A power operation area periphery pedestrian safety early warning method

Abstract

The present application relates to a kind of electric power operation area peripheral pedestrian safety early warning method, comprising the following steps: step S1: the step of obtaining video stream data, which extracts the video stream data of work site from the video monitoring device of electric power operation site, and the video stream data extracted is decoded and analyzed to obtain the image data information to be detected;Step S2: the step of image data segmentation, the image data to be detected obtained in step S1 is input into the isolation device segmentation model trained, and the segmentation image output by isolation device segmentation model is morphologically processed, to obtain the outer contour data of isolation device;Step S3: the step of obtaining pedestrian target frame position data, step S4: the step of obtaining the human body key point data of pedestrian, step S5: the step of predicting the running direction of pedestrian, step S6: the step of pedestrian trajectory prediction.

Description

Technical Field

[0001] This invention belongs to the field of power safety early warning technology, and relates to an early warning method and system, particularly a method for early warning of pedestrian safety outside a power operation area. Background Technology

[0002] In existing technologies, isolation nets, warning lines, fences, and other devices should be set up in power operation sites such as high-altitude operations and high-voltage equipment construction to prevent pedestrians from approaching or entering.

[0003] If the work site is complex and crowded, pedestrians may accidentally enter the work area, posing a significant safety hazard.

[0004] Currently, pedestrian access to the work area is managed using either dedicated personnel or unmanned operation. Dedicated personnel typically require multiple staff members to ensure the safety of the entire perimeter of the work area. If a pedestrian approaches or attempts to enter the work area, the staff will stop them. However, this method is labor-intensive. Unmanned operation usually uses infrared sensors to monitor for approaching pedestrians. Because this method cannot distinguish between workers, pedestrians, or general heat sources, it is prone to false detections. This is a shortcoming of existing technology.

[0005] In view of this, it is very necessary to provide a method for pedestrian safety warning outside the power operation area to solve the above-mentioned technical problems existing in the prior art. Summary of the Invention

[0006] The purpose of this invention is to address the deficiencies of the existing technology by providing a method for designing a safety early warning system for pedestrians outside a power operation area, thereby solving the aforementioned technical problems.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for pedestrian safety warning outside a power operation area includes the following steps:

[0009] Step S1: The step of acquiring video stream data. This step extracts video stream data of the power operation site from the video monitoring device at the operation site, and decodes and analyzes the extracted video stream data to obtain the image data information to be detected.

[0010] Step S2: The image data segmentation step involves inputting the image data to be detected obtained from the decoding and analysis in step S1 into the trained isolation device segmentation model, and performing morphological processing on the segmented image output by the isolation device segmentation model to obtain the outer contour data of the isolation device.

[0011] Step S3: Obtaining pedestrian bounding box location data. In this step, the image data to be detected obtained from the decoding and analysis in step S1 is input into the trained worker inspection model to check for workers and pedestrians in the image data. If a pedestrian is present, the pedestrian bounding box location data [x] is output. min ,y min ,x max ,y max ];

[0012] Step S4: Obtaining pedestrian keypoint data. In this step, the pedestrian's bounding box position data [x] from step S3 is used as the basis for the data. min ,y min ,x max ,y max The image data to be detected in step S1 is cropped, and the cropped image data is input into the trained human keypoint detection model to obtain human keypoint data, which is represented as: p0={(x0,y0),(x1,y1)……(x 24 ,y 24 The human body location points are obtained by weighted averaging, and the positions of the human body location points in the image to be detected described in step S1 are calculated.

[0013] Step S5: The step of predicting the direction of pedestrian movement. In this step, a behavioral movement direction prediction model is constructed using the human body position points in step S4 to make a rough prediction of the pedestrian movement trend.

[0014] Step S6: Pedestrian trajectory prediction step. In this step, a pedestrian trajectory prediction model is established using Long Short-Term Memory Networks (LSTM) to make detailed predictions of pedestrian positions within a preset time period in the future.

[0015] Preferably, in step S1:

[0016] Video stream data is decoded using OpenCV tools.

