Substation operator safety monitoring method based on video analysis

Abstract

The invention provides a transformer substation operator safety monitoring method based on video analysis. A monitoring system comprises a video acquisition module, a deep learning identification detection module, a frame difference module and a subsequent logic module. The video acquisition module is used for acquiring video data of substation field operation personnel, then inputting video images into the deep learning identification detection module for data identification, and transmitting detection and identification results into the rear frame difference module for data comparison. A comparison result is transmitted to a follow-up logic module to judge whether the behavior of the substation field operation personnel is safe or not. According to the substation operation personnel safety monitoring system based on video analysis, a frame difference method and a deep learning identification detection method are combined; the transformer substation environment is monitored in real time to analyze whether field operators correctly wear safety helmets, whether safety belts are hung at low positions and used at high positions, whether construction behaviors are compliant, whether dangerous behaviors exist and the like, and the construction standard and safety of the transformer substation operators are ensured.

Description

Technical Field

[0001] The present invention relates to the field of video image technology recognition technology, and in particular to a method for safety monitoring of transformer substation operators based on video analysis. Background Art

[0002] In the power industry, workers often face potentially dangerous working environments such as high altitudes and high voltages, which may pose a serious threat to their safety. In order to reduce these risks, the industry has formulated a series of detailed safety operating specifications to ensure that workers can follow safety standards when performing their tasks. Despite these specifications, in actual operations, some construction workers may ignore safety measures for various reasons, such as not wearing safety helmets, safety belts, or not using safety hooks as required, all of which may lead to serious safety accidents.

[0003] In order to strengthen on-site safety management, construction units usually assign special patrol personnel to supervise the operation and ensure that safety regulations are followed; however, due to the particularity of high-altitude operations, there are certain limitations in supervision from the ground, because high-altitude workers may not be within the line of sight of ground supervisors, which makes supervision more difficult. It is difficult for supervisors to fully understand the dynamics of construction workers, and the construction environment of substations is a relatively dangerous area. Small construction errors can lead to serious accidents. For this reason, the applicant proposes a substation operator safety monitoring method based on video analysis according to the monitoring needs of substation operators. It conducts real-time monitoring and analysis of the substation environment to determine whether on-site workers wear safety helmets correctly, whether safety belts are hung low and used high, whether construction behaviors are compliant, and whether there are any dangerous behaviors, so as to ensure the safety of substation operators' construction specifications. Summary of the invention

[0004] To solve the above technical problems, the present invention proposes a safety monitoring method for substation workers based on video analysis. The monitoring system is equipped with a video acquisition module, a deep learning recognition and detection module, a frame difference module and a subsequent logic module; the video data of the substation on-site workers are collected by the video acquisition module, and then the video image is input into the deep learning recognition and detection module for data recognition, and the detection and recognition result is passed to the post-frame difference module for data comparison, and the comparison result is passed to the subsequent logic module to determine whether the behavior of the substation on-site workers is safe. The system combines the frame difference method with the deep learning recognition and detection method to perform real-time monitoring and analysis on the substation environment to determine whether the on-site workers wear safety helmets correctly, whether the safety belts are hung low and used high, whether the construction behavior is compliant and whether there is any dangerous behavior, etc., to ensure the safety of the construction specifications of the substation workers.

[0005] To achieve the above object, the technical solution adopted by the present invention is:

[0006] A method for monitoring the safety of substation workers based on video analysis, characterized in that: the substation worker safety monitoring system based on video analysis includes a video acquisition module, a deep learning recognition detection module, a frame difference module and a subsequent logic module; the system working steps are:

[0007] Step 1: Install a surveillance video camera with a video acquisition module in the substation operator area that needs to be monitored;

[0008] Step 2: The video acquisition module acquires video images in real time and transmits the video images to the deep learning recognition and detection module;

[0009] Step 3: The deep learning recognition and detection module uses the Yolov5 model, CSPDarknet53 as the backbone network, SPPF as the additional module of Neck, CSP-PAN as the feature fusion module of the Neck network, and uses YOLOv3 as the head network to perform data recognition processing on the video image;

[0010] Step 4: The data processed by the deep learning recognition and detection module is transferred to the frame difference module for comparison processing;

[0011] Step 5: The data processed by the frame difference module is transmitted to the subsequent logic module to determine whether the substation site workers wear safety helmets correctly, whether the safety belts are hung low and used high, whether the construction behavior is compliant and whether there are any dangerous behaviors, etc., and an alarm message is issued in case of violation;

[0012] Step 6: After receiving the alarm information, the management personnel will make rectifications on site and eliminate the alarm.

