Track operator cross-track safety action identification method based on attitude estimation
Through the improved human posture estimation algorithm YOLOv8n-Pose, railway staff can identify key points and safe movements of human bodies, solving the problems of inefficiency of traditional manual monitoring and safety hazards, and achieving automatic supervision with high accuracy and high speed.
Patent Information
- Application Number
- CN202411664497.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-20
- Publication Date
- 2025-03-18
AI Technical Summary
During the railway maintenance process, traditional methods rely on manual monitoring of railway workers' cross-rail safety actions, which poses inefficiency and safety hazards.
The improved human posture estimation calculation method YOLOv8n-Pose is used to identify human key points for track workers, and their safety actions are automatically identified based on the key point position information. This method improves recognition accuracy and efficiency through technologies such as knowledge transfer, knowledge distillation and lightweight network structure.
It realizes automatic identification of high accuracy and high speed of cross-rail safety actions of railway workers, reduces the cost and safety risks of manual participation, and improves supervision efficiency.
Smart Images
Figure CN119649448A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of posture estimation and deep learning, and in particular to a method for identifying safe cross-track actions of track workers based on posture estimation. Technical Background
[0002] During railway maintenance, railway workers must abide by strict standards when crossing the rails, that is, they can only cross the rails after performing the standard safety actions of "one stop, two look, three confirm, and four pass" at specific crossings monitored by cameras. In order to supervise this standard safety action, the traditional method of manual monitoring and management is inefficient, prone to omissions, and poses safety hazards.
[0003] To this end, people use cameras to monitor and identify railway staff, then track their travel paths and save the video, but no human posture or movement estimation is performed. Specific supervision of cross-track safety actions still requires manual review of the saved videos. Although this improves the efficiency of supervision to a certain extent, manual participation is still required, and the accuracy and efficiency are low, making it impossible to fully automate real-time and accurate supervision. Summary of the invention
[0004] In response to the problems existing in the background technology, the present application proposes a method for identifying the safe cross-track actions of track workers based on posture estimation.
[0005] Technical solution:
[0006] A method for identifying cross-track safety actions of track workers based on posture estimation, using an improved human posture estimation algorithm YOLOv8n-Pose to identify key points of the human body of the track workers, and identifying safety actions based on the position information of the key points of the human body, and the specific steps include: a training step of an improved YOLOv8n-Pose network model, a step of identifying key points of the human body based on the improved YOLOv8n-Pose network model, and a step of identifying safety actions based on the key point position information; the training step of the improved YOLOv8n-Pose network model specifically includes:
[0007] 1) Using knowledge transfer technology, the YOLOv8m-Pose and improved YOLOv8n-Pose models trained based on the COCO key point dataset are used as the training initialization parameters of the teacher and student networks respectively;
[0008] 2) Collect m images of track workers crossing the track;
[0009] 3) Annotate cross-track images to create a dataset;
[0010] 4) Add n random images from the COCO val 2017 key point dataset to the dataset;
[0011] 5) Use the prepared dataset to train the YOLOv8m-Pose teacher network;
[0012] 6) The improved YOLOv8n-Pose is used as the student network, and the network is trained using the prepared data set. The knowledge distillation technology is used in the training. The cross entropy loss calculated by the output of the teacher network and the output of the student network is added as a soft loss to the total loss of the student network, so that the student network can learn the knowledge of the teacher network;
[0013] 7) After training is completed, save the improved YOLOv8n-Pose student network weights.
[0014] Preferably, the step of performing human key point recognition based on the improved YOLOv8n-Pose network model specifically includes:
[0015] 1) Obtain real-time images from on-site monitoring of track operations;
[0016] 2) Detect whether there is a human body in the image;
[0017] 3) If there is a human body, the image is sent to the trained improved YOLOv8n-Pose network model for human key point recognition;
[0018] 4) Safe action recognition based on the position information of key points of the human body;
[0019] 5) Output the safety action recognition results.
