOLOv8 is redesigned using efficient feature extraction modules HGStem and HGBlock; and in order to deal with the image size transformation problem in the feature extraction process, a deep separable convolution is used to complete the image upsampling operation. In the feature interaction network part, YOLOv8-E expands the original three-layer feature interaction to a four-layer feature interaction structure, and uses a new efficient attention unit to replace the original C2F attention unit to achieve accurate detection of small targets in high-altitude monitoring images. YOLOv8-E retains the model structure of YOLOv8, including a feature extraction network, a feature interaction network, and a target detection network, and makes corresponding improvements to the problem of small target detection. The feature extraction network is used to extract the features of the monitoring image and obtain a feature map; the feature interaction network is used to interact and fuse the feature map to generate a four-layer interactive feature map; the target detection network is used to output the recognition result of the wearing status of the construction personnel's protective equipment according to the four-layer interactive feature map.

[0047] The network structure of YOLOv8-E is as follows Figure 3 shown. Figure 3 The numbers 0 to 9 in the figure constitute the feature extraction network, the numbers 10 to 28 constitute the feature interaction network, and Detect corresponds to the target detection network.

[0048] In another exemplary embodiment of the present application, the feature extraction network includes: a HGStem module, a first HGBlock module, a first depth-separable convolutional layer, a second HGBlock module, a second depth-separable convolutional layer, a third HGBlock module, a fourth HGBlock module, a fifth HGBlock module, a third depth-separable convolutional layer and a sixth HGBlock module connected in sequence.

[0049] The feature extraction network is introduced in detail below.

[0050] When identifying and detecting small objects in high-resolution images, traditional target detection models often face accuracy issues. To address this challenge, YOLOv8-E introduces the feature extraction modules HGStem and HGBlock in the RT-DETR (Real-Time Detection Transformer) method, and modifies the convolutional network to a deep separable convolution DWConv, redesigning the feature extraction network to improve the ability to detect the status of construction workers.

[0051] (1)HGStem module

[0052] In the method of the present application, the HGStem module is used as the initial feature extraction layer of the network to preprocess and reduce the dimension of the input image data, providing effective basic features for subsequent network layers. Compared with the standard convolution module used by YOLOv8, the HGStem module can retain and fuse more local feature information while reducing the size of the feature map through a combination of multi-layer convolution and pooling operations; in addition, the multi-branch design of the HGStem module allows multi-path information flow, which helps to capture features of different scales and complexities, rather than relying solely on a single convolution kernel. Such a design can bring better feature expression capabilities, allowing the model to extract more meaningful features when processing complex inputs. The network structure of the HGStem module is shown below. Figure 4 shown.

[0053] In HGStem, the input feature map size is H×W×C. It first passes through a convolution layer with a convolution kernel of 3×3 and a stride of 2 to reduce the spatial size of the feature map by half, and adjust the number of feature channels. The output feature map size is The Padding operation fills the right and bottom borders of the feature map with pixels, making the size of the feature map become Then, the convolution branch reuses two convolutional layers with a stride of 1 and a size of 2×2 to adjust the size of the feature map, and adds a boundary pixel padding operation between the two convolutional layers to obtain a size of The pooling branch uses only one maximum pooling operation (MaxPooling) to perform 2×2 maximum pooling to extract the local spatial maximum of the feature map, which helps to further centralize the features and highlight the local important information. The rounding up calculation method is used during the pooling operation, which ensures that the spatial size of the feature map remains unchanged after the pooling operation. The output results of the two branches will be spliced ​​in the channel dimension, and the result is The concatenated feature map is processed by a 3×3 convolution to integrate the features of the pooling path and the convolution path to enhance the expressiveness of the features. Finally, a 1×1 convolution is used to adjust the number of channels to provide appropriate feature distribution and channel configuration for subsequent modules.

[0054] (2)HGBlock module

[0055] The traditional YOLOv8 model uses C2F units to extract content from the feature maps of each layer. However, since the C2F unit uses conventional convolution operations, the computational complexity is relatively high. At the same time, since C2F only uses simple splicing operations to complete the fusion of multi-gradient flow features, this may be at a disadvantage in capturing targets of different scales in the feature map. The HGBlock module has obvious advantages in its overall structure. It uses lightweight convolution operations to improve the overall efficiency of the model. In order to enhance the ability to capture important features, HGBlock uses compression-excitation operations and more matrix merging operations. Therefore, in the feature extraction network of YOLOv8-E, HGBlock is used as the base layer for feature extraction. Its network structure is as follows: Figure 5 As shown in the figure, HGBlock contains multiple convolutional modules connected in series, and each layer performs feature extraction and fusion. (It is very fast to use this module).

