A power engineering construction safety early warning method and system

Abstract

The application belongs to the technical field of action recognition, and particularly relates to a power engineering construction safety early warning method and system, which comprises the following steps: carrying out threshold segmentation and Hough straight line detection on monitored images of a power engineering construction site to obtain a rigid support structure set, obtaining a scene instantaneous motion potential index according to an optical flow vector of visible key points, combining the distance between the visible key points and the rigid support structure in a possible contact set to construct a human body and rigid structure interaction stability index, calculating a posture risk bias factor, using a graph convolution network to complete the key points, correcting a preliminary complete posture estimation result according to the posture risk bias factor and a vector representing the influence of gravity, and determining the position of a human body center of mass key point according to the correction result, which is used to calculate the vertical acceleration of the human body center of mass key point and to judge the personnel falling risk. The application solves the problem of generating a "pseudo-true" posture under key point occlusion by using a graph convolution network model.

Description

Technical Field

[0001] This invention relates to the field of motion recognition technology. More specifically, this invention relates to a method and system for early warning of safety during power engineering construction. Background Technology

[0002] Power engineering construction, especially the erection of transmission towers and the maintenance of substations, is usually carried out in complex outdoor high-altitude environments. The site is filled with scaffolding, steel beams, towers and other metal structures, forming a complex scene that looks like a "steel jungle". Construction workers face an extremely high risk of falling when working in this environment.

[0003] Therefore, using on-site surveillance cameras and computer vision technology to identify the risk of personnel falling in real time and automatically is of great significance for ensuring life safety and preventing production accidents.

[0004] To address the frequent visual occlusion problem at construction sites, where workers' bodies are partially obscured by steel structures, existing technologies are beginning to employ human pose completion algorithms based on Graph Convolutional Networks (GCNs). This algorithm abstracts key points of the human skeleton as nodes in a graph, and skeleton connections as edges. By using GCNs to learn the inherent anatomical constraints of the human body, when some key points cannot be detected due to occlusion, the algorithm can infer and complete the occluded key points based on the positions of visible key points, thereby restoring the complete human pose.

[0005] However, when such technology is applied to the specific scenario of power engineering, it exposes a deep-seated defect: the risk of generating "pseudo-real" poses under the occlusion of key nodes. That is, when core nodes that are crucial to the judgment of human posture stability (such as the hip and the midpoint of the torso) are occluded by the steel structure for a long time, the graph convolutional network model relies solely on general, prior knowledge of human anatomy to complete the task. This can lead to the model generating a "pseudo-real" pose that is anatomically correct but incorrect in the current scenario.

[0006] Therefore, how to overcome the limitation of graph convolutional network models relying solely on anatomical priors when key nodes are occluded, and how to integrate the physical constraints and dynamic information of the scene into the pose completion process to avoid the generation of "pseudo-real" poses, is a problem that current technology urgently needs to solve. Summary of the Invention

[0007] To address the technical problem of generating "pseudo-real" poses caused by the limitation of the above-mentioned graph convolutional network model relying solely on anatomical priors when key nodes are occluded, this invention provides solutions in the following aspects.

[0008] In a first aspect, the present invention provides a method for early warning of safety during power engineering construction, comprising: extracting all visible key points for each frame of a monitoring video of a power engineering construction site using a human pose estimation model, and forming a set of visible key points; obtaining a scene instantaneous motion potential index by weighting and averaging the optical flow vectors of all visible key points between two adjacent frames using the motion weights of the visible key points; performing threshold segmentation and Hough line detection on the images to obtain a set of rigid support structures; and constructing a relationship between the human body and the rigid support structure based on the distance between each visible key point in the possible contact set and the pixels in the set of rigid support structures, as well as the scene instantaneous motion potential index. The system employs a rigid structure interaction stability index, where the potential contact set is a subset of the visible keypoint set. Based on the relationship between the human body and the rigid structure interaction stability index and the stability threshold, a posture risk bias factor is calculated. A graph convolutional network is used to complete the visible keypoint set, yielding a preliminary complete posture estimation result. This preliminary complete posture estimation result is then corrected based on the posture risk bias factor and a vector representing the influence of gravity. Based on the final complete posture estimation result, the location of the human body's center of mass keypoint is determined. Finally, based on the location of the human body's center of mass keypoint in two images, the vertical acceleration of the human body's center of mass keypoint is calculated for assessing the risk of a person falling.

