AC-YOLOX identification model and safety helmet wearing identification method
By using the AC-YOLOX recognition model and multi-scale feature fusion and feature optimization techniques, the accuracy of safety helmet detection at construction sites has been improved, solving the problems of low detection accuracy and small target recognition in complex environments.
Patent Information
- Application Number
- CN202310058537.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-01-18
- Publication Date
- 2025-12-23
- Estimated Expiration
- 2043-01-18
AI Technical Summary
Existing safety helmet detection algorithms have low accuracy in complex and ever-changing construction site environments, especially for small targets and occluded targets. Furthermore, traditional methods are labor-intensive and costly.
The AC-YOLOX recognition model is adopted, and features at four scales are extracted through the DarkNet53 backbone and SPP layer architecture. Combined with FocusNet, CBS module, CBAM attention module and AC-SR super-resolution module, multi-scale feature fusion and feature optimization are performed to enhance adaptability to complex environments.
It improves the accuracy of helmet wearing detection, solves the problem of low detection accuracy in complex environments, and enables effective identification of small targets and occluded targets.
Smart Images
Figure CN116682135B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of image recognition, in particular to a safety helmet wearing recognition model, a recognition method and a recognition system. BACKGROUND
[0002] As a high labor intensity and high risk intensive industry, safety control is an issue that people need to pay attention to at all times. In order to avoid injuries to construction workers, construction sites require workers to use safety protection equipment according to safety regulations. However, due to the weak safety protection awareness of construction workers, they often ignore the importance of wearing safety helmets. Therefore, construction sites often use surveillance cameras to monitor workers and ensure that they wear safety helmets. This video monitoring method requires workers to concentrate for a long time and observe multiple cameras capturing different positions in the construction site. This method not only consumes cost, manpower and energy, but also the observer cannot monitor the safety protection situation at all times and easily overlooks critical safety violations. Therefore, it is of great significance to use target detection technology to realize safety supervision of whether workers wear safety helmets in real scenes.
[0003] In recent years, researchers have tried to automatically detect safety helmets in surveillance system or drone videos. In 2014, Liu Xiaohui et al. used skin color detection to locate the face area, and determined whether the worker wore a safety helmet according to the color difference between the safety helmet and the hair. In 2015, Shrestha et al. first detected the face, then converted it to a grayscale image, and used an edge recognition algorithm to determine whether there was a semicircular profile of a safety helmet above the face. In 2017, Jie Li et al. used HOG features and SVM to classify pedestrians, and then used color feature recognition to realize safety helmet detection. However, these traditional target detection methods are greatly affected by the environment. When the target environment is affected by external light intensity, noise and interference color, etc., the detection speed of the target is slow and the detection accuracy is low. In 2019, X Long et al. used a deep learning algorithm SSD to detect safety helmet wearing. In 2020, Shanshan Huang et al. proposed a computer vision safety helmet automatic detection method based on YOLOv3 algorithm, which introduced a new attention mechanism to enhance the convolutional features.
[0004] Most existing safety helmet wearing detection algorithms are affected by the actual complex and variable scene, such as noise, uneven illumination, target occlusion, etc., resulting in low detection accuracy. In addition, there are often small targets and false targets similar to safety helmets in construction site images, such as Figure 1 As shown in the figure, the target in the uppermost red box contains less feature information and has low resolution, Figure 2 There is a white fisherman's hat similar to a safety helmet (red box) in the middle, which will affect the accuracy of detection.
[0005] Therefore, in view of the problems of interference information of actual images and difficulty in detecting small targets in safety helmets, a new safety helmet recognition method needs to be proposed. SUMMARY
[0006] The present application aims at the deficiencies of the prior art, and provides a safety helmet wearing recognition model, a recognition method and a recognition system.
[0007] To achieve the above-mentioned purpose, the technical scheme adopted by the present application is:
[0008] The present application provides an AC-YOLOX recognition model in the first aspect, comprising:
[0009] The Backbone module is used for feature extraction of the input image into four scale features.