[0056] First, the input feature map size is H×W×C. HGBlock constructs a series of N convolutional layers in a cyclic manner. These convolutional layers can be standard convolution (Conv) or light convolution (LightConv), depending on the parameter settings of each convolutional layer. For the first convolutional layer, the number of input channels is C and the number of output channels is C. m ; For each subsequent convolutional layer, the number of input and output channels is C m , and the convolution kernel size is k×k. These convolution modules do not change the spatial size of the feature map, so after passing through all n convolution modules, the output feature map size of each module remains H×W×C m During the forward propagation of the network, the initial input feature map x and the output features of each convolutional layer are stored in a list y. These features will be concatenated to form a new feature map with a size of H×W×(C+N×C m ). Then, the concatenated feature map passes through a 1×1 compressed convolution layer to reduce the number of channels from C+N×C m Compressed to C2 / 2. Subsequently, this feature map is subjected to an excitation convolution with a convolution kernel of 1×1 to further adjust the number of channels of the feature map to C2. Finally, when the number of input channels matches the set number of output channels, HGBlock will perform a residual operation, that is, directly adding the input feature map to the final output feature map to obtain the final output feature map H×W×C2. This residual connection helps to retain the original input features while enhancing the gradient flow ability of the network and improving the training effect of the model.

[0057] (3) Depthwise Separable Convolution DWConv

[0058] In order to achieve multi-scale feature extraction, YOLOv8-E uses deep separable convolution (DWConv) to adjust the spatial dimension of the input feature map. Deep separable convolution reduces the computational complexity of the convolution layer while retaining the effectiveness of feature extraction by decomposing the standard convolution into two steps (i.e., deep convolution and point-by-point convolution). In the deep convolution stage, each input channel is convolved with a separate convolution kernel, and each channel is independent of each other, without cross-channel convolution calculations. This method retains the spatial information of each channel in the input feature map, but does not model the correlation between channels. Next, by using point-by-point convolution (1*1 convolution), all the output channels of the deep convolution are linearly combined to achieve information interaction and fusion between channels.

[0059] In another exemplary embodiment of the present application, the feature interaction network includes: an SPPF module, a first Upsample module, a first Concat module, a first VoVGSCSP module, a second Upsample module, a second Concat module, a second VoVGSCSP module, a third Upsample module, a third Concat module, a third VoVGSCSP module, a first Conv layer, a fourth Concat module, a fourth VoVGSCSP module, a second Conv layer, a fifth Concat module, a fifth VoVGSCSP module, a third Conv layer, a sixth Concat module and a sixth VoVGSCSP module.

[0060] The SPPF module is connected to the sixth HGBlock module, the first Upsample module and the sixth VoVGSCSP module respectively. The first Upsample module is connected to the first Concat module; the first Concat module is connected to the fifth HGBlock module and the first VoVGSCSP module respectively; the first VoVGSCSP module is connected to the second Upsample module and the fifth Concat module respectively.

[0061] The second Upsample module is connected to the second Concat module, and the second Concat module is respectively connected to the second HGBlock module and the second VoVGSCSP module; the second VoVGSCSP module is respectively connected to the third Upsample module and the fourth Concat module.

[0062] The third Upsample module is connected to the third Concat module; the third Concat module is connected to the first HGBlock module and the third VoVGSCSP module respectively; the third VoVGSCSP module is connected to the first Conv layer, and the interactive feature map output by the third VoVGSCSP module is used as the first layer interactive feature map.

[0063] The first Conv layer is connected to the fourth Concat module; the fourth Concat module is connected to the fourth VoVGSCSP module; the fourth VoVGSCSP module is connected to the second Conv layer, and the interactive feature map output by the fourth VoVGSCSP module is used as the second layer interactive feature map.

[0064] The second Conv layer is connected to the fifth Concat module; the fifth Concat module is connected to the fifth VoVGSCSP module; the fifth VoVGSCSP module is connected to the third Conv layer, and the interactive feature map output by the fifth VoVGSCSP module is used as the interactive feature map of the third layer.

