Training method of illegal fire detection model and detection method using the model

Abstract

This invention discloses a training method for a non-compliant hot work detection model and a detection method using the model. The training method includes the following steps: S1 forming a training set of images; S2 obtaining the non-compliant hot work detection model; the detection method includes the following steps: S1 obtaining the category of each target in the image and its detection box result; S2 determining whether non-compliant hot work behavior exists. Beneficial effects: The training method for the non-compliant hot work detection model disclosed in this invention overcomes the problems of missed detections and false detections, significantly improving the accuracy of on-site hot work image detection results; the non-compliant hot work detection method disclosed in this invention significantly improves the accuracy of judging non-compliant hot work behavior.

Description

Technical fields:

[0001] This invention patent belongs to the field of detection technology for illegal hot work at hot work sites, specifically involving the training method of an illegal hot work detection model and the detection method using the model. Background technology:

[0002] Hot work refers to temporary work that may generate sparks, flames, and hot surfaces. Common hot work includes electric welding and gas welding (thermal cutting). If hot work is not properly managed, it can easily lead to fires or even explosions. Therefore, safety is the primary concern. Safety assurance must protect not only the personal safety of hot work personnel and supervisors but also the safety of property. In hot work management, "utilizing advanced science and technology to improve safety early warning" has become a trend. Among these measures, using video images captured at the work site and artificial intelligence to automatically analyze violations and identify violators during hot work, and issuing timely warnings, can, to a certain extent, ensure the standardization and safety of hot work and prevent accidents.

[0003] Currently, there are several advanced methods for automatically detecting flames or smoke, such as single-stage deep learning-based target detection algorithms like the YOLO series and two-stage deep learning algorithms like the R-CNN series, which can achieve intelligent detection of flames or smoke. Two-stage target detection algorithms generate a series of candidate bounding boxes using region candidate box methods, and then classify the samples using convolutional neural networks, resulting in high detection accuracy. Single-stage target detection algorithms integrate the tasks of extracting candidate bounding boxes and classification into a single network, directly transforming the target bounding box localization problem into a regression problem, thus improving the algorithm's real-time performance.

[0004] However, the detection of unauthorized hot work primarily assesses whether the hot work site meets the requirements and whether the hot work operation is in violation of regulations. It requires not only detecting flames of varying sizes and shapes, but also simultaneously checking whether personnel are wearing protective measures (such as face shields), whether there is a supervisor present, and whether fire extinguishers are available. The two deep learning-based detection algorithms mentioned above cannot effectively detect unauthorized hot work at the work site. During the detection process, the large differences in the size and shape of the targets—for example, the varied postures of hot work operators (some standing, some crouching); the occasional bright spark during welding (sometimes a single bright spot, sometimes a large area of ​​bright sparks); and the influence of the work site background and lighting—all contribute to missed and false detections, thus affecting the accuracy of judging unauthorized hot work. Summary of the Invention:

[0005] One objective of this invention is to provide a training method for a non-compliant hot work detection model that improves the accuracy of on-site hot work image detection results.

[0006] Another objective of this invention is to provide a method for detecting unauthorized hot work that accurately determines such behavior.

[0007] The technical solution of this invention discloses, in one aspect, a training method for a detection model of unauthorized hot work, which includes the following steps:

[0008] S1: Organize the images of the hot work site into an image dataset, and annotate the flames, fire extinguishers, personnel, and protective masks in the images with ground truth bounding boxes and ground truth class labels; the annotated images form a training set;

[0009] S2: Train the pre-built Faster R-CNN object detection model using the training set described in step S1 to obtain the unauthorized hot work detection model; wherein the loss function L used in the pre-built Faster R-CNN object detection model is:

[0010] L=λ1L RPN +λ2L Fast RCNN +λ3L dist

[0011] Where L RPN It is the loss generated by the region recommendation network in recommending candidate regions; L Fast RCNN It is the loss caused by the detection; L dist This invention introduces constraints on the position of the detection boxes for the flame and protective mask based on prior knowledge of the positions of the hot work operator, protective mask, and flame in a hot work scenario. These constraints guide the model during training to extract features of the flame and protective mask within a more reasonable area, while simultaneously optimizing the position of the detection boxes to improve the weights of the trained model. λ1, λ2, and λ3 are three hyperparameters used to adjust L... RPN L Fast RCNN and L dist The weights of the three functions are adjusted.

