A method for monitoring height-climbing safety of transformer substation workers

Abstract

The present invention relates to a method for monitoring the height-climbing safety of transformer substation workers, and belongs to the field of target detection technology. The method comprises the following steps: extracting human body joint points from a video screen, and then connecting each pair of associated body parts through a posture line, and the posture line will be combined into the whole-body posture of the human body detected in the image; detecting and judging the climbing action of the worker; if it is detected that the worker has the climbing behavior, the protective measures for the climbing worker shall be identified: judging whether the safety belt and safety helmet are correctly worn according to the requirements, and specifically, using the YOLO algorithm to detect whether there is a corresponding object in the human body image area. When the present invention monitors the on-site video, the response time to the safety protection problem of personnel during the climbing period is about 3.4 seconds. The behavior of inadequate protection in climbing work can be alarmed in a very short time, which improves the detection accuracy and reaction speed, thereby greatly improving the safety problem in the climbing work of the transformer substation.

Description

Technical Field

[0001] The invention relates to a method for monitoring the height climbing safety of transformer substation workers, belongs to the technical field of target detection, and is suitable for identifying the action and position of a person or object in a scene under different scenes. Background Art

[0002] Although traditional monitoring is a mature solution with high accuracy and real-time performance, it requires too much equipment to be installed during the operation, and the backend staff needs to monitor attentively, which requires a large number of professional staff and consumes manpower and material resources. Traditional monitoring solutions have a low degree of automation and can only be deployed in key substations. For substations and branch substations in remote areas, the cost is too high. In addition, due to the complex environment of substations, the installation of monitoring equipment is restricted by the location of the main equipment in the substation. The location of each device needs to be planned in advance. It is difficult to upgrade old substations without monitoring equipment, which is not suitable for large-scale applications.

[0003] 1. Traditional monitoring solution The traditional monitoring solution refers to the use of various sensors and cameras to integrate into a complete set of monitoring solutions to detect the operating status of substation equipment, monitor the substation environment, and monitor and protect the operation and maintenance of substation staff. The posture estimation of the human body mainly uses multiple monitoring cameras to shoot the on-site operations of the staff, and then transmits the video stream to the background server. The rear monitoring personnel can watch the operation video in real time to find abnormal conditions in time. There will be an alarm device on site to receive the alarm signal from the background to remind the staff of dangerous actions. Another technology is electronic fence, which surrounds the work area by setting up vibration sensing cables and vibration sensors. When the staff touches the boundary of the work area, the signal is sensed and the alarm is triggered; buried cables and intrusion detectors are buried in dangerous areas. When the staff enters the blind spot of monitoring or enters the dangerous area, the sensor will send an alarm to the control center. At the same time, the monitoring personnel will regularly inspect the operation area and record the operation status.

[0004] 2. Existing Attitude Estimation Algorithms

[0005] There are two main ideas in traditional posture estimation algorithms: generative methods (model-based) and discriminative methods (model-free).

[0006] (1) Generation method (model-based)

[0007] Based on the structure of the graph and the deformable component model, a 2D human body component detector is designed. The graph model is used to build the connectivity of each component, and then the pairing relationship of key nodes is combined to restore the posture of the human body. The algorithm steps are as follows: Use the HOG method for human body detection. By manually specifying a rectangular statistical area to include the target human head and shoulders as initial information, the image is divided into several areas (foreground, background); then the foreground is highlighted and the human body parts are searched in the foreground area; single-frame image analysis is performed, and the image processed in the previous step is used to obtain the human body edge information and posture information, and the appearance model is learned at the same time; finally, considering that the clothing and appearance of the target human body in the same video sequence will not change significantly, the subsequent human body posture can be estimated using the model learned in the previous step and the spatial geometric continuity using the maximum entropy method.

[0008] (2) Discrimination method (no model)

[0009] The discriminant method directly learns the mapping from the input source to the human posture space (learning-based) or searches for non-existent examples (example-based) without using a human body model. The process is: given a reference posture, remove the unmatched postures in the image according to the marker distribution; implement circular convolution based on fast Fourier transform, and then use it to implement a fast histogram matching algorithm, which calculates the rotation angle between the posture in the image and the reference posture; finally, use nonlinear least squares to correct the final result. The discriminant method is usually faster than the generative method. At the same time, it can be supplemented by sensor detection (camera, infrared sensor, etc.), and the detection accuracy can be improved through manual observation and re-inspection.

