A method, apparatus and device for worker behavior compliance screening

Abstract

The application discloses a kind of job personnel behavior compliance discrimination method, device and equipment, for discriminating the compliance of field job personnel behavior.The application includes: using historical operation image is combined with preset double-ring optimization algorithm training preset initial limb feature extraction network, generates target limb feature extraction network, extracts target limb feature in real-time scene image by target limb feature extraction network, whether the behavior of job personnel corresponding to target limb feature is compliant based on the comparison result of target limb feature and preset feature range, wherein the training improvement of initial limb feature extraction network using double-ring optimization algorithm improves the robustness and acuteness of initial limb feature extraction network, so that the recognition rate of job personnel behavior compliance discrimination is improved in the case of complex environment.

Description

Technical Field

[0001] This invention relates to the field of personnel identification technology, and in particular to a method, apparatus, and equipment for identifying the compliance of operational personnel behavior. Background Technology

[0002] With the rapid development of the social economy and the continuous improvement of technology, my country's power industry has experienced rapid growth in recent years. In the power production process, strict requirements are placed on the clothing and safety equipment used by on-site workers. For example, throughout the work process, sleeves must be rolled up, trouser legs rolled up, and safety helmets must be worn properly; the chest and back must not be exposed. When using ladders to climb, other personnel must provide proper support. When multiple people are lifting power poles, they must be shoulder-to-shoulder and move in unison; and when lifting power poles, they should coordinate with each other. Therefore, it is necessary to detect and identify the compliance of workers' behavior, clothing, and equipment use during on-site operations. If non-compliance with safety measures is detected, an alarm should be issued immediately, and the personnel's images should be recorded. The identification of worker behavior compliance requires feature extraction from clothing, tools, and limbs.

[0003] However, in the power production process, most of the work is done in teams and in the field. The limbs of the workers are often obscured, interlocked, or penetrated. In addition, the surrounding residents walk back and forth, and vehicles and objects pass by from time to time, which will seriously interfere with the recognition of the workers' limb movements. The existing methods for identifying compliance of workers' behavior have low accuracy in limb feature extraction and low recognition rate of workers' behavior compliance, which poses a safety hazard. Summary of the Invention

[0004] This invention provides a method, apparatus, and equipment for identifying compliance of worker behavior, which solves the technical problem of low identification rate in complex environments.

[0005] The first aspect of this invention provides a method for identifying the compliance of worker behavior, comprising:

[0006] In response to the model training request, historical job scene videos are acquired and preprocessed to generate multiple historical image matrices;

[0007] A preset initial limb feature extraction network is trained using multiple historical image matrices combined with a preset double-loop optimization algorithm to generate a target limb feature extraction network.

[0008] When an identification request is received from the power operation monitoring system, the image of the operation area corresponding to the identification request is obtained and preprocessed to generate an image matrix;

[0009] The target limb features are extracted from the image matrix using the target limb feature extraction network.

[0010] Based on the comparison results between the target limb features and the preset feature range, it is determined whether the behavior of the operator corresponding to the target limb features is compliant.

[0011] Optionally, in response to the model training request, historical work scene videos are acquired and preprocessed to generate multiple historical image matrices, including:

[0012] Respond to model training requests and retrieve videos of historical assignment scenarios;

[0013] The historical operation scene video was cropped and scaled to obtain multiple historical human images;

[0014] Digitize the pixels in multiple historical human body images to generate multiple first historical image matrices;

[0015] The local brightness and contrast of multiple first historical image matrices are adjusted according to preset brightness and preset contrast, and the tilt angle of multiple first historical image matrices is corrected to a preset angle to generate multiple second historical image matrices.

[0016] Multiple second historical image matrices are standardized according to preset size, preset pixels and preset ratio to generate multiple historical image matrices.

[0017] Optionally, the step of training a preset initial limb feature extraction network using multiple historical image matrices combined with a preset dual-loop optimization algorithm to generate a target limb feature extraction network includes:

[0018] Select one of the multiple historical image matrices as the target historical image matrix according to a preset rule;

[0019] Historical limb feature data are extracted from the target historical image matrix using the initial limb feature extraction network.

[0020] The comprehensive optimization value is calculated using the historical limb feature data, the actual feature data corresponding to the target historical image matrix, and a preset double-loop optimization algorithm.

[0021] The parameters in the initial limb feature extraction network are adjusted using the comprehensive optimization value to generate an updated initial limb feature extraction network.

[0022] The process jumps to the step of selecting one of the multiple historical image matrices as the target historical image matrix according to a preset rule, until the comprehensive optimization value converges, thereby generating the target limb feature extraction network.

[0023] Optionally, the dual-loop optimization algorithm includes an automatic filter parameter adaptation function, an outer loop optimization function, an inner loop optimization function, and a comprehensive optimization function; the step of calculating the comprehensive optimization value using the historical limb feature data, the actual feature data corresponding to the target historical image matrix, and the preset dual-loop optimization algorithm includes:

[0024] Using the historical limb feature data and the actual feature data, the automatic adaptation filter parameters are calculated through the automatic adaptation filter parameter function;

[0025] Using the automatic filter parameters and the historical limb feature data, the outer loop optimization value is calculated through the outer loop optimization function;

[0026] Using the automatic filter parameters, the historical limb feature data, and the actual feature data, the inner loop optimization value is calculated through the inner loop optimization function;

[0027] Using the outer loop optimization value and the inner loop optimization value, the comprehensive optimization value is calculated through the comprehensive optimization function.

[0028] Optionally, the historical limb feature data includes a historical limb pixel estimation array, the actual feature data includes a historical limb pixel actual array, and the automatic adaptation filter parameter function includes a first adder, a first subtractor, a second subtractor, a third subtractor, a fourth subtractor, a first logarithmic calculation module, a second logarithmic calculation module, a first exponential calculation module, a first setting module, a second setting module, a third setting module, a fourth setting module, a fifth setting module, a first divider, a first multiplier, a second multiplier, and a third multiplier; the step of using the historical limb feature data and the actual feature data to calculate the automatic adaptation filter parameters through the automatic adaptation filter parameter function includes:

[0029] The actual array of historical limb pixels is sent to the first input of the first adder, and the estimated array of historical limb pixels is sent to the second input of the first adder for addition.