[0012] Furthermore, in step S2, L RPN Defined as:

[0013]

[0014] Where N cls λ represents the total number of samples in a minimum batch size. 11 N represents the balancing weight parameter. regIndicates the number of pixels in each feature map; i represents the index of the anchor point in a batch; p i Predict the probability that the anchor point is a foreground target; Labels representing actual values; The logarithmic loss between the predicted and actual labeled values ​​is represented by... To calculate; It is regression loss, using To calculate, R is the smooth L1 loss function:

[0015]

[0016] t i ={t x ,t y ,t w ,t h} is a vector representing the four parameters of the predicted detection box, (t x ,t y ) represents the coordinates of the center point of the predicted detection box, t w and t h These represent the width and height of the predicted detection box, respectively; Is with t i Vectors of the same dimension This represents the center point coordinates, width, and height values ​​of the actual bounding box.

[0017] Furthermore, in step S2, L Fast RCNN Defined as:

[0018] L Fast RCNN =L cls (p,u)+λ 22 [u≥1]L IOC (t u ,v)

[0019] Where p is the predicted probability of all target classes; u is the true probability of all target classes. λ 22 L represents the balancing weight parameter. cls (p,u) is the cross-entropy loss function for multi-class classification:

[0020]

[0021] Where N is the number of detected instances, M is the number of categories, and u ij p represents the true probability that the i-th instance belongs to class j, and is a binary index; ij The probability that the i-th instance predicted by the model belongs to class j;

[0022] L IOC Representing regression loss, using To calculate, where where t u The predicted position value of the target box is represented by ; v represents the actual position value of the target box; (x, y) are the coordinates of the center point of the target box, and w and h are the width and height of the target box, respectively.

[0023] Furthermore, in step S2, L dist Defined as:

[0024] L dist =u k1 ·u m2 [(x k -x m ) 2 +(y k -y m ) 2 ]+u k1 ·u h3 [(x k -x h ) 2 +(y k -y h ) 2 ]

[0025] The categories are: personnel (value 1), protective mask (value 2), flame (value 3), and fire extinguisher (value 4); k1 u represents the true probability value that the k-th detected instance belongs to the personnel category and is a hot work operator. m2 u represents the true probability that the m-th detected instance is a protective face mask. h3 This represents the true probability that the h-th detected instance is of the flame category; (x k ,y k ), (x m ,y m ), (x h ,y h ) represent the predicted center point coordinates of the hot work operator, protective mask, and flame instance, respectively; where u k1 ·u m2 [(x k -x m ) 2 +(y k -y m ) 2 [This refers to the distance between the hot work operator and the location of the protective mask detection frame.] k1 ·u h3 [(x k -x h )2 +(y k -y h ) 2 This refers to the distance between the hot work operator and the flame detection frame.

[0026] Furthermore, Where u i1 This represents the true probability value that the i-th detected instance belongs to the personnel category, (x i ,y i ), (x m ,y m ) represent the predicted personnel and the center point coordinates of the protective mask, respectively.

[0027] Another aspect of the technical solution of the present invention discloses a method for detecting unauthorized hot work, which includes the following steps:

[0028] S1: Use the illegal hot work detection model to detect the input hot work site video image and obtain the category of each target in the image and its detection box result;

[0029] S2: A comprehensive judgment is made on the category of each target in the image in step S1 and the results of its detection box to determine whether there is any illegal hot work.

[0030] Furthermore, in step S1, the real-time hot work site video image is input into the Faster RCNN detection algorithm using the illegal hot work detection model to detect personnel, fire extinguishers, protective masks and flames in the hot work site video image, and obtain the category of each target in the image and its detection box result.