[0010] 3. Posture Estimation Based on Deep Learning

[0011] Compared with traditional posture estimation algorithms, deep learning can extract higher-level and better expressive features of the target, and can also integrate the extracted features, classification and other information into one model.

[0012] (1) CPM

[0013] CPM uses neural networks to learn image features and spatial information at the same time. Compared with the previous use of CNN to extract image features, neural networks can learn these two types of information at the same time, which can achieve better results and can achieve end-to-end learning. CPM uses a multi-layer structure. The input of each layer is the original image features and the output of the previous layer. The current layer extracts information through the convolution layer based on these two types of information to produce new results. After multiple layers of iteration, a more accurate result is produced.

[0014] (2)MSPN

[0015] First, for each layer of the model, MSPN uses GlobalNet in CPN (backbone is ResNet). Secondly, MSPN adds cross-stage feature fusion. Since information loss is inevitable after repeated downsampling-upsampling, the two feature maps of the corresponding resolution in the previous layer downsampling and upsampling process are connected in MSPN and fused with the feature map of the current layer downsampling. In this way, the feature map obtained in the current layer contains more high-level information and reduces the information loss between layers. In addition, this residual structure-like design also helps to alleviate the gradient disappearance problem. Finally, MSPN also uses coarse-to-fine supervision. Intuitively, for a multi-layer network, as the number of layers increases, the estimation of key points is a coarse-to-fine process. In this way, when performing intermediate supervision, the ground truth can also be set to a coarse-to-fine heatmap, that is, the Gaussian kernel of the front layer is larger, and the Gaussian kernel of the back layer is smaller. As the number of layers increases, the position of the key points is required to be more and more accurate.

[0016] In traditional machine learning algorithms, although the generation method has high time efficiency, it extracts artificial shallow features (such as HOG, SIFT, etc.), which loses most of the image information, resulting in the algorithm being constrained by image quality. At the same time, due to the single component model, when the human body posture changes in a large range, the accuracy of the component model cannot meet such a large number of deformations, and the result of posture estimation has multiple feasible solutions, and the scope of application is narrow. In the substation, due to the presence of a large number of devices, the received image data cannot completely contain the key points of the model: in single-frame image recognition, in order to obtain a complete human body posture, the target human body image has certain requirements on the angle, usually the front or offset angle is relatively small, to ensure that the various limb joints of the human body are visible, which makes a considerable part of the image cannot be used for recognition. The recognition accuracy is low and not practical. Therefore, the generation method can achieve good results in single-frame images or sequence images, but it has inherent deficiencies for complex scenes such as substations. The present invention improves the target tracking and target recognition algorithms in the substation environment, and improves the detection accuracy and reaction speed of the algorithm. Summary of the invention

[0017] The present invention provides a method for monitoring the height climbing safety of transformer substation workers, which is used to improve the detection accuracy and reaction speed.

[0018] The technical solution of the present invention is: a method for monitoring the height-climbing safety of substation workers, and the specific steps of the method for monitoring the height-climbing safety of substation workers are as follows:

[0019] Step 1: Extract human joints from the video screen, and then connect each pair of related body parts through a posture line. The posture lines will be combined into the whole body posture of the human body detected in the image. The Openpose algorithm is used to extract joints of people in the video screen, and human body detection and human body region fusion steps are added.

[0020] Step 2: Detection and judgment of the worker's climbing action; if the worker is detected to have climbed, proceed to the next step;

[0021] Step 3, Identification of protective measures for climbing staff: Determine whether the safety belt and helmet are worn correctly according to the requirements. Specifically, use the YOLO algorithm to detect whether there is a corresponding object in the human image area.

[0022] As a further solution of the present invention, the specific steps of Step 1 are:

[0023] Step 1.1, first extract the input image frame by frame and use the YOLO algorithm to detect the position of the human body in the image, and extract the detection frame for subsequent posture estimation; the detection frame has fault tolerance;

[0024] When the predicted human body detection frame is too large, the algorithm will have a higher fault tolerance. However, when the human body detection frame is too small to capture the entire human body, it will affect the results of subsequent posture estimation. Therefore, the human body detection frame of the previous step is enlarged. The formula is as follows:

[0025] x_lar=x*0.9

[0026] y_lar=y*0.9

[0027] w_lar=x_lar+w*1.2<inuige.cols?