[0010] The architecture of DarkNet53 backbone and SPP layer is adopted, the input image is scaled first, then three scale image features are obtained through FocusNet, and four scale feature maps are obtained through residual convolution and SPP pooling operation.
[0011] The Focus module, two CBS modules, the Csplayer module, one CBS module, one Csplayer module, one CBS module, one Csplayer module, one CBS module, SPP pooling operation and one Csplayer module are sequentially connected.
[0012] The features are output in stages C3, C4, C5 and C6.
[0013] The neck module is used for multi-scale feature fusion of the features extracted by the Backbone module.
[0014] The neck module is used for multi-scale feature fusion of the features extracted by the Backbone module.
[0015] The first Conv convolution module, the first UpSampling2D upsampling module, the first CBAM attention module, the first Concat+Csplayer module, the second Conv convolution module, the second UpSampling2D upsampling module, the second Concat+Csplayer module, the third Conv convolution module, the AC-SR super-resolution module and the third Concat+Csplayer module are sequentially connected.
[0016] The output of the third Concat+Csplayer module is connected to the input of the first Downsample module;
[0017] The first Downsample module, the fourth Concat+Csplayer module, the sixth CBAM attention module, the second Downsample module, the fifth Concat+Csplayer module, the seventh CBAM attention module, the third Downsample module and the sixth Concat+Csplayer module are sequentially connected;
[0018] The features output by the Backbone module at stages C3, C4, C5 and C6 correspond to the inputs of the second, third, fourth and fifth CBAM attention modules, respectively; the outputs of the second, third and fourth CBAM attention modules correspond to the inputs of the first Conv module, the first Concat+Csplayer module, the second Concat+Csplayer module and the third Concat+Csplayer module, respectively;
[0019] The first Conv module is connected to the sixth Concat+Csplayer module, the second Conv module is connected to the fifth Concat+Csplayer module, and the second Concat+Csplayer module is connected to the fourth Concat+Csplayer module;
[0020] The third Concat+Csplayer module, the fourth Concat+Csplayer module, the fifth Concat+Csplayer module and the sixth Concat+Csplayer module are connected to each other as the output of the neck module, and correspond to the output of four scale feature layers P2, P3, P4 and P5;
[0021] The head module includes an FNP feature classification module and a regression module.
[0022] The FNP feature classification module performs decoupling operations on the four scale feature layers P2, P3, P4 and P5 output by the neck module, and decomposes them into three branch operations of judging the category, judging the foreground and background, and positioning the target;
[0023] The regression module maps the judging category branch operation and the judging foreground and background branch operation to the number of target categories and the number of target categories, respectively, through convolution, and then converts them into prediction probabilities through Sigmoid operation;
[0024] The positioning target branch operation is also mapped to the required regression center point horizontal coordinate, vertical coordinate, width and height through convolution.
[0025] The second aspect of the present application provides a safety helmet wearing recognition model, which is trained by the following steps:
[0026] The safety helmet video data collection and shooting step: shooting the safety helmet wearing video data of the site staff and intercepting picture data at different frame numbers;
[0027] The data set labeling step: using the labelme tool to label the picture data, dividing it into two categories of hat and person, and converting the label data into a VOC format data set;
[0028] The data set expansion and processing step: using the Mosaic and / or Mixup data enhancement method to expand the data set, wherein the Mosaic method is to randomly splice four images and then scale to the same input size, and the Mixup method is to simulate the effect of occluded targets by superimposing different images;
[0029] The model construction step: constructing the recognition network according to the AC-YOLOX recognition model;
[0030] The model file training and testing step: inputting the data set into the AC-YOLOX recognition model for training and testing, refining it into a training set, a test set and a validation set according to a 7:1:2 ratio, and finally obtaining the model file of the AC-YOLOX recognition model.
[0031] The third aspect of the present application provides a non-transitory computer readable storage medium, which stores a computer program, and the program is executed by a processor to realize the safety helmet wearing recognition model.