[0009] This invention incorporates instantaneous motion potential indices for the scene and stability indices for the interaction between the human body and rigid structures, integrating the physical laws and dynamic information of the scene into posture analysis. This effectively solves the problem of traditional graph convolutional network models generating "pseudo-real" postures under occlusion of key nodes, greatly reducing the false negative rate of safety risks caused by model misjudgment. Simultaneously, by actively analyzing and quantifying the interaction state between the person and the environment's steel structure, the algorithm can still make reasonable inferences consistent with physical reality even when key information is missing, demonstrating stronger environmental adaptability and robustness. Finally, through continuous quantification of the stability indices for the interaction between the human body and rigid structures, this invention can calculate a higher posture risk bias factor in the "instability" stage before a clear fall occurs, such as poor contact between the person and the support or violent swaying of the body. This allows for earlier identification of potential fall risks, buying time for intervention measures.

[0010] Preferably, the step of extracting all visible keypoints using a human pose estimation model includes: a total of 25 complete keypoints, and each keypoint contains coordinates and a confidence score; therefore, the number of visible keypoints obtained is no more than 25, and each visible keypoint contains coordinates and a confidence score.

[0011] Preferably, the step of obtaining the instantaneous motion potential index of the scene by weighting the optical flow vectors of all visible keypoints between two adjacent frames using the motion weights of the visible keypoints includes: In the formula, This is an indicator of the instantaneous motion potential of the scene; This represents the number of all visible keypoints in the set of visible keypoints. For the first The visible keypoints are the optical flow vectors between the current frame image and the previous frame image; The L2 norm of a vector; For the first The motion weights of each visible keypoint.

[0012] This invention introduces temporal dynamic features, namely optical flow velocity, into the completion process, solving the problem that human pose completion algorithms based on graph convolutional networks rely solely on the anatomical structure of a single frame, and overcoming the limitations of static pose.

[0013] Preferably, the step of performing threshold segmentation and Hough line detection on the image to obtain a set of rigid support structures includes: performing threshold segmentation on the image using the Otsu method to obtain a binary image; including pixels belonging to rigid support structures in the foreground pixels of the binary image; performing Hough line detection on all foreground pixels in the binary image, and retaining only line segments whose length (i.e., the number of pixels constituting the line segment) is greater than a threshold; and for all retained line segments, using all pixels constituting these line segments as a set of rigid support structures.

[0014] Preferably, the step of constructing a stability index for human-rigid structure interaction based on the distance between each visible key point in the possible contact set and the pixel points in the rigid support structure set, and the instantaneous motion potential index of the scene, includes: In the formula, As an indicator of the stability of the interaction between the human body and rigid structures; The number of all visible keypoints in the potentially accessible set; For the first in the possible contact set One visible key point; A collection of rigidly supported structures; Indicates the first Visible key points The minimum Euclidean distance to the pixels in the set of rigid support structures; It is a natural exponential function; This is an indicator of the instantaneous motion potential of the scene.

[0015] This invention integrates motion characteristics and spatial location characteristics when constructing a stability index for the interaction between the human body and a rigid structure, achieving dual verification of physical support. This effectively reflects the real physical interaction state and provides a physical basis for subsequent risk correction.

[0016] Preferably, the step of calculating the posture risk bias factor based on the relationship between the human body and the rigid structure interaction stability index and the stability threshold includes: In the formula, This is the attitude risk bias factor; As an indicator of the stability of the interaction between the human body and rigid structures; This is a preset stability threshold; It is a natural exponential function; Let be the coefficient, and =20.