[0028] w*1.2:image.cols-x_lar

[0029] h_lar=y_lar+h*1.2<image.rows?

[0030] h*1.2:image.rowws-y_lar

[0031] Among them, x and x_lar are the horizontal coordinates of the upper left corner before and after magnification, y and y_lar are the vertical coordinates of the upper left corner before and after magnification, w and w_lar are the image widths before and after magnification, h and h_lar are the image heights before and after magnification, image.cols is the width of the image, and image.rows is the height of the image;

[0032] Step 1.2: The human body regions that are close to each other are fused through the human body region fusion step to improve the recognition efficiency of the posture estimation algorithm and define the IOU of the two human body regions. MAx for When the IOU of two volume regions MAX When the threshold is exceeded, the two regions are fused, otherwise they are not fused; the fusion is performed according to the formula:

[0033] x=minx i , x j )

[0034] y=min(y i ,y j )

[0035] w=max{x i +w i , x j +w j}-x

[0036] h=max{y i +h i ,y j +h j}-y

[0037] Among them, (x, y, w, h) is the bounding box of the human body area after fusion, (x i ,y i , w i ,h i ), (x j ,y j , w j ,h j ) are the borders of the two human body regions to be fused, (x, y) represents the coordinates of the upper left corner of the border, and (w, h) represents the width and height of the border;

[0038] Step 1.3, extract features from the fused image using the VGG19 convolutional network to obtain a 128-dimensional feature map of the image, and pass the feature map to two parallel convolutional layer branches. The first branch is used to predict 18 joint point confidence maps PartConfidence Maps, each of which represents a joint in the human skeleton; the second branch predicts a set of 38 joint point association degree PartAffinity Fields;

[0039] Step 1.4, use Bipartite Matching in graph theory to find the Part Association and connect the joints of the same person; the correctness of this step depends on the vector nature of the PAF itself, and finally merge into the overall skeleton of a person;

[0040] Step 1.5. Finally, based on the PAFs method for Multi-Person Parsing, the Multi-person parsing problem is converted into a graphs problem and the partial graph matching algorithm Hungarian Algorithm is used to find the augmenting path, and the augmenting path is used to find the maximum matching of the bipartite graph.

[0041] As a further solution of the present invention, the specific steps of Step 2 are:

[0042] Step 2.1, compare the human posture information obtained from the video with the stored climbing posture in the database, then use a histogram to describe the distribution of 10 bone angles in the motion sequence, and use the distance between the bone angle histograms of the two motion sequences as the similarity measure of the motion sequences; and use the cosine distance to measure the similarity between the two high-dimensional feature vectors, specifically by calculating the cosine of the angle between the two vectors;

[0043] Step 2.2, if the comparison coincidence of 5 consecutive frames of images exceeds 80%, proceed to the next step of detection;

[0044] Step 2.3, use the improved R-CNN algorithm to detect whether there are ladders or other climbing equipment in the picture information; specifically, the sliding window method is changed to the region proposal network to generate the detection frame. The region proposal network is a fully convolutional network. Its core idea is to slide the sliding window on the last convolutional layer. Due to the candidate region mechanism and border regression, multi-scale and multi-aspect ratio region proposals are obtained; the region proposal network is trained end-to-end for the task of generating region proposals, and can predict the bounding box and score of the target at the same time;

[0045] Step 2.4: If the confidence level of ladder climbing is above 80% in the image, it is considered as climbing behavior.

[0046] As a further solution of the present invention, in step Step 3: first use the target detection algorithm to extract the human body image in the picture, and then input the human body image into the YOLO algorithm model. If the detected objects include a seat belt and a safety helmet, and the confidence level exceeds 80%, it is considered that the staff has correctly complied with the relevant protective measures.