[0030] The first adjustment coefficient output by the first setting module is sent to the first input terminal of the first subtractor, and the output of the first adder is connected to the second input terminal of the first subtractor to perform subtraction operation;

[0031] The output of the first subtractor is connected to the input of the first logarithmic calculation module to perform logarithmic operations;

[0032] The output of the first adder is also connected to the input of the second logarithmic calculation module to perform logarithmic operations;

[0033] The output of the first logarithmic calculation module is connected to the first input terminal of the second subtractor, and the output of the second logarithmic calculation module is connected to the second input terminal of the second subtractor to perform subtraction operations;

[0034] The output of the second subtractor is connected to the first multiplier in two paths to perform multiplication operations;

[0035] The output of the first multiplier is connected to the first input of the second multiplier, and the second adjustment coefficient output by the second setting module is sent to the second input of the second multiplier to perform multiplication.

[0036] The output of the second multiplier is connected to the first input of the first divider, and the fourth adjustment coefficient output by the fourth setting module is sent to the second input of the first divider to perform the division operation.

[0037] The output of the first divider is connected to the input of the first exponent calculation module to perform exponentiation.

[0038] The third adjustment coefficient output by the third setting module is sent to the second input terminal of the third subtractor, and the fifth adjustment coefficient output by the fifth setting module is sent to the first input terminal of the third subtractor for subtraction operation.

[0039] The output of the first exponent calculation module is connected to the second input of the third multiplier, and the result of the third subtractor is sent to the first input of the third multiplier for multiplication.

[0040] The output of the third multiplier is connected to the second input of the fourth subtractor. The fifth adjustment coefficient output by the fifth setting module is also sent to the first input of the fourth subtractor to perform subtraction and output the automatic adaptation filter parameters.

[0041] Optionally, the historical limb feature data includes a first estimated limb joint array, a second estimated limb joint array, and estimated limb vector data corresponding to the first estimated limb joint array and the second estimated limb joint array; the actual feature data includes an actual historical limb joint array; the outer loop optimization function includes a fourth multiplier, a fifth multiplier, a seventh multiplier, a second adder, a first square root extractor, a fifth subtractor, a sixth subtractor, a second divider, a third divider, a sixth setting module, a first absolute value calculation module, a first accumulator, and a second accumulator; the step of using the automatic filter parameters and the historical limb feature data to calculate the outer loop optimization value through the outer loop optimization function includes:

[0042] The first estimated limb joint point array is sent to the fourth multiplier in two paths for multiplication.

[0043] The second estimated limb joint point array is sent to the fifth multiplier in two paths for multiplication.

[0044] The output of the fourth multiplier is connected to the first input of the second adder, and the output of the fifth multiplier is connected to the second input of the second adder to perform addition operations.

[0045] The output of the second adder is connected to the input of the first square root extractor to perform square root operations.

[0046] The first estimated limb joint point array is also sent to the first input terminal of the fifth subtractor, and the second estimated limb joint point array is also sent to the second input terminal of the fifth subtractor for subtraction operation;

[0047] The output of the first square root is connected to the second input of the second divider, and the output of the fifth subtractor is connected to the first input of the second divider to perform division operations.

[0048] The output of the second divider is connected to the input of the first accumulator for accumulation.

[0049] The first adjustment coefficient output by the sixth setting module is sent to the second input terminal of the third divider, and the output of the first accumulator is connected to the first input terminal of the third divider to perform the division operation.

[0050] The output of the third divider is connected to the second input of the sixth subtractor, and the estimated limb vector data is sent to the first input of the sixth subtractor for subtraction.

[0051] The output of the sixth subtractor is connected to the input of the first absolute value calculation module to perform absolute value calculation;

[0052] The output of the first absolute value calculation module is connected to the first input terminal of the seventh multiplier, and the automatic adaptation filter parameters are sent to the second input terminal of the seventh multiplier for multiplication.

[0053] The output of the seventh multiplier is connected to the input of the second accumulator to perform accumulation operations and output the outer loop optimized value.

[0054] Optionally, the historical limb feature data includes an array of historical joint point estimated confidence values ​​and an array of historical limb pixel detections; the actual feature data includes an array of n actual historical joint point pixels; the inner loop optimization function includes n+1 subtractors, n+1 absolute value calculation modules, n dividers, n exponent calculation modules, n setting modules, a large value comparison module, an eighth multiplier, and a third accumulator, where n is the number of limb joint points and n is a positive integer; the step of using the automatic filter parameters, the historical limb feature data, and the actual feature data to calculate the inner loop optimization value through the inner loop optimization function includes:

[0055] The historical limb pixel detection arrays are respectively sent to the first input terminals of the 1st to nth subtractors, and the n actual historical joint pixel arrays are respectively sent to the second input terminals of the corresponding subtractors for subtraction operations.

[0056] The outputs of the first to nth subtractors are connected to the input terminals of the corresponding absolute value calculation modules to perform absolute value calculations.

[0057] The outputs of the 1st to nth absolute value calculation modules are respectively connected to the first input terminal of the corresponding divider, and the adjustment coefficients output by the 1st to nth setting modules are respectively sent to the second input terminal of the corresponding divider to perform division operations.

[0058] The outputs of the first to nth dividers are connected to the inputs of the corresponding exponent calculation modules to perform exponentiation operations.

[0059] The outputs of the first to nth exponent calculation modules are respectively connected to the input of the large value comparison module to calculate the maximum value;

[0060] The output of the large value comparison module is connected to the second input of the (n+1)th subtractor, and the historical key point prediction confidence value array is sent to the first input of the (n+1)th subtractor for subtraction operation.

[0061] The output of the (n+1)th subtractor is connected to the input of the (n+1)th absolute value calculation module to perform absolute value calculation;

[0062] The output of the (n+1)th absolute value calculation module is connected to the first input of the eighth multiplier, and the automatic adaptation filter parameters are sent to the second input of the eighth multiplier for multiplication.

[0063] The output of the eighth multiplier is connected to the input of the third accumulator to perform accumulation operations and output the inner loop optimized value.

[0064] Optionally, determining whether the behavior of the worker corresponding to the target limb feature is compliant based on the comparison results between the target limb feature and a preset feature range includes:

[0065] Compare the target limb features with the preset feature range;

[0066] If the target limb feature is within a preset feature range, then the behavior of the staff member corresponding to the target limb feature is determined to be compliant.

[0067] If the target limb feature is not within the preset feature range, the behavior of the staff member corresponding to the target limb feature is determined to be non-compliant.

[0068] A second aspect of the present invention provides a device for verifying the compliance of worker behavior, comprising:

[0069] The historical image preprocessing module is used to respond to model training requests, acquire historical operation scene videos, perform preprocessing, and generate multiple historical image matrices.

[0070] The network training module is used to train a preset initial limb feature extraction network by combining multiple historical image matrices with a preset double-loop optimization algorithm, and to generate a target limb feature extraction network.

