Night edge operation dangerous behavior early warning system, identification method and identification device

Abstract

The invention provides a night limb operation dangerous behavior early warning system, a recognition method and a recognition device, and the early warning system comprises a camera which is used for obtaining an image of a night limb operation scene to be recognized and a daytime image of the same limb operation scene; an alarm which is installed on the to-be-identified night limb operation scene and is used for sending out a reminding signal; and the control system is in signal connection with the camera and the alarm and is used for inputting a trained dangerous behavior recognition model according to the image of the night limb operation scene to be recognized and the daytime image of the same limb operation scene to obtain a dangerous behavior recognition result, and sending the dangerous behavior recognition result to the camera. The alarm is controlled to send out a reminding signal according to the dangerous behavior recognition result; the dangerous behavior comprises leaning on a near-edge protective rail; the method can be used for carrying out dangerous action recognition and timely reminding on the near-edge operation at night.

Description

Technical Field

[0001] The invention relates to a nighttime edge operation dangerous behavior early warning system, an identification method and an identification device, and belongs to the technical field of construction management. Background Art

[0002] At present, in the construction industry, in order to prevent the risk of workers falling from heights, measures such as edge guardrails and safety ropes need to be set up at construction sites. However, many workers lean directly on the edge guardrails for convenience, which is very dangerous. Many edge guardrails are not as strong as they look, which can easily cause safety accidents. Supervisor inspections cannot achieve full coverage and timely reminders.

[0003] Most construction sites are now equipped with real-time cameras to achieve timely on-site monitoring. In the prior art, there are methods for intelligently identifying dangerous actions through on-site scene photos.

[0004] Chinese patent CN112396017B discloses a method and system for identifying engineering safety hazards based on image recognition. The method includes: obtaining video stream data of construction site images based on fixed cameras and drones; performing identification based on the pre-processed construction site images and the image detection and recognition model to obtain construction site safety hazard results, which can be used to improve the production efficiency of supervisors and reduce costs.

[0005] However, existing technologies are unlikely to solve the problem of nighttime edge construction. Due to insufficient light sources in the nighttime environment, image recognition is relatively difficult, and supervisors and workers are prone to drowsiness, making it difficult to identify and promptly remind dangerous actions (mainly those that rely on edge protection facilities). Summary of the invention

[0006] The purpose of the present invention is to overcome the deficiencies in the prior art and to provide a nighttime dangerous behavior warning system, identification method and identification device for edge work, which are used to identify dangerous actions and provide timely reminders for edge work at night.

[0007] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0008] In a first aspect, the present invention provides a nighttime edge operation dangerous behavior warning system, comprising:

[0009] A camera, used to obtain an image of a nighttime edge operation scene to be identified and a daytime image of the same edge operation scene;

[0010] An alarm is installed at the scene of the nighttime edge operation to be identified, and is used to send out a warning signal;

[0011] A control system is connected to the camera and the alarm signal respectively, and is used to input the trained dangerous behavior recognition model according to the image of the nighttime edge working scene to be identified and the daytime image of the same edge working scene, obtain the dangerous behavior recognition result, and control the alarm to send a reminder signal according to the dangerous behavior recognition result; the dangerous behavior includes leaning against the edge protection railing;

[0012] The dangerous behavior recognition model includes a preprocessing module, a day and night data fusion module and a danger detection module;

[0013] The preprocessing module is used to preprocess the image of the nighttime edge working scene and the daytime image of the same edge working scene;

[0014] The day and night data fusion module is used to perform self-attention fusion on the image of the nighttime edge working scene to be identified and the daytime image of the same edge working scene to obtain a fusion feature.

[0015] The danger detection module is used to perform identification based on the fusion features to obtain a dangerous behavior identification result.

[0016] Furthermore, the daytime image of the same edge operation scene is an image of the edge operation scene at 12 noon on the same day or an image of the edge operation scene at 12 noon on the previous day.

