An intelligent monitoring method for violation of control room personnel

Abstract

The present application belongs to the technical field of visual monitoring, and particularly relates to an intelligent monitoring method for illegal behavior of personnel in a control room. The present application uses a U-Net-GNN light compensation model to obtain a preliminary light compensation image using the encoder and decoder structure of the U-Net network. Then, the pixels are modeled as graph nodes using GNN, the edge connection is constructed based on the spatial adjacent relationship, the light correction information is transmitted through three-layer graph convolution, the local light consistency is optimized, the fusion weight is finally generated by the full connection layer, and the first target image is output. The improved YOLOv6 model is used to identify the first target image. The head and limb feature recognition branch network are added and connected with the Backbone. The network is trained and verified by combining pre-training and using the training data set generated by the generative adversarial network (GAN). The network is optimized based on the bounding box regression loss function, which can more accurately identify and locate the staff, obtain the bounding box information of the staff to obtain the second target image, and further improve the visual recognition accuracy.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of visual monitoring, and particularly relates to an intelligent monitoring method for personnel rule violation in a control room. BACKGROUND

[0002] The problem of uneven illumination in the control room, as a key scene of fire safety, has long plagued the stability of the personnel rule violation recognition system. Specifically, uneven illumination is usually caused by power fluctuations, lamp aging, unreasonable installation layout, or environmental obstruction, and its direct consequence is the decline in image quality: local over-bright or over-dark areas will produce shadows or highlights, covering the details of personnel actions (such as gestures, operation steps), resulting in limited performance of target detection and behavior analysis algorithms. For example, in fire safety monitoring, uneven illumination may cause the system to mistakenly identify device shadows as personnel rule violation, or miss the action of unauthorized closing of fire-fighting equipment due to high light reflection.

[0003] Some solutions retrain the target monitoring model by incorporating a large amount of training data obtained in complex light environments to improve the generalization ability and environmental light adaptation ability of the target monitoring model, so as to improve the identification ability of the target monitoring model in different light environments of the fire control room for personnel rule violation. However, this usually leads to an increase in model complexity, thereby increasing the demand for computing resources, and cannot improve the performance of the rule violation monitoring algorithm in complex light conditions while ensuring real-time performance. SUMMARY

[0004] To solve the above problems in the prior art, the application provides an intelligent monitoring method for personnel rule violation in a control room, which solves the problem of reduced identification accuracy of visual recognition algorithms for personnel rule violation in the control room due to uneven illumination.

[0005] The purpose of the application can be achieved by the following technical solution: an intelligent monitoring method for personnel rule violation in a control room, comprising the following steps:

[0006] S1: acquiring real-time images of the control room and pre-processing the real-time images, inputting the real-time images into a pre-set U-Net-GNN light compensation model, and outputting a first target image after completing light compensation, wherein the U-Net-GNN light compensation model uses a U-Net network to extract a preliminary light compensation image, uses a GNN network to optimize local light consistency through pixel space relationship, and generates a first target image by fusing weights of U-Net output and GNN output;

[0007] S2: identifying the first target image using a YOLOv6 model and identifying and positioning the frame of the staff to obtain a second target image;

[0008] S3: identifying the first target image and the second target image, extracting the pose information of the staff as the violation criterion information, using a pre-set violation behavior judgment algorithm to identify the violation criterion information, and if there is a violation, the current frame image is intercepted and an alarm is given;

[0009] S4: controlling the camera to capture the third target image of the continuous time frame.

[0010] Preferably, the structure of the U-Net-GNN light compensation model comprises:

[0011] The encoder of the U-Net network is reduced by four levels, the number of channels is doubled, the decoder also restores the image resolution through four levels of upsampling structure, each level splices and fuses the deep features with the corresponding encoder features, and adopts transposed convolution to realize 2 times upsampling to obtain a preliminary light compensation image.

[0012] Optimizing the light using a graph neural network GNN: including modeling the pixels of the preliminary light compensation image as graph nodes, the RGB values of the pixels as features, constructing edge connections based on spatial adjacency relationships, and transmitting light correction information through three layers of graph convolution;

[0013] Using a weight ratio to calculate the weight distribution of the outputs of the U-Net network and the GNN.