[0071] The image preprocessing module is used to acquire and preprocess the image of the work area corresponding to the identification request when a recognition request is received from the power operation monitoring system, and generate an image matrix.

[0072] The feature extraction module is used to extract target limb features from the image matrix through the target limb feature extraction network;

[0073] The compliance judgment module is used to determine whether the behavior of the operator corresponding to the target limb feature is compliant based on the comparison results between the target limb feature and the preset feature range.

[0074] A third aspect of the present invention provides an electronic device, including a memory and a processor;

[0075] The memory is used to store program code and transmit the program code to the processor;

[0076] The processor is configured to execute the operator behavior compliance identification method according to any one of the first aspects of the present invention based on the instructions in the program code.

[0077] As can be seen from the above technical solutions, the present invention has the following advantages:

[0078] This invention trains a preset initial limb feature extraction network using historical work images and a preset dual-loop optimization algorithm to generate a target limb feature extraction network. The target limb feature extraction network extracts target limb features from real-time scene images. Based on the comparison results between the target limb features and a preset feature range, it determines whether the worker's behavior corresponding to the target limb features is compliant. The dual-loop optimization algorithm improves the robustness and sensitivity of the initial limb feature extraction network, thereby increasing the recognition rate of worker behavior compliance identification in complex environments. Attached Figure Description

[0079] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0080] Figure 1 This is a flowchart illustrating the steps of a method for identifying compliance of worker behavior according to Embodiment 1 of the present invention.

[0081] Figure 2 This is a schematic diagram of the structure of the power operation supervision and training platform provided in Embodiment 1 of the present invention;

[0082] Figure 3 This is a flowchart illustrating the steps of a method for identifying compliance of worker behavior according to Embodiment 2 of the present invention.

[0083] Figure 4 This is a flowchart illustrating the calculation of the automatic filter parameter adaptation function provided in Embodiment 2 of the present invention.

[0084] Figure 5 This is a flowchart illustrating the calculation of the outer loop optimization function provided in Embodiment 2 of the present invention;

[0085] Figure 6 This is a flowchart illustrating the calculation of the inner loop optimization function provided in Embodiment 2 of the present invention;

[0086] Figure 7 This is a flowchart illustrating the calculation of the comprehensive optimization function provided in Embodiment 2 of the present invention;

[0087] Figure 8 This is a structural block diagram of a worker behavior compliance identification device provided in Embodiment 3 of the present invention. Detailed Implementation

[0088] This invention provides a method, apparatus, and equipment for identifying compliance of worker behavior, which addresses the technical problem of low identification rate in complex environments.

[0089] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0090] Please see Figure 1 , Figure 1 The flowchart illustrates the steps of a method for identifying the compliance of worker behavior according to Embodiment 1 of the present invention.

[0091] This invention provides a method for identifying compliance of worker behavior, comprising:

[0092] Step 101: Respond to the model training request, acquire historical job scene videos and preprocess them to generate multiple historical image matrices.

[0093] Understandably, preprocessing may include operations such as cropping human images from historical work scene videos, correcting human angles, adjusting the image, and standardizing, in order to form a data matrix that meets detection accuracy requirements and is suitable for image recognition, which will then be output to subsequent steps.

[0094] Step 102: Train a preset initial limb feature extraction network using multiple historical image matrices combined with a preset double-loop optimization algorithm to generate a target limb feature extraction network.

[0095] Understandably, the initial limb feature extraction network is built based on CNNs (Convolutional Neural Networks) such as VD-CNN and F-CNN. It flexibly utilizes data augmentation techniques to apply appropriate enhancement parameters to different groups of limb key nodes (hereinafter referred to as joints), eliminating differences in the resolution of different limb joints and improving the accuracy of detection under various scales, backgrounds, and lighting conditions. This allows for the detection, enhancement, and extraction of human limb feature data from images. The quality of the optimization and training results of the generalized CNN system is crucial for improving the robustness and sensitivity of the method for identifying compliance issues related to worker behavior.

[0096] It should be noted that step 102 uses a dual-loop optimization algorithm to improve the initial limb feature extraction network through offline / online training and rolling optimization adjustments, so as to improve the robustness and sensitivity of the initial limb feature extraction network and improve the recognition rate of worker behavior compliance identification in complex environments.

[0097] Optionally, step 102 can be performed on the power operation supervision training platform; please refer to [link / reference]. Figure 2 , Figure 2 This is a schematic diagram of the power operation supervision and training platform. The power operation supervision and training platform includes a physical resource layer 201, a dispatch management layer 202, a training environment layer 203, and a business application layer 204. It ensures that the intelligent recognition algorithm can run normally and stably and is ported to the edge terminal, providing users with full lifecycle management functions for model training, prediction, evaluation, and deployment.

[0098] The physical resource layer 201 includes heterogeneous computing hardware (CPU, GPU), storage, network devices, and security protection devices.

[0099] The scheduling management layer 202 is developed based on Kubernetes and Docker, and includes cluster management, resource virtualization, and task scheduling.

[0100] The training environment layer 203 is provided as a Docker service, including mainstream learning frameworks such as TensorFlow, PyTorch, Caffe, scikit-learn, and XGBoost, as well as interactive code debugging notebooks like JupyterHub and MPI parallel programming interfaces. The system runtime environment and learning environment are managed through version iteration via a Docker repository.

[0101] The business application layer 204 includes data processing, data annotation, model training, and model deployment. The model training module utilizes various machine learning and deep learning training environments, suspending training after configuring pre-written training scripts and parameters. The entire training process is automated by a backend pipeline, revolving around data processing, data annotation, training, and model management for model production. Model training uses Docker to pre-configure learning environments such as TensorFlow, PyTorch, Caffe, scikit-learn, and XGBoost. Utilizing a task scheduling system, users can submit learning task code to the cluster. The task management system allocates resources according to the user's quota, creates the user-specified environment, and adds the learning task to the task queue. The learning program runs when resources are available. Users can submit code with a single click to generate distributed tasks, significantly reducing development costs and resource consumption.

[0102] Step 103: When an identification request is received from the power operation monitoring system, the image of the operation area corresponding to the identification request is obtained and preprocessed to generate an image matrix.

[0103] Understandably, the power operation monitoring system is used to monitor and manage power operation sites in real time, and can perform functions such as verifying the compliance of personnel behavior, intelligent operation and maintenance, intelligent security alarms, and intelligent supervision.

[0104] Step 104: Extract target limb features from the image matrix using a target limb feature extraction network.

[0105] It should be noted that the target limb features extracted by the target limb feature extraction network can include model data, feature map data, confidence data, joint data, and limb vector data of each limb in the image.