[0017] This invention constructs a pose risk bias factor by quantitatively analyzing the interaction between personnel and the surrounding environment, which is used to correct the pose completion results of graph convolutional networks, thereby effectively avoiding the generation of "false" poses and improving the accuracy of risk identification in complex occlusion environments.

[0018] Preferably, the step of correcting the preliminary complete attitude estimation result based on the attitude risk bias factor and the vector representing the influence of gravity includes: In the formula, The coordinates of the corrected key points; The coordinates of key points in the preliminary complete attitude estimation results; This is the attitude risk bias factor; The vector representing the influence of gravity is vertically downwards and its magnitude is the scene mapping value of gravitational acceleration, which is equal to the product of gravitational acceleration and image resolution. It is equal to the time interval between the acquisition of two adjacent frames.

[0019] This invention corrects the initial complete pose estimation results, which forces the output of the graph convolutional network to be more consistent with the physical scene, thereby eliminating the possibility of filling in "falling" as "squatting" or other "pseudo-real" poses.

[0020] Preferably, determining the position of the human body's center of mass keypoint based on the obtained final complete attitude estimation result includes: normalizing the confidence scores of all keypoints in the obtained final complete attitude estimation result, using the normalized confidence scores as the position weights of the keypoints; and obtaining the position of the human body's center of mass keypoint by weighted summation of the coordinates of all keypoints using the position weights.

[0021] Preferably, the step of calculating the vertical acceleration of the human body's center of gravity key point based on the position of the key point in the two frames of images for assessing the risk of a person's fall includes: calculating the difference in the ordinate of the final human body's center of gravity key point in the two frames of images, and calculating the difference in the ordinate and the acquisition time interval between the two adjacent frames of images. The quotient of the vertical velocity of the human body's center of mass keypoint is used as the vertical velocity. This vertical velocity is then converted into its position in scene space using image resolution. The vertical velocity of the human body's center of mass keypoint is calculated and... The quotient of the value is used as the vertical acceleration of the key point of the human body's center of gravity; when the vertical acceleration of the key point of the human body's center of gravity is greater than or equal to the value of the vertical acceleration of the key point of the human body's center of gravity, the vertical acceleration of the key point of the human body's center of gravity is greater than or equal to the vertical acceleration of the key point of the human body's center of gravity. When the system determines that a person is at risk of falling, it immediately triggers an alarm mechanism. This is the acceleration due to gravity.

[0022] Secondly, the present invention provides a power engineering construction safety early warning system, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned power engineering construction safety early warning method is implemented.

[0023] By adopting the above technical solution, a computer program for the above-mentioned power engineering construction safety early warning method is generated and stored in a memory so that it can be loaded and executed by a processor. A terminal device can then be made based on the memory and processor for convenient use.

[0024] The beneficial effects of this invention are as follows:

[0025] This invention incorporates instantaneous motion potential indices for the scene and stability indices for the interaction between the human body and rigid structures, integrating the physical laws and dynamic information of the scene into posture analysis. This effectively solves the problem of traditional graph convolutional network models generating "pseudo-real" postures under occlusion of key nodes, greatly reducing the false negative rate of safety risks caused by model misjudgment. Simultaneously, by actively analyzing and quantifying the interaction state between the person and the environment's steel structure, the algorithm can still make reasonable inferences consistent with physical reality even when key information is missing, demonstrating stronger environmental adaptability and robustness. Finally, through continuous quantification of the stability indices for the interaction between the human body and rigid structures, this invention can calculate a higher posture risk bias factor in the "instability" stage before a clear fall occurs, such as poor contact between the person and the support or violent swaying of the body. This allows for earlier identification of potential fall risks, buying time for intervention measures. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating a power engineering construction safety early warning method according to the present invention;

[0027] Figure 2 This is a flowchart illustrating step S2. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0030] This invention discloses a method for early warning of construction safety in power engineering, referring to... Figure 1 This includes steps S1-S4:

[0031] S1. For each frame of the monitoring video of the power engineering construction site, a human pose estimation model is used to extract all visible key points.

[0032] First, monitoring videos are acquired through monitoring equipment deployed at the power engineering construction site. These videos consist of continuous image frames.

