Image recognition-based monitoring target automatic detection method for power distribution network non-power-off operation

Abstract

This application discloses an automatic target detection method for live-line power distribution monitoring based on image recognition. The method includes: acquiring images of the work site and preprocessing them to obtain standardized images; extracting multi-scale feature maps using a shared backbone feature extraction network; detecting candidate regions for protective equipment and key human body points using a dual-branch network; constructing a human posture chain functionally associated with the protective equipment based on the key points; calculating an initial association score between the candidate regions and key points; analyzing the structural integrity of the protective equipment to obtain structural feature vectors; calculating a posture consistency index based on the posture chain; and finally fusing the above parameters to obtain functional association credibility, thereby determining the protective equipment wearing status. This method achieves accurate detection of the protective equipment wearing status, adapts to complex work scenarios, reduces false positives and false negatives, and improves the automation and intelligence level of safety monitoring for live-line power distribution operations.

Description

Technical Field

[0001] This application relates to the field of image recognition technology, specifically to an automatic detection method for monitoring targets during live-line power distribution operations based on image recognition. Background Technology

[0002] Safety management of live-line work on power distribution networks is a core aspect of power operation and maintenance. The proper wearing of insulating protective equipment by workers is a key measure to avoid personal safety risks. Image recognition-based automatic detection technology for the wearing status of protective equipment has become the mainstream application technology in this scenario due to its advantages of intelligence and automation.

[0003] Current detection technologies for protective equipment used in live-line work on power distribution networks mostly rely on basic target detection and simple rule-based judgment. They can only identify the presence of protective equipment, but cannot accurately determine whether it is being worn effectively, or differentiate between situations with significantly different safety risks, such as improper wearing and complete absence of protective equipment. In complex field environments such as strong obstructions, multiple people working simultaneously, non-standard working postures, and tools obstructing protective equipment, these technologies are prone to misjudgments due to accidental overlap between protective equipment and body parts, resulting in high false positive and false negative rates. Furthermore, they struggle to identify situations where changes in working posture cause structural or functional failures in the protective equipment.

[0004] Meanwhile, the judgment logic of existing detection technologies is highly dependent on manual settings, lacks end-to-end learning capabilities, has poor engineering adaptability, and has not established a reasonable uncertainty state modeling mechanism. In scenarios such as blurred images or temporary obstruction by protective equipment, it is forced to make rigid judgments that are either one or the other, which can easily lead to false alarms and is difficult to adapt to the actual needs of power safety supervision. Summary of the Invention

[0005] The purpose of this application is to provide an automatic detection method for monitoring targets of power distribution network uninterrupted operation based on image recognition, so as to solve the problems mentioned in the background art.

[0006] According to a first aspect of this application, an automatic detection method for monitoring targets during live-line work in power distribution networks based on image recognition is provided, comprising the following steps:

[0007] S1, acquire images of the power distribution network live-line work site, preprocess the images to obtain standardized images;

[0008] S2, input the standardized image into the shared backbone feature extraction network to extract multi-scale feature maps;

[0009] S3, input the multi-scale feature map into the protective equipment detection branch and the human body key point detection branch respectively to obtain the protective equipment candidate region set and the human body key point set accordingly;

[0010] S4. Based on the set of key human body points, construct a human body posture chain according to a preset topological connection relationship, wherein the topological connection relationship is related to the function of the protective equipment;

[0011] S5, calculate the initial correlation score between each candidate area of ​​protective equipment and each key point of the human body based on the set of candidate areas of protective equipment and the set of key points of the human body;

[0012] S6. Based on the initial association score, analyze the structural integrity of the wearing structure of each candidate area of ​​protective equipment to obtain a structural feature vector;

[0013] S7. Based on the human posture chain, the structural feature vector, and the protective equipment candidate region, calculate the posture consistency index between each protective equipment candidate region and the corresponding human posture chain.

[0014] S8, the initial association score, the structural feature vector and the attitude consistency index are fused to obtain the functional association credibility of each protective equipment candidate region;

[0015] S9. Determine the wearing status of protective equipment based on the functional association confidence level.

[0016] Preferably, the human posture chain includes a head and neck chain, an arm chain, and a torso chain; the head and neck chain is connected in the order of head key points, neck key points, and shoulder key points, the arm chain is connected in the order of shoulder key points, elbow key points, and hand key points, and the torso chain is connected in the order of waist key points, shoulder key points, and knee key points.

[0017] Preferably, the initial correlation score between each candidate area of ​​protective equipment and each key point on the human body is calculated, including:

[0018] Extract the intersection-over-union (IoU) feature, distance attenuation feature, and positional consistency feature between each candidate region of protective equipment and each key point of the human body;

[0019] The initial association score is obtained by weighting and summing the intersection-union ratio feature, distance decay feature, and location consistency feature.

[0020] Preferably, the structural integrity of the wearing structure of each candidate area of ​​protective equipment is analyzed to obtain a structural feature vector, including:

[0021] Analyze the corresponding functional area structure based on the category of protective equipment;

[0022] For candidate helmet regions, analyze their coverage of key points on the head and the degree of chin strap closure to generate corresponding structural feature vectors.

[0023] For candidate regions of insulating clothing, analyze their coverage continuity of key points of arm chains and torso chains, and generate corresponding structural feature vectors.

[0024] For candidate areas of seat belts, analyze the integrity of their connection structure with key points of the waist and the effectiveness of the stress structure of the shoulder belt and waist belt, and generate corresponding structural feature vectors.

[0025] Preferably, the posture consistency index between each candidate region of protective equipment and the corresponding human posture chain is calculated, including:

[0026] Calculate the displacement change rate between the center of the core functional area of ​​the candidate area of ​​protective equipment and the corresponding key point of the human posture chain to obtain the displacement consistency.

[0027] The coverage retention rate is obtained by calculating the coverage retention rate of the core functional area of ​​the candidate area of ​​protective equipment to the corresponding human posture chain key points as the posture changes.

[0028] The attitude consistency index is obtained by fusing the displacement consistency and the coverage retention.

[0029] Preferably, the initial association score, the structural feature vector, and the attitude consistency index are fused to obtain the functional association confidence level of each candidate region of protective equipment, including:

[0030] The initial association score, the structural feature vector, and the pose consistency index are weighted and fused according to preset weight coefficients to obtain the functional association credibility. The weight coefficients are determined through training.

[0031] For multiple candidate associations formed by associating the same candidate area of ​​protective equipment with multiple key points of different human bodies, the functional association credibility of each candidate association is ranked.