[0020] Preferably, the step of identifying the safety action based on the key point position information specifically includes:
[0021] 1) Determine whether it is a safe left-hand action based on the key point position information;
[0022] 2) Determine whether it is a right-hand safety action based on the key point position information;
[0023] 3) Determine whether it is a forward hand-raising action based on the key point position information;
[0024] 4) Determine whether it is a cross-track action based on the key point position information;
[0025] 5) Output the recognition results.
[0026] Preferably, the YOLOv8n-Pose network consists of a backbone network, a neck network and a head network. The method for improving the YOLOv8n-Pose network model is to add an EMA attention module between the connection between the neck network and the head network of the network. Specifically, an EMA attention module is added between the connection between the first, third and fourth C2F_Faster modules from top to bottom and from left to right in the neck network and the head network, and α-InnerSIoU is used as the improved human frame loss, and the human key point loss function is improved, while the network structure is lightweight; the EMA module uses a method of encoding global information to readjust the channel weights in each parallel branch, and captures pixel-level relationships through cross-dimensional interactions.
[0027] Preferably, the specific calculation formula of α-InnerSIoU is:
[0028] L α-InnerSIoU =1-(1-α)IoU-αIoU inner +Δ+Ω
[0029] Among them, α is a learnable parameter, limited to between 0 and 1, which is used to adjust the weight of the intersection of the predicted box and the true box, as well as the weight of the intersection of the auxiliary predicted box and the auxiliary true box;
[0030] IoU represents the intersection-over-union ratio between the predicted box and the real box, and the calculation formula is:
[0031]
[0032] The intersection represents the area where the real box and the predicted box intersect, and the union represents the area where the real box and the predicted box are combined.
[0033] IoU inner It represents the intersection-over-union ratio of the auxiliary prediction box and the auxiliary real box. The calculation formula is:
[0034]
[0035] The intersection inner Represents the intersection area of the auxiliary real box and the auxiliary prediction box, union inner Represents the area of the auxiliary real box and the auxiliary prediction box;
[0036] Δ represents the normalized penalty term of the distance between the center point of the predicted box and the center point of the real box, and the calculation formula is:
[0037]
[0038] in:
[0039]
[0040] In the formula, Λ represents the angle mapping value between the real box and the predicted box, γ is the angle penalty term between the predicted box and the real box, and c w is the x-axis distance between the center point of the predicted box and the real box, c h is the y-axis distance between the center point of the predicted box and the real box, σ is the Euclidean distance between the center point of the predicted box and the real box, and ρ x , ρ y are the x-axis and y-axis distance penalty items between the center point of the predicted box and the center point of the real box respectively; the coordinates of the center points of the real box and the auxiliary real box are The coordinates of the center points of the prediction box and the auxiliary prediction box are (x c ,y c ),
[0041] Ω represents the normalized penalty term for the shape similarity between the predicted box and the true box, and the calculation formula is:
[0042]
[0043] in:
[0044]
[0045] θ is a hyperparameter that controls the network's calculation weight for the penalty term of the shape similarity between the predicted box and the true box; w gt 、h gt are the width and height of the real frame, w and h are the width and height of the predicted frame, are the width and height of the auxiliary real frame, w inner 、h inner are the width and height of the auxiliary prediction box respectively, and ratio is the ratio hyperparameter of the auxiliary box.
[0046] Preferably, the improved human body key point loss function calculation formula is:
[0047]
[0048] in x i ,y i represents the predicted position of the i-th human key point, represents the true position of the i-th human key point; μ is a trainable parameter; s is the area of the human body prediction box; k i is the weight of the i-th human key point; v i Indicates the type of annotation of the i-th human key point, 0, 1, and 2 represent unlabeled, invisible annotation, and visible annotation, respectively; δ(*) is a conditional function, which takes 1 when the conditions in the brackets are met, otherwise it takes 0.