[0013] Furthermore, the Yolov5 model training adopted by the deep learning recognition and detection module in step 3 of the substation operator safety monitoring system based on video analysis is specifically as follows:

[0014] 1) Use a camera to collect pictures of the work site and mark them according to five categories: people, helmets, safety belts, poles and safety hooks;

[0015] 2) Construct a Yolov5 target detection network model, which mainly includes input end, backbone, and loss function;

[0016] 3) The input end uses the Mosaic data augmentation method, which increases the diversity of the dataset and the generalization ability of the model by randomly cropping and splicing four different pictures into a new picture. This method not only enriches the background of the picture, but also indirectly increases the batch size, so that more pictures can be calculated during batch normalization, thereby reducing the dependence on a large batch size;

[0017] 4) Using the backbone function, the backbone of YOLOv5 adopts the New CSP-Darknet53 structure, which is an improvement on the Darknet architecture used in previous versions. At the beginning of the backbone, YOLOv5 introduces the Focus structure, which achieves downsampling of feature maps and increases the number of channels through slicing and splicing operations, thereby improving the model's ability to detect small targets; the CSP structure reduces the amount of calculation and the number of parameters while maintaining the efficiency of feature extraction; the Bottleneck CSP module, which is a key component in the backbone, consists of multiple Conv modules and a Bottleneck module, and enhances the feature extraction capability of the network through residual connections and feature fusion; the SPPF (Spatial Pyramid Pooling with Feature map Fusion) module is an improved version of the SPP (Spatial Pyramid Pooling), which is used for multi-scale feature fusion. Through pooling operations at different scales, feature maps of different scales are fused into feature maps of unified scale, thereby improving the robustness of the network; PANet enhances the representation capability of features by aggregating features at multiple scales;

[0018] 5) Use the loss function of YOLOv5, which mainly includes: classification loss, target confidence loss and bounding box regression loss;

[0019] 6) Using the classification loss function, YOLOv5 uses binary cross-entropy loss to handle multi-label classification problems, allowing each detection box to predict multiple categories. The binary cross-entropy loss formula can be expressed as:

[0020]

[0021] Where: N is the total number of samples in the data set;

[0022] M is the total number of categories;

[0023] yic is a binary indicator (0 or 1), which is 1 if sample i belongs to category c; otherwise, it is 0;

[0024] Pic is the probability that the model predicts that sample i belongs to category c;

[0025] 7) Use the target confidence loss function. In YOLOv5, the calculation of the target confidence loss may also consider the IoU (intersection over union) between the predicted box and the true box to adjust the loss value; if the IoU between the predicted box and the true box is high, the model's confidence prediction of the predicted box is more accurate, and the loss is smaller; conversely, if the IoU is low, the loss is greater. The formula for the target confidence loss can be expressed as:

[0026]

[0027] Where: Npos is the number of positive samples;

[0028] Nneg is the number of negative samples;

[0029] pi is the confidence prediction of the i-th prediction box;

[0030] λneg is the weight of negative sample loss, which is used to balance the loss of positive and negative samples;

[0031] 8) Using the bounding box regression loss function, YOLOv5 uses CIoU (Complete Intersection over Union) as the loss function for bounding box regression; the CIoU loss function is an improved IoU loss function, which not only considers the overlapping area of ​​the bounding box, but also the size, scale and direction of the bounding box, thereby improving the accuracy of the bounding box prediction; the calculation formula of the CIoU loss function is as follows:

[0032]

[0033] Where: IoU represents the intersection over union ratio between the predicted bounding box and the true bounding box;

[0034] ρcio represents the Euclidean distance from the center point of the predicted bounding box to the center point of the true bounding box;

[0035] c represents the diagonal length of the true bounding box, which is used to normalize the center point distance;

[0036] α and β are weights used to balance the size and scale factors, which are usually adjusted dynamically according to the size of IoU;

[0037] 9) Input the labeled data in 1) into the model in 2) and train the model using stochastic gradient descent until the model converges.