[0032] Candidate associations with functional association confidence scores below a preset threshold are removed, and the candidate associations with the highest confidence scores are retained as valid associations for the candidate area of ​​the protective equipment. The corresponding functional association confidence scores are used as the final functional association confidence scores.

[0033] Preferably, determining the wearing status of protective equipment based on the functional association confidence level includes:

[0034] Based on the visibility of key points on the human body, the detection confidence of candidate areas for protective equipment, and the credibility of the aforementioned functional association, it is determined whether there is an uncertain state.

[0035] If there is uncertainty, the wearing status is further determined by combining the detection confidence level, either the protective equipment is present but the wearing status is uncertain or the detection of the protective equipment is unreliable.

[0036] If there is no uncertainty, the wearing status is determined as either clearly valid wearing or clearly invalid wearing based on the comparison between the functional association confidence level and the preset high confidence threshold and low confidence threshold.

[0037] Preferably, the determination of whether an uncertainty exists specifically means that an uncertainty exists when any of the following conditions are met:

[0038] The detection confidence level of the candidate region for protective equipment is within a preset confidence level fluctuation range;

[0039] The proportion of invisible key points in human body key points is greater than the preset occlusion judgment threshold.

[0040] The attitude consistency index is within a preset middle range;

[0041] The credibility of the functional association is within a preset middle range.

[0042] In a second aspect, this application also provides an automatic detection system for monitoring targets during live-line work in power distribution networks based on image recognition, comprising:

[0043] The image preprocessing module is used to acquire images of the power distribution network live-line work site, preprocess the images to obtain standardized images;

[0044] The feature extraction module, connected to the image preprocessing module, is used to input the standardized image into the shared backbone feature extraction network to extract multi-scale feature maps.

[0045] A multi-branch detection module, connected to the feature extraction module, is used to input the multi-scale feature map into the protective equipment detection branch and the human body key point detection branch respectively, and obtain the protective equipment candidate region set and the human body key point set accordingly.

[0046] A human posture construction module, connected to the multi-branch detection module, is used to construct a human posture chain based on the set of human key points and according to a preset topological connection relationship, wherein the topological connection relationship is related to the function of the protective equipment.

[0047] The association scoring module is connected to the multi-branch detection module and the human posture construction module respectively, and is used to calculate the initial association score between each protective equipment candidate region and each human key point based on the protective equipment candidate region set and the human key point set.

[0048] The structural analysis module, connected to the association scoring module, is used to analyze the structural integrity of each candidate area of ​​protective equipment based on the initial association score, and obtain a structural feature vector.

[0049] The posture consistency analysis module is connected to the human posture construction module, the structure analysis module and the multi-branch detection module respectively, and is used to calculate the posture consistency index between each protective equipment candidate region and the corresponding human posture chain based on the human posture chain, the structural feature vector and the protective equipment candidate region.

[0050] The credibility fusion module is connected to the association scoring module, the structural analysis module and the posture consistency analysis module respectively, and is used to fuse the initial association score, the structural feature vector and the posture consistency index to obtain the functional association credibility of each protective equipment candidate region;

[0051] The status determination module, connected to the credibility fusion module, is used to determine the wearing status of protective equipment based on the functional association credibility.

[0052] This application constructs a human posture chain associated with the function of protective equipment, realizing the feature transformation from discrete key points to continuous posture structure, which can adapt to the detection needs of non-standard working postures. It combines multi-dimensional feature calculation to obtain an initial association score, analyzes the structural integrity of the protective equipment and calculates the posture consistency index, and integrates multiple parameters to obtain functional association credibility. This upgrades the protective equipment wearing status from static judgment to dynamic structure-posture consistency analysis, accurately identifying protective failures caused by structural defects and posture changes. Simultaneously, it establishes an uncertainty state judgment mechanism to achieve a four-level classification of wearing status, avoiding rigid judgments in ambiguous scenarios and reducing false alarm rates. This method is suitable for complex work sites with strong obstructions and multiple people working together, significantly improving the automation and intelligence level of power distribution network live-line work safety monitoring. Attached Figure Description

[0053] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0054] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0055] Figure 1 A schematic diagram of an automatic target detection method for live-line power distribution monitoring based on image recognition provided in this application embodiment;

[0056] Figure 2 This is a schematic diagram of the deep learning model structure provided in the embodiments of this application;

[0057] Figure 3 The initial association score calculation process provided for the embodiments of this application is illustrated in the figure;

[0058] Figure 4 This is a schematic diagram of the attitude consistency index calculation process provided in an embodiment of this application;

[0059] Figure 5 A schematic diagram illustrating the calculation process of the functional association credibility index provided in this application embodiment;

[0060] Figure 6 This is a schematic diagram of an automatic detection system for monitoring targets of power distribution network uninterrupted operation based on image recognition, provided in an embodiment of this application. Detailed Implementation

[0061] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0062] It should be noted that all user information (including but not limited to user device information, user personal information, object information corresponding to device usage data, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, device usage data, etc.) involved in all embodiments of this application are information and data authorized by the user or fully authorized by all parties.

[0063] This method is suitable for safety monitoring scenarios of live-line work in power distribution networks. The executing entities can include video acquisition terminals, edge computing nodes or servers, and processing units that run target detection and semantic analysis models. Image data is transmitted between the executing entities via a network, without relying on additional sensors. This method can be deployed in video monitoring systems for live-line work in power distribution networks, work safety supervision platforms, mobile inspection terminals, and edge computing node collaborative systems. It analyzes single frames or continuous image sequences collected at the work site to automatically detect the wearing status of insulating protective equipment such as safety helmets, insulating suits, and safety belts. It is adaptable to complex live-line work environments in power distribution networks, such as strong obstructions, multiple people working simultaneously, non-standard working postures, and tools obstructing protective equipment.

[0064] For example, the deep learning model implementing this method is a multi-level semantic constraint detection model, such as... Figure 2 As shown, Figure 2This is a schematic diagram of the deep learning model structure provided in an embodiment of this application. The model includes an image input module, a shared backbone feature extraction network module, a protective gear detection branch module, a human keypoint detection branch module, a pose chain construction module, an association scoring calculation module, a structural integrity analysis module, a pose consistency calculation module, an association credibility fusion module, and a wearing status decision module. Specifically, the output of the image input module is unidirectionally connected to the input of the shared backbone feature extraction network module, and the output of the shared backbone feature extraction network module is also unidirectionally connected to the inputs of the protective equipment detection branch module and the human key point detection branch module. The output of the human key point detection branch module is unidirectionally connected to the input of the posture chain construction module, and the outputs of the protective equipment detection branch module and the human key point detection branch module are also unidirectionally connected to the input of the association scoring calculation module. The output of the association scoring calculation module is unidirectionally connected to the input of the structural integrity analysis module, and the outputs of the posture chain construction module, the structural integrity analysis module, and the protective equipment detection branch module are also unidirectionally connected to the input of the posture consistency calculation module. The outputs of the association scoring calculation module, the structural integrity analysis module, and the posture consistency calculation module are also unidirectionally connected to the input of the association credibility fusion module, and the output of the association credibility fusion module is unidirectionally connected to the input of the wearing status decision module.