[0049] Preferably, the method of lightweight network structure is:
[0050] Replace the BottleNeck of the YOLOv8n-Pose C2F structure with the FasterBlock module in FasterNet.
[0051] Specifically, the method for determining whether it is a safe action based on the key point position information is as follows:
[0052] According to the key point position information of the human body, the facial orientation, the relationship between the left arm and the body, and the relationship between the right arm and the body are analyzed. If the angle between the wrist, elbow, shoulder and shoulder and the shoulder and hip on the same side reaches or exceeds the set threshold, it indicates a safe action. The specific methods for judging the safe actions for the left and right directions are as follows:
[0053] 1) Determine whether the person is turning his head to the left by connecting the key points of the ears, eyes, and nose on both sides. When the person turns his head to the left, the left ear and left eye can be identified. If the person is performing a safety action and is turning his head to the left, it is determined that he is performing a left-facing safety action.
[0054] 2) Determine whether the person is turning his head to the right by connecting the key points of the ears, eyes, and nose on both sides. When the person turns his head to the right, the right ear and right eye can be identified. If the person is performing a safety action and is turning his head to the right, it is determined that he is performing a right-turn safety action.
[0055] Specifically, the method for determining whether it is a forward hand-raising action or a cross-track action according to the key point position information is as follows:
[0056] 1) The method for judging whether it is a forward hand-raising action is as follows: judging whether the person is in a frontal face state by the connection state of the key points of the ears, eyes, and nose on both sides. When the person is in a frontal face state, only the ears can be recognized; if the person is in a frontal face state and the angle between the connection line of the elbow, shoulder, and hip on one side exceeds 150 degrees, it is judged that the person is performing a forward hand-raising action;
[0057] 2) The method for judging whether it is a cross-track action is to judge by the position information of key points: determine it by judging the relative position relationship between the line connecting the key points of the two ankles of the person and the preset track line. If the line connecting the person's ankles is always on the same side of the track line, it is considered that the person has not crossed the line; if it appears on both sides of the track line, it is considered that the person has crossed the line.
[0058] Beneficial Effects
[0059] 1) Use the improved YOLOv8n-Pose model to identify key points of track workers and identify their safety actions based on the key point position information, which has the advantages of high recognition accuracy, fast speed, reduced labor costs and effectively improved detection efficiency;
[0060] 2) Adding the EMA attention module to the neck position of the YOLOv8n-Pose model effectively improves the network's ability to recognize the position of the human body and its key points;
[0061] 3) Using α-InnerSIoU as the human position loss function can improve the recognition accuracy of the human position and the network's ability to detect the human body. At the same time, it can speed up the network convergence speed during training and reduce the training time.
[0062] 4) Using an improved key point loss function makes the human key point regression prediction more accurate and speeds up the network training;
[0063] 5) Replace BottleNeck of the YOLOv8n-Pose C2F structure with the FasterBlock module, which lightweights the network structure, greatly reduces the number of parameters while keeping the network performance almost unchanged, improves the model reasoning speed, and facilitates network deployment on edge computing devices;
[0064] 6) In the training process of the improved YOLOv8n-Pose model, the transfer learning technology is used to use the model parameters obtained by training based on the COCO human key point dataset as the initialization parameters of the model training, thereby improving the performance and generalization ability of the model and speeding up the training speed when the actual human body data of track workers is limited;
[0065] 7) In the training process of the improved YOLOv8n-Pose model, the knowledge distillation technology is used, YOLOv8m-Pose is used as the teacher network, and the improved YOLOv8n-Pose is used as the student network. During the training, the cross entropy loss calculated between the output of the teacher network and the output of the student network is added to the total loss as a soft loss, so that the student network can learn the knowledge of the teacher network and improve the network recognition accuracy and robustness while keeping the network parameters as small as possible. BRIEF DESCRIPTION OF THE DRAWINGS
[0066] Figure 1 This is a safety action recognition flow chart of the present invention
[0067] Figure 2 The improved YOLOv8n-Pose network structure of the present invention
[0068] Figure 3 This is the improved IoU calculation description diagram of the present invention
[0069] Figure 4 This is the Faster Block structure diagram of the present invention
[0070] Figure 5The improved YOLOv8n-Pose recognition result of the cross-track safety action of the rail operator in this invention is DETAILED DESCRIPTION
[0071] The present invention will be further described below in conjunction with embodiments, but the protection scope of the present invention is not limited thereto:
[0072] The invention discloses a method for identifying cross-track safety actions of track workers based on posture estimation. An improved human posture estimation algorithm YOLOv8n-Pose is used to identify key points of the human body of the track workers, and safety actions are identified according to the position information of the key points of the human body. The specific steps include: a training step of an improved YOLOv8n-Pose network model, a step of identifying key points of the human body based on the improved YOLOv8n-Pose network model, and a step of identifying safety actions based on the key point position information.