[0017] Preferably, in step S2, the isolation device is a fence, isolation net, telescopic guardrail, or other device with isolation and warning functions set up at the work site. The trained isolation device segmentation model is trained based on the DeepLabV3 semantic segmentation network. The purpose of using the isolation device segmentation model is to segment the isolation devices at the work site, thereby obtaining the outer contour of the isolation devices in the image.

[0018] The specific training process of the isolation device segmentation model is as follows:

[0019] Preparing the training set: Images of the work site with isolation devices are taken. The isolation devices in the training set images are labeled using the Pixel AnnotationTools tool to obtain image labels. The images and their labels are matched one-to-one as the training set for the isolation device segmentation model. The dataset is set up with a training set and a validation set at an 8:2 ratio. The training set is input into the DeepLabV3 keypoint detection algorithm and trained on an NVIDIA GPU server using the algorithm's default parameters. The model performance is evaluated using IOU as the evaluation criterion until the model meets the usage requirements, thus obtaining the isolation device segmentation model.

[0020] Preferably, in step S3, the workers wear different clothing than pedestrians; the workers wear electrical grid work clothes and safety helmets. The target box is a shortened version of the result output by the YOLOv5 model, including the minimum bounding box [x]. min ,y min ,x max ,y max The information consists of three parts: category, confidence level, and other information. The coordinates of the upper left corner of the minimum rectangular bounding box are (x...). min ,y min The coordinates of the lower right corner are (x max ,y max );

[0021] The worker detection model trained in step S3 is based on the YOLOv5 object detection network. The specific training process of the worker detection model is as follows:

[0022] Preparing the training set: Images of the work site are taken, and then LabLE Img annotation tools are used to annotate the workers and pedestrians in the training set images to obtain image labels. The images and their corresponding labels are used as the training set for the worker detection model. The dataset is set up with a training set and a validation set at an 8:2 ratio. The training set is input into the YOLOv5 object detection algorithm and trained on an NVIDIA GPU server using the algorithm's default parameters. The model performance is evaluated using mAP as the evaluation criterion until the model meets the usage requirements, thus obtaining the worker detection model.

[0023] Preferably, in step S4, the trained human keypoint detection model is trained based on the OpenPose keypoint detection network; the specific training process of the human keypoint detection model is as follows:

[0024] Preparing the training set: The work site with isolation devices was photographed, and then the Labellme annotation tool was used to annotate 25 key points of the human body in the training set images to obtain image labels. The photographed images and the generated labels were matched one by one as the dataset for human key point detection.

[0025] Set the training set and validation set in a ratio of 8:2;

[0026] The training set is input into the OpenPose keypoint detection algorithm and trained on an NVIDIA GPU server using the algorithm's default parameters.

[0027] The performance of the model is evaluated using AP as the evaluation criterion. Based on engineering experience, this invention sets AP > 0.79, and the trained model meets the usage requirements, thus obtaining the human key point detection model.

[0028] like Figure 2 The diagram shows 25 key points of the human body. The key points for the torso are numbered a1 = {0, 1, 2, 5, 8, 9, 12, 15, 16, 17, 18}, with coordinate set p1 = {(x0, y0), (x1, y1)……(x18, y18)}. The key points for the limbs are set a2 = {3, 4, 6, 7, 10, 11, 13, 14, 19, 20, 21, 22, 23, 24}, with coordinate set p2 = {(x3, y3), (x4, y4)……(x24, y24)}. Since the key points of the torso position better reflect the true position of the human body during movement, based on engineering experience, the weight of the key points for the torso is k1 = 0.8, while the weight of the key points for the limbs is k2 = 0.2. The values ​​of k1 and k2 can be adjusted according to different implementation scenarios, but the principle that k1 is greater than k2 should be followed. Human body position point p3 = (x l y l )=p1*k1 / 11+p2*k2 / 14, the position of the human body relative to the image to be inspected described in step S1 is p=(x,y)=((x l +x min ),(y l +y min )).