[0031] Furthermore, the specific method for determining step S2 is as follows:

[0032] S21: If both personnel and flame targets are detected in the image, it is determined to be a hot work site;

[0033] S22: Based on step S21, determine whether a target of the protective face mask category is detected simultaneously. If not, an alarm is triggered indicating that a protective face mask is not being worn. If it is, further determine whether the protective face mask is being worn on the face of the detected personnel instance. If not, it is considered a violation and an alarm is triggered. If it is, it indicates that a personnel instance is wearing a protective face mask.

[0034] S23: Based on the detection in step S22 that a person is wearing a protective mask, determine whether there are other persons besides that person. If not, it means that there is no monitoring personnel on site, which is also a violation of regulations and an alarm will be triggered. If there are, it means that there are monitoring personnel on site.

[0035] S24: Based on the detection of a monitoring personnel in step S23, determine whether the target category of the fire extinguisher has been detected. If not, it is considered an unauthorized hot work and an alarm is triggered; if so, it is considered a normal hot work.

[0036] Furthermore, in step S22, the method for determining whether the protective mask is worn on the face of the detected personnel instance is as follows: the intersection area φ is calculated sequentially based on the detection position of the protective mask and the upper part of the detection frame of all detected personnel instances. If each φ = 0, it means that none of the personnel instances are wearing protective masks, which is a violation and an alarm is triggered. If any φ > 0, it means that a personnel instance is detected wearing a protective mask.

[0037] Furthermore, φ = S A ∩S B ;where S A S represents the area of ​​the upper half of the detection bounding box for a given person instance. B This refers to the area of ​​the detected protective face shield's bounding box region.

[0038] Advantages of this invention:

[0039] 1. The training method for the unauthorized hot work detection model disclosed in this invention incorporates prior knowledge of the positions of the hot work operator, protective mask, and flame at the hot work site into the model's loss function, based on prior knowledge of these positions. dist This guides the training process of the Faster R-CNN target detection model, enabling the unauthorized hot work model in this invention to effectively detect hot work operators, protective masks, flames, and fire extinguishers, thereby overcoming the problems of missed detections and false detections and significantly improving the accuracy of on-site hot work image detection results.

[0040] 2. The illegal hot work detection method disclosed in this invention makes full use of the categories of each target in the image detected by the illegal hot work detection model and the results of its detection boxes to make logical judgments, thereby determining whether there is any illegal hot work behavior at the hot work site, which greatly improves the accuracy of the judgment of illegal hot work behavior. Attached image description:

[0041] Figure 1 This is a flowchart of the training method for the unauthorized hot work detection model in Embodiment 1 of the present invention.

[0042] Figure 2 This is a training framework diagram of the Faster R-CNN object detection model in Embodiment 1 of the present invention.

[0043] Figure 3 This is a flowchart of the unauthorized hot work detection method according to Embodiment 2 of the present invention.

[0044] Figure 4This is a logic diagram of the unauthorized hot work detection method according to Embodiment 2 of the present invention. Detailed implementation method:

[0045] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0046] Example 1: Figure 1-2 As shown, the training method for the unauthorized hot work detection model includes the following steps:

[0047] S1: Extract image frames from the monitoring video data of the hot work site at preset time intervals and save them as hot work site scene images; organize the hot work site scene images into an image dataset, and use Labelme to annotate the flames, fire extinguishers, personnel, and protective mask targets in the images with real bounding boxes and real category labels; the above-annotated images form a training set.

[0048] S2: Train the pre-built Faster R-CNN target detection model using the training set described in step S1 to obtain the unauthorized hot work detection model.

[0049] S21: Use the ResNet-50 network model to extract features from the training set in step S1 and generate feature maps.

[0050] S22: Input the feature map from step S21 into the RPN network to obtain a set of region proposals. This includes sliding a 3×3 window point by point on the feature map. Each point in the feature map, i.e., the center point of each sliding window position, is mapped to the original image to generate candidate regions of interest with three different scales (128, 256, 512) and three different aspect ratios (1:1, 0.5:1, 1:0.5). Then, the anchors are determined to be foreground or background by a softmax layer. Finally, the anchors are corrected using bounding boxes to obtain accurate candidate region proposals.