[0065] This model employs a supervised end-to-end training approach. The training dataset consists of real image data collected from live-line power distribution operations and manually annotated augmented datasets. Annotations include protective equipment detection boxes, human keypoint coordinates, protective equipment wearing status, and pose chain topology. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio. First, the model is pre-trained using publicly available object detection and human keypoint detection datasets to initialize weight parameters. Then, the model is fine-tuned using a dedicated live-line power distribution operation dataset, with synchronized updates to the weight parameters of each module. Finally, threshold optimization is performed using the validation set, and model performance is validated using the test set until it meets the requirements of practical applications.

[0066] The following detailed description, in conjunction with specific embodiments, illustrates the implementation process of the automatic detection method for monitoring targets of live-line power distribution network based on image recognition, as described in this application. It should be noted that these embodiments are merely for explaining this application and are not intended to limit the scope of protection of this application. Any conventional adjustments or substitutions made by those skilled in the art to the steps without departing from the concept of this application should be included within the scope of protection of this application.

[0067] like Figure 1 As shown in the figure, this application discloses a schematic diagram of an automatic target detection method for live-line power distribution monitoring based on image recognition, which includes the following method steps:

[0068] S1, acquire images of the power distribution network live-line work site, preprocess the images to obtain standardized images;

[0069] S2, input the standardized image into the shared backbone feature extraction network to extract multi-scale feature maps;

[0070] S3, input the multi-scale feature map into the protective equipment detection branch and the human body key point detection branch respectively to obtain the protective equipment candidate region set and the human body key point set accordingly;

[0071] S4. Based on the set of key human body points, construct a human body posture chain according to a preset topological connection relationship, wherein the topological connection relationship is related to the function of the protective equipment;

[0072] S5, calculate the initial correlation score between each candidate area of ​​protective equipment and each key point of the human body based on the set of candidate areas of protective equipment and the set of key points of the human body;

[0073] S6. Based on the initial association score, analyze the structural integrity of the wearing structure of each candidate area of ​​protective equipment to obtain a structural feature vector;

[0074] S7. Based on the human posture chain, the structural feature vector, and the protective equipment candidate region, calculate the posture consistency index between each protective equipment candidate region and the corresponding human posture chain.

[0075] S8, the initial association score, the structural feature vector and the attitude consistency index are fused to obtain the functional association credibility of each protective equipment candidate region;

[0076] S9. Determine the wearing status of protective equipment based on the functional association confidence level.

[0077] In some implementations, for step S1, visual data of the power distribution network live-line work site is collected through a video acquisition terminal. The acquisition format is a single-frame RGB image or a continuous RGB image sequence, and the acquired images are formed into an image set. The data structure of a single frame image is as follows: ,in The pixel height of the image. 3 represents the pixel width of the image, and 3 represents the number of three channels in the RGB image. For example, the frame rate of the video acquisition terminal can be configured according to the monitoring needs of the work site, and the resolution of the acquired image can be adjusted according to the processing capability of the edge computing node to ensure that the acquired image clearly presents the human characteristics and protective equipment features of the workers.

[0078] For image sets Each frame of the image undergoes three preprocessing operations in sequence: scale normalization, brightness equalization, and noise suppression. These operations eliminate the influence of external factors such as illumination, scale, and noise on the model's detection performance, and output a standardized image. .

[0079] In the scale normalization operation, the shorter side of the image is scaled to a fixed length. The longer side of the image is scaled proportionally to maintain the aspect ratio and avoid feature extraction bias caused by differences in image scale. Optionally, a fixed length is used. The value can be calibrated according to the model's input requirements and detection accuracy, ranging from 320 pixels to 1024 pixels. The brightness equalization operation uses a histogram equalization algorithm to adjust the brightness of the RGB channels of the image separately, improving the visual effect of the image under extreme lighting conditions such as strong backlighting and low light, and enhancing the feature contrast of the human body and protective gear in the image. The noise suppression operation uses a Gaussian filtering algorithm to smooth the image after brightness equalization, filtering out Gaussian noise and salt-and-pepper noise generated during image acquisition. Optionally, the kernel size of the Gaussian filter can be configured to an odd number, and the kernel parameters are calibrated according to the type of noise in the image.

[0080] In some embodiments, for step S2, a standardized image is input into a shared backbone feature extraction network to extract multi-scale feature maps. The standardized image... The input is a shared backbone feature extraction network module, which employs an improved convolutional backbone network based on residual networks. This module consists of three layers: an initial convolutional layer, a residual feature extraction layer, and a multi-scale feature pyramid. Through layer-by-layer feature extraction and fusion, it outputs a multi-scale feature map. .

[0081] The initial convolutional layer is used for normalized images The process involves performing convolution, batch normalization, and activation function processing to map the image from pixel space to a low-dimensional feature space, outputting a primary feature map. The residual feature extraction layer consists of multiple residual blocks connected in series. By using residual connections, it solves the gradient vanishing problem in deep neural networks, extracts deep features from the primary feature map, and outputs intermediate feature maps with different receptive fields. The multi-scale feature pyramid performs upsampling, downsampling, and feature fusion operations on the intermediate feature maps to construct a multi-scale feature hierarchy, adapting to the feature extraction needs of protective gear of different sizes and key parts of the human body.

[0082] The final output multi-scale feature map The multi-scale feature map structure is as follows: F1 is the large-scale feature map, corresponding to the detailed features of the image, used to detect small-sized protective equipment components (e.g., F1 is 1 / 4 of the original image size); F2 is the medium-scale feature map, corresponding to the mid-level features of the image, used to detect key human body parts (e.g., F2 is 1 / 4 of the original image size); and F3 is the small-scale feature map, corresponding to the global features of the image, used to detect large-sized protective equipment as a whole (e.g., F3 is 1 / 6 of the original image size). This multi-scale feature map provides a shared feature base for subsequent protective equipment detection and human body key point detection, enabling feature reuse and reducing model computation.

[0083] In some embodiments, for step S3, the multi-scale feature map is input into the protective equipment detection branch and the human body key point detection branch respectively, thereby obtaining the protective equipment candidate region set and the human body key point set. Specifically, the multi-scale feature map... Simultaneously input the protective equipment detection branch module and the human body key point detection branch module. The two branches execute the detection operations in parallel and output the candidate region set for protective equipment respectively. Set of key points of the human body .