[0073] The training steps for the improved YOLOv8n-Pose network model include:
[0074] 1) Using knowledge transfer technology, the YOLOv8m-Pose and improved YOLOv8n-Pose models trained based on the COCO key point dataset are used as the training initialization parameters of the teacher and student networks respectively;
[0075] 2) Collected 1,911 cross-track images of track workers;
[0076] 3) Annotate cross-track images to create a dataset;
[0077] 4) Add 1346 random images from the COCO val 2017 key point dataset to the dataset;
[0078] 5) Use the prepared dataset to train the YOLOv8m-Pose teacher network;
[0079] 6) The improved YOLOv8n-Pose is used as the student network, and the network is trained using the prepared data set. The knowledge distillation technology is used in the training. The cross entropy loss calculated by the output of the teacher network and the output of the student network is added as a soft loss to the total loss of the student network, so that the student network can learn the knowledge of the teacher network;
[0080] 7) After training is completed, save the improved YOLOv8n-Pose student network weights.
[0081] The improved YOLOv8n-Pose network model is to add an EMA attention module to the neck position of the original YOLOv8n-Pose model, use α-InnerSIoU as the improved human frame loss, and improve the human key point loss function. Finally, the BottleNeck of the YOLOv8n-Pose C2F structure is replaced by the FasterBlock module in FasterNet to lightweight the network structure.
[0082] The improved YOLOv8n-Pose network structure is as follows Figure 2 As shown in the figure, the parallel substructure and cross-space feature fusion of the EMA attention mechanism improve the calculation weight of the human body and its key point areas in the network prediction process, enabling the network to better utilize the feature information related to the human body and its key points, and reduce the impact of uneven sample distribution and background on recognition.
[0083] Combination Figure 3 , the specific calculation formula of α-InnerSIoU is:
[0084] L α-InnerSIoU =1-(1-α)IoU-αIoU inner +Δ+Ω
[0085] Where α is a learnable parameter, limited to between 0 and 1, used to adjust the prediction box ( Figure 3 Green solid line frame) and the real frame ( Figure 3 Black solid line frame) IoU and auxiliary prediction frame ( Figure 3 Green dashed box) and auxiliary ground truth box ( Figure 3 Black dashed box) Intersection-over-Union weight.
[0086] IoU represents the intersection-over-union ratio between the predicted box and the real box, and the calculation formula is:
[0087]
[0088] The intersection represents the area where the real box and the predicted box intersect, and the union represents the area where the real box and the predicted box are combined.
[0089] IoU inner It represents the intersection-over-union ratio of the auxiliary prediction box and the auxiliary real box. The calculation formula is:
[0090]
[0091] The intersection inner Represents the intersection area of the auxiliary real box and the auxiliary prediction box, union inner Represents the area of the auxiliary true box and the auxiliary predicted box.