[0065] The third Conv layer is connected to the sixth Concat module; the sixth Concat module is connected to the sixth VoVGSCSP module, and the interactive feature map output by the sixth VoVGSCSP module is used as the fourth-layer interactive feature map.

[0066] The feature interaction network is introduced in detail below.

[0067] Although the improvement of the feature extraction network can improve the accuracy of the target detection task to a certain extent, there is still room for further improvement in the model performance. In order to achieve better detection performance in small target detection tasks such as high-altitude multi-angle construction personnel status recognition, the method of this application also improves and optimizes the target detection network in terms of feature interaction and fusion capabilities, adding a new feature fusion and output layer on the basis of the traditional YOLOv8 model's three-layer feature output; at the same time, a more efficient attention unit VoVGSCSP is used in the feature interaction part to replace the C2F unit in the baseline model.

[0068] (1) Network structure improvement

[0069] In the feature interaction network, the main difference between the improvement of the network structure by the method of this application and the baseline model is reflected in the depth and complexity of feature fusion and processing. Compared with the simple upsampling and concatenation operations of the baseline model, YOLOv8-E achieves more complex feature transfer and optimization by introducing more feature interaction layers and other intermediate steps, thereby better supporting detection tasks.

[0070] like Figure 3As shown in Figure 1, the network structure from the 10th to the 28th layer of the network demonstrates the process of multi-scale feature fusion and optimization. The SPPF module (Spatial Pyramid Pooling-Fast) in the 10th layer integrates contextual information of different scales through spatial pyramid pooling, and then upsamples in the 11th layer to improve the spatial resolution of the feature map. In the 12th layer, the feature map from the 7th layer is spliced ​​with the upsampled feature map to provide rich multi-scale features for subsequent processing. The VoVGSCSP module in the 13th layer effectively separates redundant gradient flows and reduces the amount of calculation through the combination of CSPNet and VoVNet, and further optimizes the features. Then, the feature map is upsampled again in the 14th layer and spliced ​​with the feature map of the 3rd layer in the 15th layer. The VoVGSCSP module in the 16th layer continues to optimize the features, passes them to the 17th layer for upsampling, and fuses them with the feature map of the 1st layer in the 18th layer. The VoVGSCSP module in the 19th layer deeply optimizes the fused features, and then further extracts features through convolution operations in the 20th layer. In the 21st layer, the feature maps from the 20th and 16th layers are concatenated and passed to the VoVGSCSP module in the 22nd layer for processing. The processing from the 23rd to the 28th layer repeats the feature concatenation, convolution, and optimization of the VoVGSCSP module. Finally, the 19th, 22nd, 25th, and 28th layers generate high-level feature maps for object detection. The whole process ensures that the network can effectively fuse feature information of different scales through multiple feature concatenation, upsampling, and optimization operations, thereby improving the detection accuracy and generalization ability of the model.

[0071] (2) Efficient Attention Unit VoVGSCSP

[0072] By combining GSConv and CSPNet strategies, VoVGSCSP can effectively reduce redundant gradient flows and improve learning ability during feature interaction. The network structure of the GSConv module is as follows: Figure 6 As shown in the figure, it mainly includes two convolution operations and a channel shuffle operation. GSConv first compresses the number of channels to half of the input through a 1*1 convolution layer (Conv1), reducing the complexity of subsequent calculations. Then, the second convolution layer (Conv2) uses a 5*5 convolution kernel to process the output of the Conv1 layer. During the processing, the number of channels of the feature map remains unchanged. Then, GSConv uses a splicing operation to splice the outputs of Conv2 and Conv1 in the channel dimension to form a new feature map. Finally, GSConv uses a reshape operation to divide the channels of the feature map into two groups, and then uses permute and reshape operations to achieve the redistribution of feature information and enhance the network's expression ability and information flow.

[0073] The GSBottleneck module used in VoVGSCSP is a bottleneck structure based on GSConv. Its network structure is as follows: Figure 7 As shown in the figure. This module contains two consecutive GSConv layers. The convolution kernel size of the first layer is 1*1. The number of channels of the input feature map is compressed without changing the size of the feature map, so as to reduce the amount of calculation and extract important features. The convolution kernel size of the second layer is 3*3, which is mainly used for feature learning and adjusts the number of channels to the target number of channels. In addition, the module also contains a residual connection operation constructed using a 1*1 convolution layer. The output of the residual connection operation is added to the output of the second GSConv layer to form a new feature representation.

