An unsafe behavior identification method for distribution network non-power-off operation site by fusing openpose and at-stgcn

Abstract

The application relates to a distribution network non-power-off operation site unsafe behavior recognition method fusing OpenPose and AT-STGCN, and the method comprises the following steps: collecting distribution network non-power-off operation site video data, and constructing an operation personnel behavior image data set; using an OpenPose model to extract human body skeleton key points from video frames, adopting partial affinity field connection key points, and constructing a human body skeleton space-time graph; constructing an AT-STGCN model introducing a space-time attention mechanism, wherein the time attention mechanism is used for extracting time motion features of the same joint between different frames, and the space attention mechanism is used for extracting space structure correlation features between different joints; inputting the human body skeleton space-time graph into the AT-STGCN model to perform behavior classification recognition; and designing a distribution network non-power-off operation unsafe behavior judgment logic, matching the behavior category recognized by the model with a preset unsafe behavior rule library, judging whether unsafe behavior occurs or not, and performing alarm. The application can realize accurate recognition and real-time early warning of the unsafe behavior of operation personnel, and effectively improves the safety control level of the distribution network non-power-off operation site.

Description

Technical Field

[0001] This invention relates to the fields of safety monitoring and computer vision technology for live-line work in power distribution networks, and particularly to a method and system for identifying unsafe behaviors at live-line work sites in power distribution networks that integrates OpenPose and AT-STGCN. Background Technology

[0002] Live-line work on distribution networks is a crucial means of ensuring continuous power supply, and the safety of its operation directly impacts the lives of workers and the stable operation of the power grid. Accidents resulting in personal injury or death due to inadequate safety precautions and improper operation are frequent, causing serious social impact and economic losses. Therefore, implementing systematic and intelligent safety monitoring of workers' on-site behavior has become a vital cornerstone in building a safe defense line for live-line work on distribution networks.

[0003] In recent years, deep learning-based human behavior recognition technology has made significant progress. OpenPose, as a bottom-up multi-person pose estimation method, can effectively extract human skeletal key points from images and has been widely used in the field of human behavior analysis. Spatial Temporal Graph Convolutional Networks (STGCN) extract skeletal features through graph convolution, which can effectively achieve the recognition of human actions.

[0004] However, the application of existing technologies in live-line work scenarios for distribution networks still faces many challenges. First, the working environment in distribution networks is complex, with interference factors such as shading and changes in lighting, leading to significant errors in traditional algorithms. Second, existing research largely focuses on general action datasets, lacking specific identification methods for the unique hazardous behaviors inherent in live-line work. Furthermore, while the STGCN model can extract spatiotemporal features, its ability to focus on important frames and key points is insufficient, requiring improvement in recognition accuracy.

[0005] Therefore, researching a method that can adapt to complex working environments and accurately identify unsafe behaviors of personnel working on live distribution networks is a problem that urgently needs to be solved by people in this field. Summary of the Invention

[0006] In view of this, it is necessary to provide a method for identifying unsafe behaviors in live-line power distribution operations that integrates OpenPose and AT-STGCN, in order to solve the problems of low accuracy in identifying unsafe behaviors of workers and difficulty in adapting to complex working environments in the existing technology.

[0007] This invention provides a method for identifying unsafe behaviors at power distribution network live-line work sites that integrates OpenPose and AT-STGCN, including:

[0008] (1) Collect video data of the power distribution network live-line operation site and construct a dataset of images of the operation personnel's behavior;

[0009] (2) Use the OpenPose model to extract human skeleton key points from video frames of the image dataset, and use partial affinity fields to connect the key points to construct a spatiotemporal map of human skeleton;

[0010] (3) Construct an AT-STGCN model that incorporates a spatiotemporal attention mechanism, wherein the AT-STGCN model includes a temporal attention mechanism module and a spatial attention mechanism module;

[0011] (4) Input the spatiotemporal map of the human skeleton into the AT-STGCN model for behavior classification and recognition, and output the behavior category;

[0012] (5) Design a logic for judging unsafe behaviors during live-line work in power distribution networks. Match the identified behavior categories with the preset unsafe behavior rule library to determine whether unsafe behaviors have occurred and issue an alarm.