[0051] S23: Using the feature map from step S21 and the candidate region proposal from step S22 as input, map the candidate region proposal to the corresponding position of the feature map, and perform RoI Pooling on the mapped feature map to obtain the proposed feature map.

[0052] The RoI Pooling operation is implemented by dividing the mapped feature map into blocks of the same size (the number of blocks is the same as the dimension of the output feature map), and then performing max pooling on each block to transform a set of candidate region feature maps of different sizes into feature maps of a fixed and uniform size, i.e., proposal feature maps.

[0053] S24: Using the proposed feature maps from step S23, calculate the category (e.g., personnel, fire extinguisher, protective mask, and flame) of each proposed feature map through the fully connected layer and Softmax function of the detection network, and output the predicted category probability; at the same time, use bounding box regression again to obtain the position offset of the proposed region to obtain a more accurate predicted detection box.

[0054] The loss function L used in the pre-built Faster R-CNN object detection model is:

[0055] L=λ1L RPN +λ2L Fast RCNN +λ3L dist

[0056] Where L RPN It is the loss generated by the region recommendation network in recommending candidate regions; L Fast RCNN It is the loss caused by the detection; L dist This invention introduces constraints on the position of the detection boxes for flames and protective masks based on prior knowledge of the positions of the hot work operator, protective mask, and flame in a hot work scenario. These constraints guide the model during training to extract features of the flame and protective mask within a more reasonable area, while simultaneously optimizing the position of the detection boxes, resulting in a more optimized weighting of the trained unauthorized hot work detection model. λ1, λ2, and λ3 are three hyperparameters used to adjust L... RPN L Fast RCNN and L dist The weights of the three functions are adjusted; in this embodiment, the specific values ​​of the three hyperparameters λ1, λ2 and λ3 are determined based on experience, then judged by experiments and modified, and finally λ1 = λ2 = 1 and λ3 = 0.1 are obtained.

[0057] Where L RPN Defined as:

[0058]

[0059] Where N cls N represents the total number of samples in a minimum batch size, in this embodiment. cls =256, λ 11 λ represents the balancing weight parameter, in this embodiment 11 =10, N reg N represents the number of pixels in each feature map; in this embodiment, N is the number of pixels in each feature map. reg ~2400; i represents the index of an anchor point in a batch; p i Predict the probability that the anchor point is a foreground target; Labels representing the actual values; when the label is the background. The value is 0 when it is labeled as foreground. The value is 1; The logarithmic loss between the predicted and actual labeled values ​​is represented by... To calculate; It is regression loss, using To calculate, R is the smooth L1 loss function:

[0060]

[0061] t i ={t x ,t y ,t w ,t h} is a vector representing the four parameters of the predicted detection box, (t x ,t y ) represents the coordinates of the center point of the predicted detection box, t w and t h These represent the width and height of the predicted detection box, respectively; Is with t i Vectors of the same dimension This represents the center point coordinates, width, and height values ​​of the actual bounding box.

[0062] Where L Fast RCNN Defined as:

[0063] L Fast RCNN =L cls (p,u)+λ 22 [u≥1]L IOC (t u ,v)

[0064] Where p is the predicted probability of all target classes; u is the true probability of all target classes. λ 22 λ represents the balancing weight parameter, in this embodiment 22 =10; L cls (p,u) is the cross-entropy loss function for multi-class classification:

[0065]

[0066] Where N is the number of detected instances, M is the number of categories, and u ij U represents the true probability that the i-th instance belongs to class j. It is a binary indicator. Based on the annotations in the training set, when the i-th instance belongs to class j, u ij =1, otherwise u ij=0; p ij The probability that the i-th instance predicted by the model belongs to class j is calculated by using a fully connected layer and the Softmax function to determine the class to which each proposed feature map belongs, and the output predicted class probability is then calculated.

