A human posture detection method and system in a complex power scenario

Abstract

A human posture detection method and system in a complex power scene. The ROI of the power scene image is extracted, and the preliminary features, human detection frame and key point information are obtained through the YOLOv8-Pose model; the human normalized position field, equipment structure field and skeleton topology field are constructed based on the position offset, power equipment structure response and human skeleton topology, and the preliminary features are refined; further, the human evidence energy field, equipment counter-evidence energy field and time sequence consistent energy field are constructed to generate enhanced features; the node feature matrix is constructed based on the skeleton node features and confidence mask, the global posture features are obtained through graph convolution and Transformer modeling and aggregation, the action category probability is output through the classification head, and the accurate identification of human posture in a complex power scene is realized.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of computer vision and power safety technology, specifically relating to a method and system for human posture detection in complex power scenarios. Background Technology

[0002] Intelligent technologies such as computer vision can be used to detect the posture of workers in real time in power operation scenarios, identify abnormal behaviors such as falling from heights, climbing illegally, losing balance, and approaching live parts, and achieve early warning of safety risks.

[0003] The existing technologies mainly adopt the following two types of schemes: (1) the method based on target detection and human key point detection, such as using target detection models such as YOLO to locate the human body, and then combining the pose estimation algorithm to extract the human body key points, and judging whether there is an abnormal pose by the positional relationship of the key points. (2) the method based on human skeleton temporal modeling, such as using graph convolutional networks (GCN) to analyze the spatial topological relationship between human key points and the continuous action changes, thereby realizing dynamic behavior recognition.

[0004] However, in complex power operation scenarios, the above mainstream solutions still have the following shortcomings: (1) Outdoor operation environment has shading factors, and in the case of long-distance shooting, some key joints are prone to missed detection or false detection; (2) Most of them focus on key point positioning in a single image, lacking analysis of the deep spatial relationship between the joints of the human body; (3) Most of them are aimed at detecting single violations, or only focus on equipment status monitoring, lacking comprehensive analysis of the relationship between personnel posture, action process and safety risks.

[0005] The aforementioned problems result in low accuracy of human posture detection, insufficient dynamic behavior recognition capabilities, and a lack of safety closed-loop management in complex power operation environments. There is a need for an intelligent posture detection method that can integrate the spatial relationships of key human points, temporal dynamic features, and global semantic information to improve the ability to identify abnormal operation behaviors in complex scenarios and achieve proactive early warning of power operation safety risks. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method and system for human posture detection in complex power scenarios. By constructing a normalized human position field, an equipment structure field, and a human skeleton topological field, and further integrating human evidence, equipment counter-evidence, and time-series consistency energy fields, multi-field constraint enhancement is applied to the initial human features. Based on the enhanced features, skeleton node features with confidence masks are constructed, and global posture features are obtained through joint modeling using graph convolution and Transformer, achieving accurate human posture recognition in complex power scenarios.

[0007] The first aspect of this application discloses a method for human posture detection in complex power scenarios, which adopts the following technical solution: Images of power field operation scenarios are collected and preprocessed. After extracting ROI images, they are input into the YOLOv8-Pose model to initially extract feature maps, human detection boxes, key point coordinates, and confidence scores. A normalized human body position field is constructed based on the positional offset between the human body detection box and the preliminary feature map; a power equipment structure field is constructed by extracting multiple structural responses from the ROI image; a skeleton map is established based on the anatomical connection relationship of human body key points and the position inversion of suspected occluded key points is performed to form a human body skeleton topological field; the preliminary feature map is processed based on the three field output responses to obtain refined features. The strength of human evidence is calculated based on the effective region of the detection box, the key point occlusion inversion response, and the human skeleton connection region response in the refined features, and normalized to obtain the human evidence energy field; the strength of device counter-evidence is calculated based on the multi-type structural responses in the refined features, and normalized to obtain the device counter-evidence energy field; the temporal consistency response is calculated using the key point positions of adjacent frames of the refined features, and normalized to obtain the temporal consistency energy field; the enhanced features are obtained by processing the refined features based on the three field responses. Local node features are obtained from the skeleton graph animation enhancement features. Confidence masks are set for the nodes and a node feature matrix is ​​constructed. After graph convolution calculation and Transformer modeling, the node features are output. After aggregation through the confidence mask, the global pose features are output. A classification head is used to map the global pose features to the action category probability distribution to identify the human pose type.

[0008] Furthermore, the process of constructing the normalized human body position field includes: Calculate any position in the preliminary feature map Normalized coordinates relative to the center of the human detection box; The normalized coordinates are concatenated with the initial feature map, and after being mapped by the network, the normalized human body position field coordinates in the feature map space are output. The output response value at that location.

[0009] Furthermore, the construction process of the power equipment structure field includes: Extract multiple structural responses from ROI images, including linear convolution response, Sobel edge gradient magnitude response, and HSV brightness channel highlight reflection response; By fusing multiple structural responses, the coordinates of the power equipment structural field in the feature map space can be obtained. The output response value at that location.

[0010] Furthermore, the construction process of the human skeleton topological field includes: The key points are constructed into a human skeleton diagram according to their anatomical adjacency; key points with a confidence level lower than a set threshold are identified as suspected occluded key points; For suspected occluded keypoints, define a set of adjacent visible keypoints; Based on the positions of all adjacent visible keypoints in the set, the inversion positions of suspected occluded keypoints are calculated by combining the standard bone vector and the current human bounding box width and height; the standard bone vector is calculated using historical human pose data. For the inversion location, the credibility is calculated based on the scores of reasonable human body parts, structural consistency, and equipment counter-evidence, and a key point occlusion inversion response map is generated. For each valid skeleton edge, a Gaussian response is generated and the corresponding maximum value is taken to form the topological field of the human skeleton.

[0011] Furthermore, the score for the reasonable part of the human body is obtained by taking the normalized position field response of the human body at the inverted position; The structural consistency score is calculated as the average Gaussian similarity of the distances between adjacent visible keypoints and their inversion locations. The equipment counter-proof score is obtained by taking the structural field response of the power equipment at the inversion location.