[0084] For the protective equipment detection branch module, which is an anchor-frame-based target detection head, it employs a detection logic that combines regression and classification. It performs detection frame regression and category classification for three types of protective equipment: safety helmets, insulated clothing, and safety belts, outputting the bounding box coordinates of each candidate region. ,category Detection confidence This forms an initial candidate region set. A detection threshold for protective equipment is then set. This parameter is the effective judgment threshold for candidate areas of protective equipment, determined by the validation set through a threshold optimization method, and its value ranges from 0.5 to 0.8. When the detection confidence level... If the candidate region is deemed a valid candidate region for protective equipment, it is then determined; otherwise, it is discarded. The final output is a set of valid candidate regions for protective equipment. ,in This is the sequence number of the valid candidate region.

[0085] The human keypoint detection branch module is a heatmap-based human keypoint detection network. It uses heatmap regression to locate key human body parts related to the function of protective equipment. The selected keypoints include the head, neck, shoulders, elbows, hands, waist, and knees. A corresponding heatmap is generated for each keypoint, and the peak position of the heatmap is the pixel coordinate of the keypoint. The visibility of the keypoint is also output. , The value of is 0, 1, or 2, representing that the keypoint is invisible, partially visible, or fully visible, respectively, indicating the reliability of the keypoint detection. The final output is a set of human keypoints. ,in The key point number is used as the index. , These are the x and y coordinates of the key points, respectively.

[0086] In some embodiments, for step S4, a human posture chain is constructed based on the set of human key points and according to a preset topological connection relationship. This embodiment addresses the technical problem that existing detection technologies can only output discrete human key point coordinates, failing to characterize continuous human working postures and thus unable to analyze the dynamic consistency between protective equipment and human posture. The principle is to topologically connect discrete key points based on the physiological structure of the human body and the functional relationship of the protective equipment, forming a structured human posture representation. This representation can accurately characterize the local human posture related to the function of the protective equipment, transforming discrete key point features into continuous posture structure features. This allows the model to analyze the wearing status of protective equipment based on the overall working posture of the human body, improving the model's adaptability to non-standard working postures.

[0087] Specifically, the input for this step is a set of key points on the human body. The posture chain construction module performs the posture chain construction operation. Posture chain is a user-defined term referring to the local posture structure of the human body, formed by key points of critical parts of the body through preset topological connections, and directly related to the function of protective equipment. It is the core structural feature for characterizing human working postures. The preset topological connections are strongly correlated with the function of the protective equipment. The constructed human posture chains include three types: head and neck chains, arm chains, and torso chains. The topological connection rules for each type of posture chain are as follows: the head and neck chains are unidirectionally connected according to the order of key points on the head, neck, and shoulders, matching the function of wearing a safety helmet; the arm chains are unidirectionally connected according to the order of key points on the shoulders, elbows, and hands, matching the function of wearing insulated clothing; the torso chains are unidirectionally connected according to the order of key points on the waist, shoulders, and knees, matching the function of wearing a safety belt.

[0088] For example, the pose chain construction module represents the constructed human pose chain as a pose graph. ,in This refers to the aforementioned set of key points on the human body. This refers to the topological connections between key points. These are directed edges, representing the connection order of key points. (Position graph) The output is structured matrix data, which contains the coordinates of key points, visibility information, and topological connection information, and can fully depict the local working posture of the human body and protective equipment related to their functions.

[0089] In some embodiments, for step S5, based on the set of candidate protective equipment regions and the set of human body key points, an initial association score is calculated between each candidate protective equipment region and each human body key point. The main idea is to extract association features between candidate protective equipment regions and human body key points from three dimensions: spatial overlap, spatial distance, and location prior. An initial association score is generated through weighted fusion. These multi-dimensional features can more accurately characterize the association relationship between the two, overcoming the limitations of a single spatial overlap feature. This forms a quantitative indicator that can represent the initial association degree between protective equipment and human body key points, effectively reducing the probability of misjudgment caused by relying solely on spatial overlap.

[0090] Specifically, the input for this step is a set of candidate areas for protective equipment. Set of key points of the human body The initial correlation score calculation operation is performed by the correlation score calculation module, which first extracts the candidate areas for each protective equipment. With each key point of the human body The three association features—crossover union ratio, distance decay, and location consistency—are used to obtain the initial association score. These three association features are then weighted and summed. .

[0091] Please see Figure 3 , Figure 3 The initial association score calculation process provided for the embodiments of this application is illustrated in the diagram. In S301, the intersection-union feature is extracted. The bounding box of the candidate region for protective equipment is calculated. Key points of the human body Intersection over Union (IoU) between neighborhood boxes The neighborhood box is based on key points A square frame centered on the human body, whose side length can be determined as a fixed pixel value according to human scale, and the intersection-over-union ratio (IoU). The value ranges from 0 to 1, representing the degree of spatial overlap between the candidate area of ​​protective equipment and key points on the human body. The larger the value, the greater the spatial overlap between the two.

[0092] In S302, distance attenuation features are extracted. Candidate regions for protective equipment are calculated. Center coordinates and key points of the human body The Euclidean distance between pixel coordinates is used to obtain the distance decay feature after normalization and decay processing of the Euclidean distance. , The value ranges from 0 to 1, representing the spatial distance correlation between the candidate area of ​​protective equipment and key points on the human body. The closer the spatial distance between the two, the higher the correlation. The larger the value.

[0093] In S303, location consistency feature Extraction. Based on the functional attributes of protective equipment, a priori positional relationship between the protective equipment and key points on the human body is preset. For example, a safety helmet should be positioned above key points on the head, and insulating clothing should cover the area from the shoulders to the hands. The relative position of the candidate area center of the protective equipment to the corresponding key point is calculated to determine whether it conforms to this priori relationship and the degree of conformity, mapped to a value between 0 and 1. A higher degree of conformity indicates a better relationship. The larger the value.

[0094] In S304, the three types of association features mentioned above are weighted and summed to obtain the initial association score. The calculation formula is:

[0095]

[0096] In the formula, , , These are the weight coefficients for the three types of association features. The weight coefficients of the crossover-union ratio (CUNR) feature represent the degree of contribution of the CUNR feature to the initial association score. The weighting coefficients for the distance decay feature represent its contribution to the initial association score. The weight coefficient for the location consistency feature represents its contribution to the initial association score. This weight coefficient is determined from the training set using a grid search method. Based on the actual scenario of live-line work in distribution networks, a higher weight is assigned to the location consistency feature. , , The values ​​of all values ​​are between 0 and 1, and they satisfy the normalization constraint. .