[0033] Then, for each frame of the image, the OpenPose human pose estimation model was used to perform initial pose detection on the construction workers in the image, extracting all key points. Each key point includes coordinates and a confidence score. The complete set of 25 key points includes: nose, neck, right shoulder, right elbow, right wrist, left shoulder, left elbow, left wrist, right hip, right knee, right ankle, left hip, left knee, left ankle, right eye, left eye, right ear, left ear, left big toe, left little toe, left heel, right big toe, right little toe, right heel, and background points. Among them, 24 are actually valid human key points. Therefore, the back key points are excluded. For the scenic spot, the remaining 24 key points can be divided into 5 human body parts: head, upper limbs, lower limbs, feet, and torso. The head includes 5 key points: nose, right eye, left eye, right ear, and left ear; the upper limbs include 6 key points: right shoulder, right elbow, right wrist, left shoulder, left elbow, and left wrist; the lower limbs include 6 key points: right hip, right knee, right ankle, left hip, left knee, and left ankle; the feet include 6 key points: left big toe, left little toe, left heel, right big toe, right little toe, and right heel; and the torso includes 1 key point: the neck.

[0034] Furthermore, in the absence of occlusion, all 24 key points can be extracted. However, when occlusion exists, all 24 key points cannot be extracted. Therefore, all extracted key points are treated as visible key points and formed into a set of visible key points. The number of visible key points in the obtained set is no more than 24.

[0035] S2. Weight the optical flow vectors of all visible keypoints between two adjacent frames to obtain the instantaneous motion potential index of the scene; perform threshold segmentation and Hough line detection on the image to obtain the set of rigid support structures; construct the human body and rigid structure interaction stability index based on the distance between each visible keypoint in the possible contact set and the pixel in the rigid support structure set, as well as the instantaneous motion potential index of the scene, to calculate the posture risk bias factor.

[0036] The flowchart for step S2 is shown below. Figure 2 The process includes steps S201 to S204, specifically as follows:

[0037] S201. By using the motion weights of visible keypoints, the magnitudes of the optical flow vectors of all visible keypoints in two adjacent frames are weighted and averaged to obtain the instantaneous motion potential index of the scene.

[0038] It should be noted that optical flow is a technique used to estimate the motion of pixels in a video. It can capture the displacement of pixels between consecutive frames and represent it in the form of an optical flow field. The calculation of the optical flow field can be implemented by various algorithms, mainly divided into sparse optical flow and dense optical flow. Among them, sparse optical flow mainly focuses on key points in the image, including corner points, key points or significant feature points. Sparse optical flow includes the LucasKanade algorithm.

[0039] Specifically, the LucasKanade algorithm is used to process two adjacent frames of images to obtain the optical flow field between the two adjacent frames. The optical flow field includes the optical flow vector of the visible key points between the two adjacent frames. The optical flow vector can represent the movement of the pixel between the two adjacent frames.

[0040] Furthermore, in order to capture dynamic features that foreshadow instability or falls, rather than relying solely on static posture information, an instantaneous motion potential index is constructed using the motion weights and optical flow vectors of each visible key point. This index is used to measure the intensity of motion of visible body parts. High motion potential energy suggests possible sudden changes in state, such as slipping or tripping.

[0041] The formula for calculating the instantaneous motion potential index of the scene is:

[0042] ;

[0043] In the formula, This is an indicator of the instantaneous motion potential of the scene; This represents the number of all visible keypoints in the set of visible keypoints. For the first The optical flow vector of a visible keypoint between the current frame and the previous frame represents the instantaneous velocity and direction of the keypoint. Denotes the L2 norm of a vector, therefore, Characterizes the magnitude of the optical flow vector; For the first The motion weights of each visible keypoint.

[0044] Among them, the lower limbs, upper limbs and feet are the three extremities of the human body. Their violent movements can better reflect the overall instability. Therefore, the motion weight of the visible key points belonging to the lower limbs and feet is 0.3, the motion weight of the visible key points belonging to the upper limbs is 0.2, the motion weight of the visible key points belonging to the trunk is 0.15, and the motion weight of the visible key points belonging to the head is 0.05.