[0038] Furthermore, the data processing of the frame difference module in step 4 of the substation operator safety monitoring system based on video analysis is specifically as follows: in a video sequence, when an object moves, there will be differences in the grayscale values ​​of three consecutive frames. By calculating the absolute value of the grayscale difference between two frames of images, a stationary object will appear as zero in the difference image, while a moving object, especially its contour, will appear as a non-zero value due to the grayscale change; once these absolute values ​​exceed a preset threshold, the moving target can be identified and target detection can be achieved; the advantage of the inter-frame difference method is that the algorithm is simple and easy to implement, and the program design difficulty is low; it is insensitive to environmental factors such as lighting changes, can adapt to a variety of dynamic scenes, and exhibits high robustness. The calculation is specifically as follows:

[0039] 1) Select three consecutive frames from the video sequence, denoted as frame 1, frame 2 and frame 3;

[0040] 2) Convert the three images into grayscale images for pixel-level comparison; the pixel values ​​of grayscale images are usually between 0 and 255;

[0041] 3) Calculate the frame difference. First, calculate the pixel difference between frame 1 and frame 2 to obtain the difference image 1:

[0042] D1(x,y)=|I1(x,y)-I2(x,y)|; then calculate the pixel difference between frame 2 and frame 3 to get the difference image 2:

[0043] D2(x,y)=|I2(x,y)-I3(x,y)|where I1, I2, and I3 represent the grayscale values ​​of the pixel position (x,y) in frame 1, frame 2, and frame 3, respectively;

[0044] 4) Take the larger value between the two difference images, D final (D1(x,y),D2(x,y));

[0045] 5) Apply thresholding to the final difference image to distinguish between moving and stationary areas. Only when the pixel value of the difference image exceeds the set threshold, the pixel is considered to belong to the moving area;

[0046] 6) Erosion and then dilation of the thresholded image to remove noise and fill small spaces;

[0047] 7) Finally, the contour and position of the moving target are extracted based on the processed image.

[0048] Furthermore, the data processing of the subsequent logic module in step 5 of the substation operator safety monitoring system based on video analysis is specifically as follows:

[0049] 1) If the result of the frame difference method indicates that the worker only moves horizontally and there is no tower in the detection result, it is considered that the worker is on the ground and the worker who is not wearing a safety helmet is considered to be in violation and a warning is issued.

[0050] 2) If the result of the frame difference method indicates that a worker is moving vertically and is on the pole tower, then the worker is considered to be climbing the pole tower and marked. It is necessary to detect whether the worker is wearing a safety helmet and a safety belt. When the left and right sides of the safety belt detection frame are beyond the pole tower, it means that the worker is using the waist rope of the safety belt correctly during the climbing process. Otherwise, it is considered a violation and a warning is issued;

[0051] 3) If the marking operator has reached or is close to the top of the tower detection frame, it is considered that the operator has reached the construction area and needs to check whether the safety hook is higher than the human body. If it is not high enough, it is considered a violation and a warning is issued;

[0052] 4) If the marking operator's construction actions exceed the normal construction range or abnormal actions are detected, a warning will be issued.