[0032] The fourth aspect of the present application provides a safety helmet wearing recognition method, which comprises:
[0033] The model deployment step: deploying the model file of the safety helmet wearing recognition model to a platform server using tensorflow-serving;
[0034] The recognition step: inputting real-time monitoring video images into the platform server for safety helmet wearing recognition.
[0035] The fifth aspect of the present application provides a safety helmet wearing recognition system, which comprises:
[0036] A monitoring camera for collecting monitoring video images to be recognized;
[0037] A platform server in communication connection with the monitoring camera, used for executing the model deployment step and the recognition step in the safety helmet wearing recognition method.
[0038] The present application has outstanding substantial features and significant progress compared with the prior art, specifically:
[0039] 1、 the AC-YOLOX recognition model of the present application increases the processing of four scale features from the original YOLOX model processing of three scale features;
[0040] 2、 the design of the AC-SR super-resolution module in the AC-YOLOX recognition model of the present application introduces super-resolution technology into the target detection model, optimizes the extracted picture features in the feature upsampling process, extracts the foreground features in the image as much as possible while fusing the multi-scale features in the model, and adds a CBAM attention module to emphasize the meaningful features in the spatial and channel dimensions, realize the model's adaptive focusing on different areas, pay attention to important features and suppress unnecessary features; solve the problem of weak representation caused by insufficient detection features, thereby improve the problem of low detection accuracy and missed detection of small targets in safety hat wearing detection.
[0041] 3、 the AC-YOLOX recognition model of the present application is used for construction site safety hat wearing detection, solves the problem of low detection accuracy caused by the influence of actual complex and changeable scenes, such as noise, uneven illumination, and target occlusion;
[0042] 4、 the present application is applied to the construction site AI automatic identification platform to monitor the safety hat wearing situation of employees in the actual construction site in real time. BRIEF DESCRIPTION OF DRAWINGS
[0043] Figure 1 is a data set example 1 mentioned in the background art.
[0044] Figure 2 is a data set example 2 mentioned in the background art.
[0045] Figure 3 is the overall framework diagram of the AC-YOLOX recognition model of the present application.
[0046] Figure 4 is the framework diagram of the AC-SR super-resolution module of the AC-YOLOX recognition model of the present application.
[0047] Figure 5 is a data set example formed in the step of collecting and shooting safety hat video data in embodiment 2 of the present application.
[0048] Figure 6 is the test result when the model is tested in embodiment 2 of the present application.
[0049] Figure 7 is a flowchart of embodiment 4 of the present application. DETAILED DESCRIPTION
[0050] The technical solutions of the present application are described in further detail below through specific embodiments.
[0051] Embodiment 1
[0052] As shown in Figure 3 and Figure 4 , the present embodiment provides an AC-YOLOX recognition model, comprising:
[0053] Backbone module
[0054] for feature extraction of the input image into four scales of features;
[0055] The architecture of DarkNet53 backbone and SPP layer is adopted, first the input image is scaled, then three scale image features are obtained through FocusNet, and then four scale feature maps are obtained through residual convolution and SPP pooling operation;
[0056] comprising a Focus module, two CBS modules, a Csplayer module, a CBS module, a Csplayer module, a CBS module, a Csplayer module, a CBS module, an SPP pooling operation and a Csplayer module connected in sequence;
[0057] output features at stages C3, C4, C5 and C6;
[0058] The Backbone module first scales the input image to 640x640x3 size, obtains 320x320x3 scale image features through FocusNet, which is used to reduce the width and height of the input image and stack information into channels. Then, residual convolution and SPP pooling operation are used to obtain four scale (160x160x128, 80x80x256, 40x40x512, 20x20x1024) feature maps. Compared with the original YOLOX, AC-YOLOX outputs 160x160x128 scale feature layer C6 after the first layer CspLayer, which increases from three scale features to four scale features, providing multi-scale feature information for subsequent neck module feature fusion.