[0012] Furthermore, the process of obtaining the refined features includes: Definition 1 minus the structural field of the power equipment is the repulsion factor; The normalized human body position field, human skeleton topological field, key point occlusion inversion response map, and repulsion factor are concatenated, and refined weights are output through convolutional mapping and activation function. The initial feature map is weighted with refined weights to output refined features.

[0013] Furthermore, the process of constructing the human evidence energy field includes: For any position in the refining feature The strength of human evidence is obtained by weighting the corresponding effective area indicator of the detection box, the maximum value of the key point occlusion inversion response map, and the response of the human skeleton connection area. when The set of pixels located within the boundary of the detection box has an indicator value of 1, otherwise it has a value of 0; The response value of the human skeleton connection area is The topological field response of the human skeleton.

[0014] Furthermore, the process of constructing the energy field by the device includes: For any position in the refining feature The corresponding power equipment structural field response, linear convolution response, HSV brightness channel high brightness reflective response and skeleton connection penalty term are weighted and normalized to obtain the equipment's reverse energy field. The skeleton connection penalty term is defined as 1 minus the response of the human skeleton connection area.

[0015] Furthermore, the construction process of the time-consistent energy field is as follows: The consistency of key points is calculated based on the displacement deviation of key points in two adjacent frames in the refined features; the occlusion inversion response map is weighted using the key point consistency to obtain the temporal consistency response; The timing consistency response is calculated point by point along the key points, and the maximum value is taken for normalization to obtain the timing consistency energy field.

[0016] Furthermore, the process of obtaining the enhanced features includes: The competitive enhancement weights of the human evidence energy field, the device counter-evidence energy field, and the temporal consistency energy field are calculated through a competitive normalization mechanism. The refined features are weighted by competition enhancement weights to generate enhanced features.

[0017] The second aspect of this application discloses a human posture detection system for complex power scenarios, which implements the human posture detection method described in the first aspect of this application. The system includes: Basic data acquisition module: Collects and preprocesses images of power field operation scenarios, extracts ROI images, and inputs them into the YOLOv8-Pose model to initially extract feature maps, human detection boxes, key point coordinates, and confidence scores; Key point feature refinement module: Constructs a normalized human body position field based on the positional offset between the human body detection box and the preliminary feature map; extracts multiple structural responses from the ROI image to construct a power equipment structural field; establishes a skeleton map based on the anatomical connection relationship of human body key points and performs position inversion on suspected occluded key points to form a human skeleton topological field; processes the preliminary feature map based on the three-field output response to obtain refined features. The key point feature enhancement module calculates the strength of human evidence based on the effective region of the detection box, the key point occlusion inversion response, and the human skeleton connection region response in the refined features, and normalizes it to obtain the human evidence energy field; it calculates the strength of device counter-evidence based on the multi-class structural responses in the refined features, and normalizes it to obtain the device counter-evidence energy field; it calculates the temporal consistency response using the key point positions of adjacent frames of the refined features, and normalizes it to obtain the temporal consistency energy field; and it processes the refined features based on the three-field responses to obtain the enhanced features. The human pose recognition module obtains local node features from the skeleton graph motion enhancement features, sets confidence masks for the nodes and constructs a node feature matrix; outputs node features after graph convolution calculation and Transformer modeling, and outputs global pose features after aggregation through confidence masks; and uses a classification head to map the global pose features to the action category probability distribution to identify the human pose type.

[0018] The beneficial effects of this invention are that, compared with the prior art, 1. General coordinate attention mechanisms primarily rely on spatial encoding in the horizontal and vertical directions, lacking explicit modeling of occlusion sources, equipment structures, and human skeleton topology. However, this invention, through three-domain coupling of the human body region, power equipment region, and skeleton topology region, enables low-confidence keypoints to undergo occlusion inversion based on adjacent visible keypoints, and suppresses pseudo-limb responses such as conductors, towers, and insulators through equipment counter-evidence terms. Therefore, this module can simultaneously capture the combined effects from human body spatial location, equipment interference structures, and skeleton connection relationships, improving the completeness and stability of keypoint detection in complex power operation scenarios.

[0019] 2. Traditional parameter-free spatial energy enhancement methods primarily calculate significance based on the relationship between neurons within the same channel and local means and variances, which can highlight regions with strong local responses to a certain extent. However, in power operation scenarios, regions with strong local responses do not necessarily belong to the human body area; they may also originate from conductors, tower edges, insulator outlines, safety rope reflections, or tool bag edges. Therefore, this invention no longer judges feature importance solely based on a single spatial energy value, but generates enhancement weights through the competitive relationship between human evidence, equipment counter-evidence, and temporal consistency. This module does not require additional trainable attention parameters and can effectively enhance realistic human posture regions while suppressing equipment structural interference regions and transient noise regions, while maintaining a lightweight design. Attached Figure Description

[0020] Figure 1 This is an algorithm framework diagram of a human pose detection method based on the improved YOLOv8-Pose. Figure 2 This is a diagram showing the algorithmic logic relationship between key point inversion and competition enhancement. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0022] The overall technical architecture of this invention is an end-to-end human posture detection system, including a preprocessing module, an improved YOLOv8-Pose posture backbone detection module, a human-equipment-skeleton three-domain coupled key point occlusion inversion feature refinement module, a human evidence-equipment counter-evidence-temporal consistency ternary competitive energy field parameter-free enhancement module, a graph convolution feature fusion module, a Transformer global feature enhancement module, and a posture recognition module. Each module is linked in sequence to realize the detection of abnormal postures from the acquisition of images at the power operation site. The specific innovations mainly include (1) a human-equipment-skeleton three-domain coupled key point occlusion inversion feature refinement mechanism, (2) a human evidence-equipment counter-evidence-temporal consistency ternary competitive energy field parameter-free enhancement strategy, and (3) a graph convolution fusion Transformer decision posture.

[0023] Example 1 As an embodiment of this application, refer to Figure 1 This paper presents a specific implementation method for human posture detection in complex power scenarios.