[0014] Preferably, the edge connections are constructed based on spatial adjacency relationships, wherein the edge weight of the edge connection The formula for calculating is:

[0015] ;

[0016] wherein, and are the coordinates of the pixel points, is the influence range of the control neighborhood, which is determined using a four-neighborhood marking method.

[0017] Preferably, the calculation of the weight ratio comprises:

[0018] The preliminary light compensation image output by the U-Net and the light correction image output by the GNN ;

[0019] The fusion weight value is calculated using an attention mechanism, wherein the calculation formula of the fusion weight value is:

[0020] ;

[0021] wherein, MLP is a multi-layer perception mechanism, which limits the output range of the weight value to between.

[0022] obtaining an output first target image of the light compensation model based on the fusion weight proportion value The calculation formula is:

[0023] .

[0024] Preferably, the YOLOv6 model is an improved YOLOv6 model, and the improved YOLOv6 model is a head and limb feature recognition branch network added to the basic architecture of the YOLOv6 model, and the branch network is connected with the Backbone feature extraction layer of the YOLOv6 model.

[0025] Preferably, the determination of the improved YOLOv6 model comprises:

[0026] Step 1: Obtain an image data set, and divide the image data set into training images and verification images according to a certain proportion, and use the training images to pre-train the YOLOv6 network;

[0027] Step 2: Take the pre-trained YOLOv6 network as a basic framework, add a head and a limb feature recognition branch network to generate an improved YOLOv6 network, and connect the branch network to the Backbone feature extraction layer;

[0028] Step 3: The improved YOLOv6 network is trained and verified using the training method of step 1, and the training data set used is generated by a generative adversarial network;

[0029] Step 4: Calculate the loss function value based on the target recognition result through a bounding box regression loss function;

[0030] Step 5: Use the loss function value to optimize the improved YOLOv6 network.

[0031] Preferably, step 3 comprises: the improved YOLOv6 network frames the predicted bounding box of the predicted staff according to a pre-set threshold;

[0032] The bounding boxes with a confidence lower than the set confidence threshold are removed using the confidence level;

[0033] The bounding boxes after screening are processed again by a set non-maximum suppression algorithm to remove redundant bounding boxes with high overlap;

[0034] Output the staff bounding box recognition result.

[0035] Preferably, the calculation formula of the loss function value is:

[0036] ;

[0037] wherein, and are weight coefficients, is predicted bounding box information, is real bounding box information, is predicted part location confidence, is real part location confidence, and N is the total number of elements of the bounding box.

[0038] Preferably, step five comprises: using gradient descent method to update network parameters with the goal of minimizing the loss function value, and using cosine annealing learning rate strategy to dynamically adjust the learning rate of the improved YOLOv6 model in each round of training until the loss function value is lower than the set threshold, and outputting the model parameters after optimization.

[0039] The beneficial effects of the present application are:

[0040] Firstly, the present application uses the U-Net-GNN light compensation model to obtain a preliminary light compensation image by using the encoder and decoder structure of the U-Net network; then uses GNN to model the pixels as graph nodes, constructs edge connections based on spatial adjacency, transmits light correction information through three-layer graph convolution to optimize local light consistency, finally generates fusion weights by the full connection layer, and outputs the first target image, laying a foundation for subsequent accurate identification.

[0041] Secondly, the improved YOLOv6 model is used to identify the first target image, and the head and limb feature recognition branch network is added and connected with the Backbone, combined with pre-training, using the training data set generated by the generative adversarial network (GAN) to train and verify, and optimizing the network based on the bounding box regression loss function, which can more accurately identify and locate the staff, obtain the bounding box information to obtain the second target image, and further improve the identification accuracy. Thus, the personnel pose information can be better extracted as the violation criterion, and when it is judged that there is a violation, the current frame image is intercepted in time to alarm, and the camera is controlled to capture continuous time frame images, realizing real-time and accurate monitoring of the violation behavior of the personnel in the control room.