[0059] module
[0060] for multi-scale feature fusion of the features extracted by the Backbone module;
[0061] The backbone module comprises three Conv convolution modules, two UpSampling2D up-sampling modules, six Concat+Csplayer modules, three Downsample down-sampling modules, seven CBAM attention modules and one AC-SR super-resolution module;
[0062] The first Conv convolution module, the first UpSampling2D up-sampling module, the first CBAM attention module, the first Concat+Csplayer module, the second Conv convolution module, the second UpSampling2D up-sampling module, the second Concat+Csplayer module, the third Conv convolution module, the AC-SR super-resolution module and the third Concat+Csplayer module are sequentially connected;
[0063] The output of the third Concat+Csplayer module is connected to the input of the first Downsample down-sampling module;
[0064] The first Downsample down-sampling module, the fourth Concat+Csplayer module, the sixth CBAM attention module, the second Downsample down-sampling module, the fifth Concat+Csplayer module, the seventh CBAM attention module, the third Downsample down-sampling module and the sixth Concat+Csplayer module are sequentially connected;
[0065] The features output by the Backbone module at stages C3, C4, C5 and C6 correspond to the inputs of the second, third, fourth and fifth CBAM attention modules, respectively; the outputs of the second, third and fourth CBAM attention modules correspond to the inputs of the first Conv convolution module, the first Concat+Csplayer module, the second Concat+Csplayer module and the third Concat+Csplayer module, respectively;
[0066] The first Conv convolution module is connected with the sixth Concat+Csplayer module, the second Conv convolution module is connected with the fifth Concat+Csplayer module, and the second Concat+Csplayer module is connected with the fourth Concat+Csplayer module;
[0067] The third Concat+Csplayer module, the fourth Concat+Csplayer module, the fifth Concat+Csplayer module and the sixth Concat+Csplayer module are connected with each other as the output of the neck module, and correspondingly output four scale feature layers P2, P3, P4 and P5;
[0068] The AC-SR super-resolution module comprises an ACmix module, a Conv convolution and up-sampling module.
[0069] Module
[0070] The FNP feature classification module and the regression module are included.
[0071] The FNP feature classification module performs decoupling operations on the P2, P3, P4 and P5 scale feature layers output by the neck module respectively, and decomposes them into three branch operations of judging the category, judging the foreground and background and positioning the target.
[0072] The regression module respectively maps the judging category branch operation and the judging foreground and background branch operation to the target category number and the target category number through convolution, and then converts them into prediction probabilities through Sigmoid operation.
[0073] The positioning target branch operation is also mapped to the required regression center point horizontal coordinate, vertical coordinate, width and height through convolution.
[0074] The head module is used for classification and regression. First, decoupling operations are performed on the four different scale features output by the neck module, wherein each decoupling operation is first compressed through a 1x1 channel convolution to reduce the number of channels and the amount of calculation, and then decomposed into three branch operations of judging the category, judging the foreground and background and positioning the target. Among them, judging the category and judging the foreground belong to the classification task, which is mapped to the target category number and the target category number through two 3x3 and one 1x1 convolution, and then converted into prediction probability through Sigmoid operation; positioning the target belongs to the regression operation, which is also mapped to the required regression 4 coordinate values (center point horizontal coordinate, vertical coordinate, width and height) through two 3x3 and one 1x1 convolution.
[0075] Embodiment 2
[0076] The safety helmet wearing recognition model provided in this embodiment is trained by the following steps:
[0077] The safety helmet video data collection and shooting step: shooting the safety helmet wearing video data of the construction site staff and intercepting the picture data under different frame numbers; shooting the pictures of the real personnel imitating the safety helmet wearing situation in the construction site, respectively collecting pictures of different scales, different angles and different light intensities, a total of 5411 pictures, and the data set example is as follows Figure 5 as shown;
[0078] The data set labeling step: using the labelme tool to label the picture data, dividing it into two categories of hat and person, and converting the label data into a VOC format data set;
[0079] The data set is expanded and processed by using data augmentation: the data set is expanded by using Mosaic and / or Mixup data augmentation methods, wherein the Mosaic method is to rescale to the same input size by random splicing of four images, and the Mixup method is to simulate the effect of occluded targets by superimposing different images;
[0080] The model is constructed: the AC-YOLOX recognition model is constructed according to the method described in Embodiment 1.