[0024] S1. In one embodiment of Example 1, images / video streams of the work scene are collected by power field monitoring equipment, and noise reduction, size normalization, and illumination equalization are performed in sequence. The key ROI area of ​​the work is extracted, irrelevant background is filtered out, and the amount of computation is reduced.

[0025] 1.1 Image / video stream acquisition; One or more high-definition surveillance cameras (such as CCD industrial cameras or network cameras) deployed at the power operation site continuously capture raw images or video streams of the operation scene at a fixed frame rate (e.g., 25 frames / second). The captured images are in RGB color format with a resolution of no less than 1920×1080 pixels.

[0026] 1.2 Image denoising processing; The acquired raw images undergo denoising processing to eliminate image noise introduced by factors such as equipment vibration, light fluctuations, and sensor thermal noise in the power operation site environment. Specifically, one or a combination of the following methods is used: (a) Median filtering: Using a 3×3 or 5×5 sliding window, the gray values ​​of each pixel in the image are sorted, and the median value is used to replace the current pixel value, effectively removing salt and pepper noise; (b) Gaussian filtering: A Gaussian kernel (σ=1.5, kernel size 5×5) is used to perform convolution operation on the image to smooth the image and suppress Gaussian noise; (c) Bilateral filtering: It preserves image edge information while denoising, and is suitable for preserving the edge contours of power equipment.

[0027] Preferably, median filtering is used when there is obvious salt-and-pepper noise in the image; Gaussian filtering is used when there is Gaussian noise; and bilateral filtering is used when both noise reduction and edge preservation are required.

[0028] 1.3 Dimensional normalization processing; The denoised images are uniformly scaled to a preset standard size to ensure consistent input size for subsequent processing modules (such as deep learning recognition models). Specifically, bicubic or bilinear interpolation algorithms are used for image scaling, with the target size set according to the input requirements of the subsequent recognition model (e.g., uniformly scaled to 640×640 pixels or 512×512 pixels). Size normalization eliminates image scale differences caused by different monitoring devices and shooting distances, ensuring consistency in subsequent feature extraction.

[0029] 1.4 Illumination equalization processing; The image after size normalization is subjected to illumination equalization processing to eliminate uneven lighting at power work sites caused by factors such as weather changes, time of day, and shadow occlusion. Specifically, one or a combination of the following methods is used: (a) Histogram equalization: stretching the grayscale histogram of the image to make the pixel grayscale values ​​evenly distributed across the entire grayscale range, thereby improving the overall contrast of the image; (b) Adaptive Histogram Equalization with Limited Contrast (CLAHE): The image is divided into several small blocks (such as an 8×8 grid), and histogram equalization is performed on each block. By limiting the contrast amplitude (clipLimit parameter is set to 2.0), local over-enhancement is avoided, which effectively improves the detail performance of unevenly lit areas. (c) Retinex algorithm: Based on Retinex theory, the image is decomposed into illumination component and reflectance component. The illumination component is estimated and equalized by guided filtering, and then fused with the reflectance component to achieve illumination equalization.

[0030] Preferably, for outdoor power operation scenarios with significant changes in lighting conditions, the CLAHE algorithm is used for lighting equalization.

[0031] 1.5 Extraction of key ROI regions for the task; After image equalization, extract the Region of Interest (ROI) and filter out irrelevant background areas to reduce the computational burden of subsequent processing. Specifically, one of the following methods can be used: (a) Static ROI extraction: Based on the fixed monitoring perspective of the power operation site, the coordinate range of the operation area in the image is pre-defined (such as the pixel coordinate range [y1:y2, x1:x2]), and the image area within this coordinate range is directly cropped as the ROI; (b) Dynamic ROI extraction: Target detection algorithms (such as YOLO series, Faster-RCNN, etc.) are used to detect targets such as workers and power equipment in the image, and dynamic ROI regions are generated based on the detection boxes; or background modeling and foreground segmentation methods are used to extract moving target regions as ROIs; (c) ROI extraction based on semantic segmentation: The image is segmented at the pixel level using a pre-trained semantic segmentation network (such as U-Net) to generate a mask of the working region, and the ROI is extracted based on the mask.

[0032] Preferably, for power operation scenarios with a fixed monitoring perspective, a static ROI extraction method is used, where the coordinate range of the operation area is pre-defined and the ROI is directly cropped; for scenarios with a variable monitoring perspective or an unfixed operation area, a dynamic ROI extraction method is used. The extracted ROI regions serve as input images for subsequent processing modules such as behavior recognition and violation detection.

[0033] Through the sequential processing described in 1.1 to 1.5 above, the acquisition and preprocessing of power field monitoring images are completed, and the ROI image, which has been denoised, normalized in size, and equalized in illumination, and is focused on the key operation area, is output, providing high-quality input data for subsequent intelligent recognition and analysis.

[0034] S2. Input the preprocessed image (S1 output) from S1 into the improved YOLOv8-Pose model to extract the coordinates and confidence scores of 13 core joint key points of the human body, and complete the preliminary detection of human targets.

[0035] In a preliminary step, the ROI image output from S1 is input into YOLOv8-Pose, which outputs a high-dimensional feature tensor. Human body detection bounding box, coordinates of 13 key points and their confidence scores.

[0036] S2 preprocessing: Performs regular nonmaximum suppression on the human detection bounding boxes output by YOLOv8-Pose, and decodes the bounding boxes and keypoint coordinates to the ROI image pixel coordinate system. The keypoint confidence threshold is only used to generate visibility markers; keypoints below the threshold are marked as suspected occlusions but are not deleted, and their coordinates and confidence scores are still used for occlusion judgment and location inversion in S2.1.

[0037] The ROI image output by S1 is scaled and pixel normalized according to the model input size and then input into YOLOv8-Pose. The detection head and keypoint prediction head output high-dimensional feature tensors, human detection boxes and keypoint sets, which serve as inputs for subsequent three-domain coupled feature refinement and ternary competitive energy field enhancement.