[0074] The network structure of the VoVGSCSP module used in the method of the present application is as follows: Figure 8 As shown in the figure, the module uses two parallel 1*1 convolutional layers to perform channel dimension reduction. The output of branch 1 is then passed to a sequence of n GSBottleneck layers to complete efficient feature processing, and the number of channels remains unchanged during the processing; at the same time, branch 2 directly outputs a processed reduced dimension feature map. The outputs of the two branches are merged through a splicing operation, and the merged feature map is passed through a third 1*1 convolutional layer to adjust the number of channels and generate the final output. The VoVGSCSP module combines parallel paths with residual connections to achieve efficient feature extraction and fusion, improving the computational efficiency and feature expression capabilities of the model.

[0075] In another exemplary embodiment of the present application, the process of training the YOLOv8-E model in step 104 can be replaced by the following steps 301 to 302:

[0076] Step 301: Divide the construction worker status dataset into a model training set, an effect verification set, and an application test set in a ratio of 7:2:1.

[0077] Step 302: Use the model training set, effect verification set and application test set to train, verify and test the YOLOv8-E model respectively to obtain an unsafe state recognition model.

[0078] At the construction site, equipment such as drones and high-altitude cameras are used to complete the collection of training, verification and test data in the method of identifying unsafe conditions of personnel during infrastructure construction. The collected image format is 1920×1080. In order to balance the speed of model training and detection accuracy, the original image format will be compressed to 640*640 to train the YOLOv8-E model. This application uses Windows system and NVIDIA GeForce RTX 4090D graphics card to build a high-performance computing platform to improve the training speed of the model. During the training process, the batch size is 16, the number of training cycles is 1000, Adam is selected as the optimizer for this model training, and the initial learning rate is 0.001; in the loss gain setting, the box gain loss, category (cls) loss gain, and dfl loss gain are set to 7.5, 0.5, and 1.5 respectively. In addition, the early stopping mode of training is turned on during the training process, and the number of cycles waiting for no obvious improvement for early stopping is set to 50.

[0079] The working framework of the unsafe behavior detection method for infrastructure construction workers in this application is as follows Fig. 9 The unsafe state identification model can be applied to infrastructure construction sites to detect unsafe states of construction workers and provide warning information to on-site supervisors.

[0080] In order to more intuitively analyze and evaluate the detection performance of the YOLOv8-E model proposed in the method of this application for identifying unsafe conditions of construction workers during infrastructure construction, the method of this application uses confusion matrix, PR curve, F1 score curve, accuracy, recall rate and all-category average precision (mAP) and other indicators to evaluate the advancement and effectiveness of the method of this application compared with the baseline model from multiple aspects.

[0081] (1) Confusion Matrix Analysis

[0082] The confusion matrix is ​​a commonly used method in target detection and recognition tasks to evaluate the performance of classification models. It uses an n*n matrix (where n represents the number of categories) to show the correspondence between the predicted results (columns) and the actual labels (rows).

[0083] The confusion matrix of the YOLOv8 model (baseline model) reflects the performance of the model in the task of identifying unsafe conditions of construction workers, such as Fig.10As shown. From the prediction results of the Helmet category, the model correctly classified 0.81 samples as helmets, but 0.09 of the No Helmet samples were still misclassified as helmets, and 0.17 of the Background was also misclassified as a helmet. For the No Helmet category, the baseline model performed poorly, with only 0.49 of the No Helmet samples correctly classified, while 0.09 of the Helmet samples and 0.40 of the Background samples were misclassified as No Helmets. In the Vest category, the model correctly classified 0.85 of the samples as vests, but at the same time misclassified 0.02 of the No Vest samples and 0.12 of the Background as vests. On the No Vest category, the accuracy of the baseline model was 0.73, but 0.03 of the vest samples and 0.23 of the background samples were misclassified as No Vests. Overall, the baseline model performs poorly in distinguishing the background from the target object, especially in the no-helmet and no-vest categories, where more backgrounds are misclassified as these categories, indicating that the model has weak capabilities in background suppression and inter-category distinction.