[0013] Furthermore, the method for collecting video data from live-line power distribution operations includes:

[0014] 1) High-definition cameras deployed on MR glasses and fixed ground monitoring cameras are used to collect video of the work site under different lighting conditions and weather conditions from multiple angles;

[0015] 2) Use FFmpeg to extract video frames at a fixed frame rate to form the original image dataset. The dataset covers normal behavior categories and unsafe behavior categories (including insufficient safety distance, improper insulation shielding, etc.).

[0016] 3) Perform data augmentation operations on the acquired images, including random rotation, random scaling, random cropping, and brightness adjustment, to enhance the robustness and generalization ability of the model.

[0017] Furthermore, the method for constructing the spatiotemporal map of the human skeleton includes:

[0018] 1) Human Skeletal Keypoint Extraction: The OpenPose bottom-up human pose estimation method is adopted. First, the first 10 layers of the VGG network are used to extract features from the input image, generating feature maps F. Then, a multi-stage convolutional neural network is used for iterative prediction. Each stage contains two branches: one branch outputs a 2D confidence map S of the keypoints. t The other branch output part is the affinity field L. t The calculation formula is as follows:

[0019]

[0020] In the formula, ρt For the prediction network of 2D confidence graphs, ϕ t This is a prediction network for partial affinity fields. Through multi-stage progressive optimization, the keypoint detection results are made more accurate. The keypoints extracted in this embodiment include 14 keypoints: nose, neck, left and right shoulders, left and right elbows, left and right wrists, left and right hips, left and right knees, and left and right ankles.

[0021] 2) Partial Affinity Field Connection: For detected joints, it is necessary to determine which joints belong to the same human body and form the correct skeletal connections. Partial Affinity Field L c (p) is a 2D vector field that stores the position and orientation information of the limb. For the limb c of the k-th person, its PAF value at pixel p is defined as:

[0022]

[0023] v represents a unit vector in the direction of the limb, and its expression is:

[0024]

[0025] In the formula, x j1,k and x j2,k Let p represent the true positions of body nodes j1 and j2 for the limb c of the figure k in the figure. Point p is located on the plane vector pointing from j1 to j2. Therefore, the joint affinity field value of point p is the vector sum of j1 and j2.

[0026] 3) By using the joints to pair d j1 and d j2 The PAF (Proof-of-Flight) method determines whether the connection before the key point is correct; its expression is:

[0027]

[0028] In the formula, d j1 and d j2 E represents two consecutive pixels; p(u) represents the pixel between consecutive pixels. The larger the E value, the higher the consistency between the connection direction and the PAF field direction, and the higher the probability of key point connection. When the E value exceeds a preset threshold, it is determined that there is a valid connection between two key points, thus forming a complete human skeleton map.

[0029] Furthermore, the cross-frame spatiotemporal graph G=(V,E) includes:

[0030] 1) Node set V = {v ti | t=1,…,T; i=1,…,N} represents all the joints in the skeleton sequence, where T is the number of video frames and N is the total number of human joints in each frame. Each node v tIt contains a feature vector consisting of the (x,y) coordinates of the key point in the image and the confidence score.

[0031] 2) The marginal set E consists of two subsets, namely E S and E F Intra-frame connection edge E S ={v ti v tj | v ti v tj The (i,j)∈H} describes the skeleton-internal connections of each frame, where H is the set of naturally connected human joints; inter-frame connections EF={v ti v (t+1)i} contains inter-frame edges, connecting the same joints between adjacent frames, and characterizing the motion trajectory of that joint point over time.

[0032] Furthermore, the spatiotemporal graph convolution of the AT-STGCN network includes: layers 1-3 with 64 output channels, layers 4-6 with 128 output channels, and layers 7-9 with 256 output channels. Each layer includes a graph convolutional network and a temporal convolutional network, and integrates spatial attention and temporal attention mechanisms in each spatiotemporal graph convolutional block.

[0033] 1) Spatial Attention Mechanism: The spatial attention mechanism is used to extract spatial structural association features between different joints, enabling the network to automatically focus on key joints that contribute more to the recognition of the current action (e.g., when recognizing a "climbing" action, joints in the arms and legs are more important than joints in the torso). The core idea of ​​the spatial attention mechanism is: for each joint, calculate its correlation weight with all other relevant nodes; the joint with the larger the weight, the greater its contribution to the feature representation of the current joint.