[0106] Step 105: Based on the comparison results between the target limb characteristics and the preset characteristic range, determine whether the behavior of the operator corresponding to the target limb characteristics is compliant.

[0107] Understandably, the preset feature range can be set according to the corresponding standard behaviors and actions of the workers' safety behavior specifications.

[0108] In Embodiment 1 of the present invention, a preset initial limb feature extraction network is trained by combining historical work images with a preset dual-loop optimization algorithm to generate a target limb feature extraction network. The target limb feature extraction network extracts target limb features from real-time scene images. Based on the comparison results between the target limb features and a preset feature range, it is determined whether the behavior of the worker corresponding to the target limb features is compliant. The training improvement of the initial limb feature extraction network by the dual-loop optimization algorithm improves the robustness and sensitivity of the initial limb feature extraction network, thereby improving the recognition rate of worker behavior compliance identification in complex environments.

[0109] Please see Figure 3 , Figure 3 The flowchart illustrates the steps of a method for identifying the compliance of worker behavior as provided in Embodiment 2 of the present invention.

[0110] This invention provides a method for identifying compliance of worker behavior, comprising:

[0111] Step 301: Respond to the model training request, acquire historical job scene videos and preprocess them to generate multiple historical image matrices.

[0112] Optionally, step 301 includes the following sub-steps:

[0113] Respond to model training requests and retrieve videos of historical assignment scenarios;

[0114] Cropping and scaling historical operation scene videos yielded multiple historical human images;

[0115] Digitize the pixels in multiple historical human images to generate multiple first historical image matrices;

[0116] The local brightness and contrast of multiple first historical image matrices are adjusted according to preset brightness and preset contrast, and the tilt angle of multiple first historical image matrices is corrected to a preset angle to generate multiple second historical image matrices.

[0117] Multiple second historical image matrices are standardized according to preset size, preset pixels and preset ratio to generate multiple historical image matrices.

[0118] Understandably, after cropping and scaling the portion containing the human body from each image frame of a historical work scene video to obtain historical human body images, a convolutional neural network can be used to convert these historical human body images into an image matrix for subsequent image processing.

[0119] The tilt angles of multiple first historical image matrices are corrected to a preset angle to adjust the tilt angle of the human body in the images, facilitating recognition by the initial limb feature extraction network in subsequent steps. The preset angle is a pre-defined human body tilt angle.

[0120] Step 302: Select one historical image matrix from multiple historical image matrices according to preset rules as the target historical image matrix.

[0121] In this embodiment of the invention, the preset rule can be a random sampling rule such as simple random sampling, stratified sampling, systematic sampling, cluster sampling, or multi-stage sampling.

[0122] Step 303: Extract historical limb feature data from the target historical image matrix through the initial limb feature extraction network.

[0123] It should be noted that the human body in the image can be divided into multiple limb parts according to its limb structure, such as the head, torso, arms, hands, legs, and feet, and each limb part can include more than one joint. Since limb parts in the target historical image matrix are often partially occluded, the initial limb feature extraction network will generate historical limb feature data representing the entire human body by calculating and enhancing the detected uncrowded limb parts in the image. This historical limb feature data can include model data, feature map data, confidence data, joint point data, limb vector data, and pixel data for each limb in the image. Based on the occlusion status, it can be divided into detection data for uncrowded limb parts and estimated data for occluded limb parts.

[0124] Step 304: Calculate the comprehensive optimization value using historical limb feature data, actual feature data corresponding to the target historical image matrix, and a preset double-loop optimization algorithm.

[0125] It should be noted that the actual feature data refers to the actual limb feature data corresponding to the target historical image matrix, which can be obtained by manually annotating the actual location and region of the limbs and joints in the target historical image matrix.

[0126] Optionally, the dual-loop optimization algorithm includes an automatic filter parameter adaptation function, an outer loop optimization function, an inner loop optimization function, and a comprehensive optimization function; step 304 includes the following sub-steps:

[0127] S1. Using historical limb feature data and actual feature data, the automatic adaptation filter parameters are calculated through the automatic adaptation filter parameter function;

[0128] S2. Using automatic filter parameters and historical limb feature data, calculate the outer loop optimization value through the outer loop optimization function;

[0129] S3. Using automatic filter parameters, historical limb feature data, and actual feature data, calculate the inner loop optimization value through the inner loop optimization function;

[0130] S4. Using the outer loop optimization value and the inner loop optimization value, calculate the comprehensive optimization value through the comprehensive optimization function.

[0131] Optionally, the historical limb feature data includes a historical limb pixel prediction array, and the actual feature data includes a historical limb pixel actual array.

[0132] Please see Figure 4 , Figure 4 The flowchart of the calculation of the automatic filter parameter adaptation function provided in Embodiment 2 of the present invention is as follows: The automatic filter parameter adaptation function includes a first adder 401, a first subtractor 403, a second subtractor 406, a third subtractor 414, a fourth subtractor 417, a first logarithm calculation module 404, a second logarithm calculation module 405, a first exponent calculation module 412, a first setting module 402, a second setting module 409, a third setting module 413, a fourth setting module 411, a fifth setting module 415, a first divider 410, a first multiplier 407, a second multiplier 408, and a third multiplier 416; Sub-step S1 includes:

[0133] The actual array of historical limb pixels is sent to the first input terminal of the first adder 401, and the estimated array of historical limb pixels is sent to the second input terminal of the first adder 401 for addition operation;

[0134] The first adjustment coefficient output by the first setting module 402 is sent to the first input terminal of the first subtractor 403, and the output of the first adder 401 is connected to the second input terminal of the first subtractor 403 to perform subtraction operation;

[0135] The output of the first subtractor 403 is connected to the input of the first logarithmic calculation module 404 to perform logarithmic operations;

[0136] The output of the first adder 401 is also connected to the input of the second logarithmic calculation module 405 to perform logarithmic operations;

[0137] The output of the first logarithmic calculation module 404 is connected to the first input terminal of the second subtractor 406, and the output of the second logarithmic calculation module 405 is connected to the second input terminal of the second subtractor 406 to perform subtraction operations;

[0138] The output of the second subtractor 406 is connected to the first multiplier 407 in two paths to perform multiplication operations;

[0139] The output of the first multiplier 407 is connected to the first input terminal of the second multiplier 408, and the second adjustment coefficient output by the second setting module 409 is sent to the second input terminal of the second multiplier 408 to perform multiplication operations.

[0140] The output of the second multiplier 408 is connected to the first input terminal of the first divider 410, and the fourth adjustment coefficient output by the fourth setting module 411 is sent to the second input terminal of the first divider 410 to perform division operation.