[0017] Furthermore, the pretreatment method includes:

[0018] Decoding the image of the nighttime edge working scene to be identified and the daytime image of the same edge working scene to obtain RGB image data;

[0019] The RGB image data is subjected to Gaussian filtering, image normalization and image specific size scaling processing to obtain a preprocessed image of a nighttime edge working scene and a daytime image of the same edge working scene.

[0020] Furthermore, self-attention fusion is performed on the image of the nighttime edge working scene to be identified and the daytime image of the same edge working scene to obtain fusion features, including:

[0021] The day and night data fusion module includes a daytime input branch, a nighttime input branch and a channel fusion module;

[0022] The daytime input branch and the nighttime image input branch each include two self-attention modules;

[0023] The daytime input branch is used to perform two self-attention module processes on the input daytime images to obtain a daytime processing result;

[0024] The night input branch is used to perform two self-attention module processes on the input night images to obtain night processing results;

[0025] The channel fusion module is used to perform channel connection fusion on the daytime processing results and the nighttime processing results of the same edge operation scene to obtain a fused daytime and nighttime fusion feature map.

[0026] Furthermore, the alarm device includes a speaker installed at the edge operation scene.

[0027] Furthermore, the camera also includes a camera installed on a drone.

[0028] Furthermore, the danger detection module is built based on the CNN model.

[0029] Furthermore, the training method of the dangerous behavior recognition model includes:

[0030] Obtain nighttime images and daytime images of the same edge operation scene and corresponding nighttime dangerous behavior annotations to obtain a training data set;

[0031] The detection model is trained using the training data set, and the network is trained using the Adam optimizer.

[0032] In a second aspect, the present invention provides a method for identifying dangerous behaviors in nighttime edge operations, comprising the following steps:

[0033] Acquire an image of the nighttime edge work scene to be identified and a daytime image of the same edge work scene;

[0034] According to the image of the edge working scene to be identified at night and the daytime image of the same edge working scene, a trained dangerous behavior recognition model is input to obtain a dangerous behavior recognition result; the dangerous behavior includes leaning on the edge guardrail;

[0035] The dangerous behavior recognition model includes a preprocessing module, a day and night data fusion module and a danger detection module;

[0036] The preprocessing module is used to preprocess the image of the nighttime edge working scene and the daytime image of the same edge working scene;

[0037] The day and night data fusion module is used to perform self-attention fusion on the image of the nighttime edge working scene to be identified and the daytime image of the same edge working scene to obtain a fusion feature.

[0038] The danger detection module is used to perform identification based on the fusion features to obtain a dangerous behavior identification result.

[0039] Furthermore, the daytime image of the same edge operation scene is an image of the edge operation scene at 12 noon on the same day or an image of the edge operation scene at 12 noon on the previous day.

[0040] Furthermore, the pretreatment method includes:

[0041] Decoding the image of the nighttime edge working scene to be identified and the daytime image of the same edge working scene to obtain RGB image data;

[0042] The RGB image data is subjected to Gaussian filtering, image normalization and image specific size scaling processing to obtain a preprocessed image of a nighttime edge working scene and a daytime image of the same edge working scene.

[0043] Furthermore, self-attention fusion is performed on the image of the nighttime edge working scene to be identified and the daytime image of the same edge working scene to obtain fusion features, including:

[0044] The day and night data fusion module includes a daytime input branch, a nighttime input branch and a channel fusion module;

[0045] The daytime input branch and the nighttime image input branch each include two self-attention modules;

[0046] The daytime input branch is used to perform two self-attention module processes on the input daytime images to obtain a daytime processing result;

[0047] The night input branch is used to perform two self-attention module processes on the input night images to obtain night processing results;

[0048] The channel fusion module is used to perform channel connection fusion on the daytime processing results and the nighttime processing results of the same edge operation scene to obtain a fused daytime and nighttime fusion feature map.

[0049] Furthermore, the danger detection module is built based on the CNN model.