[0084] The YOLOv8-E model proposed by the method of this application shows better performance in the same task. The confusion matrix of the YOLOv8-E model is as follows: Fig.11 As shown. From the prediction results of the helmet category, the model correctly classified 0.88 samples as helmets, reduced the proportion of no-helmet samples misclassified as helmets to 0.06, and the background false detection ratio was 0.10. In the no-helmet category, the correct classification rate of YOLOv8-E was 0.72, which was significantly better than the baseline model, and reduced the proportion of helmet samples misclassified as no-helmet to 0.01, and the background false detection ratio was also reduced to 0.21. In the vest category, the accuracy of YOLOv8-E increased to 0.92, and the no-vest samples and background samples misclassified as vests were 0.01 and 0.07 respectively, showing a significant reduction in the false detection rate in this category. The accuracy of the no-vest category increased to 0.79, the proportion of vest samples misclassified as no-vest was 0.01, and the background false detection ratio was 0.18. In general, the YOLOv8-E model proposed by the method of this application has a higher correct classification rate in multiple categories than the baseline model, especially in reducing background false detection and improving the detection accuracy of no helmet and no vest categories, indicating that the robustness of the model in target detection tasks has been significantly improved.

[0085] (2) PR curve analysis

[0086] The Precision-Recall (PR) curve is also one of the commonly used methods for evaluating state recognition models. It measures the performance of the model through two important indicators: precision and recall.

[0087] Precision refers to the proportion of positive samples predicted by the target detection model to the number of positive samples that are actually positive. The calculation formula is:

[0088]

[0089] Among them, TP (True Positive) indicates the number of samples that are actually positive and correctly predicted as positive; FP (False Positive), also known as the number of false positives, indicates the number of samples that are actually negative but are mistakenly predicted as positive. Accuracy measures how many of the model's predictions are correct.

[0090] Recall refers to the proportion of positive samples correctly identified by the model among all actual positive samples. The calculation formula is:

[0091]

[0092] Among them, FN (False Negative) is also called the number of missed detections, which indicates the number of samples that are actually positive samples but are mistakenly predicted as negative samples. The recall rate measures the recognition ability of the model in positive samples, that is, how many positive samples are successfully detected.

[0093] The PR curve of the baseline model is as follows Fig.12 As shown, the PR curve of the YOLOv8-E model proposed by the method of this application is as follows Fig.13 As shown. In the PR curve of the YOLOv8 model, its overall mAP@0.5 is 0.770, indicating that the model has a good average performance on all categories. However, the performance of each category varies greatly. In the helmet category, the mAP of YOLOv8 is 0.838, and the curve shows high precision and recall. The detection performance of the model in this category is relatively stable. For the nohelmet category, the mAP is 0.583, which is a low value, indicating that the model has a poor recall rate in the no-helmet category, and there are many missed detections or false detections. For the vest and no vest categories, the mAPs are 0.883 and 0.775, respectively, which shows that the YOLOv8 model performs relatively better in these two categories, showing high precision and recall, while the recall rate of the no-vest category is slightly insufficient.

[0094] In contrast, the PR curve of the YOLOv8-E model shows stronger overall detection performance, and its mAP@0.5 is improved to 0.852. The mAP of the helmet category is increased to 0.901, and both precision and recall are significantly improved. The curve is almost close to the coordinate axis, which proves the excellent performance of the model in detecting this type of target. The mAP of the no helmet category is also increased to 0.777, which shows that the YOLOv8-E model has successfully improved the detection ability of the no helmet category, the recall rate has been significantly improved, and the model performs better in missed detection and false detection. For the vest category, the mAP reaches 0.911, which is the highest value among all categories, indicating that the model performs particularly well in vest detection and almost reaches the optimal state. In the no vest category, the mAP is 0.820, which is an improvement over the YOLOv8 model, indicating that the detection robustness of the YOLOv8-E model in the no vest category is further enhanced.