[0141] The output of the first divider 410 is connected to the input of the first exponent calculation module 412 to perform exponentiation.

[0142] The third adjustment coefficient output by the third setting module 413 is sent to the second input terminal of the third subtractor 414, and the fifth adjustment coefficient output by the fifth setting module 415 is sent to the first input terminal of the third subtractor 414 to perform subtraction operation.

[0143] The output of the first exponent calculation module 412 is connected to the second input terminal of the third multiplier 416, and the operation result of the third subtractor 414 is sent to the first input terminal of the third multiplier 416 for multiplication operation.

[0144] The output of the third multiplier 416 is connected to the second input of the fourth subtractor 417. The fifth adjustment coefficient output by the fifth setting module 415 is also sent to the first input of the fourth subtractor 417 to perform subtraction operations and output automatically adapted filter parameters.

[0145] It should be noted that the historical limb pixel prediction array refers to the limb pixel prediction data output by the initial limb feature extraction network; the historical limb pixel actual array refers to the actual limb pixel data labeled in the target historical image matrix.

[0146] Optionally, the historical limb feature data includes a first estimated limb joint array, a second estimated limb joint array, and estimated limb vector data corresponding to the first estimated limb joint array and the second estimated limb joint array.

[0147] Please see Figure 5 , Figure 5This is a flowchart of the calculation of the outer loop optimization function provided in Embodiment 2 of the present invention. The outer loop optimization function includes a fourth multiplier 501, a fifth multiplier 502, a seventh multiplier 512, a second adder 503, a first square root extractor 504, a fifth subtractor 505, a sixth subtractor 510, a second divider 506, a third divider 509, a sixth setting module 508, a first absolute value calculation module 511, a first accumulator 507, and a second accumulator 513; sub-step S2 includes:

[0148] The first estimated limb joint point array is sent to the fourth multiplier 501 in two paths for multiplication.

[0149] The second estimated limb joint array is sent to the fifth multiplier 502 in two separate paths for multiplication.

[0150] The output of the fourth multiplier 501 is connected to the first input terminal of the second adder 503, and the output of the fifth multiplier 502 is connected to the second input terminal of the second adder 503 to perform addition operations.

[0151] The output of the second adder 503 is connected to the input of the first square root extractor 504 to perform square root operations;

[0152] The first estimated limb joint array is also sent to the first input terminal of the fifth subtractor 505, and the second estimated limb joint array is also sent to the second input terminal of the fifth subtractor 505 for subtraction operation.

[0153] The output of the first square root extractor 504 is connected to the second input terminal of the second divider 506, and the output of the fifth subtractor 505 is connected to the first input terminal of the second divider 506 to perform division operations.

[0154] The output of the second divider 506 is connected to the input of the first accumulator 507 for accumulation operation;

[0155] The first adjustment coefficient output by the sixth setting module 508 is sent to the second input terminal of the third divider 509, and the output of the first accumulator 507 is connected to the first input terminal of the third divider 509 to perform division operation;

[0156] The output of the third divider 509 is connected to the second input of the sixth subtractor 510, and the estimated limb vector data is sent to the first input of the sixth subtractor 510 for subtraction.

[0157] The output of the sixth subtractor 510 is connected to the input of the first absolute value calculation module 511 to perform absolute value calculation;

[0158] The output of the first absolute value calculation module 511 is connected to the first input terminal of the seventh multiplier 512, and the filter parameters are automatically adapted and sent to the second input terminal of the seventh multiplier 512 for multiplication.

[0159] The output of the seventh multiplier 512 is connected to the input of the second accumulator 513 to perform accumulation operations and output the outer loop optimized value.

[0160] It should be noted that the historical limb feature data extracted by the initial limb feature extraction network from the target historical image matrix includes multiple estimated limb joint point arrays corresponding to all relevant nodes and multiple estimated limb vector data between all joints;

[0161] The estimated limb joint array refers to the estimated data of each joint position output by the initial limb feature extraction network, including the first estimated limb joint array and the second estimated limb joint array. The first estimated limb joint array and the second estimated limb joint array refer to the estimated limb joint array corresponding to any two adjacent joints. The estimated limb vector data refers to the estimated data of the vector between each adjacent joint output by the initial limb feature extraction network. The estimated limb vector data corresponding to the first estimated limb joint array and the second estimated limb joint array refers to the vector data between the joint corresponding to the first estimated limb joint array and the joint corresponding to the second estimated limb joint array.

[0162] Since there are multiple joint combinations formed by selecting any two adjacent joints, the above calculation process for the outer loop optimization function only calculates the outer loop optimization function value for one joint combination. In reality, the outer loop optimization values ​​corresponding to all joint combinations should be calculated for use in the comprehensive optimization value calculation of subsequent steps.

[0163] Optionally, the historical limb feature data includes an array of historical joint point estimated confidence values ​​and an array of historical limb pixel detections, and the actual feature data includes an array of n actual historical joint point pixels;

[0164] Please see Figure 6 , Figure 6 The flowchart for calculating the inner loop optimization function provided in Embodiment 2 of the present invention is shown. The inner loop optimization function includes n+1 subtractors 601.1-601.n and 607, n+1 absolute value calculation modules 602.1-602.n and 608, n dividers 603.1-603.n, n exponent calculation modules 605.1-605.n, n setting modules 604.1-604.n, a large value comparison module 606, an eighth multiplier 609, and a third accumulator 610, where n is the number of limb joints and n is a positive integer; sub-step S3 includes:

[0165] The historical limb pixel detection arrays are sent to the first input terminals of the first to nth subtractors 601.1-601.n respectively, and the n actual historical joint pixel arrays are sent to the second input terminals of the corresponding subtractors 601.1-601.n respectively for subtraction operation;

[0166] The outputs of the first to nth subtractors 601.1-601.n are respectively connected to the input terminals of the corresponding absolute value calculation modules 602.1-602.n to perform absolute value calculations;

[0167] The outputs of the first to nth absolute value calculation modules 602.1-602.n are respectively connected to the first input terminal of the corresponding dividers 603.1-603.n, and the adjustment coefficients output by the first to nth setting modules 604.1-604.n are respectively sent to the second input terminal of the corresponding dividers 603.1-603.n for division operation.

[0168] The outputs of the first to nth dividers 603.1-603.n are respectively connected to the input terminals of the corresponding exponent calculation modules 605.1-605.n to perform exponentiation operations;

[0169] The outputs of the first to nth exponent calculation modules 605.1-605.n are respectively connected to the input of the large value comparison module 606 to calculate the maximum value;

[0170] The output of the large value comparison module 606 is connected to the second input of the (n+1)th subtractor 607, and the historical key point prediction confidence value array is sent to the first input of the (n+1)th subtractor 607 for subtraction operation.