[0050] Furthermore, the training method of the dangerous behavior recognition model includes:

[0051] Obtain nighttime images and daytime images of the same edge operation scene and corresponding nighttime dangerous behavior annotations to obtain a training data set;

[0052] The detection model is trained using the training data set, and the network is trained using the Adam optimizer.

[0053] In a third aspect, the present invention provides a device for identifying dangerous behaviors of nighttime edge operations, comprising:

[0054] Input module: used to obtain the image of the nighttime edge operation scene to be identified and the daytime image of the same edge operation scene;

[0055] Identification module: used to input the trained dangerous behavior identification model according to the image of the nighttime edge working scene to be identified and the daytime image of the same edge working scene, to obtain the dangerous behavior identification result; the dangerous behavior includes leaning on the edge guardrail;

[0056] The dangerous behavior recognition model includes a preprocessing module, a day and night data fusion module and a danger detection module;

[0057] The preprocessing module is used to preprocess the image of the nighttime edge working scene and the daytime image of the same edge working scene;

[0058] The day and night data fusion module is used to perform self-attention fusion on the image of the nighttime edge working scene to be identified and the daytime image of the same edge working scene to obtain a fusion feature.

[0059] The danger detection module is used to perform identification based on the fusion features to obtain a dangerous behavior identification result.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] 1. The present invention uses a self-attention module (Transformer module) to fuse daytime and nighttime data, so that the network can better capture the background information in the image, thereby better improving the recognition effect.

[0062] 2. The present invention uses a dangerous behavior recognition model to more accurately identify dangerous behavior categories. It combines global scene features with individual character features to obtain more representative fusion features, which are used to construct an interactive relationship graph between characters, which is then trained using a graph convolutional network, and a classifier is used to accurately predict the category of dangerous behavior. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a framework diagram of the early warning system of the present invention;

[0064] Figure 2 is a flow chart of the identification method of the present invention;

[0065] Figure 3 It is a structural diagram of the Transformer module. DETAILED DESCRIPTION

[0066] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0067] Embodiment 1:

[0068] like Figure 1 As shown, this embodiment provides a nighttime edge operation dangerous behavior warning system, including:

[0069] The camera is used to obtain images of the nighttime edge work scene to be identified and daytime images of the same edge work scene; the camera is preferably an infrared camera, which can be installed next to the tower crane work light on site.

[0070] An alarm (preferably a loudspeaker arranged on site) is installed at the site of the nighttime edge operation scene to be identified, and is used to send out a warning signal;

[0071] A control system is connected to the camera and the alarm signal respectively, and is used to input the trained dangerous behavior recognition model according to the image of the nighttime edge working scene to be identified and the daytime image of the same edge working scene, obtain the dangerous behavior recognition result, and control the alarm to send a reminder signal according to the dangerous behavior recognition result; the dangerous behavior includes leaning against the edge protection railing;

[0072] The control system can be a computer with the following configuration: Intel Core i5-12400F core processor, NVIDIA GeForce RTX 3060Ti graphics processor, main frequency 2.50 GHz, memory 32GB, and operating system Windows 10.

[0073] The dangerous behavior recognition model includes a preprocessing module, a day and night data fusion module and a danger detection module;

[0074] The preprocessing module is used to preprocess the image of the nighttime edge working scene and the daytime image of the same edge working scene;

[0075] The day and night data fusion module is used to perform self-attention fusion on the image of the nighttime edge working scene to be identified and the daytime image of the same edge working scene to obtain a fusion feature.

[0076] The danger detection module is used to perform identification based on the fusion features to obtain a dangerous behavior identification result.

[0077] Specifically, the image acquisition method can acquire video stream data of construction site images based on fixed cameras and drones.