[0029] Preferably, step S5 includes the following steps:

[0030] Step S51: Define the direction of pedestrian movement and obtain the sequence of pedestrian position points P = {p1, p2, ..., p...} in n frames of images. n-1 ,p n}, where t n The direction of motion of the position point at time t is defined as t n-1 Time and tn The angle θ formed by the vector of points at time and position and the horizontal axis in a counterclockwise direction n ,like Figure 3 As shown. The range of θ is 0°-360°. The parameter n is set according to the camera frame rate f and actual engineering needs, and its range is 3f-5f. In this embodiment of the invention, n is set to 90.

[0031] Step S52: Construct a direction prediction model, and form a sample set θ = {θ1, θ2, ..., θ3} of the motion directions of all positions in step S51. n-1 ,θ n A k-means algorithm is used to construct a direction prediction model. The k value is set according to engineering requirements; in this embodiment, k is set to 12. After calculation by the k-means model, the output clustering result is C = {C1, C2, ..., C}. k}, where C k Let k be the k-th cluster. Select the cluster with the largest number of samples, and use the angle of its cluster center as the predicted direction of the pedestrian. The predicted direction means the most representative direction of movement of the pedestrian in the n-frame image sequence.

[0032] Step S53: Pedestrian movement trend judgment. During implementation, a rough warning angle range a needs to be set based on the camera angle and the outer contour of the isolation device in step S2. min -a max ,like Figure 3 As shown. If the pedestrian's direction of movement in step S52 is within the rough warning angle range, then proceed to step S6; otherwise, return to step S1.

[0033] Preferably, step S6 includes the following steps:

[0034] Step S61: Construct a dataset by obtaining the human body location point sequence of n frames from step S4, where the location point sequence of each pedestrian is P = {p1, p2, ..., p...} n-1 ,p n These serve as training or testing samples for the LSTM. In this embodiment of the invention, the number of training samples for human body location sequence is 10,000, and the number of testing samples is 5,000. In this embodiment of the invention, n is set to 90.

[0035] Step S62: LSTM network training. The sequence in step S61 is preprocessed by difference and normalization. The network is trained using the Adam optimizer with a learning rate of 0.001.

[0036] Step S63: Pedestrian position prediction, let the current time be t. n Using the sequence of human body location points in n frames of images as input to the LSTM model, the prediction of t is performed. n+1 -tn+m The location of pedestrians within a time period, such as Figure 4 As shown. Parameters n and m are set according to engineering requirements. In this embodiment of the invention, n is set to 90 and m is set to 15.

[0037] Step S64: Alarm Judgment. Based on the pedestrian trajectory predicted in Step S63, determine whether it intersects with the outer contour of the isolation device. If an intersection occurs, it indicates that the pedestrian is about to enter the work area, and an alarm needs to be issued to the pedestrian; otherwise, return to Step S1.

[0038] The beneficial effects of this invention lie in its ability to model key targets around the work site using image processing and deep learning methods. This enables rapid differentiation between workers and pedestrians, and extraction of pedestrian location information, thereby improving the monitoring capability for safety around the work site. A deep learning-based method is used to extract key human body points, and a human body location estimation model is established based on the fluctuation amplitude of different key points during human movement, improving the robustness of human body location prediction. This invention employs a two-level early warning strategy to warn pedestrians approaching or attempting to enter the work area. The first level of early warning uses a machine learning-based method to quickly identify the direction of pedestrian movement; the second level uses a deep learning-based method to predict pedestrian trajectories. This two-level early warning approach can dynamically execute the early warning strategy according to the actual scenario, improving algorithm processing efficiency.

[0039] Furthermore, the design principle of this invention is reliable, the structure is simple, and it has a very wide range of application prospects.

[0040] Therefore, it is evident that the present invention has outstanding substantive features and significant progress compared with the prior art, and the beneficial effects of its implementation are also obvious. Attached Figure Description

[0041] Figure 1 This is a flowchart of a pedestrian safety early warning method for the periphery of a power operation area provided by the present invention.