[0053] The benefits of this application are:

[0054] 1. The substation operator safety monitoring system based on video analysis uses a camera to obtain video data from the operation site, and realizes real-time monitoring of the operator's safety through analysis of the video data. Based on video data analysis, it realizes real-time monitoring of the operator's safety;

[0055] 2. Substation operator safety monitoring system based on video analysis Operator safety monitoring is automatically identified by the algorithm and does not require manual prior calibration;

[0056] 3. The substation operator safety monitoring system based on video analysis uses the results of the deep learning recognition detection model and the frame difference method module to improve the accuracy of identifying operator safety monitoring through post-processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 Schematic diagram of the working process of the system of the present invention. DETAILED DESCRIPTION

[0058] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments:

[0059] like Figure 1 As shown, a substation operator safety monitoring method based on video analysis is shown, and the substation operator safety monitoring system based on video analysis includes a video acquisition module, a deep learning recognition detection module, a frame difference module and a subsequent logic module; the system working steps are as follows:

[0060] Step 1: Install a surveillance video camera with a video acquisition module in the substation operator area that needs to be monitored; the camera can be an existing camera in the substation, or a dedicated camera can be used; the video analysis server does not refer to a specific computer model, but can be a single computer or a distributed computer system, as well as the necessary computer software system based on the computer hardware; the camera collects video stream data and transmits it to the video analysis server, which converts the video stream into a continuous image sequence;

[0061] Step 2: The video acquisition module acquires video images in real time and transmits the video images to the deep learning recognition and detection module;

[0062] Step 3: The deep learning recognition and detection module uses the Yolov5 model, CSPDarknet53 as the backbone network, SPPF as the additional module of Neck, CSP-PAN as the feature fusion module of the Neck network, and uses YOLOv3 as the head network to perform data recognition processing on the video image;

[0063] Step 4: The data processed by the deep learning recognition and detection module is transferred to the frame difference module for comparison processing;

[0064] Step 5: The data processed by the frame difference module is transmitted to the subsequent logic module to determine whether the substation site workers wear safety helmets correctly, whether the safety belts are hung low and used high, whether the construction behavior is compliant and whether there are any dangerous behaviors, etc., and an alarm message is issued in case of violation;

[0065] Step 6: After receiving the alarm information, the management personnel will make rectifications on site and eliminate the alarm.

[0066] The Yolov5 model training steps used in the deep learning recognition and detection module in step 3 of the substation operator safety monitoring system based on video analysis are as follows:

[0067] 1) Use a camera to collect pictures of the work site and mark them according to five categories: people, helmets, safety belts, poles and safety hooks;

[0068] 2) Construct a Yolov5 target detection network model, which mainly includes input end, backbone, and loss function;

[0069] 3) The input end uses the Mosaic data augmentation method, which increases the diversity of the dataset and the generalization ability of the model by randomly cropping and splicing four different pictures into a new picture. This method not only enriches the background of the picture, but also indirectly increases the batch size, so that more pictures can be calculated during batch normalization, thereby reducing the dependence on a large batch size;

[0070] 4) Using the backbone function, the backbone of YOLOv5 adopts the New CSP-Darknet53 structure, which is an improvement on the Darknet architecture used in previous versions. At the beginning of the backbone, YOLOv5 introduces the Focus structure, which achieves downsampling of feature maps and increases the number of channels through slicing and splicing operations, thereby improving the model's ability to detect small targets; the CSP structure reduces the amount of calculation and the number of parameters while maintaining the efficiency of feature extraction; the Bottleneck CSP module, which is a key component in the backbone, consists of multiple Conv modules and a Bottleneck module, and enhances the feature extraction capability of the network through residual connections and feature fusion; the SPPF (Spatial Pyramid Pooling with Feature map Fusion) module is an improved version of the SPP (Spatial Pyramid Pooling), which is used for multi-scale feature fusion. Through pooling operations at different scales, feature maps of different scales are fused into feature maps of unified scale, thereby improving the robustness of the network; PANet enhances the representation capability of features by aggregating features at multiple scales;

[0071] 5) Use the loss function of YOLOv5, which mainly includes: classification loss, target confidence loss and bounding box regression loss;

[0072] 6) Using the classification loss function, YOLOv5 uses binary cross-entropy loss to handle multi-label classification problems, allowing each detection box to predict multiple categories. The binary cross-entropy loss formula can be expressed as:

[0073]

[0074] Where: N is the total number of samples in the data set;