[0034] 2) Temporal Attention Mechanism: The temporal attention mechanism is used to extract the temporal motion features of the same key point across different frames, enabling the network to automatically focus on key frames in the action sequence (such as the start frame, transition frame, and end frame of the action). The core idea of ​​the temporal attention mechanism is: for each key point, calculate its relevance weight across different time frames; the frame with the larger the weight, the greater its contribution to the feature representation of the current frame.

[0035] Furthermore, the input to the AT-STGCN network includes: a constructed spatiotemporal graph of the human skeleton G=(V,E), with node feature dimensions of N×T×C, where N is the number of joints, T is the number of input frames, and C is the feature dimension. After multiple layers of spatiotemporal graph convolution, the spatiotemporal dimension is pooled through a global average pooling layer, and finally, the probability distribution of the behavior category is output through a fully connected layer and a softmax layer.

[0036] Furthermore, the neighborhood node space partitioning strategy of the AT-STGCN model includes:

[0037] 1) Move node v ti The neighborhood set is divided into three subsets: the root node subset, the centripetal node subset, and the centrifugal node subset;

[0038] 2) The label division function is defined as follows:

[0039]

[0040] In the formula, d j d represents the average distance from the joint to the center of gravity of the skeleton. i This represents the average distance from the joints to the centroid of the skeleton in all skeleton spacetime graphs.

[0041] Furthermore, the spatiotemporal graph convolution includes:

[0042] 1) Building upon spatial graph convolution, the temporal dimension is incorporated into the convolutional neighborhood, forming spatiotemporal graph convolution. For node v ti The mathematical expression for its graph convolution operation is:

[0043]

[0044] In the formula, f in Given the input feature map, f out To output the feature map, Z ti (v tj p(v) is the normalization term. ti ,v tj ) is the sampling function, w(v) ti ,v tj ) is the weight function.

[0045] 2) Based on spatial graph convolution, spatiotemporal modeling of the image's skeletal sequence is performed. In graph construction, the temporal domain of the graph is formed by connecting the same nodes of adjacent frames. Therefore, to extend spatial graph convolution to the spatiotemporal domain, the concept of neighborhood needs to be expanded. That is, the range of the neighborhood in spatiotemporal graph convolution is as follows:

[0046]

[0047]

[0048] In the formula, the parameter ϕ controls the time range of the graph to be included in the neighborhood.

[0049] 3) To perform convolution operations on the spatiotemporal graph, the weight function needs to be redefined. Since the time axis is ordered, v needs to be modified when defining the weight function. ti Spatiotemporal neighborhood tags:

[0050]

[0051] In the formula ς ti (v tj ) is node v in the single-frame case. ti The neighborhood of the location.

[0052] Furthermore, the rule library for unsafe behaviors at the power distribution network live-line operation site includes: insufficient safety distance, failure to wear a safety helmet, and non-standard insulation shielding.

[0053] Furthermore, the unsafe behavior judgment logic includes:

[0054] 1) For unsafe behaviors based on action categories: When the behavior category identified by the model is consistent with the behavior category defined in the rule base and the confidence level exceeds the threshold, an alarm is triggered directly;

[0055] 2) For unsafe behaviors that require spatial distance judgment: Combine the spatial distance information calculated by monocular visual SLAM technology, and trigger an alarm when the spatial distance is less than the preset safety threshold.

[0056] Furthermore, the unsafe behavior identification system for live-line power distribution operations integrating OpenPose and AT-STGCN includes: a memory, a processing unit, and a display. The memory stores a computer program for storing video data and model parameters. The processing unit executes the above-mentioned method. The display outputs and displays the unsafe behavior identification results and alarm information.

[0057] Compared with the prior art, the beneficial effects of the present invention include:

[0058] 1. The proposed method for identifying unsafe behaviors in live-line power distribution operations combines OpenPose skeleton extraction technology with the AT-STGCN model, which incorporates a spatiotemporal attention mechanism. The temporal attention mechanism enhances the network's ability to capture key motion frames, while the spatial attention mechanism enhances the network's ability to focus on key nodes, significantly improving the accuracy of behavior recognition in complex scenarios of live-line power distribution operations.