[0171] The output of the (n+1)th subtractor 607 is connected to the input of the (n+1)th absolute value calculation module 608 to perform absolute value calculation;

[0172] The output of the (n+1)th absolute value calculation module 608 is connected to the first input of the eighth multiplier 609, and the filter parameters are automatically adapted and sent to the second input of the eighth multiplier 609 for multiplication.

[0173] The output of the eighth multiplier 609 is connected to the input of the third accumulator 610 to perform accumulation operations and output the inner loop optimized value.

[0174] It should be noted that the n joints in this step belong to the same limb part. The calculation process of the inner loop optimization function above only calculates the inner loop optimization function value for one limb part. In fact, the inner loop optimization values ​​corresponding to all limb parts should be calculated for the comprehensive optimization value calculation in subsequent steps. The historical joint estimated confidence value array refers to the estimated data of the confidence values ​​of all relevant nodes at each pixel output by the initial limb feature extraction network. The historical limb pixel detection array refers to the limb pixel detection data output by the initial limb feature extraction network. The actual historical joint pixel array refers to the actual joint pixel array labeled in the target historical image matrix.

[0175] Optionally, please refer to Figure 7 , Figure 7 This is a flowchart of the calculation of the comprehensive optimization function provided in Embodiment 2 of the present invention. The comprehensive optimization function includes a fourth accumulator 701, a fifth accumulator 704, a third adder 702, a fourth adder 705, a fifth adder 707, a seventh setting module 703, and an eighth setting module 706; sub-step S4 includes:

[0176] The optimized value of the outer loop is sent to the input of the fourth accumulator 701 for accumulation.

[0177] The output of the fourth accumulator 701 is connected to the first input terminal of the third adder 702, and the seventh adjustment coefficient output by the seventh setting module 703 is sent to the second input terminal of the third adder 702 for addition operation.

[0178] The inner loop optimized value is sent to the input of the fifth accumulator 704 for accumulation.

[0179] The output of the fifth accumulator 704 is connected to the first input of the fourth adder 705, and the eighth adjustment coefficient output by the eighth setting module 706 is sent to the second input of the fourth adder 705 for addition operation.

[0180] The output of the third adder 702 is connected to the first input of the fifth adder 707, and the output of the fourth adder 705 is connected to the second input of the fifth adder 707 to perform addition operations and output a comprehensive optimized value.

[0181] It should be noted that the inputs in the calculation process of the above comprehensive optimization function are the outer loop optimization values ​​corresponding to all joint combinations and the inner loop optimization values ​​corresponding to all limb parts.

[0182] Step 305: Adjust the parameters in the initial limb feature extraction network using the comprehensive optimization value to generate the updated initial limb feature extraction network.

[0183] It should be noted that after calculating the comprehensive optimized value, the parameters in the initial limb feature extraction network can be adjusted through backpropagation using the Adaptive Estimates method.

[0184] Step 306: Proceed to step 302 until the comprehensive optimization value converges, generating the target limb feature extraction network.

[0185] It is understandable that steps 302 to 306 are the training process for obtaining the target limb feature extraction network from the initial limb feature extraction network. A dual-loop optimization algorithm is used to improve the initial limb feature extraction network through offline / online training and rolling optimization. The inner loop optimization value represents the maximum confidence between the output of the initial limb feature extraction network and the actual limb joint pixel data, while the outer loop optimization value represents the limb vector matrix between joints. By adjusting the parameters of the initial limb feature extraction network, the recognition error of the initial limb feature extraction network is optimized.

[0186] Step 307: When an identification request is received from the power operation monitoring system, the image of the operation area corresponding to the identification request is obtained and preprocessed to generate an image matrix.

[0187] Step 308: Extract target limb features from the image matrix using a target limb feature extraction network;

[0188] Step 309: Based on the comparison results between the target limb characteristics and the preset characteristic range, determine whether the behavior of the operator corresponding to the target limb characteristics is compliant.

[0189] Optionally, step 309 includes the following sub-steps:

[0190] Compare the target limb features with the preset feature range;

[0191] If the target body features fall within the preset feature range, the behavior of the staff member corresponding to the target body features is deemed compliant.

[0192] If the target body features are not within the preset feature range, the behavior of the staff member corresponding to the target body features is determined to be non-compliant.

[0193] In Embodiment 2 of the present invention, a preset initial limb feature extraction network is trained using historical work images combined with a preset dual-loop optimization algorithm to generate a target limb feature extraction network. The target limb feature extraction network extracts target limb features from real-time scene images. Based on the comparison results between the target limb features and a preset feature range, it is determined whether the behavior of the worker corresponding to the target limb features is compliant. The dual-loop optimization algorithm is used to train and improve the initial limb feature extraction network. The outer loop optimization value represents the maximum confidence of the output of the initial limb feature extraction network with the actual limb joint data, and the inner loop optimization value represents the limb vector matrix between joints. By adjusting the parameters of the initial limb feature extraction network, the recognition error of the initial limb feature extraction network is optimized, and the robustness and sensitivity of the initial limb feature extraction network are improved, thereby improving the recognition rate of worker behavior compliance identification in complex environments.

[0194] Please see Figure 8 , Figure 8 This is a structural block diagram of a worker behavior compliance identification device provided in Embodiment 3 of the present invention.

[0195] The present invention provides a device for identifying compliance of worker behavior, comprising:

[0196] The historical image preprocessing module 801 is used to respond to model training requests, acquire historical operation scene videos and preprocess them to generate multiple historical image matrices.

[0197] The network training module 802 is used to train a preset initial limb feature extraction network by combining multiple historical image matrices with a preset double-loop optimization algorithm, and to generate a target limb feature extraction network.

[0198] Image preprocessing module 803 is used to acquire and preprocess the image of the work area corresponding to the identification request when the power operation monitoring system receives the identification request, and generate an image matrix.

[0199] Feature extraction module 804 is used to extract target limb features from the image matrix through a target limb feature extraction network;

[0200] The compliance judgment module 805 is used to determine whether the behavior of the operator corresponding to the target limb characteristics is compliant based on the comparison results between the target limb characteristics and the preset characteristic range.

[0201] Optionally, the historical image preprocessing module 801 is specifically used for:

[0202] Respond to model training requests and retrieve videos of historical assignment scenarios;

[0203] Cropping and scaling historical operation scene videos yielded multiple historical human images;

[0204] Digitize the pixels in multiple historical human images to generate multiple first historical image matrices;

[0205] The local brightness and contrast of multiple first historical image matrices are adjusted according to preset brightness and preset contrast, and the tilt angle of multiple first historical image matrices is corrected to a preset angle to generate multiple second historical image matrices.