[0047] The beneficial effect of the present invention is that when the present invention monitors the on-site video, the response time to the safety protection problem of personnel climbing is about 3.4 seconds. Due to the 3-second continuous monitoring threshold, the present invention can alarm the behavior of insufficient protection during climbing work in a very short time, improve the detection accuracy and response speed, and thus greatly improve the safety problem of climbing work in substations. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a CPM structure diagram in the prior art of the present invention;

[0049] Figure 2 is a flow chart of the present invention;

[0050] Figure 3 It is a schematic diagram of the calculation area of ​​the entire object system in the present invention. DETAILED DESCRIPTION

[0051] Example 1: Figure 1-3 As shown, a method for monitoring height-climbing safety of substation workers is provided, and the specific steps of the method for monitoring height-climbing safety of substation workers are as follows:

[0052] Step 1: Extract human joints from the video screen, and then connect each pair of related body parts through a posture line. The posture lines will be combined into the whole body posture of the human body detected in the image. The Openpose algorithm is used to extract joints of people in the video screen, and human body detection and human body region fusion steps are added.

[0053] Step 2: Detection and judgment of the worker's climbing action; if the worker is detected to have climbed, proceed to the next step;

[0054] Step 3, Identification of protective measures for climbing staff: Determine whether the safety belt and helmet are worn correctly according to the requirements. Specifically, use the YOLO algorithm to detect whether there is a corresponding object in the human image area.

[0055] As a further solution of the present invention, the specific steps of Step 1 are:

[0056] Step 1.1, first extract the input image frame by frame and use the YOLO algorithm to detect the position of the human body in the image, and extract the detection frame for subsequent posture estimation; the detection frame has fault tolerance;

[0057] When the predicted human body detection frame is too large, the algorithm will have a higher fault tolerance. However, when the human body detection frame is too small to capture the entire human body, it will affect the results of subsequent posture estimation. Therefore, the human body detection frame of the previous step is enlarged. The formula is as follows:

[0058] x_lar=x*0.9

[0059] y_lar=y*0.9

[0060] w_lar=x_lar+w*1.2<image.cols?

[0061] w*1.2:image.cols-x_lar

[0062] h_lar=y_lar+h*1.2<imagee.rows?

[0063] h*1.2:image.rous-y_lar

[0064] Among them, x and x_lar are the horizontal coordinates of the upper left corner before and after magnification, y and y_lar are the vertical coordinates of the upper left corner before and after magnification, w and w_lar are the image widths before and after magnification, h and h_lar are the image heights before and after magnification, image.cols is the width of the image, and image.rows is the height of the image;

[0065] Step 1.2: The human body regions that are close to each other are fused through the human body region fusion step to improve the recognition efficiency of the posture estimation algorithm and define the IOU of the two human body regions. MAx for When the IOU of two volume regions MAX When the threshold is exceeded, the two regions are fused, otherwise they are not fused; the fusion is performed according to the formula:

[0066] x=min(x i , x j )

[0067] y=min(y i ,y j )

[0068] w=max{x i +w i , x j +w j}-x

[0069] h=max{y i +h i ,y j +h j}-y

[0070] Among them, (x, y, w, h) is the bounding box of the human body area after fusion, (x i ,y i , w i ,h i ), (x j ,y j , w j ,h j ) are the borders of the two human body regions to be fused, (x, y) represents the coordinates of the upper left corner of the border, and (w, h) represents the width and height of the border;

[0071] Step 1.3, extract features from the fused image using the VGG19 convolutional network to obtain a 128-dimensional feature map of the image, and pass the feature map to two parallel convolutional layer branches. The first branch is used to predict 18 joint confidence maps PartConfidence Maps, each of which represents a joint in the human skeleton; the second branch predicts a set containing 38 joint affinity PartAffinityFields;

[0072] Step 1.4, use Bipartite Matching in graph theory to find the Part Association and connect the joints of the same person; the correctness of this step depends on the vector nature of the PAF itself, and finally merge into the overall skeleton of a person;

[0073] Step 1.5. Finally, based on the PAFs method for Multi-Person Parsing, the Multi-person parsing problem is converted into a graphs problem and the partial graph matching algorithm Hungarian Algorithm is used to find the augmenting path, and the augmenting path is used to find the maximum matching of the bipartite graph.

[0074] After the joint points are extracted, we can obtain the joint point locations of each person in the video, including the nose, neck, eyes, shoulders, elbows, wrists, hips, knees, and ankles. Next, each pair of associated body parts is connected by a posture line. These posture lines will be combined into the whole body posture of the human body detected in the image. Next, climbing action detection is performed based on the combined human posture information to determine whether the wearing status of the staff needs to be further detected.