[0045] The dynamic activity level of a person is quantified by weighted averaging the movement speed of all visible keypoints: when a person is in a non-steady state, about to fall, their visible extremities tend to exhibit faster movement speeds, thus increasing the instantaneous motion potential index of the scene. The value increases significantly, for example, when a worker misses a step, the visible optical flow velocity of the hand / foot increases. A dramatic increase Significantly elevated; while during normal squatting, limb movement is smooth. Keep it low; and, through By assigning higher motion weights to the extremities, the characteristic signals of a fall can be amplified, especially when the hip is frequently obscured. While the severity of falls may decrease, extremity movement remains the primary indicator of a fall, and motion information from remaining visible key points can still drive the movement. Changes should be made to avoid the loss of dynamic information due to the absence of key nodes.

[0046] It should be noted that this invention introduces temporal dynamic features, namely optical flow velocity, into the completion process, which solves the problem that human pose completion algorithms based on graph convolutional networks rely solely on the anatomical structure of a single frame, thus overcoming the limitations of static pose.

[0047] S202. Perform threshold segmentation and Hough line detection on the image to obtain a set of rigid support structures.

[0048] It should be noted that in the power engineering construction safety early warning system, accurately identifying rigid support structures such as steel structures, scaffolding, and platforms based on the material properties and geometric structural rules of metals is a prerequisite for subsequent calculation of the stability index of human interaction with rigid structures.

[0049] Specifically, the image is segmented using the Otsu method to obtain a binary image. In industrial scenarios, rigid support structures such as steel structures, scaffolding, and platforms are composed of metal materials, and the metal surface has high reflectivity, easily producing strong reflections and specular reflections. Therefore, the foreground pixels in the binary image include pixels belonging to rigid support structures, and these rigid support structures are composed of a large number of parallel or perpendicular straight lines. Therefore, Hough line detection is performed on all foreground pixels in the binary image, retaining only line segments whose length (i.e., the number of pixels constituting the line segment) is greater than a threshold. For all retained line segments, all pixels constituting these line segments are used as the set of rigid support structures.

[0050] Among them, the Otsu method and Hough line detection are well-known techniques and will not be elaborated here.

[0051] The specific value of the quantity threshold can be set according to the actual application scenario and needs, and the value range of the quantity threshold is [10, 35]. In this invention, the quantity threshold is set to 15.

[0052] S203. Based on the distance between each visible key point in the possible contact set and the pixel points in the rigid support structure set, as well as the instantaneous motion potential index of the scene, construct a stability index for the interaction between the human body and the rigid structure.

[0053] It should be noted that, in order to illustrate whether a person is in a stable state supported by a rigid support structure in the environment, a stability index for the interaction between the human body and the rigid structure was further constructed by combining motion information, namely the instantaneous motion potential index of the scene, and the spatial relationship between the person and the rigid support structure in the environment.

[0054] Specifically, based on the distance between each visible key point in the possible contact set and the pixel points in the rigid support structure set, as well as the instantaneous motion potential index of the scene, an interaction stability index between the human body and the rigid structure is constructed, where the possible contact set is a subset of the visible key point set.

[0055] The formula for calculating the stability index of the interaction between the human body and a rigid structure is:

[0056] ;

[0057] In the formula, As an indicator of the stability of the interaction between the human body and rigid structures; The number of all visible keypoints in the potentially accessible set; For the first in the possible contact set One visible key point; A collection of rigidly supported structures; Indicates the first Visible key points The minimum Euclidean distance to the pixels in the set of rigid support structures; It is a natural exponential function; This is an indicator of the instantaneous motion potential of the scene.

[0058] Furthermore, the set of all visible key points that may come into support contact with the rigid structure in the power construction scenario is considered as the potential contact set. Therefore, the potential contact set is a subset of the visible key point set. The key points that may come into support contact with the rigid structure in the power construction scenario mainly include key points such as feet, ankles, hands, and wrists, specifically: left big toe, left little toe, left heel, right big toe, right little toe, right heel, left ankle, right ankle, left knee, right knee, left elbow, left wrist, right elbow, right wrist, left hip, and right hip.