[0097] The correlation score calculation module calculates the corresponding initial correlation score for each combination of candidate areas of protective equipment and key points of each human body, and finally generates an initial correlation score set.

[0098] In some embodiments, for step S6, based on the initial association score, the structural integrity of the wearing structure of each candidate region of protective equipment is analyzed to obtain a structural feature vector. According to the category of protective equipment, the structural integrity of each core functional region of the protective equipment is analyzed in conjunction with the initial association score. By quantifying the degree of structural integrity using high-dimensional vectors, structural defects of the protective equipment can be accurately identified. Upgrading the wearing status of protective equipment from a binary judgment of presence or absence to a continuous value judgment of structural integrity effectively identifies scenarios where the equipment appears to be worn correctly but actually has structural safety defects, improving the model's accuracy in depicting the wearing status of protective equipment.

[0099] Specifically, the inputs for this step are the initial association score set and the protective equipment candidate region set. The structural integrity analysis module performs the wearing structural integrity analysis operation. The structural feature vector is a custom term that refers to a high-dimensional feature vector that characterizes whether the core functional areas of the protective equipment are complete after wearing. Each dimension of the vector corresponds to a core functional area of ​​the protective equipment, and the magnitude of the dimension value represents the degree of integrity of the corresponding functional area.

[0100] The analysis is performed separately according to the category of protective equipment. For the core functional areas of three types of protective equipment—safety helmets, insulating suits, and safety belts—the structural integrity is analyzed in conjunction with the initial correlation score, and corresponding structural feature vectors are generated. Each dimension of the structural feature vector is normalized, with values ​​ranging from 0 to 1. Higher values ​​indicate greater structural integrity of the corresponding functional area. The specific analysis logic is as follows:

[0101] For the candidate helmet region, the coverage of the helmet to the head key points is analyzed by combining its initial association score with the head key points. At the same time, it is detected whether the chin strap region of the helmet forms a closed structure, such as whether the left and right ends of the chin strap are connected in the image and whether they cover the chin key point region. Two dimensions of the structural feature vector are generated, which correspond to the head coverage and the chin strap closure, respectively.

[0102] For candidate regions of insulating clothing, the initial correlation scores with key points of the arm chain and torso chain are combined to analyze the coverage continuity of the core insulating regions of the insulating clothing, such as cuffs, shoulders, and torso, to key points of the arm chain and torso chain. This generates four dimensions of the structural feature vector, which correspond to the coverage continuity of the cuffs, shoulders, upper torso, and lower torso, respectively.

[0103] For candidate areas of the seat belt, the initial correlation score with the lumbar key point is used to analyze whether the connection structure between the seat belt body and the lumbar key point is complete. Simultaneously, the shoulder straps and lap belts of the seat belt are checked to see if they form an effective load-bearing structure. Three dimensions of the structural feature vector are generated, corresponding to the integrity of the lumbar connection structure, the effectiveness of the shoulder strap load-bearing structure, and the effectiveness of the lap belt load-bearing structure, respectively. The structural integrity analysis module generates a corresponding structural feature vector for each candidate area of ​​protective equipment. .

[0104] In some embodiments, for step S7, based on the human posture chain, structural feature vector, and candidate regions of protective equipment, a posture consistency index is calculated between each candidate region of protective equipment and the corresponding human posture chain. By jointly modeling the spatial distribution characteristics of the protective equipment and the topological characteristics of the human posture chain, the synchronization degree of the protective area of ​​the protective equipment with changes in the human posture chain is calculated. The consistency between the two is characterized by a quantitative index, enabling accurate identification of protective structural failures caused by posture changes. This upgrades the protective equipment wearing status from static judgment to dynamic structure-posture consistency analysis, effectively identifying functional failure scenarios of protective equipment under non-standard working postures and reducing the false negative rate under complex working postures.

[0105] Specifically, the input for this step is the pose diagram of the human pose chain. Structural feature vectors Collection of candidate areas for protective equipment The posture consistency calculation module performs the posture consistency index calculation operation. The posture consistency index is a comprehensive quantitative index that characterizes the displacement synchronization and coverage maintenance between the core protection area of ​​the protective equipment and the corresponding human posture chain. It is the core parameter for judging whether the protective equipment maintains functional coverage under changes in working posture.

[0106] Please see Figure 4 , Figure 4 This is a schematic diagram of the posture consistency index calculation process provided in this application embodiment. In S401, the protective equipment and the human posture chain are jointly modeled. The posture consistency calculation module first extracts the spatial distribution features of the candidate region of the protective equipment. These features include the bounding box coordinates of the candidate region of the protective equipment, the pixel position of the core functional region, and the spatial mapping information of the structural feature vector, etc. Then, the spatial distribution features are spliced ​​and fused with the topological features of the human posture chain and the coordinate features of key points to generate a joint feature matrix. This matrix simultaneously contains the structural features of the protective equipment and the topological features of the human posture, realizing the joint modeling of the protective equipment and the human posture chain.

[0107] In S402, calculate the attitude consistency index. Attitude consistency index Based on displacement consistency With coverage retention The result of the fusion is calculated using the following formula:

[0108]

[0109] In the formula, displacement consistency With coverage retention Both are normalized quantitative indicators, with values ​​ranging from 0 to 1. They are calculated in parallel without any order of priority. The specific calculation logic is as follows:

[0110] Displacement consistency Calculate the rate of change of displacement between the center coordinates of the core functional area of ​​the candidate protective equipment and the coordinates of the corresponding key points in the human posture chain. If the core functional area of ​​the protective equipment moves synchronously with the key points of the human posture chain, the rate of change of displacement is 0. Take the maximum value of 1; if the core functional area of ​​the protective equipment is inconsistent with the displacement of the key points of the human posture chain, resulting in offset or separation, the displacement change rate increases. The value decreases accordingly.

[0111] Coverage retention Calculate the retention rate of the core functional area of ​​the candidate protective equipment region on the corresponding key points of the human posture chain as the human posture changes. If the protective equipment still maintains effective coverage of the corresponding key points after the human posture changes, the retention rate is 1. Take the maximum value of 1; if the human body posture changes, the protective equipment may experience abnormal stretching or displacement, leading to a decrease in the coverage of the corresponding key points and a reduction in the retention rate. The value decreases accordingly.