[0075] As a further solution of the present invention, the specific steps of Step 2 are:

[0076] Step 2.1, compare the human posture information obtained from the video with the stored climbing posture in the database, then use a histogram to describe the distribution of 10 bone angles in the motion sequence, and use the distance between the bone angle histograms of the two motion sequences as the similarity measure of the motion sequences; and use the cosine distance to measure the similarity between the two high-dimensional feature vectors, specifically by calculating the cosine of the angle between the two vectors;

[0077] Step 2.2, if the comparison coincidence of 5 consecutive frames of images exceeds 80%, proceed to the next step of detection;

[0078] Step 2.3, use the improved R-CNN algorithm to detect whether there are ladders or other climbing equipment in the picture information; specifically, the sliding window method is changed to the region proposal network to generate the detection frame. The region proposal network is a fully convolutional network. Its core idea is to slide the sliding window on the last convolutional layer. Due to the candidate region mechanism and border regression, multi-scale and multi-aspect ratio region proposals are obtained; the region proposal network is trained end-to-end for the task of generating region proposals, and can predict the bounding box and score of the target at the same time;

[0079] Step 2.4: If the confidence level of ladder climbing is above 80% in the image, it is considered as climbing behavior.

[0080] As a further solution of the present invention, in step Step 3: first use the target detection algorithm to extract the human body image in the picture, and then input the human body image into the YOLO algorithm model. If the detected objects include a seat belt and a safety helmet, and the confidence level exceeds 80%, it is considered that the staff has correctly complied with the relevant protective measures.

[0081] The specific steps of the YOLO algorithm in Step 3 are as follows:

[0082] (1) First, the input image is evenly divided into SxS grid cells. If the center of an object falls within this grid, then this grid is responsible for predicting the object.

[0083] (2) The image is input into the CNN feature extraction and prediction network. The convolutional layer is responsible for extracting the features of the image. After the features of the image pass through the fully connected layer, the output is a (S*S)*30 dimension. Each 1*1*30 dimension corresponds to one of the S*S cells in the original image. Each cell needs to predict the coordinates of two bounding boxes (bbox) (x center ,y center, w, h) and confidence, where the x of the center coordinate center ,y center Relative to the corresponding grid normalized to between 0-1, w, h are normalized to between 0-1 using the width and height of the image. In addition to regressing the coordinates themselves, a confidence value is also predicted. This confidence represents the confidence that the predicted bbox contains the obiect and how accurate the bbox prediction is.

[0084] (3) In addition to the bbox, each grid also needs to predict category information and perform three detections on the objects in the grid. The anchor size used for each detection is different. The anchor with 32x downsampling is the largest and suitable for detecting large targets, the anchor with 16x downsampling is suitable for objects of average size, and the anchor with 8x downsampling is the smallest and suitable for detecting small targets. After the detection is completed, logistic regression is used to give a target score to the part surrounded by the anchor, that is, how likely this position is to be the target. If it does not exceed the set threshold, it will be discarded.

[0085] (4) Finally, non-maximum suppression is used to select the bbox with the highest confidence and the most likely representation of the target. First, the bbox with the highest confidence is taken as the target, and then the intersection area between the remaining bbox and the target bbox is compared. If the intersection area is greater than the set threshold, the bbox is removed from the remaining bboxes, and the process is repeated.

[0086] (5) Finally, the location, category and confidence score of each object are output.

[0087] If the target detection confidence of the safety belt and safety helmet in the human body area is lower than 80%, it is considered that the staff did not wear the safety belt and safety rope when climbing according to the safety regulations. At this time, the system will automatically alarm and save the collected monitoring images, and record the time and place of the illegal operation in the background to achieve the function of intelligent monitoring.

[0088] In order to verify the effect of the algorithm proposed in the present invention, 2000 pictures collected from the substation site were identified, including 500 pictures of climbing work, of which 100 pictures showed that the safety measures were not worn properly. The monitoring results of the present invention are shown in the following table:

[0089] Table 1 is the test results of the climbing safety monitoring system

[0090]

[0091]

[0092] At the same time, when the present invention monitors the on-site video, the response time to the safety protection problem of personnel climbing is about 3.4 seconds. Since the present invention has a 3-second continuous monitoring threshold, it can be judged that the present invention can alarm the behavior of inadequate protection during climbing work in a very short time, thereby greatly improving the safety problem of climbing work in substations.