[0038] 2.1 To address the challenges of varying scales, equipment interference, and partial human occlusion at power operation sites, this paper constructs a normalized human position field based on the positional offset between the human detection box and the preliminary feature map, building upon the initial detection using YOLOv8-Pose. A power equipment structure field is constructed by extracting multiple structural responses from the ROI image. A skeleton map is established based on the anatomical connections of key points, and the position of occluded key points is inverted to form a human skeleton topological field. Refined features are obtained by processing the preliminary feature map based on the three-field output responses. This improves the accuracy of human key point localization, reduces equipment false detection rate, and enhances the recoverability and reliability of occluded key points.

[0039] Organize the input data for this step: Let the input three-dimensional feature tensor (S1 output) be... ,in, Indicates the number of channels. and These represent the height and width of the feature map, respectively.

[0040] Let the human detection bounding box obtained from the initial detection by YOLOv8-Pose be... ,in, Indicates the coordinates of the center of the human body frame. This indicates the width and height of the human body frame.

[0041] Let the set of key points of the human body be... ,in, Indicates the first The coordinates of the key points This indicates the confidence level of the key point. Indicates the number of key points.

[0042] (1) Constructing a normalized human body position field; For feature maps any position in Calculate its relative to the center of the human body frame. Normalized coordinates :

[0043] in, and These represent the horizontal and vertical offsets of the current feature point relative to the center of the human target, respectively. Unlike directly using absolute image coordinates, this normalized representation can eliminate scale differences caused by different shooting distances, different human body sizes, and different image resolutions, allowing the model to focus more on the relative structural positions within the human target.

[0044] The normalized position field of the human body can then be expressed as: ; ; in, Represents the coordinates of the normalized position field of the human body in the feature map space. The output response value at that point, which is the pixel value of the final generated location field feature map at that point; For position field mapping function; This represents the channel features of the current high-dimensional feature vector at its current position. Indicates feature splicing, , For length is A vector of all 1s; , , and These are the network's learnable parameters; This is a sigmoid function, with the output normalized to [0,1]. This location field is used to characterize the relative location of the current feature point within the human bounding box, such as the head and shoulder region, torso region, upper limb region, and lower limb region.

[0045] (2) Constructing the power equipment structural field; For any location in the ROI region image output by S1 The grayscale image is processed by linear convolution kernels in multiple directions (such as 0°, 45°, 90°, 135°, or 8 directions). Perform convolution, then take the maximum (or root of square) value of the corresponding absolute values ​​in each direction, and normalize to obtain the linear structure response. ; For grayscale images The gradient magnitude is calculated using the Sobel operator, and the edge direction response is obtained after normalization. ; For the ROI region image, convert it from RGB to HSV, and for the luminance channel V, calculate its mean. and standard deviation Let the brightness threshold be... The brightness in the luminance channel V Exceeding The portion is normalized to obtain the high-brightness reflective response. The normalization is expressed as: ; in, It is the global maximum value of V.

[0046] The structural field representation of the power equipment is then: ; ; in, To represent the structural field of power equipment in the coordinates of the feature map space The output response value at that location, Indicates the response of a linear structure. Indicates the response in the edge direction. Indicates a reflective or highlighting response. This represents the device structure mapping function. This is the truncation function. The device structure field is used to identify external human interference areas that may cause false detections or occlusion of key points. The three types of structural responses are first calculated in the pixel coordinate system of the ROI image. Then, a horizontal and vertical ratio mapping is established based on the ROI image size and the initial feature map size. Bilinear interpolation is used to resample the response map into an H×W map, aligning each feature map position with the corresponding receptive field center in the ROI before participating in the device structure field calculation.

[0047] In power operation sites, structures such as conductors, tower edges, insulator outlines, crossarms, safety ropes, and tool bag edges often exhibit elongated linear features, regular edges, high-brightness reflections, or repetitive textures. Their local morphology resembles human limbs, making them easily misidentified by models as key areas such as elbows, wrists, knees, and ankles. Therefore, this invention aligns the linear structure responses, edge direction responses, and high-brightness reflection responses from the ROI image to the feature map space to construct the equipment structure field.

[0048] (3) Construct the topological field of the human skeleton; Based on the anatomical adjacency relationships of the human body, the topology of the human skeleton is defined by the numbering and connecting edges of 13 core joints of the human body. When the 13 key points are the nose, both shoulders, elbows, wrists, hips, knees and ankles, they are connected sequentially to the nose and both shoulders, shoulder-elbow-wrist, ipsilateral shoulder-hip, hip-knee-ankle, and both shoulders and both hips. For example, head (0)-neck (1), neck (1)-left shoulder (2), neck (1)-right shoulder (3), left shoulder (2)-left elbow (4), ...

[0049] Let the human skeleton be... ,in, Represents the set of key points on the human body. This represents the set of edges connecting the human skeleton. For any keypoint... If its confidence level is lower than the set threshold If so, the key point is determined to be in a suspected occlusion state, as shown below: ; For a binary function, when Indicates the first The key points are suspected to be obstructed key points, meaning that they may be affected by equipment obstruction, reflection interference, or human body obstruction. This indicates that the key point is visible.

[0050] In a further implementation, for The points are inverted and compensated based on adjacent visible keypoints in the human skeleton topology. This includes the following steps: For suspected obstruction of key points Define its set of adjacent visible keypoints. for: ; If the set If empty, continue selecting. The second-order visible keypoints are keypoints that are separated by two skeleton edges and satisfy the confidence threshold.

[0051] Based on the set of adjacent visible key points Suspected of obscuring key points inversion position for: ; in, Adjacent visible keypoints Location; This indicates the confidence level of the key point; The relative displacement of adjacent joints is determined by the proportion of the human skeleton's side length, the direction of joint connections, and the current human frame size. To prevent stable terms with a denominator of zero, in a further implementation, the relative displacement between adjacent joints is calculated as follows: both the key point position and the inversion position are represented by (x,y) two-dimensional coordinates, and the displacements in the x and y directions are calculated separately during inversion.