[0206] Multiple second historical image matrices are standardized according to preset size, preset pixels and preset ratio to generate multiple historical image matrices.

[0207] Optionally, the network training module 802 includes the following sub-modules:

[0208] The target selection submodule is used to select one historical image matrix from multiple historical image matrices according to preset rules;

[0209] The feature extraction submodule is used to extract historical limb feature data from the target historical image matrix through the initial limb feature extraction network;

[0210] The dual-ring optimization calculation submodule is used to calculate the comprehensive optimization value using historical limb feature data, actual feature data corresponding to the target historical image matrix, and a preset dual-ring optimization algorithm.

[0211] The parameter adjustment submodule is used to adjust the parameters in the initial limb feature extraction network using comprehensive optimization values, and to generate an updated initial limb feature extraction network.

[0212] The network training submodule is used to jump to the target selection submodule to perform the corresponding steps until the comprehensive optimization value converges, generating the target limb feature extraction network.

[0213] Optionally, the dual-loop optimization algorithm includes an automatic filter parameter adaptation function, an outer-loop optimization function, an inner-loop optimization function, and a comprehensive optimization function; the dual-loop optimization calculation submodule includes the following units:

[0214] The automatic adaptation filter parameter calculation unit is used to calculate the automatic adaptation filter parameters by using historical limb feature data and actual feature data through the automatic adaptation filter parameter function.

[0215] The outer loop optimization value calculation unit is used to calculate the outer loop optimization value using automatic filter parameters and historical limb feature data through the outer loop optimization function;

[0216] The inner loop optimization value calculation unit is used to calculate the inner loop optimization value using automatic filter parameters, historical limb feature data and actual feature data, through the inner loop optimization function.

[0217] The comprehensive optimization value calculation unit is used to calculate the comprehensive optimization value by using the outer loop optimization value and the inner loop optimization value through the comprehensive optimization function.

[0218] Optionally, the compliance assessment module 805 includes the following sub-modules:

[0219] The data comparison submodule is used to compare the target limb features with preset feature ranges;

[0220] The first judgment submodule is used to determine that the behavior of the staff member corresponding to the target limb feature is compliant if the target limb feature is within the preset feature range.

[0221] The second judgment submodule is used to determine that the behavior of the staff member corresponding to the target limb feature is non-compliant if the target limb feature is not within the preset feature range.

[0222] This invention also provides an electronic device, including a memory and a processor. The memory stores program code and transmits the program code to the processor. The processor executes a method for verifying the compliance of operator behavior according to the instructions in the program code.

[0223] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, modules, sub-modules and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0224] In the several embodiments provided in this application, it should be understood that the disclosed apparatus, modules, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0225] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0226] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0227] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0228] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying the compliance of worker behavior, characterized in that, include: In response to the model training request, historical job scene videos are acquired and preprocessed to generate multiple historical image matrices; A preset initial limb feature extraction network is trained using multiple historical image matrices combined with a preset double-loop optimization algorithm to generate a target limb feature extraction network. When an identification request is received from the power operation monitoring system, the image of the operation area corresponding to the identification request is obtained and preprocessed to generate an image matrix; The target limb features are extracted from the image matrix using the target limb feature extraction network. Based on the comparison results between the target limb features and the preset feature range, it is determined whether the behavior of the operator corresponding to the target limb features is compliant; The process of training a preset initial limb feature extraction network using multiple historical image matrices combined with a preset double-loop optimization algorithm to generate a target limb feature extraction network includes: Select one of the multiple historical image matrices as the target historical image matrix according to a preset rule; Historical limb feature data are extracted from the target historical image matrix using the initial limb feature extraction network. The comprehensive optimization value is calculated using the historical limb feature data, the actual feature data corresponding to the target historical image matrix, and a preset double-loop optimization algorithm. The parameters in the initial limb feature extraction network are adjusted using the comprehensive optimization value to generate an updated initial limb feature extraction network. The process jumps to the step of selecting one of the multiple historical image matrices as the target historical image matrix according to a preset rule until the comprehensive optimization value converges, thereby generating the target limb feature extraction network. The dual-loop optimization algorithm includes an automatic filter parameter adaptation function, an outer loop optimization function, an inner loop optimization function, and a comprehensive optimization function; the step of calculating the comprehensive optimization value using the historical limb feature data, the actual feature data corresponding to the target historical image matrix, and the preset dual-loop optimization algorithm includes: Using the historical limb feature data and the actual feature data, the automatic adaptation filter parameters are calculated through the automatic adaptation filter parameter function; Using the automatically adaptable filter parameters and the historical limb feature data, the outer loop optimization value is calculated through the outer loop optimization function; Using the automatic adaptation filter parameters, the historical limb feature data, and the actual feature data, the inner loop optimization value is calculated through the inner loop optimization function; Using the outer loop optimization value and the inner loop optimization value, the comprehensive optimization value is calculated through the comprehensive optimization function.

2. The method for verifying the compliance of worker behavior according to claim 1, characterized in that, The response model training request acquires historical job scene videos and preprocesses them to generate multiple historical image matrices, including: Respond to model training requests and retrieve videos of historical assignment scenarios; The historical operation scene video was cropped and scaled to obtain multiple historical human images; Digitize the pixels in multiple historical human body images to generate multiple first historical image matrices; The local brightness and contrast of multiple first historical image matrices are adjusted according to preset brightness and preset contrast, and the tilt angle of multiple first historical image matrices is corrected to a preset angle to generate multiple second historical image matrices. Multiple second historical image matrices are standardized according to preset size, preset pixels and preset ratio to generate multiple historical image matrices.