[0075] M is the total number of categories;

[0076] yic is a binary indicator (0 or 1), which is 1 if sample i belongs to category c; otherwise, it is 0;

[0077] Pic is the probability that the model predicts that sample i belongs to category c;

[0078] 7) Use the target confidence loss function. In YOLOv5, the calculation of the target confidence loss may also consider the IoU (intersection over union) between the predicted box and the true box to adjust the loss value; if the IoU between the predicted box and the true box is high, the model's confidence prediction of the predicted box is more accurate, and the loss is smaller; conversely, if the IoU is low, the loss is greater. The formula for the target confidence loss can be expressed as:

[0079]

[0080] Where: Npos is the number of positive samples;

[0081] Nneg is the number of negative samples;

[0082] pi is the confidence prediction of the i-th prediction box;

[0083] λneg is the weight of negative sample loss, which is used to balance the loss of positive and negative samples;

[0084] 8) Using the bounding box regression loss function, YOLOv5 uses CIoU (Complete Intersection over Union) as the loss function for bounding box regression; the CIoU loss function is an improved IoU loss function, which not only considers the overlapping area of ​​the bounding box, but also the size, scale and direction of the bounding box, thereby improving the accuracy of the bounding box prediction; the calculation formula of the CIoU loss function is as follows:

[0085]

[0086] Where: IoU represents the intersection over union ratio between the predicted bounding box and the true bounding box;

[0087] ρcio represents the Euclidean distance from the center point of the predicted bounding box to the center point of the true bounding box;

[0088] c represents the diagonal length of the true bounding box, which is used to normalize the center point distance;

[0089] α and β are weights used to balance the size and scale factors, which are usually adjusted dynamically according to the size of IoU;

[0090] 9) Input the labeled data in 1) into the model in 2) and train the model using stochastic gradient descent until the model converges.

[0091] The data processing of the frame difference module in step 4 of the substation operator safety monitoring system based on video analysis is as follows: in the video sequence, when an object moves, the grayscale values ​​of three consecutive frames will be different. By calculating the absolute value of the grayscale difference between two frames, the stationary object will appear as zero in the difference image, while the moving object, especially its contour, will appear as a non-zero value due to the grayscale change; once these absolute values ​​exceed the preset threshold, the moving target can be identified and the target detection can be achieved; the advantage of the inter-frame difference method is that the algorithm is simple and easy to implement, and the program design difficulty is low; it is insensitive to environmental factors such as lighting changes, can adapt to a variety of dynamic scenes, and show high robustness. Its calculation is as follows:

[0092] 1) Select three consecutive frames from the video sequence, denoted as frame 1, frame 2 and frame 3;

[0093] 2) Convert the three images into grayscale images for pixel-level comparison; the pixel values ​​of grayscale images are usually between 0 and 255;

[0094] 3) Calculate the frame difference. First, calculate the pixel difference between frame 1 and frame 2 to obtain the difference image 1:

[0095] D1(x,y)=|I1(x,y)-I2(x,y)|; then calculate the pixel difference between frame 2 and frame 3 to get the difference image 2:

[0096] D2(x,y)=|I2(x,y)-I3(x,y)|where I1, I2, and I3 represent the grayscale values ​​of the pixel position (x,y) in frame 1, frame 2, and frame 3, respectively;

[0097] 4) Take the larger value between the two difference images, D final (D1(x,y),D2(x,y));

[0098] 5) Apply thresholding to the final difference image to distinguish between moving and stationary areas. Only when the pixel value of the difference image exceeds the set threshold, the pixel is considered to belong to the moving area;

[0099] 6) Erosion and then dilation of the thresholded image to remove noise and fill small spaces;

[0100] 7) Finally, the contour and position of the moving target are extracted based on the processed image.

[0101] The data processing of the subsequent logic modules in step 5 of the substation operator safety monitoring system based on video analysis is specifically as follows:

[0102] 1) If the result of the frame difference method indicates that the worker only moves horizontally and there is no tower in the detection result, it is considered that the worker is on the ground and the worker who is not wearing a safety helmet is considered to be in violation and a warning is issued.