[0067] L IOC Representing regression loss, using To calculate, where where t u The predicted position value of the target box is represented by ; v represents the actual position value of the target box; (x, y) are the coordinates of the center point of the target box, and w and h are the width and height of the target box, respectively.

[0068] Where L dist Defined as:

[0069] L dist =u k1 ·u m2 [(x k -x m ) 2 +(y k -y m ) 2 ]+u k1 ·u h3 [(x k -x h ) 2 +(y k -y h ) 2 ]

[0070] The personnel category is set to 1, the protective mask category to 2, the flame category to 3, and the fire extinguisher category to 4.

[0071] u k1 This represents the true probability that the k-th detected instance is a personnel operator. In this instance, since there are supervisors and other personnel present at the hot work site besides the hot work operator, it is necessary to identify which instance is the hot work operator. Based on prior knowledge, the location of the hot work operator should coincide with the detection location frame of the protective mask. Therefore, the minimum distance between the detection location of the personnel instance and the detection location of the protective mask is used to determine which personnel instance is the hot work operator. To exclude the influence of guardianship instances, where u i1 This represents the true probability value of the i-th detected instance being a person. In this embodiment, based on the true categories labeled in the training set, when the i-th detected instance is a person, u i1 =1, otherwise u i1 =0; (x i ,y i ), (x m,y m ) represent the predicted coordinates of the center point of the personnel and the protective mask, respectively; distance u i1 ·u m2 [(x i -x m ) 2 +(y i -y m ) 2 When the minimum value is reached, k = i. At this time, the i(k)th detection instance is a hot work operator, therefore u k1 =1.

[0072] u m2 This represents the true probability that the m-th detected instance is a face shield. In this embodiment, based on the true categories labeled in the training set, when the m-th detected instance is a face shield, u m2 =1, otherwise u m2 =0.

[0073] u h3 This represents the true probability that the h-th detected instance is of the flame category. In this embodiment, based on the true categories labeled in the training set, when the h-th detected instance is of the flame category, u h3 =1, otherwise u h3 =0.

[0074] (x k ,y k ), (x m ,y m ), (x h ,y h ) represent the predicted center point coordinates of the hot work operator, protective mask, and flame instance, respectively; where u k1 ·u m2 [(x k -x m ) 2 +(y k -y m ) 2 [This refers to the distance between the hot work operator and the location of the protective mask detection frame.] k1 ·u h3 [(x k -x h ) 2 +(y k -y h ) 2 This refers to the distance between the hot work operator and the flame detection frame.

[0075] The training set is used to perform 100,000 iterations of training according to the method in step S2. During each iteration, the error of the hidden layer is calculated using the backpropagation algorithm based on the model loss function L, and the parameters of the Faster R-CNN model (i.e., the parameters of the RPN network in step S22 and the parameters of the fully connected layer of the detection network in step S24) are updated using the gradient descent algorithm. The specific update formula is as follows: in w in the model j The updated result, w j η is the j-th weight parameter in the model; η is the learning rate, initialized to 0.0004, which decays to half of the original learning rate after every 10,000 iterations. This indicates the differentiation operation.

[0076] The training method for the unauthorized hot work detection model disclosed in this invention incorporates prior knowledge of the positions of the hot work operator, protective mask, and flame at the hot work site into the model's loss function, based on prior knowledge of these positions. dist This guides the training process of the Faster R-CNN target detection model, enabling the unauthorized hot work model in this invention to effectively detect hot work operators, protective masks, flames, and fire extinguishers, thereby overcoming the problems of missed detections and false detections and significantly improving the accuracy of on-site hot work image detection results.

[0077] Example 2: As Figure 3 As shown, the method for detecting unauthorized hot work using the unauthorized hot work detection model of Example 1 includes the following steps:

[0078] S1: Input the real-time hot work site video image into the Faster RCNN detection algorithm using the illegal hot work detection model to detect personnel, fire extinguishers, protective masks and flames in the hot work site video image, and obtain the category of each target in the image and its detection box result.