[0112] For example, when the worker makes a gesture of raising their hand, the model analyzes the displacement change rate and coverage retention rate of the core functional area of ​​the cuff area of ​​the insulating clothing and the arm chain to determine whether the cuff area maintains synchronous displacement and effective coverage with the key points of the forearm; when the worker makes a gesture of bending over, the model analyzes the displacement change rate and coverage retention rate of the core functional area of ​​the safety belt and the torso chain to determine whether the relative geometric relationship between the safety belt and the key points of the waist is disrupted.

[0113] The attitude consistency calculation module calculates the corresponding attitude consistency index for each candidate area of ​​protective equipment. , The initial value range is 0 to 2, and after normalization, it is mapped to the interval of 0 to 1.

[0114] In some embodiments, for step S8, the initial association score, structural feature vector, and posture consistency index are fused to obtain the functional association credibility of each candidate region of protective equipment. The three core features—initial association score, structural feature vector, and posture consistency index—are weighted and fused to generate a quantitative index that characterizes the degree of functional matching between protective equipment and human body parts. Simultaneously, a competition suppression mechanism is introduced to eliminate low-credibility associations and retain valid associations. This comprehensive approach, characterizing the functional matching degree between protective equipment and human body parts from three dimensions—initial association, structural integrity, and posture consistency—avoids erroneous binding of protective equipment to human body parts through the competition suppression mechanism, effectively reducing the false detection rate in complex work scenarios.

[0115] Specifically, the input for this step is the initial association score. Normalized structural feature vectors Normalized attitude consistency index The feature fusion and competition suppression operations are performed by the association credibility fusion module. Functional association credibility refers to the quantitative indicator that characterizes whether the association between the candidate area of ​​protective equipment and key points of the human body has real wearing significance, and is the core basis for judging the effective wearing of protective equipment.

[0116] Please see Figure 5 , Figure 5 This is a schematic diagram illustrating the calculation process of the functional association credibility index provided in the embodiments of this application. In S501, the initial association score is... Structural feature vectors Posture consistency index The functional association credibility is obtained by weighting and fusing according to preset weight coefficients. The calculation formula is:

[0117]

[0118] In the formula, , , These are the weight coefficients for the three types of features. The weighting coefficients of the initial association score represent the contribution of the initial association score to the credibility of the functional association. These are the weight coefficients of the structural feature vector, representing the degree to which structural integrity contributes to the credibility of functional association. The weighting coefficient for the posture consistency index represents the contribution of posture consistency to functional association credibility. This weighting coefficient is obtained by self-learning from the training set using the gradient descent method. During the training process, the model automatically updates the weighting coefficient based on the loss value of the wearing state classification, so that the functional association credibility can optimally represent the degree of functional matching between protective equipment and human body parts. , , The values ​​of all values ​​are between 0 and 1, and they satisfy the normalization constraint. In the scenario of uninterrupted power supply work in the distribution network, and The sum of the values ​​is greater than This highlights the core role of structural integrity and posture consistency in functional matching.

[0119] In S502, multiple candidate association competition and suppression are implemented. In complex scenarios with multiple people working together and dense tools, a candidate area for protective equipment may overlap with multiple different key points on the human body, forming multiple candidate associations. At this time, the association credibility fusion module performs competition and suppression operations to eliminate candidate associations with low credibility and retain valid associations. The specific operation logic is as follows:

[0120] Association Competition: Initial value of the functional association confidence of all candidate associations corresponding to the same candidate area of ​​protective equipment. Sort the associations in descending order to obtain a sequence of associations with the highest to lowest credibility.

[0121] Association suppression: Set the association suppression threshold This parameter is the effective threshold for judging candidate associations. It is determined by the validation set using a threshold optimization method, and its value ranges from 0.3 to 0.5. Functional associations with a confidence level lower than 0.5 in the association sequence are removed. Candidate association relationships; if the highest functional association confidence in the association sequence is still lower than If no valid human body is associated with the candidate area of ​​the protective equipment, then it is determined that there is no valid human body associated with the candidate area.

[0122] The correlation confidence fusion module will retain the highest functional correlation confidence as the final functional correlation confidence of the candidate region for the protective equipment. The value ranges from 0 to 1. The larger the value, the higher the degree of functional matching between the candidate area of ​​the protective equipment and the key points of the human body.

[0123] In some embodiments, for step S9, the wearing status of protective equipment is determined based on functional association credibility. By setting a quantification threshold to determine uncertain scenarios, a semantic hierarchy of uncertain states is introduced. Combined with functional association credibility, a four-level classification of the wearing status of protective equipment is achieved, while also structurally expressing the source of uncertainty. This enables explicit modeling and labeling of uncertain scenarios, giving the model's output a clear semantic hierarchy and interpretability, avoiding rigid judgments in ambiguous scenarios, and meeting the practical need in power safety supervision to "avoid arbitrarily reporting violations."

[0124] Specifically, the input for this step is the final functional association confidence level. Human body key point set Collection of candidate areas for protective equipment The wearing status decision module performs the wearing status determination operation. First, it determines whether there is an uncertain state based on preset conditions. Then, it combines the determination result of the uncertain state with the functional correlation credibility to determine the final wearing status of the protective equipment.

[0125] Set occlusion detection threshold Confidence fluctuation threshold , The threshold for determining the proportion of human body key points that are occluded is determined by the validation set and ranges from 0.3 to 0.6. The fluctuation threshold for the confidence level of protective equipment testing can be configured according to the testing accuracy, with a value ranging from 0.05 to 0.15. The wearing status of the candidate area of ​​the protective equipment is determined to be uncertain when any of the following conditions are met:

[0126] Detection confidence level of candidate regions for protective equipment In Within the confidence level fluctuation range, the images have quality problems such as blurriness and strong backlighting, resulting in low reliability of the test results for protective equipment.

[0127] The proportion of invisible keypoints in the total number of keypoints in the human body is greater than Or the candidate area for protective equipment is temporarily obscured by work tools or other personnel, making the core functional area of ​​the protective equipment invisible.

[0128] Attitude Consistency Index When the value is in the middle range of 0.4 to 0.6, the worker's posture is in a transitional state, and it is impossible to accurately judge the consistency between the protective equipment and the human posture.

[0129] Functional association credibility Within the middle range of 0.4 to 0.6, it is impossible to accurately distinguish between the effective and ineffective association between protective equipment and human body parts.

[0130] In one embodiment, a high confidence threshold is set. With low confidence threshold Both are threshold values ​​for classifying the wearing status of protective equipment, determined by optimization of the validation set using a precision-recall curve. The value range is from 0.2 to 0.4. The value range is from 0.6 to 0.8, and satisfies... Based on the determination results of the uncertainty state, the wearing status of protective equipment is classified into four levels. The determination logic is executed serially, first judging the uncertainty state, and then judging the functional correlation confidence level.