[0052] (a) Using a large amount of unoccluded and fully visible human pose data as training samples, key points in all samples are analyzed. and Calculate the normalized bone vector between edges: Based on the width of the human body frame ,high and visible key points and Location and Calculate the normalized bone vector , The mean of the normalized bone vectors for all samples is taken as the standard bone vector.

[0053] (b) Furthermore, during inference, the current human body frame width is considered. and height calculate ,in and These are the horizontal and vertical components of the standard bone vector, respectively. Therefore, Given adjacent visible keypoints Estimated Candidate positions. In the formula, For the first The two-dimensional position of each adjacent visible key point It only represents the confidence level of the key point and serves as the weight for the aforementioned weighted inversion.

[0054] In this way, when key points such as the elbow, wrist, knee, and ankle are partially obscured by conductors, towers, insulators, safety ropes, or tool bags, the position can be inverted using relatively stable visible key points such as the shoulder and hip.

[0055] In a further implementation, to prevent the electrical equipment structure from being incorrectly used as the basis for human body key point compensation, a counter-evidence constraint is introduced. This constraint is applied to the key point locations obtained through inversion. If it is located in a reasonable area of ​​the human body and conforms to the skeletal connection relationship, its credibility is increased; if it falls into the high-response region of the equipment's structural field, its credibility is decreased. This credibility can be expressed as: ; in, This indicates the score for the point being located in a reasonable area of ​​the human body. This indicates the structural consistency score between the point and its adjacent skeleton edges. This indicates the score for the proof by contradiction that the point is located within the area of ​​the power equipment structure. , and For adjustment coefficients, This represents the Sigmoid activation function. All scores are normalized to [0,1]. The adjustment coefficients for the reasonable human body parts and structural consistency items are both set to 1. The equipment disproven items use the following equipment disprovenance adjustment coefficients.

[0056] Scoring of reasonable parts of the human body Take the normalized position field response of the human body at the inversion location; Structural consistency score Pick Candidate positions and final inversion positions independently calculated for each adjacent visible keypoint in the middle The average Gaussian similarity of the distance, i.e., the distance between... Find the mean; where, It is the set of adjacent visible keypoints; The scale is Gaussian, representing the allowable inversion position deviation.

[0057] Equipment counter-evidence score Take the inversion position Response. The equipment's counter-evidence adjustment coefficient can be determined as follows: ,in This is the average field value of the equipment structure within the human body frame, therefore Located in [1,2]; for example When =0.4, =1.4. This coefficient corresponds to the equipment counter-proof term adjustment coefficient in the previous formula.

[0058] After obtaining the occlusion inversion keypoints and their credibility, a keypoint occlusion inversion response map is constructed, represented as follows: ; Among them, credibility It is a scalar; when the feature map position When traversing all H×W positions, the Gaussian distance term changes with position. Multiplying the response value by the distance term yields the response value for each location; all response values ​​together constitute the keypoint. Instead of directly using a single numerical value as the response map, a two-dimensional occlusion inversion response map is generated. Thus, keypoint information is converted into H×W point-by-point responses, consistent with the fusion granularity of the aforementioned two fields.

[0059] in, Based on an adaptive Gaussian scale, the scale is adaptively determined according to the human bounding box scale, keypoint confidence, and occlusion level. For distant, small-scale, or severely occluded targets, Increase the size appropriately to expand the search range for key points; for close-range, clear targets, Reduce the size appropriately to improve positioning accuracy.

[0060] In one implementation, the human body bounding box scale is converted to the current feature map coordinates and denoted as... ,make .

[0061] in, For reference to the Gaussian scale, and These are the minimum and maximum Gaussian scales, respectively.

[0062] All units are feature map pixels; the smaller the human bounding box or the lower the confidence level of key points. The larger the scale, the greater the value. Among them, the adaptive Gaussian scale is the keypoint search scale; the base scale and the reference scale are the baseline search scale and the human bounding box normalization baseline, respectively; the minimum scale and the maximum scale define the lower limit and the upper limit of the search scale, respectively.

[0063] The topological field of the human skeleton can generate a Gaussian response around the line segment corresponding to each effective skeleton edge, and takes the maximum value over all skeleton edges, i.e. .

[0064] In this context, effective skeleton edges are connections within a predefined human skeleton where both endpoints can be used for topology calculations. When an endpoint is a visible keypoint, the detected position is used directly. When an endpoint is a suspected occluded keypoint, the inverted position is used only if its position has been inverted and the inversion confidence is not lower than a set threshold. Other connections are not included in the calculation. Therefore, effective skeleton edges are not limited to cases where both endpoints are visible keypoints. The line segment distance is the shortest Euclidean distance from the current position to that line segment; the Gaussian scale represents the skeleton response bandwidth.

[0065] Finally, the normalized human body position field, the power equipment structure field, the human skeleton topology field, and the keypoint occlusion inversion response map are fused to obtain the refined weights of the three-domain coupled features, expressed as: ; , , , These are the normalized human body position field, the human skeleton topological field, the keypoint occlusion inversion response map, and the power equipment structural field, respectively. In this embodiment, Defined as a rejection factor for high-response regions of equipment structure. Features of splicing; This represents the fusion mapping function, which uses a 1×1 convolution for mapping. This is the Sigmoid activation function.

[0066] With refined weight For the original high-dimensional feature vector Weighted summaries are used to ultimately output refined features. :

[0067] in, This indicates element-wise multiplication.

[0068] 2.2 By linking human evidence, equipment counter-evidence, and temporal consistency to model feature importance, this clearly reflects the competitive relationship between the real human posture region and the equipment interference region in power operation scenarios. This enhancement process is achieved by constructing human evidence energy fields, equipment counter-evidence energy fields, and temporal consistency energy fields. Here, human evidence energy is used to represent the probability that the current position belongs to a real human key point or skeleton region; equipment counter-evidence energy is used to represent the probability that the current position belongs to a conductor, tower, insulator, crossarm, safety rope, tool edge, or highly reflective area; and temporal consistency energy is used to represent whether the current position conforms to the human motion pattern of consecutive frames. After the ternary competition, the output of this energy function represents the credibility of the current position as a human posture feature.