[0092] Δ represents the normalized penalty term of the distance between the center point of the predicted box and the center point of the real box, and the calculation formula is:
[0093]
[0094] in:
[0095]
[0096] Where Λ represents the angle penalty term between the true box and the predicted box.
[0097] Ω represents the normalized penalty term for the shape similarity between the predicted box and the true box, and the calculation formula is:
[0098]
[0099] in:
[0100]
[0101] θ is a hyperparameter that controls the calculation weight of the penalty term for the shape similarity between the predicted box and the true box.
[0102] Figure 3 In the figure, the center coordinates of the real frame and the auxiliary real frame are The coordinates of the center points of the prediction box and the auxiliary prediction box are (x c ,y c ), w gt 、h gt are the width and height of the real frame, w and h are the width and height of the predicted frame, are the width and height of the auxiliary real frame, w inner 、h inner are the width and height of the auxiliary prediction box respectively, and ratio is the ratio hyperparameter of the auxiliary box.
[0103] The improved human key point loss function calculation formula is:
[0104]
[0105] in x i ,y i represents the predicted position of the i-th human key point, represents the true position of the i-th human key point; μ is a trainable parameter; s is the area of the human body prediction box; k i is the weight of the i-th human key point; v i Indicates the type of annotation of the i-th human key point, 0, 1, and 2 represent unlabeled, invisible annotation, and visible annotation, respectively; δ(*) is a conditional function, which takes 1 when the conditions in the brackets are met, otherwise it takes 0.
[0106] Combination Figure 2 , use FasterBlock module to replace the Bottle Neck of YOLOv8n-Pose C2F structure to lighten the network structure. FasterBlock module innovatively proposes partial convolution (PCONV), which only performs partial convolution operation on the input to reduce redundant features to achieve efficient spatial feature extraction, effectively reduce the number of network parameters, calculation amount and memory access, and speed up the network reasoning speed. The FasterBlock module structure is as follows Figure 4 shown.
[0107] like Figure 1 As shown in the figure, the steps for using the improved YOLOv8n-Pose network model to perform safe human action recognition are as follows:
[0108] 1) Obtain real-time images from on-site monitoring of track operations;
[0109] 2) Detect whether there is a human body in the image;
[0110] 3) If there is a human body, the image is sent to the trained improved YOLOv8n-Pose network model for human key point recognition;
[0111] 4) Safe action recognition based on the position information of key points of the human body;
[0112] 5) Output the safety action recognition results.
[0113] After the key points of the human body are identified, the safe action is judged based on the key point position information. The specific method for judging the safe action is: according to the key point position information of the human body, the facial orientation, the relationship between the left arm and the body, and the relationship between the right arm and the body are analyzed. If the angle between the wrist, elbow, shoulder and shoulder and the shoulder and hip on the same side reaches or exceeds the set threshold, it indicates a safe action. The angle is smaller when walking normally.
[0114] The specific left and right safety action judgment methods are as follows:
[0115] 1) The connection status of the key points of the ears, eyes, and nose on both sides is used to determine whether the person is turning his head to the left. When turning his head to the left, the left ear and left eye can be identified; if the person is performing a safety action and is turning his head to the left, it is determined that he is performing a left-facing safety action.
[0116] 2) Determine whether the person is turning his head to the right by connecting the key points of the ears, eyes, and nose on both sides. When turning his head to the right, the right ear and right eye can be identified; if the person is performing a safety action and is turning his head to the right, it is determined that he is performing a right-hand safety action.
[0117] After the safety action is judged, the forward hand-raising action and the track-crossing action are judged according to the key point position information. The forward hand-raising action judgment method is as follows:
[0118] The system determines whether a person is facing forward by the connection status of the key points of the ears, eyes, and nose on both sides. In the case of a frontal face, only the ears can be recognized. If a person is facing forward and the angle between the connection lines of the elbow, shoulder, and hip on one side exceeds 150 degrees, the person is judged to be raising his hand forward.