[0059] In industrial settings, personnel stability stems from "effective contact"; effective contact must simultaneously meet two conditions: (1) sufficient spatial proximity, i.e. (2) It is relatively small; Smaller; if a key foot point is close to the steel beam, but the overall momentum is smaller. The interaction stability index is very high, for example, when a person is sliding down rapidly. It will also decrease accordingly; therefore, through Ensure that critical points are in close contact with the steel structure only. It is an exponentially decaying function, indicating that the closer the key point is to the supporting structure, the better. When smaller, The closer the value is to 1, the greater the stability index; by... As the denominator, dynamic decay stability: when Higher, for example, when the body sways, even The value is relatively small, and the stability index is still close to 0; this addresses the problem of contact illusion, where a worker steps on the edge with one foot, Smaller but instantaneous motion potential index of the scene High, If the value approaches 0, it is correctly identified as high risk.

[0060] It should be noted that when constructing the stability index of human interaction with rigid structures, the integration of motion characteristics and spatial location characteristics provides dual verification physical support, which can effectively reflect the real physical interaction state and provide a physical basis for subsequent risk correction.

[0061] S204. Calculate the attitude risk bias factor based on the relationship between the human body and the stability index of the rigid structure and the stability threshold.

[0062] The obtained stability index of human interaction with rigid structure is quantified to obtain a factor that can directly intervene in the graph convolutional network completion process. Specifically, the posture risk bias factor is calculated based on the relationship between the stability index of human interaction with rigid structure and the stability threshold.

[0063] The formula for calculating the attitude risk bias factor is:

[0064] ;

[0065] In the formula, This is the attitude risk bias factor; As an indicator of the stability of the interaction between the human body and rigid structures; The preset stability threshold represents the baseline of the stable state during normal operation. It is a natural exponential function; The coefficient is used to control the steepness of the sigmoid function, determining the sensitivity of the risk factor to increasing as stability decreases. =20.

[0066] In power construction sites, delayed early warnings can have fatal consequences. Therefore, safety warnings at power construction sites must adhere to the principle of "better to give a false alarm than a true one." Thus, the Sigmoid function is used to measure continuous stability indicators. Mapped to a non-linear risk probability: when the stability index Much higher than the stability threshold At that time, attitude risk bias factor A value approaching 0 indicates no risk; when the stability index... Decrease to stability threshold When near, attitude risk bias factor It will grow rapidly and approach 1, indicating high risk; the obtained attitude risk bias factor takes the value in the range of (0,1), and the closer the value is to 1, the higher the risk.

[0067] It should be noted that by continuously quantifying the stability index of the interaction between the human body and rigid structures, a higher posture risk bias factor can be calculated in the "instability" stage before a clear fall occurs, such as poor contact between the person and the support or violent shaking of the body. This allows for earlier identification of potential fall risks and provides more time for intervention.

[0068] S3. Use a graph convolutional network to complete the set of visible key points to obtain a preliminary complete attitude estimation result; based on the attitude risk bias factor and the vector representing the influence of gravity, correct the preliminary complete attitude estimation result to obtain the final complete attitude estimation result.

[0069] First, the incomplete set of visible key points is completed using a graph convolutional network to obtain a preliminary complete pose estimation result, including: (1) inputting the set of visible key points and the confidence score of each visible key point into the graph convolutional network; (2) constructing a graph structure based on adjacency relationships defined in human anatomy; (3) propagating information from visible key points to occluded key points through multi-layer graph convolution to obtain and output a preliminary complete pose estimation result, which includes 24 complete key points.

[0070] Then, based on the attitude risk bias factor and the vector representing the influence of gravity, the preliminary complete attitude estimation result is corrected to obtain the final complete attitude estimation result. The correction process simulates the instability effect in physics.