[0095] (3) F-1 score curve analysis

[0096] The F1 score is the harmonic mean of precision and recall, which is used to balance the performance of the two in model detection. Especially when the difference between precision and recall is large, the F1 score provides a comprehensive indicator. The calculation formula of the F1 score is:

[0097]

[0098] The F1-Confidence curve can show the overall performance of the model at different confidence levels, and the highest point is often the best balance point of the model at that confidence level. Fig.14 is the F-1 score curve of the baseline model, Fig.15The F-1 score curve of the YOLOv8-E model proposed for the method of the present application. In the F-1 score curve of the YOLOv8 model, when the confidence level is 0.260, the overall highest F1 value is 0.74. From the perspective of different categories, the F1 value of the helmet category reaches a maximum of about 0.838, which shows that at this confidence level, the model maintains a good balance between the precision and recall rate of helmet targets, and the overall performance is relatively stable. In contrast, the F1 value of the no helmet category is only 0.583, indicating that the recall rate of the YOLOv8 model in no-helmet detection is low, and there may be more missed detections, resulting in a low F1 score. This is related to the model's ability to distinguish between positive and negative samples at different confidence levels, indicating that the no-helmet category is a category that is difficult to accurately capture in the model. The highest F1 value of the vest category reached 0.883, indicating that the model has high precision and recall when detecting vest targets, while the F1 value of the novest category was 0.775, slightly lower than the vest category, but still able to provide relatively reliable target detection performance. Overall, the YOLOv8 model performed relatively well in the detection of helmets and vests, but performed poorly in the detection of no helmets, reflecting the performance differences of the model in different categories.

[0099] The YOLOv8-E model reached a peak of the overall F1 value of about 0.84 at a confidence level of 0.462. In specific categories, the F1 value of the helmet category increased to 0.901, indicating that the precision and recall of the YOLOv8-E model in the helmet detection task have been significantly optimized, and the model can still effectively balance precision and recall at a higher confidence level. The F1 value of the nohelmet category increased to 0.777, which is significantly better than the YOLOv8 model, indicating that YOLOv8-E has improved the recall and precision in the no-helmet category, reduced missed detections and false detections, and improved the robustness of the model. For the vest category, the F1 value reached 0.911, which is the highest value of all categories in the two models, indicating that YOLOv8-E's performance in the vest detection task is almost the best. Finally, in the no vest category, the F1 value is improved to 0.820, which is an improvement over the YOLOv8 model, indicating that the model is more stable for detecting no vest targets.

[0100] In summary, the YOLOv8-E model performs better than the YOLOv8 model in all categories of F1 scores, especially in the detection of helmet and no-helmet categories, the model's precision and recall have been significantly improved.

[0101] (4) Quantitative indicator evaluation

[0102] In the method of the present application, precision (P), recall (R), and average precision (mAP) of all categories are selected as indicators for quantitative evaluation of the model designed by the method of the present application and other comparative models, as shown in Table 1. From the quantitative evaluation results in the table, it can be seen that different versions of the YOLOv8-E model have achieved significant improvements in multiple indicators compared with the basic YOLOv8 model. First, the YOLOv8-E-1 model improves its precision (Precision) from 0.78 to 0.88, the recall rate (Recall) from 0.71 to 0.73, and the mAP@0.5 and mAP@50-95 are also improved to 0.80 and 0.44 respectively by introducing HGStem and HGBlock to improve the feature extraction network on the basis of YOLOv8. This shows that by optimizing the feature extraction network, the model has significant improvements in detection accuracy and recall. The YOLOv8-E-2 model further improves the feature interaction network and adopts a four-layer feature interaction and fusion architecture to further improve the model performance. Although its accuracy is slightly reduced to 0.87, the recall rate increases to 0.79, mAP@0.5 increases to 0.84, and mAP@50-95 increases to 0.48, which shows the significant advantages of the improved feature interaction network in complex target detection. Finally, the YOLOv8-E model introduces the efficient attention unit VoVGSCSP to replace the C2f unit on the basis of YOLOv8-E-2, achieving the best performance. The accuracy of the model is further improved to 0.90, and the recall rate is maintained at 0.79. At the same time, mAP@0.5 reaches 0.85, and mAP@50-95 increases to 0.48, indicating that the improvement significantly improves the model's ability to accurately locate and identify targets. In the process of designing and verifying the construction worker unsafe state recognition model, the method of this application also uses a variety of other improved methods to optimize the benchmark model. YOLOv8-FasterNet and YOLOv8-GhostNet use FasterNet and GhostNet to optimize the structure of the feature extraction network respectively. EMA and CBAM are two attention mechanisms. However, in the final quantitative index evaluation, the best accuracy of the above improved method only reached 0.86, the recall rate remained at around 0.76, and the changes in the two indicators mAP@0.5 and mAP@50-95 were not significantly improved compared with the baseline model, which reflects that the performance improvement of the other models cannot reach the performance level of the improved model proposed by the method of this application, further verifying the advancement and effectiveness of the method proposed in this application.