[0119] The method for judging cross-track action is as follows:
[0120] The relative position relationship between the line connecting the key points of the two ankles of the person and the preset track line is determined. If the line connecting the person's ankles is always on the same side of the track line, it is considered that the person has not crossed the line; if it appears on both sides of the track line, it is considered that the person has crossed the line.
[0121] The results of cross-track safety action recognition of rail workers based on the improved YOLOv8n-Pose are as follows Figure 5 As shown, left_check indicates a left safety action, right_check indicates a right safety action, front_check indicates a forward hand-raising action, and cross indicates a cross-track action.
[0122] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.
Claims
1. A method for identifying safe cross-track actions of track workers based on posture estimation, characterized in that The improved human posture estimation algorithm YOLOv8n-Pose is used to identify the key points of the human body of the track operator, and the safety action is identified according to the position information of the key points of the human body. The specific steps include: a training step of the improved YOLOv8n-Pose network model, a human key point identification step based on the improved YOLOv8n-Pose network model, and a safety action identification step based on the key point position information; the training step of the improved YOLOv8n-Pose network model specifically includes: 1) Using knowledge transfer technology, the YOLOv8m-Pose and improved YOLOv8n-Pose models trained based on the COCO key point dataset are used as the training initialization parameters of the teacher and student networks respectively; 2) Collect m images of track workers crossing the track; 3) Annotate cross-track images to create a dataset; 4) Add n random images from the COCO val 2017 key point dataset to the dataset; 5) Use the prepared dataset to train the YOLOv8m-Pose teacher network; 6) The improved YOLOv8n-Pose is used as the student network, and the network is trained using the prepared data set. The knowledge distillation technology is used in the training. The cross entropy loss calculated by the output of the teacher network and the output of the student network is added as a soft loss to the total loss of the student network, so that the student network can learn the knowledge of the teacher network; 7) After training is completed, save the improved YOLOv8n-Pose student network weights.
2. The method according to claim 1, characterized in that The steps of identifying key points of a human body based on the improved YOLOv8n-Pose network model specifically include: 1) Obtain real-time images from on-site monitoring of track operations; 2) Detect whether there is a human body in the image; 3) If there is a human body, the image is sent to the trained improved YOLOv8n-Pose network model for human key point recognition; 4) Safe action recognition based on the position information of key points of the human body; 5) Output the safety action recognition results.
3. The method according to claim 1, characterized in that The step of identifying the safety action based on the key point position information specifically includes: 1) Determine whether it is a safe left-hand action based on the key point position information; 2) Determine whether it is a right-hand safety action based on the key point position information; 3) Determine whether it is a forward hand-raising action based on the key point position information; 4) Determine whether it is a cross-track action based on the key point position information; 5) Output the recognition results.
4. The method according to claim 1, characterized in that The YOLOv8n-Pose network consists of a backbone network, a neck network and a head network. The method for improving the YOLOv8n-Pose network model is to add an EMA attention module between the connection between the neck network and the head network of the network. Specifically, the EMA attention module is added between the connection between the first, third and fourth C2F_Faster modules from top to bottom and from left to right in the neck network and the head network, and α-InnerSIoU is used as the improved human frame loss, and the human key point loss function is improved, while the network structure is lightweight; the EMA module uses a method of encoding global information to readjust the channel weights in each parallel branch, and captures pixel-level relationships through cross-dimensional interactions.