[0071] Specifically, for any keypoint in the preliminary complete attitude estimation result, the corrected coordinates of the keypoint are:

[0072] ;

[0073] In the formula, The coordinates of the corrected key points; The coordinates of key points in the preliminary complete attitude estimation results; This is the attitude risk bias factor; The vector representing the influence of gravity is vertically downward and its magnitude is the scene mapping value of gravitational acceleration, which is the product of gravitational acceleration and image resolution. Through image resolution, gravitational acceleration is converted into acceleration in image space. It equals the time interval between the acquisition of two adjacent image frames, in seconds.

[0074] Among them, when the attitude risk bias factor When the object is raised, a displacement bias in the same direction as gravity is applied to the keypoints of the completed graph convolutional network. This bias simulates the stability index of the interaction between the human body and a rigid structure when a person loses support. Low value and attitude risk bias factor When the value is high, the body will inevitably experience a downward displacement under the influence of gravity, rather than remaining in its original position. For example, when a convolutional network miscompletes a downward fall as a squat, the position of the completed keypoints is too high, resulting in a higher posture risk bias factor. This will drive the corrected key points to move down to a reasonable position, so that the attitude conforms to the gravity constraint.

[0075] It should be noted that by correcting the output of the graph convolutional network, it is possible to force the output of the graph convolutional network to be more consistent with the physical scene, thereby eliminating the possibility of filling in "falling" as "squatting" or other "pseudo-realistic" postures.

[0076] S4. Based on the final complete attitude estimation result, obtain the position of the human body's center of gravity key point; based on the position of the human body's center of gravity key point in the two frames of images, calculate the vertical acceleration of the human body's center of gravity key point, which is used to judge the risk of personnel falling.

[0077] Specifically, for the key points in the final complete attitude estimation result, the confidence scores of all key points are normalized, and the normalized confidence scores are used as the position weights of the key points. The coordinates of all key points are weighted and summed using the position weights to obtain the position of the human body's center of mass key point.

[0078] Furthermore, for two adjacent image frames, based on the positions of the human body's center of mass keypoints in the two images, the vertical acceleration of the human body's center of mass keypoints is calculated, including: calculating the difference in the ordinates of the final human body's center of mass keypoints in the two images, and calculating the difference in ordinates relative to the acquisition time interval between the two adjacent image frames. The quotient of the vertical velocity of the human body's center of mass keypoint is used as the vertical velocity. This vertical velocity is then converted into its position in scene space using image resolution. The vertical velocity of the human body's center of mass keypoint is calculated and... The quotient is used as the vertical acceleration of the key point of the human body's center of mass.

[0079] Furthermore, the vertical acceleration at the key point of the human body's center of mass is closer to the acceleration due to gravity. The greater the probability of a person falling, the more likely they are to fall. Therefore, when the vertical acceleration at the critical point of the human body's center of gravity is greater than or equal to... When this occurs, the system determines that there is a risk of a person falling and immediately triggers an alarm mechanism.

[0080] It should be noted that this invention incorporates the physical laws and dynamic information of the scene into posture analysis by introducing instantaneous motion potential indexes of the scene and stability indexes of human interaction with rigid structures. This effectively solves the problem of generating "pseudo-real" postures under the occlusion of key nodes in traditional graph convolutional network models, greatly reducing the false negative rate of safety risks caused by model misjudgment. At the same time, by actively analyzing and quantifying the interaction state between humans and the environment steel structure, the algorithm can still make reasonable inferences that conform to physical reality even when key information is missing, demonstrating stronger environmental adaptability and robustness.

[0081] This invention also discloses a power engineering construction safety early warning system, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, a power engineering construction safety early warning method according to the present invention is implemented.