[0069] Let the input feature map of the current frame be the refined feature map. The previous frame input feature map is The set of human key points in the current frame is ,in, Indicates the current frame number One key point, This indicates its confidence level.

[0070] (1) Constructing the energy field of human evidence.

[0071] The strength of human evidence is jointly determined by the effective area of ​​the human detection bounding box, the keypoint occlusion inversion response map, and the human skeleton connection area. For refined features... any position in The strength of human evidence can be expressed as: ; in, This indicates whether the current location is within the valid area of ​​the human detection box. Indicates the first Occlusion inversion response diagram of key points This indicates the response of the human skeleton connection area. These are the non-training weight coefficients. The maximum value in the formula represents the maximum value taken point-by-point along the effective keypoint index at a fixed position. This step directly reuses the response maps generated in section 2.1 for the same frame and aligned with the refined feature space, without recalculating.

[0072] The effective area of ​​a human detection bounding box refers to the set of pixels within the boundary of the detection box; location Indicator value within the time frame region of this set Select 1 otherwise select 0. The response of the human skeleton connection region adopts the aforementioned human skeleton topological field. The three non-training weight coefficients in the human evidence are set to 1 / 3, 1 / 3, and 1 / 3 respectively, and remain unchanged during training and inference.

[0073] The energy field of human evidence can be represented as: ; in, and represents the mean and standard deviation of the strength of human evidence within the human candidate region, respectively. For stability. Human candidate region. Defined as the location of all feature maps within the current human detection bounding box. The mean and standard deviation are calculated for each human bounding box as follows: ; ; in The strength of human evidence is determined by the number of human bodies in the bounding boxes; when multiple human bodies are present, they are counted separately, and they are not mixed across human bodies.

[0074] (2) Constructing the equipment counter-evidence energy field. The equipment counter-evidence strength is jointly determined by the electrical equipment structural field, linear edge response, reflective response, and consistency with the non-human skeleton:

[0075] in, Indicates the structural field response of power equipment. Indicates the response of a linear structure. Indicates reflective response, This indicates a penalty item that does not conform to the connection area of ​​the human skeleton. These are non-training weight coefficients.

[0076] The structural response, linear structural response, and reflective response of power equipment are respectively adopted using the aforementioned methods. , and Penalty items that do not conform to the connection area of ​​the human skeleton are removed. The four non-training weight coefficients are all set to 1 / 4 and kept fixed.

[0077] Equipment interference area is defined as The mean and standard deviation of the equipment's counter-evidence strength are both within Internal evidence is calculated in the same way as human evidence.

[0078] The energy field of the device can be represented as: ; in, and These represent the mean and standard deviation of the counter-evidence intensity within the interference area of ​​the equipment, respectively.

[0079] (3) Construct a time-consistent energy field.

[0080] In power operation videos, key points of real human bodies typically exhibit continuous motion trajectories, while background equipment such as poles, conductors, and insulators mostly remain relatively stationary. Strong reflections and localized noise responses often exhibit instantaneous changes. Therefore, this invention utilizes the displacement and feature response changes of key points in adjacent frames to determine whether the current response possesses the continuity of real human body motion. Key point consistency is calculated based on the displacement deviation of key points in refined features between adjacent frames; let the first... The positions of the key points in the previous frame and the current frame are as follows: and Then the motion consistency of key points can be expressed as:

[0081] in, This represents the allowed keypoint displacement scale between adjacent frames, adaptively determined based on the human bounding box scale and historical motion velocity. In the 25 frames / second embodiment, the historical velocity is calculated first. , Let be the median of the displacements of all visible keypoints in the previous adjacent frame; then let ,in This represents the current human bounding box scale, in pixels. The larger the human bounding box or the faster the historical movement, the greater the allowed inter-frame displacement. The keypoint displacement scale is taken from the current frame's displacement scale; that is, all keypoints within the same frame share this scale.

[0082] Further, the timing consistency response is obtained: The maximum value is calculated point by point along the key point index at a fixed position.

[0083] The time-consistent energy field can then be expressed as: ; in, and These represent the mean and standard deviation of the time-consistent response, respectively.

[0084] The statistical range of the temporal consistency response is consistent with the human candidate region, that is, within the current human detection box. Internal calculation: , Each detection box is statistically analyzed separately; when there is no historical information in the first frame, the temporal consistency response is initialized to a human evidence response. The statistical domain is the human candidate region corresponding to the current human detection box.

[0085] In a further embodiment, after obtaining the human evidence energy, the device counter-evidence energy, and the temporal consistency energy, the present invention constructs a ternary competitive energy function:

[0086] in, These are preset non-training coefficients, used to control the influence intensity of human evidence, temporal consistency, and device-based counter-evidence, respectively.

[0087] In one implementation, the non-training coefficients corresponding to human evidence, temporal consistency, and device counter-evidence are set to 1, 1, and 1, respectively. The three inputs are standardized before entering the competition function, so they are set with equal weights. The coefficients are fixed and not updated during training and inference.

[0088] To further improve the ability to distinguish between real human body regions and device interference regions, this invention directly uses the aforementioned ternary competitive energy function as the input logit for competitive normalization, and obtains parameter-free enhancement weights through Sigmoid equivalent normalization, which are expressed as follows: ; in, This indicates the competition enhancement weights for the current location belonging to the real human pose region. When a region simultaneously possesses both human evidence and temporal consistency, and the device's counter-evidence is weak, Larger; when a certain area has a strong local response, but it mainly originates from conductors, towers, insulators, reflective edges, or does not conform to the movement patterns of the human skeleton, The weighting is reduced. Under the above equal weighting setting, this fraction is equivalent to performing a Sigmoid transformation on the ternary competitive energy function. Therefore, the ternary competitive energy function directly participates in the calculation of the enhancement weights.

[0089] Finally, the enhanced features are represented as:

[0090] in, This indicates element-wise multiplication.