[0078] There are two main ways to obtain video stream data: pre-installed fixed cameras and drones equipped with high-definition cameras. The video stream data acquisition method of pre-installed fixed cameras is technically realized by mainly controlling and adjusting the camera to obtain the required video stream data in real time. For camera adjustment and control, Python programming language can be used to automate programming for specific scenarios. The video stream data acquisition method of drones equipped with high-definition cameras is technically realized by building an rtmp server to receive drone streams, and the drone uploads the acquired video data to the rtmp stream server in real time through image transmission technology.

[0079] Specifically, the daytime image of the same edge operation scene is an image of the edge operation scene at 12 noon on the same day or an image of the edge operation scene at 12 noon on the previous day.

[0080] Cloud-free images at noon are preferred to improve resolution, and the image time and recognition time should not exceed 36 hours to prevent inaccurate recognition results due to changes in the construction environment.

[0081] Specifically, the pretreatment method includes:

[0082] The image of the nighttime edge operation scene to be identified and the daytime image of the same edge operation scene are decoded to obtain RGB image data; that is, for the video stream data collected by a fixed camera or a drone, each frame of the video stream is decoded based on the H.264 codec technology and converted into RGB image data.

[0083] Gaussian filtering, image normalization and image specific size scaling are performed on the RGB image data to obtain preprocessed video stream data.

[0084] Specifically, self-attention fusion is performed on the image of the nighttime edge working scene to be identified and the daytime image of the same edge working scene to obtain fusion features, including:

[0085] The day and night data fusion module includes a daytime input branch, a nighttime input branch and a channel fusion module;

[0086] The daytime input branch and the nighttime image input branch each include two self-attention modules;

[0087] The daytime input branch is used to perform two self-attention module processes on the input daytime images to obtain a daytime processing result;

[0088] The night input branch is used to perform two self-attention module processes on the input night images to obtain night processing results;

[0089] The channel fusion module is used to perform channel connection fusion on the daytime processing results and the nighttime processing results of the same edge operation scene to obtain a fused daytime and nighttime fusion feature map.

[0090] The Transformer module can adopt the common Transformer module, which is mainly composed of several encoders and decoders, such as Figure 3 As shown, no further description is given here.

[0091] The specific training data includes night images and day images of the same area and corresponding night-time dangerous behavior annotations, and the night images and day images of the same area are paired. During training, the two pictures of the night image and the day image of the same scene are taken as a group and input into the day image input branch and the night image input branch of the day and night data fusion module of the detection model respectively to train the detection model and finally obtain a trained detection model.

[0092] Specifically, the alarm device includes a loudspeaker installed at the edge operation scene, and reminders can also be given through a drone or a pager.

[0093] Specifically, the danger detection module of this embodiment can use a convolutional neural network (CNN). A convolutional neural network is a type of neural network that includes Computational and deeply structured , is one of the most popular and widely used models in the field of deep learning in recent years. It uses local connections and shared weights to obtain effective representations directly from the original data through alternating convolutional layers and pooling layers, automatically extract local features of the data, and establish dense and complete feature vectors. The typical convolutional neural network uses the back propagation (BP) algorithm to minimize the loss between the training results and the actual values, as shown in formula (1):

[0094]

[0095] The CNN that processes time series data has several channels, each of which contains several convolutional layers, excitation layers, pooling layers, and fully connected layers. The most important of these is the convolutional layer, which is used to learn spatial features at different levels. Each channel corresponds to an input data, and all channels have the same weight.

[0096] Optionally, the danger detection module can also adopt a conventional deep learning network model, such as constructing a deep learning network model based on a deep residual convolutional neural network, and training the constructed deep learning network model according to the preprocessed video stream data to obtain a backbone network model for image feature extraction.

[0097] As the core technology of the intelligent supervision mode system, a backbone network for image feature extraction is directly related to the accuracy of the entire detection and recognition. The backbone network for image feature extraction uses a deep residual convolutional neural network--DenseNet (Densely Connected Convolutional Networks), collectively referred to as the "backbone network model". Through the transfer learning method based on model weights, the existing DenseNet pre-trained model is fine-tuned in combination with the collected engineering image data to obtain an image feature extraction backbone network model suitable for the field of engineering image data.