[0079] S2: A comprehensive judgment is made based on the category of each target in the image from step S1 and its detection box results to determine whether any unauthorized hot work has occurred. Figure 4 As shown, the specific steps for judgment are as follows:

[0080] S21: If both personnel and flame targets are detected in the image, it is determined to be a hot work site.

[0081] S22: Based on step S21, determine whether a target of the protective face mask category is detected simultaneously. If not, issue an alarm indicating that a protective face mask is not being worn. If it is detected, calculate the intersection area φ based on the detection position of the protective face mask and the upper part of the detection frame of all detected personnel instances, where φ = S A ∩S B S A S represents the area of ​​the upper half of the detection bounding box for a given person instance. B The area of ​​the bounding box region of the detected protective face shield; if each φ = 0, it means that the protective face shield is not worn on the face, which is also a violation and will trigger an alarm. If any φ > 0, it means that a person has been detected wearing a protective face shield.

[0082] S23: Based on the detection in step S22 that a person is wearing a protective mask, determine whether there are other persons besides that person. If not, it means that there is no monitoring personnel on site, which is also a violation of regulations and an alarm will be triggered. If there are, it means that there are monitoring personnel on site.

[0083] S24: Based on the detection of a monitoring personnel in step S23, determine whether the target category of the fire extinguisher has been detected. If not, it is considered an unauthorized hot work and an alarm is triggered; if so, it is considered a normal hot work.

[0084] The method for detecting unauthorized hot work disclosed in this invention makes full use of the categories of each target in the image detected by the unauthorized hot work detection model and the results of their detection boxes to make logical judgments, thereby determining whether there is unauthorized hot work at the hot work site, which greatly improves the accuracy of judging unauthorized hot work.

[0085] The above are preferred embodiments of the present invention. For those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A training method for a detection model of unauthorized hot work, characterized in that, It includes the following steps: S1: Organize the images of the hot work site into an image dataset, and annotate the flames, fire extinguishers, personnel, and protective masks in the images with ground truth bounding boxes and ground truth class labels; the annotated images form a training set; S2: Train the pre-built Faster R-CNN object detection model using the training set described in step S1 to obtain the unauthorized hot work detection model; wherein the loss function L used in the pre-built Faster R-CNN object detection model is: L=λ1L RPN +λ2L Fast RCNN +λ3L dist Where L RPN It is the loss generated by the region recommendation network in recommending candidate regions; L Fast RCNN It is the loss caused by the detection; L dist It is a constraint term for the position of the detection frame of the flame and protective mask, introduced based on the prior knowledge of the position of the hot work operator, protective mask and flame in the hot work scenario; λ1, λ2 and λ3 are three hyperparameters.

2. The training method for the unauthorized hot work detection model according to claim 1, characterized in that, In step S2, L RPN Defined as: Where N cls λ represents the total number of samples in a minimum batch size. 11 N represents the balancing weight parameter. reg Indicates the number of pixels in each feature map; i represents the index of the anchor point in a batch; p i Predict the probability that the anchor point is a foreground target; Labels representing actual values; The logarithmic loss between the predicted and actual labeled values ​​is represented by... To calculate; It is regression loss, using To calculate, R is the smooth L1 loss function: t i ={t x ,t y ,t w ,t h } is a vector representing the four parameters of the predicted detection box, (t x ,t y ) represents the coordinates of the center point of the predicted detection box, t w and t h These represent the width and height of the predicted detection box, respectively; Is with t i Vectors of the same dimension This represents the center point coordinates, width, and height values ​​of the actual bounding box.

3. The training method for the unauthorized hot work detection model according to claim 1, characterized in that, In step S2, L Fast RCNN Defined as: L Fast RCNN =L cls (p,u)+λ 22 [u≥1]L IOC (t u ,v) Where p is the predicted probability of all target classes; u is the true probability of all target classes. λ 22 L represents the balancing weight parameter. cls (p,u) is the cross-entropy loss function for multi-class classification: Where N is the number of detected instances, M is the number of categories, and u ij p represents the true probability that the i-th instance belongs to class j, and is a binary index; ij The probability that the i-th instance predicted by the model belongs to class j; L IOC Representing regression loss, using To calculate, where where t u The predicted position value of the target box is represented by ; v represents the actual position value of the target box; (x, y) are the coordinates of the center point of the target box, and w and h are the width and height of the target box, respectively.