[0059] 2. The proposed method for identifying unsafe behaviors in live-line work of distribution networks, which integrates OpenPose and AT-STGCN, has designed a rule library for unsafe behaviors in live-line work scenarios of distribution networks. By combining behavior category recognition and spatial distance judgment, it can accurately identify and provide real-time alarms for typical unsafe behaviors such as insufficient safety distance, non-standard insulation shielding, and violation of bucket truck operation regulations. Attached Figure Description

[0060] Figure 1 The flowchart shows the method for identifying unsafe behaviors at the site of live-line work in power distribution networks, which integrates OpenPose and AT-STGCN, provided by this invention.

[0061] Figure 2 This is a schematic diagram of key points of the human skeleton in the OpenPose model provided by the present invention.

[0062] Figure 3 The diagram shows the network model structure of the AT-STGCN provided by this invention.

[0063] Figure 4 This is a schematic diagram of the construction of the human skeleton spatiotemporal map provided by the present invention. Detailed Implementation

[0064] This invention provides a method for identifying unsafe behaviors during live-line work in distribution networks that integrates OpenPose and AT-STGCN. Figure 1 Let's take a look. Figure 1 The flowchart of the unsafe behavior identification method for live-line work in distribution networks based on the integration of OpenPose and AT-STGCN provided by the present invention includes steps S1 to S5, wherein:

[0065] This invention provides a method for identifying unsafe behaviors at power distribution network live-line work sites that integrates OpenPose and AT-STGCN, including:

[0066] In step S1, video data of the power distribution network live-line operation site is collected to construct a dataset of images of the workers' behavior.

[0067] In step S2, the OpenPose model is used to extract human skeleton key points from video frames of the image dataset, and partial affinity fields are used to connect the key points to construct a spatiotemporal map of the human skeleton.

[0068] In step S3, an AT-STGCN model incorporating a spatiotemporal attention mechanism is constructed. The AT-STGCN model includes a temporal attention mechanism module and a spatial attention mechanism module.

[0069] In step S4, the spatiotemporal map of the human skeleton is input into the AT-STGCN model for behavior classification and recognition, and the behavior category is output.

[0070] In step S5, a logic for judging unsafe behaviors during live-line work in the distribution network is designed. The identified behavior categories are matched with a preset unsafe behavior rule base to determine whether an unsafe behavior has occurred and to issue an alarm.

[0071] Preferably, this embodiment uses a high-definition camera deployed on the insulated bucket truck and a fixed ground monitoring camera to collect video of the work site under different angles and lighting conditions. As a specific embodiment, FFmpeg is first used to extract video frames at a fixed frame rate (25 frames / second in this embodiment), forming a raw image dataset. The dataset covers normal behavior categories and unsafe behavior categories. Normal behaviors include standing, walking, bending over, climbing, reaching out, and squatting; unsafe behaviors include insufficient safety distance, not wearing a safety helmet, improper insulation shielding, improper operation of the bucket truck, dangerous climbing by personnel, and accidentally entering a live area.

[0072] Preferably, the key points of the human skeleton include 14 key points: nose, neck, left and right shoulders, left and right elbows, left and right wrists, left and right hips, left and right knees, and left and right ankles. The OpenPose model employs a bottom-up detection strategy. First, it extracts features from the input image using the first 10 layers of a VGG network, generating a feature map F. Then, it iterative prediction is performed using a multi-stage convolutional neural network, with each stage containing two branches: one branch outputs a 2D confidence map S of the joint points. t The other branch output part is the affinity field L. t The calculation process is as follows:

[0073]

[0074]

[0075] In the formula, ρ t For the prediction network of 2D confidence graphs, ϕ t This is a prediction network for partial affinity fields. Through multi-stage progressive optimization, the keypoint detection results are made more accurate.

[0076] Preferably, for detected joints, a partial affinity field is used to determine the connection validity between joints. For the limb c of the k-th person, its PAF value at pixel p is defined as:

[0077]

[0078] In the formula, v is a unit vector in the direction of the limb, and its expression is:

[0079]

[0080] Preferably, the two joint points d are calculated. j1 PAF integral on the DJ2 connection line:

[0081]

[0082]

[0083] In the formula, p(u) represents the pixel interpolation point on the line connecting two joints. As a specific embodiment, in this embodiment, when the E value exceeds 0.5, it is determined that there is a valid connection between the two joints, thereby forming a complete human skeleton map.