[0042] Through the optimization processing of the image light and the improvement of the identification model, the identification accuracy of the violation behavior of the personnel in the control room under complex light conditions is significantly improved, providing a reliable guarantee for the safety management of the control room. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to facilitate the understanding of those skilled in the art, the present application will be further described below in conjunction with the drawings.

[0044] Figure 1 is the method step block diagram of the present application;

[0045] Figure 2 The determination step of the improved YOLOv6 model used in the present application; DETAILED DESCRIPTION

[0046] In order to further illustrate the technical means and effects adopted by the present application to achieve the predetermined inventive purpose, the specific embodiments, structures, features and effects thereof according to the present application are described in detail below in combination with the drawings and preferred embodiments.

[0047] Please refer to Figure 1 - Figure 3, the embodiment provides an intelligent monitoring method for personnel violation behavior in the control room, characterized in that it comprises the following steps:

[0048] S1: acquiring real-time images of the control room and pre-processing the real-time images, inputting the real-time images into a pre-set U-Net-GNN light compensation model, and outputting a first target image after completing light compensation, the U-Net-GNN light compensation model uses a U-Net network to extract a preliminary light compensation image, uses a GNN network to optimize local light consistency through pixel space relationship, and a fully connected layer generates fusion weights of U-Net output and GNN output to obtain the first target image;

[0049] The construction of the network architecture of the light compensation model comprises:

[0050] A: U-Net network architecture:

[0051] A1: the encoder extracts features through a four-level downsampling structure, extracts visual features of different scales through hierarchical convolution operations, and adapts to detail capture under dim light:

[0052] It contains four levels of downsampling modules, each of which consists of 2 3x3 convolution + ReLU activation function, and the size of the feature map output by each level is halved in turn, and the number of channels is doubled;

[0053] The input process is:

[0054] The shallow layer (first and second levels) is used to capture the basic visual features such as edges and textures of the real-time image;

[0055] First level: input the real-time image of the control room collected in the dim environment (such as size 640x480x3), output the feature map size 320x240 after 2 times of convolution, and the number of channels increases from 3 to 64;

[0056] Second level: the size after downsampling (pooling) is 160x120, and the number of channels is 128;

[0057] Third level: size 80x60, channel number 256;

[0058] Fourth level: size 40x30, number of channels 512;

[0059] The semantic features (such as scene structure, light distribution trend) are extracted using deep layers (third and fourth levels), and different scales of light abnormal information (such as local dark area, reflective noise) are retained in the feature map output at each level;

[0060] A2: Decoder restores image architecture; the decoder also restores the image resolution through a four-level upsampling structure, each level splices and fuses deep features with corresponding encoder features, and uses transpose convolution to realize 2 times upsampling;

[0061] The output process includes:

[0062] First level: (corresponding to the fourth level of the encoder): input 512 channel feature map, transpose convolution to 80x60, splice with 256 channel features of the third level of the encoder, output channel number 512;

[0063] Second level: upsample to 160x120, splice with 128 channel features of the second level of the encoder, output channel number 256;

[0064] Third level: upsample to 320x240, splice with 64 channel features of the first level of the encoder, output channel number 128;

[0065] Fourth level: upsample to 640x480, compress to three channels through 1x1 convolution, generate preliminary light compensation image ;

[0066] The U-Net network uses skip connections, i.e., encoder-decoder splicing structure, to retain shallow details (such as dark area edges) and avoid information loss in the upsample process, and through the progressive amplification of convolution to ensure the spatial consistency of light correction, and to uniformly brighten the dark areas in the real-time image.

[0067] B: Use graph neural network GNN to model the real-time image as a graph structure, transfer light correction information through spatial topological relationship, solve the problem of local light unevenness and further optimize the real-time image, which includes:

[0068] B1: Construction of graph structure:

[0069] Define nodes: each pixel point in the real-time image is regarded as a node, and the feature is the RGB value or brightness value of the pixel point, which is used to represent the brightness of the pixel point;

[0070] Edge connection: based on spatial adjacency relationship, such as four-neighborhood labeling algorithm or eight-neighborhood labeling algorithm to determine edge weight, edge weight is determined by pixel distance, and the calculation formula of edge weight is:

[0071] ;

[0072] wherein, and are the coordinates of the pixel points, is the influence range of the control neighborhood, since the light of the dim light environment needs to be compensated, the neighborhood marking algorithm is used here, which is a four-neighborhood marking algorithm, since only the four directly adjacent pixel points are considered for information fusion when the decoder performs convolution operation or up-convolution operation, which helps to maintain the sharpness of the edge, but in some cases it may cause the loss of detailed information, so the graph structure of the GNN network is needed to calculate the weight value between the pixel points;

[0073] B2: using the graph structure to construct a GNN three-layer convolution layer, wherein the operation logic of the three-layer convolution layer is:

[0074] The first layer: used to capture the light correlation of local pixels, such as collective brightening of adjacent dark pixels;

[0075] The second layer: extending to the consistency of the regional light;

[0076] The third layer: modeling of global light trend;

[0077] B3: the output of the light correction image , the light correction coefficient of each pixel in the final output is multiplied by the original image (real-time image) :

[0078] ;

[0079] C: generating a fusion weight ratio of the U-Net network output and the GNN network output based on the weight value calculation formula, and dynamically adjusting the contribution ratio of the U-Net output and the GNN output based on the fusion weight ratio to balance the global light trend and the local details, wherein the following sub-steps are included:

[0080] C1: for the calculation of the weight ratio:

[0081] input the preliminary light compensation image of the U-Net output and the light correction image of the GNN output;

[0082] by using the attention mechanism to calculate the pixel-level fusion weight value, wherein the calculation formula of the fusion weight value is:

[0083] ;

[0084] wherein, MLP is a multi-layer perception mechanism, which limits the output range of the weight value to between;

[0085] C2: obtaining the output of the light compensation model based on the fusion weight proportion value The calculation formula of the first target image is:

[0086] ;

[0087] The U-Net-GNN light compensation model has significant advantages in image enhancement under dim light conditions through multi-dimensional architecture design. Specifically, the four-level down-sampling structure of the encoder gradually extracts multi-scale features through hierarchical convolution, which can capture basic visual details such as edges and textures, and also extract semantic features of light distribution, providing multi-level information support for subsequent light compensation; the four-level up-sampling of the decoder combined with the jump connection restores the image resolution while fusing deep light semantic features with shallow detail features, avoiding the loss of details in the traditional up-sampling process, and ensuring that the light compensated image retains the structural information of the original scene.

[0088] The graph neural network GNN module models the image pixels as graph nodes, constructs edge connections based on spatial topological relationships, and realizes global transmission of light correction information through three-layer graph convolution, effectively solving the problem of local light unevenness (such as uniformization processing of shadow and highlight areas). Compared with traditional convolution operations, it can better capture long-distance dependencies between pixels, making the light correction result more consistent with the physical scene light consistency. In addition, the model dynamically generates the fusion weights of U-Net and GNN output, which can adaptively adjust the contribution proportion of the two according to the scene light complexity, realize the global light correction ability of GNN in uniform dark area, focus on the detail retention advantage of U-Net in complex areas such as light and dark boundaries, and finally realize the balance between global light uniformity and local detail integrity.

[0089] The light compensation model through the three-level architecture of convolution feature extraction combined with graph structure light optimization and finally dynamic weight fusion can improve the image light uniformity and retain the image edge details in the scene of emergency power failure in the control room, and is more suitable for monitoring, automatic driving and other scenes with high requirements for dim image clarity compared with traditional light compensation methods.

[0090] S2: The first target image is identified using a YOLOv6 model, and the bounding box of the personnel is identified and positioned to obtain a second target image. Specifically, the YOLOv6 model can be selected according to actual identification needs to identify and position the bounding box of the staff. The Backbone part of the YOLOv6 model is usually composed of efficient convolution modules such as RepBlock structure. Through multi-layer convolution operation, the input image (first target image) is feature-extracted, the image resolution is gradually reduced, and the channel number is increased to extract the semantic information of the input image at different levels.

[0091] The Neck part adopts a path aggregation network (PAN) structure to fuse the multi-scale features output by the Backbone and enhance the expression ability of the features.