[0131] If the situation is determined to be uncertain, the detection confidence level of the candidate area for protective equipment should be further determined. Whether it is true or false: If the condition is true, the wearing status is determined to be that the protective equipment exists but the wearing status is uncertain; if the condition is false, the wearing status is determined to be that the detection of the protective equipment is unreliable.

[0132] If it is not determined to be an uncertain state, then it is based on the functional association credibility. The classification is determined as follows: if The wearing status is determined to be clearly and effectively worn; if The wearing status is determined to be clearly invalid.

[0133] The wearing status decision module ultimately outputs the wearing status of each candidate area of ​​protective equipment, and also outputs auxiliary information such as functional correlation credibility and uncertainty source identifier (if it is an uncertain status). All output results are structured data, which can be directly transmitted to the power distribution network live-line operation safety supervision platform for visualization and intelligent early warning.

[0134] This method, through the enhancement of a multi-level semantic constraint detection model, achieves accurate and reliable detection of the wearing status of insulating protective equipment during live-line work in distribution networks. It effectively solves core problems of traditional detection techniques, such as the inability to determine effective wearing, high false positive and false negative rates in complex scenarios, and lack of uncertainty modeling. This method upgrades from static wearing status judgment to structure-posture consistency analysis, from single spatial correlation judgment to multi-dimensional functional correlation credibility assessment, and from rigid binary judgment to four-level hierarchical judgment with explicit expression of uncertainty. It can accurately identify scenarios such as seemingly correct wearing but with structural defects, changes in work posture leading to protective failure, and accidental overlap between protective equipment and the human body. This method does not require changes to the original image acquisition method, does not rely on additional sensors, and is adaptable to complex work sites such as strong obstruction and multiple personnel crossing. The output results have clear semantic hierarchy and interpretability, and can be directly connected to the distribution network operation safety supervision platform, providing quantitative and accurate detection basis for safety supervision. This significantly improves the automation and intelligence level of safety monitoring for live-line work in distribution networks, reduces the manual cost and false alarm rate of safety supervision, and ensures the personal safety of workers.

[0135] It should be noted that although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. On the contrary, the steps depicted in the flowchart can be performed in a different order. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0136] Please see Figure 6 , Figure 6 This application provides a structural block diagram of an automatic target detection system for live-line power distribution monitoring based on image recognition, as part of an embodiment of the present application. The system specifically includes:

[0137] Image preprocessing module 601 is used to acquire images of the power distribution network live-line operation site, preprocess the images to obtain standardized images;

[0138] Feature extraction module 602, connected to the image preprocessing module, is used to input the standardized image into the shared backbone feature extraction network to extract multi-scale feature maps;

[0139] The multi-branch detection module 603 is connected to the feature extraction module and is used to input the multi-scale feature map into the protective equipment detection branch and the human body key point detection branch respectively, so as to obtain the protective equipment candidate region set and the human body key point set accordingly.

[0140] Human posture construction module 604 is connected to the multi-branch detection module and is used to construct a human posture chain based on the human key point set and according to a preset topological connection relationship, wherein the topological connection relationship is related to the function of the protective equipment.

[0141] The association scoring module 605 is connected to the multi-branch detection module and the human posture construction module respectively, and is used to calculate the initial association score between each protective equipment candidate region and each human key point based on the protective equipment candidate region set and the human key point set.

[0142] The structural analysis module 606, connected to the association scoring module, is used to analyze the structural integrity of each candidate area of ​​protective equipment based on the initial association score, and obtain a structural feature vector.

[0143] The posture consistency analysis module 607 is connected to the human posture construction module, the structure analysis module and the multi-branch detection module respectively, and is used to calculate the posture consistency index between each protective equipment candidate region and the corresponding human posture chain based on the human posture chain, the structural feature vector and the protective equipment candidate region.

[0144] The credibility fusion module 608 is connected to the association scoring module, the structural analysis module and the posture consistency analysis module respectively, and is used to fuse the initial association score, the structural feature vector and the posture consistency index to obtain the functional association credibility of each protective equipment candidate region;

[0145] The status determination module 609 is connected to the credibility fusion module and is used to determine the wearing status of protective equipment based on the functional association credibility.

[0146] Preferably, determining the wearing status of protective equipment based on the functional association confidence level includes:

[0147] Based on the visibility of key points on the human body, the detection confidence of candidate areas for protective equipment, and the credibility of the aforementioned functional association, it is determined whether there is an uncertain state.

[0148] If there is uncertainty, the wearing status is further determined by combining the detection confidence level, either the protective equipment is present but the wearing status is uncertain or the detection of the protective equipment is unreliable.

[0149] If there is no uncertainty, the wearing status is determined as either clearly valid wearing or clearly invalid wearing based on the comparison between the functional association confidence level and the preset high confidence threshold and low confidence threshold.

[0150] It should be noted that the working process of each module in the automatic detection system for monitoring targets of live-line work in distribution networks based on image recognition described in this embodiment can refer to the working process of the automatic detection method for monitoring targets of live-line work in distribution networks based on image recognition described in the above embodiment. The technical effects achieved are also the same as those of the automatic detection method for monitoring targets of live-line work in distribution networks based on image recognition described in the above embodiment, and will not be repeated here.

[0151] The above description represents the preferred embodiments of the present invention. It should be noted that, for those skilled in the art, various improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. An automatic target detection method for live-line working monitoring of power distribution networks based on image recognition, characterized in that, Includes the following steps: S1, acquire images of the power distribution network live-line work site, preprocess the images to obtain standardized images; S2, input the standardized image into the shared backbone feature extraction network to extract multi-scale feature maps; S3, input the multi-scale feature map into the protective equipment detection branch and the human body key point detection branch respectively to obtain the protective equipment candidate region set and the human body key point set accordingly; S4. Based on the set of key human body points, construct a human body posture chain according to a preset topological connection relationship, wherein the topological connection relationship is related to the function of the protective equipment; S5, calculate the initial correlation score between each candidate area of ​​protective equipment and each key point of the human body based on the set of candidate areas of protective equipment and the set of key points of the human body; S6. Based on the initial association score, analyze the structural integrity of the wearing structure of each candidate area of ​​protective equipment to obtain a structural feature vector; S7. Based on the human posture chain, the structural feature vector, and the protective equipment candidate region, calculate the posture consistency index between each protective equipment candidate region and the corresponding human posture chain. S8, the initial association score, the structural feature vector and the attitude consistency index are fused to obtain the functional association credibility of each protective equipment candidate region; S9. Determine the wearing status of protective equipment based on the functional association confidence level.