[0103] 2) If the result of the frame difference method indicates that a worker is moving vertically and is on the pole tower, then the worker is considered to be climbing the pole tower and marked. It is necessary to detect whether the worker is wearing a safety helmet and a safety belt. When the left and right sides of the safety belt detection frame are beyond the pole tower, it means that the worker is using the waist rope of the safety belt correctly during the climbing process. Otherwise, it is considered a violation and a warning is issued;

[0104] 3) If the marking operator has reached or is close to the top of the tower detection frame, it is considered that the operator has reached the construction area and needs to check whether the safety hook is higher than the human body. If it is not high enough, it is considered a violation and a warning is issued;

[0105] 4) If the marking operator's construction actions exceed the normal construction range or abnormal actions are detected, a warning will be issued.

[0106] The above description is only a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent change made based on the technical essence of the present invention still falls within the scope of protection required by the present invention.

Claims

1. A method for safety monitoring of substation operators based on video analysis, characterized in that: The system working steps are: Step 1: Install a surveillance video camera with a video acquisition module in the substation operator area that needs to be monitored; Step 2: The video acquisition module acquires video images in real time and transmits the video images to the deep learning recognition and detection module; Step 3: The deep learning recognition and detection module uses the Yolov5 model, CSPDarknet53 as the backbone network, SPPF as the additional module of Neck, CSP-PAN as the feature fusion module of the Neck network, and uses YOLOv3 as the head network to perform data recognition processing on the video image; Step 4: The data processed by the deep learning recognition and detection module is transferred to the frame difference module for comparison processing; Step 5: The data processed by the frame difference module is transmitted to the subsequent logic module to determine whether the substation site workers wear safety helmets correctly, whether the safety belts are hung low and used high, whether the construction behavior is compliant and whether there are any dangerous behaviors, etc., and an alarm message is issued in case of violation; Step 6: After receiving the alarm information, the management personnel will make rectifications on site and eliminate the alarm.

2. A method for safety monitoring of substation operators based on video analysis according to claim 1, characterized in that: The Yolov5 model training adopted by the deep learning recognition and detection module in step 3 of the substation operator safety monitoring system based on video analysis is specifically as follows: 1) Use a camera to collect pictures of the work site and mark them according to five categories: people, helmets, safety belts, poles and safety hooks; 2) Construct a Yolov5 target detection network model, which includes an input end, a backbone, and a loss function; 3) The Mosaic data augmentation method is used at the input end to increase the diversity of the data set and the generalization ability of the model by randomly cropping and splicing four different pictures into a new picture. This method not only enriches the background of the picture, but also indirectly increases the batch size, so that more pictures can be calculated during batch normalization, thereby reducing the dependence on large batch sizes; 4) Using the backbone function, the backbone of YOLOv5 adopts the New CSP-Darknet53 structure, which is an improvement on the Darknet architecture used in previous versions. At the beginning of the backbone, YOLOv5 introduces the Focus structure, which achieves downsampling of feature maps and increases the number of channels through slicing and splicing operations, thereby improving the model's ability to detect small targets; the CSP structure reduces the amount of calculation and the number of parameters while maintaining the efficiency of feature extraction. BottleneckCSP module, which is a key component in the backbone. It consists of multiple Conv modules and a Bottleneck module. It enhances the feature extraction capability of the network through residual connection and feature fusion. The SPPF module is an improved version of the SPP, which is used for multi-scale feature fusion. Through pooling operations at different scales, feature maps of different scales are fused into feature maps of unified scale, which improves the robustness of the network. PANet enhances the representation ability of features by aggregating features at multiple scales; 5) Use the loss functions of YOLOv5, including classification loss, target confidence loss, and bounding box regression loss; 6) Using the classification loss function, YOLOv5 uses binary cross entropy loss to handle multi-label classification problems, allowing each detection box to predict multiple categories. The binary cross entropy loss formula is expressed as: Where: N is the total number of samples in the data set; M is the total number of categories; yic is a binary indicator (0 or 1), which is 1 if sample i belongs to category c; otherwise, it is 0; Pic is the probability that the model predicts that sample i belongs to category c; 7) Use the target confidence loss function. In YOLOv5, the calculation of the target confidence loss may also consider the IoU intersection-over-union ratio between the predicted box and the true box to adjust the loss value; if the IoU between the predicted box and the true box is high, the model's confidence prediction of the predicted box is more accurate and the loss is smaller; conversely, if the IoU is low, the loss is greater. The formula for the target confidence loss is expressed as: Where: Npos is the number of positive samples; Nneg is the number of negative samples; pi is the confidence prediction of the i-th prediction box; λneg is the weight of negative sample loss, which is used to balance the loss of positive and negative samples; 8) Using the bounding box regression loss function, YOLOv5 uses CIoU as the loss function for bounding box regression; the CIoU loss function is an improved IoU loss function that takes into account the overlapping area of ​​the bounding box, as well as the size, scale, and direction of the bounding box, thereby improving the accuracy of the bounding box prediction; the calculation formula of the CIoU loss function is as follows: Where: IoU represents the intersection over union ratio between the predicted bounding box and the true bounding box; ρcio represents the Euclidean distance from the center point of the predicted bounding box to the center point of the true bounding box; c represents the diagonal length of the true bounding box, which is used to normalize the center point distance; α and β are weights used to balance the size and scale factors, which are usually adjusted dynamically according to the size of IoU; 9) Input the labeled data in 1) into the model in 2) and train the model using stochastic gradient descent until the model converges.