[0092] The original Head part is based on the fused feature map to predict the target box, judge the category, and output the confidence, realizing common image detection and recognition.

[0093] S3: The first target image and the second target image are identified to extract the pose information of the personnel, and the violation criterion information is obtained. The pre-set violation behavior judgment algorithm is used to identify the violation criterion information, and the current frame image is intercepted and an alarm is given if there is a violation.

[0094] Specifically, the first target image contains the overall environmental features of the control room, and the YOLOv6 model is used to identify the first target image and identify and position the bounding box of the personnel to obtain the bounding box information of the staff. The pose information includes the position and pose information of the staff's head and limbs, which can provide the basis for subsequent judgment of the control room violation behavior.

[0095] S4: The camera is controlled to capture the third target image of the continuous time frame.

[0096] Through the pre-set time window, the time of the intercepted current frame image is determined as the reference, and the image is continuously collected in time frames according to the set preset frequency to obtain the third target image. Specifically, when the capture mechanism is triggered, the current frame image is used as a violation signal, and the violation signal is generated as a capture trigger instruction and transmitted to the control end of the camera used to collect the control room. When the control end receives the trigger instruction, it immediately stops the current normal image collection mode and enters the capture mode instead. Under the capture mode, the control end controls the camera to continuously collect several frames of images at a higher frame rate to ensure that the clearest moment of the violation behavior can be captured. Moreover, the capture trigger instruction includes the position information of the staff's bounding box, and the control end adjusts the scanning position of the camera to align the staff's bounding box position in the maximum area of the camera's scanning range.

[0097] In the monitoring method, the U-Net-GNN light compensation model is used to process real-time images, which can effectively cope with changes in light and improve image quality, laying a foundation for subsequent recognition. With the help of the YOLOv6 model to identify the personnel frame, combined with the extraction of personnel position, state and action information in the image, the violation criterion information can be accurately obtained, the intelligent judgment and alarm of the violation behavior can be realized, and the continuous frames of images captured by the camera can be controlled to ensure the integrity and continuity of the evidence. In terms of the architecture of the light compensation model, the encoder and decoder of the U-Net network can effectively extract and restore image features through downsampling, upsampling and feature splicing and fusion, and obtain the initial light compensation image. GNN models the pixels as graph nodes, constructs edge connections based on spatial adjacency, and transmits light correction information through three layers of graph convolution, which can optimize the local light consistency. Weight allocation calculation is performed on the outputs of U-Net and GNN to realize the reasonable fusion of the outputs of the two, further improve the light compensation effect, make the processed image clearer, provide high-quality image input for subsequent personnel violation behavior recognition, and thus improve the accuracy and reliability of the entire monitoring system.

[0098] In one embodiment, the U-Net-GNN light compensation model is trained by obtaining a data set, which includes a LOL data set and a SID data set. The LOL data set includes normal light and low light paired images, and the contrast enhancement of the images is improved. The SID data set includes original sensor data and corresponding long-exposure reference images under low light conditions. The data in the obtained training set is preprocessed, including normalization processing, image enhancement (including random cropping, rotation and flipping, etc.), to improve the generalization ability of the model.

[0099] The low-light images are randomly cropped, rotated and flipped, and then supervised learning is performed to increase the generalization of the model. The model can randomly attenuate the brightness of normal light images under different light conditions, and a generative adversarial network is used to identify images under different light conditions to generate more low-light samples. The image pixel values are normalized to between to improve the convergence speed of the model and reduce the calculation time of the model.