[0103] Table 1 Quantitative evaluation of target recognition models

[0104]

[0105] In summary, by gradually improving feature extraction and interaction networks and combining efficient attention mechanisms, the YOLOv8-E series models surpass the original YOLOv8 models in all evaluation indicators, demonstrating their superiority in target recognition tasks.

[0106] (5) Actual image detection effect

[0107] The method of this application uses high-altitude multi-angle monitoring images to conduct actual tests on target recognition. The test results are as follows Fig.16 shown. Fig.16 The image in part (a) of the left column is the recognition result generated by the YOLOv8-E model. Fig.16 The image (b) in the right column is the recognition result generated by the benchmark model. The actual test results show that the YOLOv8-E method proposed by the present application has better detection performance than the benchmark model in the high-altitude multi-angle personnel status recognition task, which confirms the advancement and effectiveness of the present application's method.

[0108] Based on the same inventive concept, the embodiment of the present application also provides a device for detecting unsafe behavior of infrastructure construction personnel for implementing the above-mentioned method for detecting unsafe behavior of infrastructure construction personnel. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above-mentioned method, so the specific limitations in one or more embodiments of the device for detecting unsafe behavior of infrastructure construction personnel provided below can refer to the limitations of the method for detecting unsafe behavior of infrastructure construction personnel above, and will not be repeated here.

[0109] In an exemplary embodiment, a device for detecting unsafe behaviors of infrastructure construction workers is provided, including: a multi-view video monitoring device and a processing device. The multi-view video monitoring device is set at the infrastructure construction site; the multi-view video monitoring device is connected to the processing device; the multi-view video monitoring device is used to collect monitoring images of construction workers during the infrastructure construction process; the processing device is used to obtain the recognition result of the wearing status of the protective equipment of the construction workers based on the monitoring images and using the above-mentioned unsafe behaviors detection method for infrastructure construction workers.

[0110] As an optional implementation, the multi-view video surveillance equipment includes: a drone and a camera.

[0111] Wherein, in this implementation mode, the camera is a high-altitude camera.

[0112] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Fig.17As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the recognition results of the wearing status of the protective equipment of the construction personnel. The input / output interface of the computer device is used to exchange information between the processor and the external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a method for detecting unsafe behaviors of infrastructure construction personnel is implemented.

[0113] Those skilled in the art will understand that Fig.17 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components. In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0114] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0115] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A method for detecting unsafe behaviors of infrastructure construction workers, characterized in that: include: Obtain monitoring images of construction workers at infrastructure construction sites collected from multiple perspectives; The wearing status of the protective equipment of the construction workers in each monitoring image is annotated and together with the monitoring image, constitutes a construction worker status dataset; the protective equipment includes a safety helmet and a reflective vest; Construct a YOLOv8-E model; the YOLOv8-E model adopts the YOLOv8 model structure, uses the HGStem module and the HGBlock module as feature extraction modules, uses the depthwise separable convolution for upsampling, uses the VoVGSCSP module as an attention unit, and has a four-layer feature interaction structure; The YOLOv8-E model is trained using the construction worker status dataset to obtain an unsafe state recognition model; The monitoring images of construction workers at the infrastructure construction site collected in real time are input into the unsafe state recognition model, and the recognition results of the wearing status of the construction workers' protective equipment are output.

2. The method for detecting unsafe behaviors of infrastructure construction workers according to claim 1, characterized in that: Obtain monitoring images of construction workers at infrastructure construction sites from multiple perspectives, including: The monitoring images of construction workers during the infrastructure construction process are collected by multi-view video monitoring equipment set up at the infrastructure construction site; Expanding the collected surveillance images; the expansion includes random flipping, random translation, partial cutting and brightness adjustment; Adjust the size of the expanded surveillance image to a uniform size.

3. The method for detecting unsafe behaviors of infrastructure construction workers according to claim 1, characterized in that: The protective equipment wearing status includes: wearing a safety helmet, missing a safety helmet, wearing a reflective vest, and missing a reflective vest.