[0091] S3, Graph Convolution Feature Fusion: 3.1 GCN Input Construction: Using the human skeleton map constructed by S2.1 as the graph structure, 13 core joints are used as nodes, and the pre-defined human skeleton connection relationships are used as the edges of the graph. Visible key points are sampled using detection coordinates, while suspected occluded key points that have undergone reliable inversion are sampled using inversion coordinates. Local features of the nodes are obtained at the corresponding positions in the enhanced feature map through bilinear interpolation sampling.

[0092] For the Each node is set with a confidence mask. The key points are taken from the original detection confidence level, the key points from the confidence level are taken from the inversion confidence level, and the remaining invalid nodes are taken as 0. The node local features, normalized two-dimensional coordinates and corresponding confidence levels are concatenated and then combined with the original detection confidence level. Multiply to form the feature matrix of the GCN input nodes. .

[0093] The boundary set E is determined according to the anatomical connections of the head-shoulder, shoulder-elbow, elbow-wrist, shoulder-hip, hip-knee, and knee-ankle.

[0094] 3.2 GCN Output: The node feature matrix after confidence level masking The normalized adjacency matrix obtained from the edge set E is input into the GCN, and the fused pose features of 13 nodes are output. The output is a node-level pose feature sequence, which serves as the input for subsequent global Transformer modeling.

[0095] S4. Transformer Global Feature Enhancement and Detection: The 13 node-level pose features output by S3 are fused together. The Transformer takes a sequence of nodes as input, models the long-distance dependencies between joints through global self-attention, and outputs node features with global context. ;13 represents the number of key points on the human body, and D is the feature dimension.

[0096] Aggregation is performed using confidence mask weighted average pooling: Transformer output features of each node Multiply each value by the corresponding confidence mask, sum the results, and then divide by the sum of the masks and the sum of the stability terms to obtain the global pose features. 。 is represented as: ; This global pose feature The input classification head is used for action recognition. The classification head consists of a multi-layer perceptron, which includes linear transformation, GELU activation function, Dropout regularization and the final Softmax output layer. It maps features to the action category space and outputs the probability distribution of each category.

[0097] Based on the probability distribution of each category output by the classification head, and combined with the preset judgment threshold of the power operation safety regulations, the human posture type is identified; it is determined whether it is an abnormal posture such as not wearing a seat belt or being unbalanced, and the false alarm problem is optimized through continuous frame analysis.

[0098] The specific implementation includes: (1) Read the probability of each posture category output by the S4 classification head, and match the categories such as normal operation, not wearing a safety belt, body imbalance, and illegal climbing with the preset safety rule table; (2) When the maximum probability of the anomaly category is not less than 0.70, and there are no less than 8 key points with valid positions after S2 preprocessing (including visible key points and key points that reach the inversion confidence threshold), the current frame is recorded as an anomaly candidate frame. (3) Using 15 consecutive frames as a sliding window, if at least 10 of them are of the same anomaly category and the average probability of that category is not less than 0.70, an anomaly is confirmed; otherwise, it is suppressed as a single frame false alarm. (4) Output the personnel detection box, abnormal pose category, average confidence score and corresponding video frame, and generate early warning information. The above thresholds are determined once based on the validation set after training and remain unchanged during actual inference.

[0099] Example 2 As an embodiment of this application, a human posture detection system for complex power scenarios is disclosed, employing the specific implementation method described in Embodiment 1 above. The system includes: Basic data acquisition module: Collects and preprocesses images of power field operation scenarios, extracts ROI images, and inputs them into the YOLOv8-Pose model to initially extract feature maps, human detection boxes, key point coordinates, and confidence scores; Key point feature refinement module: Constructs a normalized human body position field based on the positional offset between the human body detection box and the preliminary feature map; extracts multiple structural responses from the ROI image to construct a power equipment structural field; establishes a skeleton map based on the anatomical connection relationship of human body key points and performs position inversion on suspected occluded key points to form a human skeleton topological field; processes the preliminary feature map based on the three-field output response to obtain refined features. The key point feature enhancement module calculates the strength of human evidence based on the effective region of the detection box, the key point occlusion inversion response, and the human skeleton connection region response in the refined features, and normalizes it to obtain the human evidence energy field; it calculates the strength of device counter-evidence based on the multi-class structural responses in the refined features, and normalizes it to obtain the device counter-evidence energy field; it calculates the temporal consistency response using the key point positions of adjacent frames of the refined features, and normalizes it to obtain the temporal consistency energy field; and it processes the refined features based on the three-field responses to obtain the enhanced features. The human pose recognition module obtains local node features from the skeleton graph motion enhancement features, sets confidence masks for the nodes and constructs a node feature matrix; outputs node features after graph convolution calculation and Transformer modeling, and outputs global pose features after aggregation through confidence masks; and uses a classification head to map the global pose features to the action category probability distribution to identify the human pose type.

[0100] As an embodiment of this application, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it employs the specific implementation method described in Embodiment 1 above for human posture detection.

[0101] As an embodiment of this application, a computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, it adopts the specific implementation method described in Embodiment 1 above for human posture detection.

[0102] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0103] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0104] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0105] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0106] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for human posture detection in complex power scenarios, characterized in that, include: Images of power field operation scenarios are collected and preprocessed. After extracting ROI images, they are input into the YOLOv8-Pose model to initially extract feature maps, human detection boxes, key point coordinates, and confidence scores. A normalized human position field is constructed based on the positional offset between the human detection bounding box and the preliminary feature map; a power equipment structure field is constructed by extracting multiple structural responses from the ROI image; A skeleton map is constructed based on the anatomical connections of key points in the human body, and the positions of suspected occluded key points are inverted to form a topological field of the human skeleton. The preliminary feature map is processed based on the three-field output response to obtain refined features. The strength of human evidence is calculated based on the effective region of the detection box, the key point occlusion inversion response, and the human skeleton connection region response in the refined features, and normalized to obtain the human evidence energy field; the strength of device counter-evidence is calculated based on the multi-type structural responses in the refined features, and normalized to obtain the device counter-evidence energy field; the temporal consistency response is calculated using the key point positions of adjacent frames of the refined features, and normalized to obtain the temporal consistency energy field; the enhanced features are obtained by processing the refined features based on the three field responses. Local node features are obtained from the skeleton graph animation enhancement features. Confidence masks are set for the nodes and a node feature matrix is ​​constructed. After graph convolution calculation and Transformer modeling, the node features are output. After aggregation through confidence masks, the global pose features are output. A classification head is used to map global pose features to action category probability distributions to identify human pose types.