5. The method according to claim 4, characterized in that The specific calculation formula of α-InnerSIoU is: L α-InnerSIoU =1-(1-α)IoU-αIoU inner +D+O Among them, α is a learnable parameter, limited to between 0 and 1, which is used to adjust the weight of the intersection of the predicted box and the true box, as well as the weight of the intersection of the auxiliary predicted box and the auxiliary true box; IoU represents the intersection-over-union ratio between the predicted box and the true box, and the calculation formula is: The intersection represents the area where the real box and the predicted box intersect, and the union represents the area where the real box and the predicted box are combined. IoU inner It represents the intersection-over-union ratio of the auxiliary prediction box and the auxiliary real box. The calculation formula is: The intersection inner Represents the intersection area of the auxiliary real box and the auxiliary prediction box, union inner Represents the area of the auxiliary real box and the auxiliary prediction box; Δ represents the normalized penalty term of the distance between the center point of the predicted box and the center point of the real box. The calculation formula is: in: In the formula, Λ represents the angle mapping value between the real box and the predicted box, γ is the angle penalty term between the predicted box and the real box, and c w is the x-axis distance between the center point of the predicted box and the real box, c h is the y-axis distance between the center point of the predicted box and the real box, σ is the Euclidean distance between the center point of the predicted box and the real box, and ρ x , ρ y are the x-axis and y-axis distance penalty items between the center point of the predicted box and the center point of the real box respectively; the coordinates of the center points of the real box and the auxiliary real box are The coordinates of the center points of the prediction box and the auxiliary prediction box are (x c ,y c ), Ω represents the normalized penalty term for the shape similarity between the predicted box and the true box, and the calculation formula is: in: θ is a hyperparameter that controls the calculation weight of the penalty term for the shape similarity between the predicted box and the true box.
6. The method according to claim 4, characterized in that The improved human key point loss function calculation formula is: in x i ,y i represents the predicted position of the i-th human key point, represents the true position of the i-th human key point; μ is a trainable parameter; s is the area of the human body prediction box; k i is the weight of the i-th human key point; v i Indicates the type of annotation of the i-th human key point, 0, 1, and 2 represent unlabeled, invisible annotation, and visible annotation, respectively; δ(*) is a conditional function, which takes 1 when the conditions in the brackets are met, otherwise it takes 0.
7. The method according to claim 4, characterized in that The method of lightweight network structure is: Replace the BottleNeck of the YOLOv8n-Pose C2F structure with the FasterBlock module in FasterNet.
8. The method according to claim 3, characterized in that The method for judging whether it is a safe action based on the key point position information is as follows: According to the key point position information of the human body, the facial orientation, the relationship between the left arm and the body, and the relationship between the right arm and the body are analyzed. If the angle between the wrist, elbow, shoulder and shoulder and the shoulder and hip on the same side reaches or exceeds the set threshold, it indicates a safe action. The specific methods for judging the safe actions for the left and right directions are as follows: 1) Determine whether the person is turning his head to the left by connecting the key points of the ears, eyes, and nose on both sides. When the person turns his head to the left, the left ear and left eye can be identified. If the person is performing a safety action and is turning his head to the left, it is determined that he is performing a left-facing safety action. 2) Determine whether the person is turning his head to the right by connecting the key points of the ears, eyes, and nose on both sides. When turning his head to the right, the right ear and right eye can be identified; if the person is performing a safety action and is turning his head to the right, it is determined that he is performing a right-hand safety action.
9. The method according to claim 3, characterized in that The specific method for judging whether it is a forward hand-raising action or a cross-track action based on the key point position information is as follows: 1) The method for judging whether it is a forward hand-raising action is as follows: judging whether the person is in a frontal face state by the connection state of the key points of the ears, eyes, and nose on both sides. When the person is in a frontal face state, only the ears can be recognized; if the person is in a frontal face state and the angle between the connection line of the elbow, shoulder, and hip on one side exceeds 150 degrees, it is judged that the person is performing a forward hand-raising action; 2) The method for judging whether it is a cross-track action is to judge by the position information of key points: determine it by judging the relative position relationship between the line connecting the key points of the two ankles of the person and the preset track line. If the line connecting the person's ankles is always on the same side of the track line, it is considered that the person has not crossed the line; if it appears on both sides of the track line, it is considered that the person has crossed the line.