[0042] Figure 2 This is a schematic diagram of key human body point annotation in an embodiment of the present invention.

[0043] Figure 3 This is a schematic diagram of pedestrian movement direction and first-level early warning in an embodiment of the present invention.

[0044] Figure 4 This is a schematic diagram of pedestrian trajectory prediction in an embodiment of the present invention. Detailed Implementation

[0045] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The following embodiments are explanations of the present invention, but the present invention is not limited to the following implementation methods.

[0046] like Figure 1-4 As shown in the figure, this embodiment provides a pedestrian safety warning method for the perimeter of a power operation area, which includes the following steps:

[0047] Step S1: The step of acquiring video stream data. This step extracts video stream data from the video monitoring device at the power work site and decodes and analyzes the extracted video stream data to obtain the image data information to be detected; in step S1:

[0048] Video stream data is decoded using OpenCV tools.

[0049] Step S2: Image data segmentation step. The image data to be detected obtained from the decoding and analysis in step S1 is input into the trained isolation device segmentation model. Morphological processing is performed on the segmented image output by the isolation device segmentation model to obtain the outer contour data of the isolation device. In step S2, the isolation device refers to a fence, isolation net, telescopic guardrail, or other device with isolation and warning functions set up at the work site. The trained isolation device segmentation model is trained based on the DeepLabV3 semantic segmentation network. The purpose of using the isolation device segmentation model is to segment the isolation devices at the work site to obtain the outer contour of the isolation devices in the image.

[0050] The specific training process of the isolation device segmentation model is as follows:

[0051] Preparing the training set: Images of the work site with isolation devices are taken. The isolation devices in the training set images are labeled using the Pixel AnnotationTools tool to obtain image labels. The images and their labels are matched one-to-one as the training set for the isolation device segmentation model. The dataset is set up with a training set and a validation set at an 8:2 ratio. The training set is input into the DeepLabV3 keypoint detection algorithm and trained on an NVIDIA GPU server using the algorithm's default parameters. The model performance is evaluated using IOU as the evaluation criterion until the model meets the usage requirements, thus obtaining the isolation device segmentation model.

[0052] Step S3: Obtaining pedestrian bounding box location data. In this step, the image data to be detected obtained from the decoding and analysis in step S1 is input into the trained worker inspection model to check for workers and pedestrians in the image data. If a pedestrian is present, the pedestrian bounding box location data [x] is output. min ,y min ,x max ,y maxIn step S3, the workers wear different clothing than pedestrians; they wear electrical grid work clothes and safety helmets. The target box is a shortened version of the YOLOv5 model output, including the minimum bounding box [x]. min ,y min ,x max ,y max The information consists of three parts: category, confidence level, and other information. The coordinates of the upper left corner of the minimum rectangular bounding box are (x...). min ,y min The coordinates of the lower right corner are (x max ,y max );

[0053] The worker detection model trained in step S3 is based on the YOLOv5 object detection network. The specific training process of the worker detection model is as follows:

[0054] Preparing the training set: Images of the work site are taken, and then LabLE Img annotation tools are used to annotate the workers and pedestrians in the training set images to obtain image labels. The images and their corresponding labels are used as the training set for the worker detection model. The dataset is set up with a training set and a validation set at an 8:2 ratio. The training set is input into the YOLOv5 object detection algorithm and trained on an NVIDIA GPU server using the algorithm's default parameters. The model performance is evaluated using mAP as the evaluation criterion until the model meets the usage requirements, thus obtaining the worker detection model.

[0055] Step S4: Obtaining pedestrian keypoint data. In this step, the pedestrian's bounding box position data [x] from step S3 is used as the basis for the data. min ,y min ,x max ,y max The image data to be detected in step S1 is cropped, and the cropped image data is input into the trained human keypoint detection model to obtain human keypoint data, which is represented as: p0={(x0,y0),(x1,y1)……(x 24 ,y 24 The human body location points are obtained through weighted averaging, and their positions in the image to be detected in step S1 are calculated. In step S4, the trained human body keypoint detection model is trained based on the OpenPose keypoint detection network. The specific training process of the human body keypoint detection model is as follows:

[0056] Preparing the training set: The work site with isolation devices was photographed, and then the Labellme annotation tool was used to annotate 25 key points of the human body in the training set images to obtain image labels. The photographed images and the generated labels were matched one by one as the dataset for human key point detection.