[0098] According to the image feature extraction backbone network model, image features are extracted from the fused day and night fusion feature map output by the day and night data fusion module.

[0099] The identification of the target object is aligned with its position. For the target objects of concern (such as workers leaning on edge guardrails, etc.), in the construction of the detection and recognition model, manual priority marking is used in the early stage to train and optimize the detection and recognition model; in the later stage, the model self-service detection and recognition, manual abnormality correction, and model self-reinforcement learning are used. Regarding the early manual marking, in the original engineering graphics collected, the target objects of concern are annotated in the form of rectangular boxes using "labelimg" and other similar tools. After the image features are extracted through the backbone network model, the manually annotated target object data and the target detection position based on the "Anchor" mechanism are regressed and fitted. Through batch data training, the detection and recognition of engineering target objects based on the backbone network model are obtained, and the position of the target object in the image is accurately located, that is, the safety hazard objects of concern in the project are identified, such as "workers leaning on edge guardrails". The trained engineering image data detection and recognition model is put into a pre-use correction stage for a period of time (until the actual project requirements are met). During this stage, the acquired real-time images (the video stream data is decoded, converted into images, and preprocessed) are used for target object detection and recognition. The recognition is corrected manually to enable the detection and recognition model to re-learn based on the correction information, so that the model can reach a level of self-enhanced learning in the later detection and recognition process of the target object.

[0100] The deep learning network model obtains video stream data of construction site images based on fixed cameras and drones; preprocesses the video stream data to obtain preprocessed video stream data; builds an image feature extraction backbone network model based on the preprocessed video stream data; extracts image features based on the image feature extraction backbone network model; trains the image detection and recognition model based on image features and manually annotated safety hazards; obtains the construction site image to be predicted; preprocesses the construction site image to obtain the preprocessed construction site image; identifies the construction site safety hazard results based on the preprocessed construction site image and the image detection and recognition model. After obtaining on-site construction photos through camera and drone shooting, deep learning is used for judgment, and the corresponding on-site safety hazard results are output for reference by supervisors, thereby achieving the purpose of improving the production efficiency of supervisors and reducing costs.

[0101] Preferably, the danger detection module includes a backbone network, a fusion module, an embedding layer, a relational reasoning module, a graph convolutional network, a global average pooling layer and a classifier.

[0102] The cropped images are assigned to the pre-trained dangerous behavior recognition model to obtain the dangerous behavior category. The dangerous behavior recognition model (i.e., global-individual features fusion network, GIFFNet, Global-Individual Features Fusion Network) includes: backbone network, fusion module, embedding layer, relational reasoning module, graph convolutional network, global average pooling layer and classifier.

[0103] The specific process of identifying dangerous behaviors is as follows:

[0104] The cropped image is input into the backbone network in RGB mode. The backbone network uses VGG-16 (Visual Geometry Group 16-layer Network) or Inception-v3 (Google Inception-v3Network) to extract the global features of the image. , ,in represents the batch size, represents the time step, represents the channel dimension of global image features, and are the width and height of the image respectively. Then, according to the global features The image is cropped and aligned using RoI Align (region of interest alignment), and the region of interest is mapped to the global features to obtain the individual (in this embodiment, the individual refers to a single person) features containing appearance details. , ,in, The channel dimension representing individual features, Represents the side length of the cropped image. Indicates the number of target persons in a single image.

[0105] Next, the global features obtained and individual characteristics Sent to the fusion module to obtain fusion features, such as Figure 3 As shown in the figure, the fusion module (GIFF, global - individual features fusion) includes: a convolutional layer, an attention mechanism module, a linear embedding layer, a Softmax layer, a layer normalization layer and a feedforward network.