4. The training method for the unauthorized hot work detection model according to claim 1, characterized in that, In step S2, L dist Defined as: L dist =u k1 ·u m2 [(x k -x m ) 2 +(y k -y m ) 2 ]+u k1 ·u h3 [(x k -x h ) 2 +(y k -y h ) 2 ] The categories are: personnel (value 1), protective mask (value 2), flame (value 3), and fire extinguisher (value 4); k1 This represents the true probability value that the k-th detected instance belongs to the personnel category and is a hot work operator; u m2 u represents the true probability that the m-th detected instance is a protective face mask. h3 This represents the true probability that the h-th detected instance is of the flame category; (x k ,y k ), (x m ,y m ), (x h ,y h ) represent the center point coordinates of the predicted hot work operator, protective mask, and flame instance, respectively.

5. The training method for the unauthorized hot work detection model according to claim 4, characterized in that, Where u i1 This represents the true probability value that the i-th detected instance belongs to the personnel category, (x i ,y i ), (x m ,y m ) represent the coordinates of the predicted personnel and the center point of the protective mask, respectively; Detection location (x) using on-site personnel examples i ,y i ) and the detection position of the protective mask (x m ,y m The minimum distance is used to identify which personnel instance i is the hot work operator k, i.e., k = i. In this case, the i(k)th detected instance is the hot work operator, therefore u k1 =1.

6. A method for detecting unauthorized hot work using the training method of the unauthorized hot work detection model obtained by any one of the claims 1-5, characterized in that, It includes the following steps: S1: Use the illegal hot work detection model to detect the input hot work site video image and obtain the category of each target in the image and its detection box result; S2: A comprehensive judgment is made on the category of each target in the image in step S1 and the results of its detection box to determine whether there is any illegal hot work.

7. The method for detecting unauthorized hot work according to claim 6, characterized in that, In step S1, the real-time hot work site video image is input into the Faster RCNN detection algorithm using the illegal hot work detection model to detect personnel, fire extinguishers, protective masks and flames in the hot work site video image, and obtain the category of each target in the image and its detection box result.

8. The method for detecting unauthorized hot work according to claim 6, characterized in that, The specific method for determining step S2 is as follows: S21: If both personnel and flame targets are detected in the image, it is determined to be a hot work site; S22: Based on step S21, determine whether a target of the protective face mask category is detected simultaneously. If not, an alarm is triggered indicating that a protective face mask is not being worn. If it is, further determine whether the protective face mask is being worn on the face of the detected personnel instance. If not, it is considered a violation and an alarm is triggered. If it is, it indicates that a personnel instance is wearing a protective face mask. S23: Based on the detection in step S22 that a person is wearing a protective mask, determine whether there are other persons besides that person. If not, it means that there is no monitoring personnel on site, which is also a violation of regulations and an alarm will be triggered. If there are, it means that there are monitoring personnel on site. S24: Based on the detection of a monitoring personnel in step S23, determine whether the target category of the fire extinguisher has been detected. If not, it is considered an unauthorized hot work and an alarm is triggered; if so, it is considered a normal hot work.

9. The method for detecting unauthorized hot work according to claim 8, characterized in that, In step S22, the method for determining whether the protective mask is worn on the face of the detected personnel instance is as follows: the intersection area φ is calculated sequentially based on the detection position of the protective mask and the upper part of the detection frame of all detected personnel instances. If each φ = 0, it means that none of the personnel instances are wearing protective masks, which is a violation and an alarm is triggered. If any φ > 0, it means that a personnel instance is detected wearing a protective mask.

10. The method for detecting unauthorized hot work according to claim 9, characterized in that, φ=S A ∩S B Where S A S represents the area of ​​the upper half of the detection bounding box for a given person instance. B This refers to the area of ​​the detected protective face shield's bounding box region.

Images