[0084] Preferably, a spatiotemporal graph structure spanning multiple frames is constructed based on a single-frame skeleton graph. The spatiotemporal graph G=(V,E) is defined, having N key points and T frame skeleton sequences, where the node set V={v... ti | t=1,…,T; i=1,…,N} represents all joints in the skeleton sequence, where T is the frame number and N is the total number of human joints; the edge set E consists of intra-frame connected edges E S ={v ti v tj | (i,j)∈H} and inter-frame connection edge E F ={ ti v (t+1)i} constitutes the set of natural connecting joints in the human body, where H is the set of natural connecting joints in the human body.

[0085] Preferably, the AT-STGCN network consists of 9 spatiotemporal graph convolutional layers: layers 1-3 contain 64 output channels, layers 4-6 contain 128 output channels, and layers 7-9 contain 256 output channels. Each layer includes a graph convolutional network and a temporal convolutional network, and spatial and temporal attention mechanisms are integrated in each spatiotemporal graph convolutional block. As a specific embodiment, the number of attention heads is set to 8, and the feature dimension d... k Set to 64. The graph convolution operation formula for the AT-STGCN model is as follows:

[0086]

[0087] In the formula, f in Given the input feature map, f out To output the feature map, Z ti (v tj p(v) is the normalization term. ti ,v tj ) is the sampling function, w(v) ti ,v tj The neighborhood set B(v) is the weight function. ti The spatiotemporal partitioning method is as follows:

[0088]

[0089] In the formula, the parameter ϕ controls the temporal range of the graph to be included in the neighborhood. As a specific embodiment, in this embodiment, the spatial neighborhood range K is set to 1, and the temporal neighborhood range ϕ is set to 3.

[0090] Preferably, in the spatial partitioning strategy for neighboring nodes, the partitioning label function is defined as follows:

[0091]

[0092] In the formula, d j For the joint point v tj The average distance to the center of gravity of the skeleton, d i This is the average distance from all relevant nodes to the center of gravity of the skeleton.

[0093] Preferably, in this embodiment, the AT-STGCN model is first pre-trained on the NTURGB+D public human motion dataset, which contains 60 motion categories and 56,880 sample video clips. Then, the pre-trained model is fine-tuned using the power distribution network uninterrupted operation behavior image dataset constructed in step S1. As a specific embodiment, the training parameters are set as follows: 50 epochs, a batch size of 64, an initial learning rate of 0.001, the Adam optimizer, and Dropout to prevent overfitting, with a Dropout rate of 0.5.

[0094] Preferably, the unsafe behavior rule base includes: insufficient safe distance, failure to wear a safety helmet, and improper insulation shielding. As a specific embodiment, this embodiment defines the judgment thresholds for various unsafe behaviors according to the "State Grid Corporation of China Power Safety Work Regulations".

[0095] Furthermore, the unsafe behavior judgment logic includes:

[0096] 1) For unsafe behaviors based on action categories: When the behavior category identified by the model matches the behavior category defined in the rule base, and the confidence level exceeds the threshold, an alarm is triggered directly. As a specific example, the confidence level threshold is set to 0.7 in this example.

[0097] 2) For unsafe behaviors requiring spatial distance judgment: Using spatial distance information calculated with monocular visual SLAM technology, an alarm is triggered when the spatial distance is less than a preset safety threshold. In this specific embodiment, the safety distance threshold is set to 0.7 meters.

[0098] Furthermore, alarm information is displayed to on-site personnel in real time on the MR intelligent auxiliary terminal's screen, and simultaneously pushed to the remote monitoring platform via wireless network.

[0099] This invention creatively combines OpenPose skeleton extraction technology with the AT-STGCN model, which incorporates a spatiotemporal attention mechanism, for the identification of unsafe behaviors during live-line work in power distribution networks. Through the dual enhancement of temporal and spatial attention mechanisms, the accuracy of behavior recognition is significantly improved. By constructing a rule base for unsafe behaviors specific to power operation scenarios, accurate warnings are issued for dangerous behaviors such as insufficient safety distance and failure to wear safety helmets, effectively ensuring on-site safety during live-line work in power distribution networks.