3. A method for safety monitoring of substation operators based on video analysis according to claim 1, characterized in that: The data processing of the frame difference module in step 4 of the substation operator safety monitoring system based on video analysis is specifically as follows: 1) Select three consecutive frames from the video sequence, denoted as frame 1, frame 2 and frame 3; 2) Convert the three images into grayscale images for pixel-level comparison; the pixel values ​​of grayscale images are usually between 0 and 255; 3) Calculate the frame difference. First, calculate the pixel difference between frame 1 and frame 2 to obtain the difference image 1: D1(x,y)=|I1(x,y)-I2(x,y)|; then calculate the pixel difference between frame 2 and frame 3 to get the difference image 2: D2(x,y)=|I2(x,y)-I3(x,y)|where I1, I2, and I3 represent the grayscale values ​​of the pixel position (x,y) in frame 1, frame 2, and frame 3, respectively; 4) Take the larger value between the two difference images, D final (D1(x,y),D2(x,y)); 5) Apply thresholding to the final difference image to distinguish between moving and stationary areas. Only when the pixel value of the difference image exceeds the set threshold, the pixel is considered to belong to the moving area; 6) Erosion and then dilation of the thresholded image to remove noise and fill small spaces; 7) Finally, the contour and position of the moving target are extracted based on the processed image.

4. A method for safety monitoring of substation operators based on video analysis according to claim 1, characterized in that: The data processing of the subsequent logic modules in step 5 of the substation operator safety monitoring system based on video analysis is specifically as follows: 1) If the result of the frame difference method indicates that the worker only moves horizontally and there is no tower in the detection result, it is considered that the worker is on the ground and the worker who is not wearing a safety helmet is considered to be in violation and a warning is issued. 2) If the result of the frame difference method indicates that a worker is moving vertically and is on the pole tower, then the worker is considered to be climbing the pole tower and marked. It is necessary to detect whether the worker is wearing a safety helmet and a safety belt. When the left and right sides of the safety belt detection frame are beyond the pole tower, it means that the worker is using the waist rope of the safety belt correctly during the climbing process. Otherwise, it is considered a violation and a warning is issued; 3) If the marking operator has reached or is close to the top of the tower detection frame, it is considered that the operator has reached the construction area and needs to check whether the safety hook is higher than the human body. If it is not high enough, it is considered a violation and a warning is issued; 4) If the marking operator's construction actions exceed the normal construction range or abnormal actions are detected, a warning will be issued.

Images