[0100] To further improve the accuracy of the YOLOv6 model in identifying and positioning the personnel frame, in one embodiment, the YOLOv6 model is an improved YOLOv6 model. The improved YOLOv6 model adds a head and a four-limb feature recognition branch network to the basic architecture of the YOLOv6 model, and connects the branch network with the Backbone feature extraction layer of the YOLOv6 model. The specific method includes:

[0101] Step one: obtain an image dataset, and divide the image dataset into training images and validation images according to a certain proportion, identify a plurality of training images through the original YOLOv6 network respectively to obtain a target identification result, and verify the YOLOv6 network after completing the training using the validation images;

[0102] The YOLOv6 model divides the input image into an S×S grid, and each grid is responsible for predicting a certain number (for example, B=3 or B=5) of bounding boxes and the class probability of the target in the bounding box. For each bounding box, the model predicts the center coordinates (offset relative to the grid unit), width w and height h (proportion relative to the entire image), and a confidence score indicating the possibility of containing a target in the bounding box. At the same time, the model also predicts the target class probability (such as the class probability of personnel violation behavior) that may exist in each grid unit;

[0103] Step two: generate an improved YOLOv6 network by adding a head and limb feature recognition branch network to the trained YOLOv6 network as a basic framework, and connecting the branch network to the Backbone feature extraction layer:

[0104] The branch network and the Backbone feature extraction layer share part of the feature information and focus on learning head and hand features. During the training process based on the YOLOv6 model architecture, more annotated data of the head and limb poses and actions of workers are added to the image dataset to enrich the training samples of the model. At the same time, the loss function is further optimized using positioning data, so that the model focuses more on identifying the positions of the heads and limbs of workers. The gradient descent algorithm or its variants (such as the Adam optimizer) are used to update the parameters of the model, and the weight parameters of the network model are updated;

[0105] Step three: train the improved YOLOv6 network using the training method of step one, and the training dataset used is the image dataset generated by the generative adversarial network (GAN), that is, the image dataset is output as the original data in the GAN network to obtain generated data, and the generated data is used as the training data of the branch network;

[0106] In multiple rounds of training, the improved model performs bounding box selection on the predicted worker bounding boxes according to a pre-set threshold;

[0107] The confidence level is used to remove the bounding boxes with low confidence, and the non-maximum suppression algorithm (NMS) is used to process the bounding boxes after screening again to remove redundant bounding boxes with high overlap;

[0108] Output the worker bounding box identification result.

[0109] Step four: calculating the loss function value based on the target recognition result through the bounding box regression loss function The bounding box regression loss function is used to measure the difference between the bounding box and the true bounding box:

[0110]

[0111] wherein, and are weight coefficients, is the predicted bounding box information, is the true bounding box information, is the predicted part position confidence, is the true part position confidence, and N is the total number of elements of the bounding box.

[0112] Step five: based on the loss function value, the improved YOLOv6 network is optimized until the loss function value is less than the set threshold;

[0113] Specifically, the gradient descent method is used to update the network parameters with the goal of minimizing the loss function value, and in each round of training, the cosine annealing learning rate strategy is used to dynamically adjust the model learning rate to speed up model convergence. Until the loss function value is lower than the set threshold, or the detection accuracy on the validation set reaches the target value, the final output is the model parameters after optimization. The newly added head and limb feature recognition branch network is connected to the Backbone feature extraction layer, the shallow feature map has high resolution and contains rich detailed information, which helps to accurately locate the specific position of the head and limbs; the deep feature map has rich semantic information, which can accurately identify the category of these parts. By connecting the branch network to the multi-scale feature extraction layer, the network can extract and analyze the head and limb features at multiple levels, preliminarily locate the target on the shallow feature map, confirm the target category and optimize the bounding box on the deep feature map, so as to realize the accurate detection of the head and limbs from coarse to fine, and improve the performance of the model in the human key part detection task.

[0114] The above is only a preferred embodiment of the present application, and is not intended to limit the present application in any form. Although the present application has been disclosed as above with a preferred embodiment, it is not intended to limit the present application. Any person skilled in the art can make some changes or modifications to the above disclosed technical content to make equivalent embodiments with equivalent changes, without departing from the scope of the technical solution of the present application. Any modification, equivalent change and modification of the above embodiments made in accordance with the technical essence of the present application, without departing from the technical solution of the present application, are still within the scope of the technical solution of the present application.