[0106] Specifically, the training method of the dangerous behavior recognition model includes:

[0107] Obtain nighttime images and daytime images of the same edge operation scene and corresponding nighttime dangerous behavior annotations to obtain a training data set;

[0108] The detection model is trained using the training data set, and the network is trained using the Adam optimizer.

[0109] The detection model is trained using the training data. The network is trained using the Adam optimizer. The maximum learning rate is set to 0.0001. When training reaches convergence, the learning rate is reduced for training. The minimum learning rate is 0.000001, batch_size is set to 4, and epoch is set to 100.

[0110] Optionally, in the method of this embodiment,

[0111] Embodiment 2:

[0112] like Figure 2 As shown, this embodiment provides a method for identifying dangerous behaviors of nighttime edge operations, comprising the following steps:

[0113] Acquire an image of the nighttime edge work scene to be identified and a daytime image of the same edge work scene;

[0114] According to the image of the edge working scene to be identified at night and the daytime image of the same edge working scene, a trained dangerous behavior recognition model is input to obtain a dangerous behavior recognition result; the dangerous behavior includes leaning on the edge guardrail;

[0115] The dangerous behavior recognition model includes a preprocessing module, a day and night data fusion module and a danger detection module;

[0116] The preprocessing module is used to preprocess the image of the nighttime edge working scene and the daytime image of the same edge working scene;

[0117] The day and night data fusion module is used to perform self-attention fusion on the image of the nighttime edge working scene to be identified and the daytime image of the same edge working scene to obtain a fusion feature.

[0118] The danger detection module is used to perform identification based on the fusion features to obtain a dangerous behavior identification result.

[0119] The danger detection module can also adopt other conventional behavior detection model architectures, such as CNN+LSTM, which can realize behavior detection and only needs to be retrained through the training set.

[0120] Embodiment three:

[0121] This embodiment provides a device for identifying dangerous behaviors of working on the edge at night, including:

[0122] Input module: used to obtain the image of the nighttime edge operation scene to be identified and the daytime image of the same edge operation scene;

[0123] Identification module: used to input the trained dangerous behavior identification model according to the image of the nighttime edge working scene to be identified and the daytime image of the same edge working scene, to obtain the dangerous behavior identification result; the dangerous behavior includes leaning on the edge guardrail;

[0124] The dangerous behavior recognition model includes a preprocessing module, a day and night data fusion module and a danger detection module;

[0125] The preprocessing module is used to preprocess the image of the nighttime edge working scene and the daytime image of the same edge working scene;

[0126] The day and night data fusion module is used to perform self-attention fusion on the image of the nighttime edge working scene to be identified and the daytime image of the same edge working scene to obtain a fusion feature.

[0127] The danger detection module is used to perform identification based on the fusion features to obtain a dangerous behavior identification result.

[0128] The device of this embodiment can be used to implement the method described in the second embodiment.

[0129] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0130] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0131] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0132] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0133] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A nighttime edge operation dangerous behavior warning system, characterized in that: include: A camera, used to obtain an image of a nighttime edge operation scene to be identified and a daytime image of the same edge operation scene; An alarm is installed at the scene of the nighttime edge operation to be identified, and is used to send out a warning signal; A control system is connected to the camera and the alarm signal respectively, and is used to input the trained dangerous behavior recognition model according to the image of the nighttime edge working scene to be identified and the daytime image of the same edge working scene, obtain the dangerous behavior recognition result, and control the alarm to send a reminder signal according to the dangerous behavior recognition result; the dangerous behavior includes leaning against the edge protection railing; The dangerous behavior recognition model includes a preprocessing module, a day and night data fusion module and a danger detection module; The preprocessing module is used to preprocess the image of the nighttime edge working scene and the daytime image of the same edge working scene; The day and night data fusion module is used to perform self-attention fusion on the image of the nighttime edge working scene to be identified and the daytime image of the same edge working scene to obtain a fusion feature. The danger detection module is used to perform identification based on the fusion features to obtain a dangerous behavior identification result.