[0057] Set the training set and validation set in a ratio of 8:2;

[0058] The training set is input into the OpenPose keypoint detection algorithm and trained on an NVIDIA GPU server using the algorithm's default parameters.

[0059] The performance of the model is evaluated using AP as the evaluation criterion. Based on engineering experience, this invention sets AP > 0.79, and the trained model meets the usage requirements, thus obtaining the human key point detection model.

[0060] like Figure 2 The diagram shows 25 key points of the human body. The key points for the torso are numbered a1 = {0, 1, 2, 5, 8, 9, 12, 15, 16, 17, 18}, with coordinate set p1 = {(x0, y0), (x1, y1)……(x18, y18)}. The key points for the limbs are set a2 = {3, 4, 6, 7, 10, 11, 13, 14, 19, 20, 21, 22, 23, 24}, with coordinate set p2 = {(x3, y3), (x4, y4)……(x24, y24)}. Since the key points of the torso position better reflect the true position of the human body during movement, based on engineering experience, the weight of the key points for the torso is k1 = 0.8, while the weight of the key points for the limbs is k2 = 0.2. The values ​​of k1 and k2 can be adjusted according to different implementation scenarios, but the principle that k1 is greater than k2 should be followed. Human body position point p3 = (x l y l )=p1*k1 / 11+p2*k2 / 14, the position of the human body relative to the image to be inspected described in step S1 is p=(x,y)=((x l +x min ),(y l +y min )).

[0061] Step S5: The step of predicting the direction of pedestrian movement. In this step, a behavioral motion direction prediction model is constructed using the human body position points from step S4 to make a rough prediction of the pedestrian movement trend. Step S5 includes the following steps:

[0062] Step S51: Define the direction of pedestrian movement and obtain the sequence of pedestrian position points P = {p1, p2, ..., p...} in n frames of images. n-1 ,p n}, where t n The direction of motion of the position point at time t is defined as t n-1 Time and t n The angle θ formed by the vector of points at time and position and the horizontal axis in a counterclockwise direction n ,like Figure 3 As shown. The range of θ is 0°-360°. The parameter n is set according to the camera frame rate f and actual engineering needs, and its range is 3f-5f. In this embodiment of the invention, n is set to 90.

[0063] Step S52: Construct a direction prediction model, and form a sample set θ = {θ1, θ2, ..., θ3} of the motion directions of all positions in step S51. n-1 ,θ n A k-means algorithm is used to construct a direction prediction model. The k value is set according to engineering requirements; in this embodiment, k is set to 12. After calculation by the k-means model, the output clustering result is C = {C1, C2, ..., C}. k}, where C k Let k be the k-th cluster. Select the cluster with the largest number of samples, and use the angle of its cluster center as the predicted direction of the pedestrian. The predicted direction means the most representative direction of movement of the pedestrian in the n-frame image sequence.

[0064] Step S53: Pedestrian movement trend judgment. During implementation, a rough warning angle range a needs to be set based on the camera angle and the outer contour of the isolation device in step S2. min -a max ,like Figure 3 As shown. If the pedestrian's direction of movement in step S52 is within the rough warning angle range, then proceed to step S6; otherwise, return to step S1.

[0065] Step S6: Pedestrian trajectory prediction. This step uses a Long Short-Term Memory Network (LSTM) to build a pedestrian trajectory prediction model, and performs detailed predictions of pedestrian positions within a preset time period. Step S6 includes the following steps:

[0066] Step S61: Construct a dataset by obtaining the human body location point sequence of n frames from step S4, where the location point sequence of each pedestrian is P = {p1, p2, ..., p...} n-1 ,p n These serve as training or testing samples for the LSTM. In this embodiment of the invention, the number of training samples for human body location sequence is 10,000, and the number of testing samples is 5,000. In this embodiment of the invention, n is set to 90.