[0093] The specific implementation modes of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above implementation modes, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. A method for monitoring the height-climbing safety of substation workers, characterized in that: The specific steps of the substation staff climbing safety monitoring method are as follows: Step 1: Extract human joints from the video screen, and then connect each pair of related body parts through a posture line. The posture lines will be combined into the whole body posture of the human body detected in the image. The Openpose algorithm is used to extract joints of people in the video screen, and human body detection and human body region fusion steps are added. Step 2: Detection and judgment of the worker's climbing action; if the worker is detected to have climbed, proceed to the next step; Step 3, Identification of protective measures for climbing staff: Determine whether the safety belt and helmet are worn correctly according to the requirements. Specifically, use the YOLO algorithm to detect whether there is a corresponding object in the human image area; The specific steps of Step 1 are: Step 1.1, first extract the input image frame by frame and use the YOLO algorithm to detect the position of the human body in the image, and extract the detection frame for subsequent posture estimation; the detection frame has fault tolerance; Step 1.2: The human body regions that are close to each other are fused through the human body region fusion step to improve the recognition efficiency of the posture estimation algorithm and define the IOU of the two human body regions. MAX for When the IOU of two human regions MAX When the threshold is exceeded, the two regions are fused, otherwise they are not fused; the fusion is performed according to the formula: x=min(x i ,x j ) y=min(y i ,and j ) w=max{x i +w i ,x j +w j }-x h=max{y i +h i ,y j +h j }-y Among them, (x, y, w, h) is the bounding box of the human body area after fusion, (x i ,y i , w i ,h i ),(x j ,y j , w j ,h j ) are the borders of the two human body regions to be fused, (x, y) represents the coordinates of the upper left corner of the border, and (w, h) represents the width and height of the border; Step 1.3, extract features from the fused image using the VGG19 convolutional network to obtain a 128-dimensional feature map of the image, and pass the feature map to two parallel convolutional layer branches. The first branch is used to predict 18 joint point confidence maps Part ConfidenceMaps, each of which represents a joint in the human skeleton; the second branch predicts a set of 38 joint point association degree Part Affinity Fields; Step 1.4, use Bipartite Matching in graph theory to find the Part Association and connect the joints of the same person; the correctness of this step depends on the vector nature of the PAF itself, and finally merge into the overall skeleton of a person; Step 1.5, finally, based on the PAFs method for Multi-Person Parsing, convert the Multi-person parsing problem into a graphs problem and use the partial graph matching algorithm Hungarian Algorithm to find the augmenting path, and use the augmenting path to find the maximum matching of the bipartite graph; The specific steps of Step 2 are: Step 2.1, compare the human posture information obtained from the video with the stored climbing posture in the database, then use a histogram to describe the distribution of 10 bone angles in the motion sequence, and use the distance between the bone angle histograms of the two motion sequences as the similarity measure of the motion sequences; and use the cosine distance to measure the similarity between the two high-dimensional feature vectors, specifically by calculating the cosine of the angle between the two vectors; Step 2.2, if the comparison coincidence of 5 consecutive frames of images exceeds 80%, proceed to the next step of detection; Step 2.3, use the improved R-CNN algorithm to detect whether there is a ladder climbing device in the picture information; specifically, the sliding window method is changed to a region proposal network to generate a detection frame. The region proposal network is a fully convolutional network. Its core idea is to slide the sliding window on the last convolutional layer. Due to the candidate region mechanism and border regression, multi-scale and multi-aspect ratio region proposals are obtained; the region proposal network is trained end-to-end for the task of generating region proposals, and can predict the bounding box and score of the target at the same time; Step 2.4: If the confidence level of ladder climbing is above 80% in the image, it is considered as climbing behavior.

2. The method for monitoring the height-climbing safety of substation workers according to claim 1 is characterized in that: In the step 3: firstly, the human body image in the picture is extracted using the target detection algorithm, and then the human body image is input into the YOLO algorithm model. If the detected objects include a safety belt and a safety helmet, and the confidence level exceeds 80%, it is considered that the staff has correctly complied with the relevant protective measures.

Images