[0081] The model file is obtained by training the recognition network: the data set is input into the AC-YOLOX recognition model for training and testing, and the data set is refined into a training set, a test set and a validation set according to a ratio of 7:1:2, a random gradient descent optimizer and a cosine annealing decay learning mechanism are used to adjust the learning rate, the initial learning rate is 0.01, the momentum of SGD is set to 0.9, the weight decay is 0.0005, and the Batchsize size is set to 16; when testing, the Batchsize size is set to 64. Finally, the model file of the AC-YOLOX recognition model is obtained. The test results are shown in Figure 6 .
[0082] Embodiment 3
[0083] The embodiment provides a non-transitory computer-readable storage medium, which stores a computer program, and the program is executed by a processor to implement the safety helmet wearing recognition model in Embodiment 2.
[0084] Embodiment 4
[0085] As Figure 7 shown, the embodiment provides a safety helmet wearing recognition method, which comprises:
[0086] The model deployment step: the model file of the safety helmet wearing recognition model in Embodiment 2 is deployed to a platform server by using tensorflow-serving;
[0087] The recognition step: real-time monitoring video images are input into the platform server for safety helmet wearing recognition.
[0088] Embodiment 5
[0089] The embodiment provides a safety helmet wearing recognition system, which comprises:
[0090] A monitoring camera is used to collect monitoring video images to be recognized;
[0091] A platform server is in communication connection with the monitoring camera, and is used to execute the model deployment step and the recognition step in the safety helmet wearing recognition method in Embodiment 4.
[0092] Those skilled in the art will appreciate that embodiments of the application can be devised for a method, a system, or a computer program product. Accordingly, the present application can be embodied in the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer non-transitory readable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer program code thereon for use by or in connection with an instruction execution system. Program code embodied on a computer readable medium can be transmitted to a transitory computer, such as a personal computer or a computer in a network, via a computer network or over a telephone line, using a modem.
[0093] The present application is described in reference to the flowchart and / or block diagrams of the method, apparatus (system) and computer program product according to embodiments of the application. It will be understood that each block of the flowchart and / or block diagrams, and combinations of blocks in the flowchart and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processing device or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.
[0094] These computer program instructions can also be stored in a computer readable medium that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable medium produce an article of manufacture including instructions which implement the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.
[0095] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.
[0096] Finally, it should be noted that the above-mentioned embodiments are only used to illustrate the technical solutions of the present application, rather than limit the technical solutions of the present application; although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the specific embodiments of the present application can be modified, or some of the technical features can be replaced by equivalent; without departing from the spirit of the technical solutions of the present application, they should be covered in the technical solutions of the present application.
Claims
1. A hard hat wear detection system, comprising: The system adopts an AC-YOLOX recognition model to recognize safety helmet wearing, and the AC-YOLOX recognition model comprises: a Backbone module for extracting features of an input image into four scales; an architecture adopting a DarkNet53 backbone and an SPP layer, which first scales the input image, then obtains three-scale image features through a Focus Net, and then obtains four-scale feature maps through residual convolution and SPP pooling operation; comprises a Focus module, a first CBS module, a second CBS module, a first Csplayer module, a third CBS module, a second Csplayer module, a fourth CBS module, a third Csplayer module, a fifth CBS module, an SPP pooling operation and a fourth Csplayer module connected in sequence; features at a stage C3 output by the fourth Csplayer module, features at a stage C4 output by the third Csplayer module, features at a stage C5 output by the second Csplayer module, and features at a stage C6 output by the first Csplayer module; a neck module for performing multi-scale feature fusion on the features extracted by the Backbone module; comprises three Conv convolution modules, two UpSampling2D up-sampling modules, six Concat+Csplayer modules, three Downsample down-sampling modules, seven CBAM attention modules and an AC-SR super-resolution module; the first Conv convolution module, the first UpSampling2D