[0067] Step S62: LSTM network training. The sequence in step S61 is preprocessed by difference and normalization. The network is trained using the Adam optimizer with a learning rate of 0.001.

[0068] Step S63: Pedestrian position prediction, let the current time be t. n Using the sequence of human body location points in n frames of images as input to the LSTM model, the prediction of t is performed. n+1 -t n+m The location of pedestrians within a time period, such as Figure 4 As shown. Parameters n and m are set according to engineering requirements. In this embodiment of the invention, n is set to 90 and m is set to 15.

[0069] Step S64: Alarm Judgment. Based on the pedestrian trajectory predicted in Step S63, determine whether it intersects with the outer contour of the isolation device. If an intersection occurs, it indicates that the pedestrian is about to enter the work area, and an alarm needs to be issued to the pedestrian; otherwise, return to Step S1.

[0070] The above-disclosed embodiments are merely preferred embodiments of the present invention, but the present invention is not limited thereto. Any non-creative variations that can be conceived by those skilled in the art, as well as any improvements and modifications made without departing from the principles of the present invention, should fall within the protection scope of the present invention.

Claims

1. A method for pedestrian safety early warning outside a power operation area, characterized in that, Includes the following steps: Step S1: The step of acquiring video stream data. This step extracts video stream data of the power operation site from the video monitoring device at the operation site, and decodes and analyzes the extracted video stream data to obtain the image data information to be detected. Step S2: The image data segmentation step involves inputting the image data to be detected obtained from the decoding and analysis in step S1 into the trained isolation device segmentation model, and performing morphological processing on the segmented image output by the isolation device segmentation model to obtain the outer contour data of the isolation device. Step S3: Obtaining pedestrian bounding box location data. In this step, the image data to be detected obtained from the decoding and analysis in step S1 is input into the trained worker inspection model to check for workers and pedestrians in the image data. If a pedestrian is present, the pedestrian bounding box location data is output. x min , y min , x max , y max ]; Step S4: Obtaining pedestrian keypoint data, in which the pedestrian's bounding box position data from step S3 is used. x min , y min , x max , y max The image data to be detected in step S1 is cropped, and the cropped image data is input into the trained human keypoint detection model to obtain human keypoint data, which is represented as follows: p 0={( x 0, y 0),( x 1, y 1)……( x 24 , y 24 The human body location points are obtained by weighted averaging, and the positions of the human body location points in the image to be detected described in step S1 are calculated. Step S5: The step of predicting the direction of pedestrian movement. In this step, a behavioral movement direction prediction model is constructed using the human body position points in step S4 to make a rough prediction of the pedestrian movement trend. Based on the camera's perspective and the outer contour data of the isolation device in step S2, a rough warning angle range is set. If the pedestrian's movement direction is within the rough warning angle range, then step S6 is executed; otherwise, step S1 is returned. Step S6: Pedestrian trajectory prediction step. In this step, a Long Short-Term Memory (LSTM) network is used to establish a pedestrian trajectory prediction model to make detailed predictions of pedestrian positions within a preset time period in the future. Based on the predicted pedestrian trajectory, determine whether it intersects with the outer contour of the isolation device; If an intersection occurs, it indicates that pedestrians are about to enter the work area, and a warning should be issued to the pedestrians. Otherwise, return to step S1.

2. The method for pedestrian safety early warning outside a power operation area according to claim 1, characterized in that, In step S1: Video stream data is decoded using OpenCV tools.

3. The method for pedestrian safety early warning outside a power operation area according to claim 2, characterized in that, In step S2, the specific training process of the isolation device segmentation model is as follows: Preparing the training set: Photographs of the work site with isolation devices are taken. PixelAnnotationTools is used to annotate the isolation devices in the training set images to obtain image labels. The images and their labels are matched one-to-one as the training set for the isolation device segmentation model. The dataset is set up with a training set and a validation set in an 8:2 ratio. The training set is input into the DeepLabV3 algorithm and trained on an NVIDIA GPU server with default parameters. The model performance is evaluated using IOU as the evaluation criterion until the model meets the usage requirements, thus obtaining the isolation device segmentation model.