4. The method for detecting unsafe behaviors of infrastructure construction workers according to claim 1, characterized in that: The YOLOv8-E model includes: a feature extraction network, a feature interaction network and a target detection network connected in sequence; The feature extraction network uses the HGStem module and the HGBlock module as feature extraction modules, and uses depthwise separable convolution for upsampling; the feature extraction network is used to extract features of the monitoring image and obtain a feature map; The feature interaction network uses the VoVGSCSP module as an attention unit and has a four-layer feature interaction structure; the feature interaction network is used to interact and fuse feature maps to generate a four-layer interactive feature map; The target detection network is used to output the recognition result of the wearing status of the protective equipment of the construction workers according to the four-layer interactive feature map.

5. The method for detecting unsafe behaviors of infrastructure construction workers according to claim 4, characterized in that: The feature extraction network includes: a HGStem module, a first HGBlock module, a first depth-separable convolutional layer, a second HGBlock module, a second depth-separable convolutional layer, a third HGBlock module, a fourth HGBlock module, a fifth HGBlock module, a third depth-separable convolutional layer and a sixth HGBlock module connected in sequence.

6. The method for detecting unsafe behaviors of infrastructure construction workers according to claim 5, characterized in that: The feature interaction network includes: an SPPF module, a first Upsample module, a first Concat module, a first VoVGSCSP module, a second Upsample module, a second Concat module, a second VoVGSCSP module, a third Upsample module, a third Concat module, a third VoVGSCSP module, a first Conv layer, a fourth Concat module, a fourth VoVGSCSP module, a second Conv layer, a fifth Concat module, a fifth VoVGSCSP module, a third Conv layer, a sixth Concat module and a sixth VoVGSCSP module; The SPPF module is connected to the sixth HGBlock module, the first Upsample module and the sixth VoVGSCSP module respectively; The first Upsample module is connected to the first Concat module; the first Concat module is respectively connected to the fifth HGBlock module and the first VoVGSCSP module; the first VoVGSCSP module is respectively connected to the second Upsample module and the fifth Concat module; The second Upsample module is connected to the second Concat module, and the second Concat module is respectively connected to the second HGBlock module and the second VoVGSCSP module; the second VoVGSCSP module is respectively connected to the third Upsample module and the fourth Concat module; The third Upsample module is connected to the third Concat module; the third Concat module is connected to the first HGBlock module and the third VoVGSCSP module respectively; the third VoVGSCSP module is connected to the first Conv layer, and the interactive feature map output by the third VoVGSCSP module is used as the first layer interactive feature map; The first Conv layer is connected to the fourth Concat module; the fourth Concat module is connected to the fourth VoVGSCSP module; the fourth VoVGSCSP module is connected to the second Conv layer, and the interactive feature map output by the fourth VoVGSCSP module is used as the second layer interactive feature map; The second Conv layer is connected to the fifth Concat module; the fifth Concat module is connected to the fifth VoVGSCSP module; the fifth VoVGSCSP module is connected to the third Conv layer, and the interactive feature map output by the fifth VoVGSCSP module is used as the third layer interactive feature map; The third Conv layer is connected to the sixth Concat module; the sixth Concat module is connected to the sixth VoVGSCSP module, and the interactive feature map output by the sixth VoVGSCSP module is used as the fourth-layer interactive feature map.

7. The method for detecting unsafe behaviors of infrastructure construction workers according to claim 1, characterized in that: The YOLOv8-E model is trained using the construction worker status dataset to obtain an unsafe state recognition model, which specifically includes: The construction worker status dataset is divided into model training set, effect verification set and application test set according to the ratio of 7:2:1; The YOLOv8-E model is trained, verified and tested using a model training set, an effect verification set and an application test set to obtain an unsafe state recognition model.

8. A device for detecting unsafe behavior of infrastructure construction workers, characterized in that: The unsafe behavior detection device for infrastructure construction workers includes: multi-view video monitoring equipment and processing equipment; Multi-view video surveillance equipment is installed at the infrastructure construction site; The multi-view video surveillance equipment is connected to the processing equipment; Multi-view video surveillance equipment is used to collect monitoring images of construction workers during infrastructure construction; The processing device is used to obtain the recognition result of the wearing status of the construction personnel's protective equipment based on the monitoring image and using the unsafe behavior detection method for infrastructure construction personnel described in any one of claims 1 to 7.

9. The unsafe behavior detection device for infrastructure construction workers according to claim 8, characterized in that: The multi-view video surveillance equipment includes: a drone and a camera.

10. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the unsafe behavior detection method for infrastructure construction workers as described in any one of claims 1-7.

Images