2. The nighttime edge operation dangerous behavior warning system according to claim 1 is characterized in that: The daytime image of the same edge operation scene is an image of the edge operation scene at 12 noon on the same day or an image of the edge operation scene at 12 noon on the previous day.

3. The nighttime edge operation dangerous behavior warning system according to claim 1 is characterized in that: The pretreatment method comprises: Decoding the image of the nighttime edge working scene to be identified and the daytime image of the same edge working scene to obtain RGB image data; The RGB image data is subjected to Gaussian filtering, image normalization and image specific size scaling processing to obtain a preprocessed image of a nighttime edge working scene and a daytime image of the same edge working scene.

4. The nighttime edge operation dangerous behavior warning system according to claim 1 is characterized in that: The image of the nighttime edge working scene to be identified and the daytime image of the same edge working scene are subjected to self-attention fusion to obtain fusion features, including: The day and night data fusion module includes a daytime input branch, a nighttime input branch and a channel fusion module; The daytime input branch and the nighttime image input branch each include two self-attention modules; The daytime input branch is used to perform two self-attention module processes on the input daytime images to obtain a daytime processing result; The night input branch is used to perform two self-attention module processes on the input night images to obtain night processing results; The channel fusion module is used to perform channel connection fusion on the daytime processing results and the nighttime processing results of the same edge operation scene to obtain a fused daytime and nighttime fusion feature map.

5. The nighttime edge operation dangerous behavior warning system according to claim 1 is characterized in that: The alarm device includes a speaker installed at the edge operation scene.

6. The nighttime edge operation dangerous behavior warning system according to claim 1 is characterized in that: The camera also includes a camera installed on a drone.

7. The nighttime edge work dangerous behavior warning system according to claim 4 is characterized in that: The danger detection module is built based on the CNN model.

8. The nighttime edge work dangerous behavior warning system according to claim 7 is characterized in that: The training method of the dangerous behavior recognition model includes: Obtain nighttime images and daytime images of the same edge operation scene and corresponding nighttime dangerous behavior annotations to obtain a training data set; The detection model is trained using the training data set, and the network is trained using the Adam optimizer.

9. A method for identifying dangerous behaviors in nighttime edge work, characterized in that: The following steps are involved: Acquire an image of the nighttime edge work scene to be identified and a daytime image of the same edge work scene; According to the image of the edge working scene to be identified at night and the daytime image of the same edge working scene, a trained dangerous behavior recognition model is input to obtain a dangerous behavior recognition result; the dangerous behavior includes leaning on the edge guardrail; The dangerous behavior recognition model includes a preprocessing module, a day and night data fusion module and a danger detection module; The preprocessing module is used to preprocess the image of the nighttime edge working scene and the daytime image of the same edge working scene; The day and night data fusion module is used to perform self-attention fusion on the image of the nighttime edge working scene to be identified and the daytime image of the same edge working scene to obtain a fusion feature. The danger detection module is used to perform identification based on the fusion features to obtain a dangerous behavior identification result.

10. A device for identifying dangerous behaviors during nighttime edge work, characterized in that: include: Input module: used to obtain the image of the nighttime edge working scene to be identified and the daytime image of the same edge working scene; Identification module: used to input the trained dangerous behavior identification model according to the image of the nighttime edge working scene to be identified and the daytime image of the same edge working scene, to obtain the dangerous behavior identification result; the dangerous behavior includes leaning on the edge guardrail; The dangerous behavior recognition model includes a preprocessing module, a day and night data fusion module and a danger detection module; The preprocessing module is used to preprocess the image of the nighttime edge working scene and the daytime image of the same edge working scene; The day and night data fusion module is used to perform self-attention fusion on the image of the nighttime edge working scene to be identified and the daytime image of the same edge working scene to obtain a fusion feature. The danger detection module is used to perform identification based on the fusion features to obtain a dangerous behavior identification result.

Images