4. The method for pedestrian safety early warning outside a power operation area according to claim 3, characterized in that, In step S3, the trained worker detection model is trained based on the YOLOv5 object detection network. The specific training process of the worker detection model is as follows: Preparing the training set: The work site was photographed, and then the LabelImg annotation tool was used to annotate the workers and pedestrians in the training set images to obtain image labels. The images and the created labels were matched one-to-one as the training set for the worker detection model. The dataset was set up as a training set and a validation set in an 8:2 ratio. The training set was input into the YOLOv5 object detection algorithm and trained on an NVIDIA GPU server with the algorithm's default parameters. The model performance is evaluated using mAP as the evaluation standard until the model meets the usage requirements, thus obtaining the operator's inspection model.

5. A method for pedestrian safety early warning outside a power operation area according to claim 4, characterized in that, In step S4, the trained human keypoint detection model is trained based on the OpenPose keypoint detection network; the specific training process of the human keypoint detection model is as follows: Prepare the training set: Take pictures of the work site with isolation devices, and then use the Labelme annotation tool to annotate 25 key points of the human body in the training set images to obtain image labels. Match the captured images and the generated labels one by one as the dataset for human key point detection. Set the training set and validation set in a ratio of 8:2; The training set was input into the OpenPose keypoint detection algorithm and trained on an NVIDIA GPU server using the algorithm's default parameters. A human body key point detection model was obtained.

6. A method for pedestrian safety early warning outside a power operation area according to claim 5, characterized in that, Step S5 includes the following steps: Step S51: Define pedestrian movement direction and obtain the sequence of pedestrian position points P={p1,p2,…,p...} in n frames of images. n-1 ,p n }, where t n The direction of motion of the position point at time t is defined as t n-1 Time and t n The angle θ formed by the vector of points at time and position and the horizontal axis in a counterclockwise direction n The range of θ is 0°-360°; the parameter n is set according to the camera frame rate f and the actual engineering needs, and its range is 3f-5f. Step S52: Construct a direction prediction model, and form a sample set θ={θ1,θ2,…,θ2} of the motion directions of all positions in step S51. n-1 ,θ n A k-means algorithm is used to construct a direction prediction model; the clustering result C={C1,C2,…,C} is output after calculation by the k-means model. k }, where C k For the k-th cluster, the cluster with the largest number of samples is selected, and the angle of its cluster center is used as the prediction result of the pedestrian direction. The meaning of the prediction result is the most representative movement direction of the pedestrian in the n-frame image sequence. Step S53: Pedestrian movement trend judgment. During implementation, a rough warning angle range a needs to be set based on the camera angle and the outer contour of the isolation device in step S2. min -a max If the pedestrian's direction of movement in step S52 is within the range of the rough warning angle, then proceed to step S6; otherwise, return to step S1.

7. A method for pedestrian safety early warning outside a power operation area according to claim 6, characterized in that, Step S6 includes the following steps: Step S61: Construct a dataset by obtaining the human body location point sequence of n frames from step S4, where the location point sequence of each pedestrian is P={p1,p2,…,p n-1 ,p n } as training or testing samples for LSTM; Step S62: LSTM network training. The sequence in step S61 is preprocessed by difference and normalization. The network is trained using the Adam optimizer with a learning rate of 0.

001. Step S63: Pedestrian position prediction, let the current time be t. n Using the sequence of human body location points in n frames of images as input to the LSTM model, the prediction of t is performed. n+1 -t n+m The location of pedestrians within a given time period is set using parameters n and m. Step S64: Alarm judgment. Based on the pedestrian trajectory predicted in step S63, determine whether it intersects with the outer contour of the isolation device. If an intersection occurs, it indicates that the pedestrian is about to enter the work area and an alarm needs to be issued to the pedestrian. Otherwise, return to step S1.

Images