Claims

1. A method for intelligent monitoring of violations by personnel in fire control rooms, characterized in that: The method comprises the following steps: S1: acquiring real-time images of the control room and preprocessing the real-time images, inputting the real-time images into a pre-set U-Net-GNN light compensation model, and outputting a first target image after completing light compensation, wherein the U-Net-GNN light compensation model uses a U-Net network to extract a preliminary light compensation image, uses a GNN network to optimize local light consistency through pixel space relationship, and generates a first target image by fusing weights of U-Net output and GNN output; S2: identifying the first target image using a YOLOv6 model and identifying and positioning the staff's bounding box to obtain a second target image; S3: identifying the first target image and the second target image, extracting the staff's pose information as a violation criterion information, identifying the violation criterion information using a pre-set violation behavior judgment algorithm, and judging whether there is a violation, and if so, intercepting the current frame image and alarming; S4: controlling the camera to capture the staff to obtain a third target image of continuous time frames; The structure of the U-Net-GNN light compensation model comprises: The encoder of the U-Net network is reduced by four levels, the number of channels is doubled, the decoder also restores the image resolution through four levels of upsampling structure, each level splices and fuses deep features with corresponding encoder features, and adopts transposed convolution to realize 2 times upsampling to obtain a preliminary light compensation image; Optimizing light using a graph neural network GNN: including modeling pixels of the preliminary light compensation image as graph nodes, taking RGB values of the pixels as features, constructing edge connections based on spatial adjacent relationship, and transmitting light correction information through three layers of graph convolution; Using weight proportion to calculate the weight distribution of the output of the U-Net network and the GNN; The calculation of the weight proportion comprises: preliminary light compensation image output by the input U-Net illumination corrected image output by the GNN ; The fusion weight value is calculated using an attention mechanism, and the calculation formula of the fusion weight value is: ； wherein MLP is a multi-layer perceptron mechanism that limits the output range of the weight values to between -1 and 1. The output first target image of the light compensation model is obtained based on the fusion weight proportion value The calculation formula is: ； The edge connection is constructed based on the spatial adjacency relationship, wherein an edge weight of the edge connection is The calculation formula is: ； wherein, and are coordinates of the pixel points, is the control neighborhood range, determined using the four-neighborhood marking method; the YOLOv6 model is an improved YOLOv6 model, the improved YOLOv6 model is a head and four-limb feature recognition branch network added on the basis architecture of the YOLOv6 model, and the branch network is connected with the Backbone feature extraction layer of the YOLOv6 model.

2. The intelligent monitoring method for personnel violation behavior in a control room according to claim 1, characterized in that: The determination of the improved YOLOv6 model comprises: Step one: obtaining an image data set, dividing the image data set into training images and validation images according to a certain proportion, and pre-training the YOLOv6 network using the training images; Step two: taking the pre-trained YOLOv6 network as a basic framework, adding a head and a four-limb feature recognition branch network to generate an improved YOLOv6 network, and connecting the branch network to the Backbone feature extraction layer; Step three: the improved YOLOv6 network is trained and verified using the training method of step one, and the training data set used is generated by a generative adversarial network from the image data set; Step four: calculating the loss function value based on the target recognition result through the boundary box regression loss function; Step five: optimizing the improved YOLOv6 network using the loss function value.

3. The intelligent monitoring method for personnel violation behavior in a control room according to claim 2, characterized in that: Step three comprises: the improved YOLOv6 network frames the predicted boundary box of the staff according to a pre-set threshold; The boundary boxes with a confidence level lower than the set confidence threshold are removed using the confidence level; The redundant boundary boxes with a high overlap degree are removed through a set non-maximum suppression algorithm for secondary processing of the screened boundary boxes. An output person bounding box recognition result.

4. The intelligent monitoring method for personnel violation behavior in a control room according to claim 2, characterized in that: The loss function value is calculated according to the following formula: ； wherein, and are weight coefficients, is predicted bounding box information, is real bounding box information, is predicted part location confidence, is real part location confidence, and N is the total number of elements of the bounding box.

5. The intelligent monitoring method for personnel violation behavior in a control room according to claim 2, characterized in that: Step five includes: using gradient descent method to update network parameters with the goal of minimizing the loss function value, and in each round of training, using cosine annealing learning rate strategy to dynamically adjust the learning rate of the improved YOLOv6 model, until the loss function value is lower than the set threshold, and output the model parameters after optimization.

Images