[0100] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying unsafe behaviors at live-line working sites in power distribution networks, integrating OpenPose and AT-STGCN, characterized in that... include: (1) Collect video data of the power distribution network live-line operation site and construct a dataset of images of the operation personnel's behavior; (2) Use the OpenPose model to extract human skeleton key points from video frames of the image dataset, and use partial affinity fields to connect the key points to construct a spatiotemporal map of human skeleton; (3) Construct an AT-STGCN model that incorporates a spatiotemporal attention mechanism, wherein the AT-STGCN model includes a temporal attention mechanism module and a spatial attention mechanism module; (4) Input the spatiotemporal map of the human skeleton into the AT-STGCN model for behavior classification and recognition, and output the behavior category; (5) Design a logic for judging unsafe behaviors during live-line work in power distribution networks, match the identified behavior categories with the preset unsafe behavior rule base, and determine whether unsafe behaviors have occurred.

2. The method for identifying unsafe behaviors at power distribution network live-line working sites that integrates OpenPose and AT-STGCN as described in claim 1, characterized in that, The method for constructing the spatiotemporal graph of the human skeleton in step S2 includes: defining a spatiotemporal graph G=(V,E), having N joints and T frame skeleton sequences, where the node set V={v ti | t=1,…,T; i=1,…,N} represents all joints in the skeleton sequence, where T is the frame number and N is the total number of human joints; the edge set E consists of intra-frame connected edges E S ={v ti v tj | (i,j)∈H} and inter-frame connection edge E F ={v ti v (t+1)i } constitutes the set of natural connecting joints in the human body, where H is the set of natural connecting joints in the human body.

3. The method for identifying unsafe behaviors at power distribution network live-line working sites that integrates OpenPose and AT-STGCN as described in claim 1, characterized in that, The method for connecting key points using partial affinity fields in step S2 includes: calculating key point pairs d. j1 and d j2 The PAF affinity E between them is given by the following formula: , In the formula, L c (p(u)) is the partial affinity field vector at pixel p(u), d j1 and d j2 Let be the position vector of two consecutive pixels.

4. The method for identifying unsafe behaviors at power distribution network live-line working sites that integrates OpenPose and AT-STGCN as described in claim 1, characterized in that, The spatiotemporal attention mechanism in step S3 includes: 1) Temporal attention mechanism: Extract the temporal motion features of the same key point between different frames and calculate the attention weights in the temporal dimension; 2) Spatial attention mechanism: Extract spatial structural association features between different key points and calculate attention weights in the spatial dimension.

5. The method for identifying unsafe behaviors at power distribution network live-line working sites that integrates OpenPose and AT-STGCN as described in claim 1, characterized in that, The graph convolution operation formula for the AT-STGCN model in step S3 is as follows: ， Among them, f in Given the input feature map, f out To output the feature map, Z ti (v tj p(v) is the normalization term. ti ,v tj ) is the sampling function, w(v) ti ,v tj ) is the weight function.

6. The method for identifying unsafe behaviors at power distribution network live-line working sites that integrates OpenPose and AT-STGCN as described in claim 5, is characterized in that... The neighborhood set B(v) ti The spatiotemporal partitioning method is as follows: ， Wherein d(v ti ,v tj ) represents the distance between nodes, K is the spatial neighborhood range, and ϕ is the temporal neighborhood range.

7. The method for identifying unsafe behaviors at power distribution network live-line working sites that integrates OpenPose and AT-STGCN as described in claim 1, characterized in that, The unsafe behavior rule base in step S5 includes: insufficient safety distance, failure to wear a safety helmet, improper insulation shielding, violation of bucket truck operation regulations, dangerous climbing behavior, and accidental entry into a live area.

8. The method for identifying unsafe behaviors at power distribution network live-line working sites that integrates OpenPose and AT-STGCN as described in claim 1, characterized in that, The OpenPose model extracts 14 key points of the human skeleton, including: nose, neck, left and right shoulders, left and right elbows, left and right wrists, left and right hips, left and right knees, and left and right ankles.

9. A system for identifying unsafe behaviors at live-line working sites in power distribution networks, integrating OpenPose and AT-STGCN, characterized in that... The device includes a memory, a processing unit, and a display. The memory stores a computer program for storing video data and model parameters. The processing unit is used to execute the method described in any one of claims 1-8. The display is used to output and show the results of unsafe behavior identification and alarm information.

Images