up-sampling module, the first CBAM attention module, the first Concat+Csplayer module, the second Conv convolution module, the second UpSampling2D up-sampling module, the second Concat+Csplayer module, the third Conv convolution module, the AC-SR super-resolution module and the third Concat+Csplayer module are connected in sequence; an output of the third Concat+Csplayer module is connected to an input of the first Downsample down-sampling module; the first Downsample down-sampling module, the fourth Concat+Csplayer module, the sixth CBAM attention module, the second Downsample down-sampling module, the fifth Concat+Csplayer module, the seventh CBAM attention module, the third Downsample down-sampling module and the sixth Concat+Csplayer module are connected in sequence; features output by the Backbone module at stages C3, C4, C5 and C6 correspond to the second, third, fourth and fifth CBAM attention modules respectively; and outputs of the second, third, fourth and fifth CBAM attention modules correspond to the first Conv convolution module, the first Concat+Csplayer module, the second Concat+Csplayer module and the third Concat+Csplayer module respectively. The first Conv convolution module is connected with the sixth Concat+Csplayer module, the second Conv convolution module is connected with the fifth Concat+Csplayer module, and the second Concat+Csplayer module is connected with the fourth Concat+Csplayer module; The third Concat+Csplayer module, the fourth Concat+Csplayer module, the fifth Concat+Csplayer module and the sixth Concat+Csplayer module are connected with each other as the output of the neck module, and correspondingly output four scale feature layers P2, P3, P4 and P5; The AC-SR super-resolution module comprises an ACmix module, a Conv convolution and an up-sampling module connected in sequence; The head module comprises an FNP feature classification module and a regression module; The FNP feature classification module performs decoupling operations on the four scale feature layers P2, P3, P4 and P5 output by the neck module, and decomposes them into three branch operations of judging the category, judging the foreground and background and positioning the target; The regression module respectively maps the judging category branch operation and the judging foreground and background branch operation to the target category number and the target category number through convolution, and then converts them into prediction probabilities through Sigmoid operation; The positioning target branch operation is also mapped to the required regression center point horizontal coordinate, vertical coordinate, width and height through convolution.
2. The safety hat wearing recognition system of claim 1, wherein: The AC-YOLOX recognition model is obtained by using the following steps for training: The safety helmet video data collection and shooting step: shooting the safety helmet wearing video data of the construction site staff and intercepting picture data at different frame numbers; The data set labeling step: using the labelme tool to label the picture data, dividing it into two categories of hat and person, and converting the label data into a VOC format data set; The data set expansion and processing step: using the Mosaic and / or Mixup data enhancement method to expand the data set, wherein the Mosaic method is to randomly splice four images and then scale to the same input size, and the Mixup method is to simulate the effect of occluded targets by superimposing different images; The model construction step: constructing a safety helmet wearing recognition network according to the AC-YOLOX recognition model; The model file training and testing step: inputting the data set into the AC-YOLOX recognition model for training and testing, refining the training set, test set and validation set in a ratio of 7:1:2, and finally obtaining the AC-YOLOX recognition model.
3. The safety hat wearing recognition system of claim 2, wherein: When the data set is input into the AC-YOLOX recognition model for training and testing, the random gradient descent optimizer and the cosine annealing decay learning mechanism are used to adjust the learning rate, the initial learning rate is 0.01, the momentum of SGD is set to 0.9, the weight decay is 0.0005, and the Batchsize size is set to 16; when testing, the Batchsize size is set to 64.
4. A non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the safety helmet wearing identification system of any one of claims 1-3.
5. A safety hat wearing recognition method characterized by comprising: Comprising: a model deployment step of deploying the model of the safety helmet wearing identification system of any one of claims 1-3 to a platform server using tensorflow-serving; an identification step of inputting real-time monitoring video images to the platform server for safety helmet wearing identification.
6. A safety helmet wearing identification system, comprising: a monitoring camera configured to collect monitoring video images to be identified; a platform server in communication connection with the monitoring camera and configured to execute the model deployment step and the identification step in the safety helmet wearing identification method of claim 5.