2. The method for human posture detection in complex power scenarios according to claim 1, characterized in that, The process of constructing the normalized human body position field includes: Calculate any position in the preliminary feature map Normalized coordinates relative to the center of the human detection box; The normalized coordinates are concatenated with the initial feature map, and after being mapped by the network, the normalized human body position field coordinates in the feature map space are output. The output response value at that location.

3. The method for human posture detection in complex power scenarios according to claim 1, characterized in that, The construction process of the power equipment structure field includes: Extract multiple structural responses from ROI images, including linear convolution response, Sobel edge gradient magnitude response, and HSV brightness channel highlight reflection response; By fusing multiple structural responses, the coordinates of the power equipment structural field in the feature map space can be obtained. The output response value at that location.

4. The method for human posture detection in complex power scenarios according to claim 1, characterized in that, The process of constructing the topological field of the human skeleton includes: The key points are constructed into a human skeleton diagram according to their anatomical adjacency; key points with a confidence level lower than a set threshold are identified as suspected occluded key points; For suspected occluded keypoints, define a set of adjacent visible keypoints; Based on the positions of all adjacent visible keypoints in the set, the inversion positions of suspected occluded keypoints are calculated by combining the standard bone vector and the current human bounding box width and height; the standard bone vector is calculated using historical human pose data. For the inversion location, the credibility is calculated based on the scores of reasonable human body parts, structural consistency, and equipment counter-evidence, and a key point occlusion inversion response map is generated. For each valid skeleton edge, a Gaussian response is generated and the corresponding maximum value is taken to form the topological field of the human skeleton.

5. The method for human posture detection in complex power scenarios according to claim 4, characterized in that: The score for the reasonable part of the human body is the normalized position field response of the human body at the inverted position. The structural consistency score is calculated as the average Gaussian similarity of the distances between adjacent visible keypoints and their inversion locations. The equipment counter-proof score is obtained by taking the structural field response of the power equipment at the inversion location.

6. The method for human posture detection in complex power scenarios according to claim 1, characterized in that, The process of obtaining the refined features includes: Definition 1 minus the structural field of the power equipment is the repulsion factor; The normalized human body position field, human skeleton topological field, key point occlusion inversion response map, and repulsion factor are concatenated, and refined weights are output through convolutional mapping and activation function. The initial feature map is weighted with refined weights to output refined features.

7. The method for human posture detection in complex power scenarios according to claim 1, characterized in that, The process of constructing the human evidence energy field includes: For any position in the refining feature The strength of human evidence is obtained by weighting the corresponding effective area indicator of the detection box, the maximum value of the key point occlusion inversion response map, and the response of the human skeleton connection area. when The set of pixels located within the boundary of the detection box has an indicator value of 1, otherwise it has a value of 0; The response value of the human skeleton connection area is The topological field response of the human skeleton.

8. The method for human posture detection in a complex power scenario according to claim 1, characterized in that, The process of constructing the energy field by the device includes: For any position in the refining feature The corresponding power equipment structural field response, linear convolution response, HSV brightness channel high brightness reflective response and skeleton connection penalty term are weighted and normalized to obtain the equipment's reverse energy field. The skeleton connection penalty term is defined as 1 minus the response of the human skeleton connection area.

9. The method for human posture detection in a complex power scenario according to claim 1, characterized in that, The construction process of the time-consistent energy field: The consistency of key points is calculated based on the displacement deviation of key points in two adjacent frames in the refined features; the occlusion inversion response map is weighted using the key point consistency to obtain the temporal consistency response; The timing consistency response is calculated point by point along the key points, and the maximum value is taken for normalization to obtain the timing consistency energy field.

10. The method for human posture detection in a complex power scenario according to claim 1, characterized in that, The process of obtaining the enhanced features includes: The competitive enhancement weights of the human evidence energy field, the device counter-evidence energy field, and the temporal consistency energy field are calculated through a competitive normalization mechanism. The refined features are weighted by competition enhancement weights to generate enhanced features.

11. A human posture detection system for complex power scenarios, executing the human posture detection method as described in any one of claims 1-10, characterized in that, The system includes: Basic data acquisition module: Collects and preprocesses images of power field operation scenarios, extracts ROI images, and inputs them into the YOLOv8-Pose model to initially extract feature maps, human detection boxes, key point coordinates, and confidence scores; Key point feature refinement module: Construct a normalized human body position field based on the positional offset between the human body detection box and the preliminary feature map; extract multiple structural responses from the ROI image to construct a power equipment structural field; establish a skeleton map based on the anatomical connection relationship of human body key points and perform position inversion on suspected occluded key points to form a human body skeleton topological field; process the preliminary feature map based on the three field output responses to obtain refined features; The key point feature enhancement module calculates the strength of human evidence based on the effective region of the detection box, the key point occlusion inversion response, and the human skeleton connection region response in the refined features, and normalizes it to obtain the human evidence energy field; it calculates the strength of device counter-evidence based on the multi-class structural responses in the refined features, and normalizes it to obtain the device counter-evidence energy field; it calculates the temporal consistency response using the key point positions of adjacent frames of the refined features, and normalizes it to obtain the temporal consistency energy field; and it processes the refined features based on the three-field responses to obtain the enhanced features. The human pose recognition module obtains local node features from the skeleton graph motion enhancement features, sets confidence masks for the nodes and constructs a node feature matrix; outputs node features after graph convolution calculation and Transformer modeling, and outputs global pose features after aggregation through confidence masks; and uses a classification head to map the global pose features to the action category probability distribution to identify the human pose type.

12. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is loaded into the processor, it implements the human posture detection method according to any one of claims 1-10.

13. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the human posture detection method according to any one of claims 1-10.

Images