2. The automatic detection method for monitoring targets of live-line power distribution operations based on image recognition according to claim 1, characterized in that, The human posture chain includes a head and neck chain, an arm chain, and a torso chain; the head and neck chain is connected in the order of head key points, neck key points, and shoulder key points; the arm chain is connected in the order of shoulder key points, elbow key points, and hand key points; and the torso chain is connected in the order of waist key points, shoulder key points, and knee key points.

3. The automatic detection method for monitoring targets of live-line work in power distribution networks based on image recognition according to claim 2, characterized in that, Calculate the initial correlation score between each candidate area of ​​protective equipment and each key point on the human body, including: Extract the intersection-over-union (IoU) feature, distance attenuation feature, and positional consistency feature between each candidate region of protective equipment and each key point of the human body; The initial association score is obtained by weighting and summing the intersection-union ratio feature, distance decay feature, and location consistency feature.

4. The automatic detection method for monitoring targets of live-line work in distribution networks based on image recognition according to claim 3, characterized in that, Analyzing the structural integrity of each candidate region for protective equipment yields a structural feature vector, including: Analyze the corresponding functional area structure based on the category of protective equipment; For candidate helmet regions, analyze their coverage of key points on the head and the degree of chin strap closure to generate corresponding structural feature vectors. For candidate regions of insulating clothing, analyze their coverage continuity of key points of arm chains and torso chains, and generate corresponding structural feature vectors. For candidate areas of seat belts, analyze the integrity of their connection structure with key points of the waist and the effectiveness of the stress structure of the shoulder belt and waist belt, and generate corresponding structural feature vectors.

5. The automatic detection method for monitoring targets of live-line power distribution operations based on image recognition according to claim 4, characterized in that, Calculate the attitude consistency index between each candidate region of protective equipment and the corresponding human posture chain, including: Calculate the displacement change rate between the center of the core functional area of ​​the candidate area of ​​protective equipment and the corresponding key point of the human posture chain to obtain the displacement consistency. The coverage retention rate is obtained by calculating the coverage retention rate of the core functional area of ​​the candidate area of ​​protective equipment to the corresponding human posture chain key points as the posture changes. The attitude consistency index is obtained by fusing the displacement consistency and the coverage retention.

6. The automatic detection method for monitoring targets of live-line work in distribution networks based on image recognition according to claim 5, characterized in that, The initial association score, the structural feature vector, and the attitude consistency index are fused to obtain the functional association credibility of each candidate region of protective equipment, including: The initial association score, the structural feature vector, and the pose consistency index are weighted and fused according to preset weight coefficients to obtain the functional association credibility. The weight coefficients are determined through training. For multiple candidate associations formed by associating the same candidate area of ​​protective equipment with multiple key points of different human bodies, the functional association credibility of each candidate association is ranked. Candidate associations with functional association confidence scores below a preset threshold are removed, and the candidate associations with the highest confidence scores are retained as valid associations for the candidate area of ​​the protective equipment. The corresponding functional association confidence scores are used as the final functional association confidence scores.

7. The automatic detection method for monitoring targets of live-line work in distribution networks based on image recognition according to claim 6, characterized in that, Determining the wearing status of protective equipment based on the functional association confidence level includes: Based on the visibility of key points on the human body, the detection confidence of candidate areas for protective equipment, and the credibility of the aforementioned functional association, it is determined whether there is an uncertain state. If there is uncertainty, the wearing status is further determined by combining the detection confidence level, either the protective equipment is present but the wearing status is uncertain or the detection of the protective equipment is unreliable. If there is no uncertainty, the wearing status is determined as either clearly valid wearing or clearly invalid wearing based on the comparison between the functional association confidence level and the preset high confidence threshold and low confidence threshold.

8. The automatic detection method for monitoring targets of live-line work in distribution networks based on image recognition according to claim 7, characterized in that, Specifically, determining whether an uncertain state exists means that an uncertain state exists when any of the following conditions are met: The detection confidence level of the candidate region for protective equipment is within a preset confidence level fluctuation range; The proportion of invisible key points in human body key points is greater than the preset occlusion judgment threshold. The attitude consistency index is within a preset middle range; The credibility of the functional association is within a preset middle range.

9. An automatic detection system for monitoring targets during live-line work in power distribution networks based on image recognition, characterized in that, include: The image preprocessing module is used to acquire images of the power distribution network live-line work site, preprocess the images to obtain standardized images; The feature extraction module, connected to the image preprocessing module, is used to input the standardized image into the shared backbone feature extraction network to extract multi-scale feature maps. A multi-branch detection module, connected to the feature extraction module, is used to input the multi-scale feature map into the protective equipment detection branch and the human body key point detection branch respectively, and obtain the protective equipment candidate region set and the human body key point set accordingly. A human posture construction module, connected to the multi-branch detection module, is used to construct a human posture chain based on the set of human key points and according to a preset topological connection relationship, wherein the topological connection relationship is related to the function of the protective equipment. The association scoring module is connected to the multi-branch detection module and the human posture construction module respectively, and is used to calculate the initial association score between each protective equipment candidate region and each human key point based on the protective equipment candidate region set and the human key point set. The structural analysis module, connected to the association scoring module, is used to analyze the structural integrity of each candidate area of ​​protective equipment based on the initial association score, and obtain a structural feature vector. The posture consistency analysis module is connected to the human posture construction module, the structure analysis module and the multi-branch detection module respectively, and is used to calculate the posture consistency index between each protective equipment candidate region and the corresponding human posture chain based on the human posture chain, the structural feature vector and the protective equipment candidate region. The credibility fusion module is connected to the association scoring module, the structural analysis module and the posture consistency analysis module respectively, and is used to fuse the initial association score, the structural feature vector and the posture consistency index to obtain the functional association credibility of each protective equipment candidate region; The status determination module, connected to the credibility fusion module, is used to determine the wearing status of protective equipment based on the functional association credibility.

10. The automatic detection system for monitoring targets of live-line power distribution operations based on image recognition according to claim 9, characterized in that, Determining the wearing status of protective equipment based on the functional association confidence level includes: Based on the visibility of key points on the human body, the detection confidence of candidate areas for protective equipment, and the credibility of the aforementioned functional association, it is determined whether there is an uncertain state. If there is uncertainty, the wearing status is further determined by combining the detection confidence level, either the protective equipment is present but the wearing status is uncertain or the detection of the protective equipment is unreliable. If there is no uncertainty, the wearing status is determined as either clearly valid wearing or clearly invalid wearing based on the comparison between the functional association confidence level and the preset high confidence threshold and low confidence threshold.

Images