[0082] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

Claims

1. A power engineering construction safety early warning method, characterized in that, The method comprises the following steps: For each frame of image in the monitoring video of the power engineering construction site, a human body posture estimation model is used to extract all visible key points and form a visible key point set; The size of the optical flow vector between all visible key points in adjacent two frames of images is weighted and averaged by the motion weight of the visible key points to obtain a scene instantaneous motion potential index; Threshold segmentation and Hough line detection are performed on the image to obtain a rigid support structure set; According to the distance between each visible key point in the possible contact set and the pixel point in the rigid support structure set, and the scene instantaneous motion potential index, a human body and rigid structure interaction stability index is constructed, which satisfies: ； stability index for human interaction with rigid structures; number of visible key points in the possible contact set; visible key point in the possible contact set; visible key point in the possible contact set; set of rigid support structures; minimum Euclidean distance of the visible key point visible key point to the pixel points in the set of rigid support structures; natural exponential function; scene instantaneous motion potential index; wherein the possible contact set is a subset of the visible key point set; According to the size relationship between the human body and rigid structure interaction stability index and the stability threshold, a posture risk bias factor is calculated, which satisfies: ； is a pose risk bias factor; is a preset stability threshold; is a coefficient, and = 20; the set of visible key points is completed by using a graph convolution network to obtain a preliminary complete pose estimation result; and the preliminary complete pose estimation result is corrected according to the pose risk bias factor and the vector representing the influence of gravity, so as to satisfy: ； coordinates of the corrected key points; coordinates of the key points in the preliminary complete pose estimation result; a vector representing the influence of gravity, the direction of which is vertically downward, and the size of which is a scene mapping value of gravitational acceleration, the scene mapping value of gravitational acceleration being equal to the product of the gravitational acceleration and the image resolution; is equal to the time interval between the acquisition of two adjacent frames of images; According to the obtained final complete posture estimation result, the position of the human body center key point is determined; according to the position of the human body center key point in the two frames of images, the vertical acceleration of the human body center key point is calculated, which is used for personnel falling risk judgment.

2. The power engineering construction safety early warning method according to claim 1, characterized in that, The method comprises the following steps: The complete key point set contains 24 key points, and each key point contains coordinates and confidence score; therefore, the number of visible key points obtained is not more than 24, and each visible key point contains coordinates and confidence score.

3. The power engineering construction safety early warning method according to claim 1, characterized in that, The method comprises the following steps: ； wherein, is a scene instantaneous motion potential indicator; is the number of all visible key points in the visible key point set; is the optical flow vector of the th visible key point between the current frame image and the previous frame image; denotes the L2 norm of a vector; is the motion weight of the th visible key point.

4. The power engineering construction safety early warning method according to claim 1, characterized in that, The method comprises the following steps: The image is threshold segmented by Otsu method to obtain a binary image; the foreground pixel points in the binary image include the pixel points belonging to the rigid support structure; Hough line detection is performed on all foreground pixel points in the binary image, and only the line segments with the number of pixel points constituting the line segment greater than the number threshold are reserved; for all the reserved line segments, all the pixel points constituting these line segments are taken as the rigid support structure set.

5. The power engineering construction safety early warning method according to claim 2, characterized in that, The method comprises the following steps: For the key points in the obtained final complete posture estimation result, the confidence scores of all key points are normalized, and the normalized result of the confidence score is taken as the position weight of the key point; The coordinates of all key points are weighted and summed by the position weight to obtain the position of the human body center key point.

6. The power engineering construction safety early warning method according to claim 1, characterized in that, The method comprises the following steps: The difference of the vertical coordinates of the final human body center key points in two frames of images is calculated, and the difference of the vertical coordinates and the time interval of adjacent two frames of images are calculated The quotient value of the time interval is taken as the vertical velocity of the human body center key point, and the vertical velocity is converted into the vertical velocity in the scene space through the image resolution; the quotient value of the vertical velocity of the human body center key point and The quotient value is taken as the vertical acceleration of the human body center key point; When the vertical acceleration of the human body's center of mass key point is greater than or equal to , the system determines that there is a risk of falling for the personnel, and immediately triggers an alarm mechanism, wherein, is the gravitational acceleration.

7. A power engineering construction safety early warning system, characterized in that, The method comprises the following steps: A processor and a memory, the memory stores computer program instructions, when the computer program instructions are executed by the processor, a power engineering construction safety warning method according to any one of claims 1-6 is realized.

Images