3. The method for verifying the compliance of worker behavior according to claim 1, characterized in that, The historical limb feature data includes a historical limb pixel estimation array, the actual feature data includes a historical limb pixel actual array, and the automatic adaptation filter parameter function includes a first adder, a first subtractor, a second subtractor, a third subtractor, a fourth subtractor, a first logarithmic calculation module, a second logarithmic calculation module, a first exponential calculation module, a first setting module, a second setting module, a third setting module, a fourth setting module, a fifth setting module, a first divider, a first multiplier, a second multiplier, and a third multiplier; the step of using the historical limb feature data and the actual feature data to calculate the automatic adaptation filter parameters through the automatic adaptation filter parameter function includes: The actual array of historical limb pixels is sent to the first input of the first adder, and the estimated array of historical limb pixels is sent to the second input of the first adder for addition. The first adjustment coefficient output by the first setting module is sent to the first input terminal of the first subtractor, and the output of the first adder is connected to the second input terminal of the first subtractor to perform subtraction operation; The output of the first subtractor is connected to the input of the first logarithmic calculation module to perform logarithmic operations; The output of the first adder is also connected to the input of the second logarithmic calculation module to perform logarithmic operations; The output of the first logarithmic calculation module is connected to the first input terminal of the second subtractor, and the output of the second logarithmic calculation module is connected to the second input terminal of the second subtractor to perform subtraction operations; The output of the second subtractor is connected to the first multiplier in two paths to perform multiplication operations; The output of the first multiplier is connected to the first input of the second multiplier, and the second adjustment coefficient output by the second setting module is sent to the second input of the second multiplier to perform multiplication. The output of the second multiplier is connected to the first input of the first divider, and the fourth adjustment coefficient output by the fourth setting module is sent to the second input of the first divider to perform the division operation. The output of the first divider is connected to the input of the first exponent calculation module to perform exponentiation. The third adjustment coefficient output by the third setting module is sent to the second input terminal of the third subtractor, and the fifth adjustment coefficient output by the fifth setting module is sent to the first input terminal of the third subtractor for subtraction operation. The output of the first exponent calculation module is connected to the second input of the third multiplier, and the result of the third subtractor is sent to the first input of the third multiplier for multiplication. The output of the third multiplier is connected to the second input of the fourth subtractor. The fifth adjustment coefficient output by the fifth setting module is also sent to the first input of the fourth subtractor to perform subtraction and output the automatic adaptation filter parameters.

4. The method for verifying the compliance of worker behavior according to claim 1, characterized in that, The historical limb feature data includes a first estimated limb joint point array, a second estimated limb joint point array, and estimated limb vector data corresponding to the first estimated limb joint point array and the second estimated limb joint point array. The outer loop optimization function includes a fourth multiplier, a fifth multiplier, a seventh multiplier, a second adder, a first square root extractor, a fifth subtractor, a sixth subtractor, a second divider, a third divider, a sixth setting module, a first absolute value calculation module, a first accumulator, and a second accumulator. The calculation of the outer loop optimization value using the automatic adaptation filter parameters and the historical limb feature data through the outer loop optimization function includes: The first estimated limb joint point array is sent to the fourth multiplier in two paths for multiplication. The second estimated limb joint point array is sent to the fifth multiplier in two paths for multiplication. The output of the fourth multiplier is connected to the first input of the second adder, and the output of the fifth multiplier is connected to the second input of the second adder to perform addition operations. The output of the second adder is connected to the input of the first square root extractor to perform square root operations. The first estimated limb joint point array is also sent to the first input terminal of the fifth subtractor, and the second estimated limb joint point array is also sent to the second input terminal of the fifth subtractor for subtraction operation; The output of the first square root is connected to the second input of the second divider, and the output of the fifth subtractor is connected to the first input of the second divider to perform division operations. The output of the second divider is connected to the input of the first accumulator for accumulation. The first adjustment coefficient output by the sixth setting module is sent to the second input terminal of the third divider, and the output of the first accumulator is connected to the first input terminal of the third divider to perform the division operation. The output of the third divider is connected to the second input of the sixth subtractor, and the estimated limb vector data is sent to the first input of the sixth subtractor for subtraction. The output of the sixth subtractor is connected to the input of the first absolute value calculation module to perform absolute value calculation; The output of the first absolute value calculation module is connected to the first input terminal of the seventh multiplier, and the automatic adaptation filter parameters are sent to the second input terminal of the seventh multiplier for multiplication. The output of the seventh multiplier is connected to the input of the second accumulator to perform accumulation operations and output the outer loop optimized value.

5. The method for verifying the compliance of worker behavior according to claim 1, characterized in that, The historical limb feature data includes an array of historical joint point estimated confidence values ​​and an array of historical limb pixel detections. The actual feature data includes an array of n actual historical joint point pixels. The inner loop optimization function includes n+1 subtractors, n+1 absolute value calculation modules, n dividers, n exponent calculation modules, n setting modules, a large value comparison module, an eighth multiplier, and a third accumulator, where n is the number of limb joint points and n is a positive integer. The calculation of the inner loop optimization value using the automatic adaptation filter parameters, the historical limb feature data, and the actual feature data through the inner loop optimization function includes: The historical limb pixel detection arrays are respectively sent to the first input terminals of the 1st to nth subtractors, and the n actual historical joint pixel arrays are respectively sent to the second input terminals of the corresponding subtractors for subtraction operations. The outputs of the first to nth subtractors are connected to the input terminals of the corresponding absolute value calculation modules to perform absolute value calculations. The outputs of the 1st to nth absolute value calculation modules are respectively connected to the first input terminal of the corresponding divider, and the adjustment coefficients output by the 1st to nth setting modules are respectively sent to the second input terminal of the corresponding divider to perform division operations. The outputs of the first to nth dividers are connected to the inputs of the corresponding exponent calculation modules to perform exponentiation operations. The outputs of the first to nth exponent calculation modules are respectively connected to the input of the large value comparison module to calculate the maximum value; The output of the large value comparison module is connected to the second input of the (n+1)th subtractor, and the historical key point prediction confidence value array is sent to the first input of the (n+1)th subtractor for subtraction operation. The output of the (n+1)th subtractor is connected to the input of the (n+1)th absolute value calculation module to perform absolute value calculation; The output of the (n+1)th absolute value calculation module is connected to the first input of the eighth multiplier, and the automatic adaptation filter parameters are sent to the second input of the eighth multiplier for multiplication. The output of the eighth multiplier is connected to the input of the third accumulator to perform accumulation operations and output the inner loop optimized value.

6. The method for verifying the compliance of worker behavior according to claim 1, characterized in that, The determination of whether the behavior of the worker corresponding to the target limb feature is compliant, based on the comparison results between the target limb feature and the preset feature range, includes: Compare the target limb features with the preset feature range; If the target limb feature is within a preset feature range, then the behavior of the staff member corresponding to the target limb feature is determined to be compliant. If the target limb feature is not within the preset feature range, the behavior of the staff member corresponding to the target limb feature is determined to be non-compliant.

7. A device for verifying the compliance of worker behavior, characterized in that, include: The historical image preprocessing module is used to respond to model training requests, acquire historical operation scene videos, perform preprocessing, and generate multiple historical image matrices. The network training module is used to train a preset initial limb feature extraction network by combining multiple historical image matrices with a preset double-loop optimization algorithm, and to generate a target limb feature extraction network. The image preprocessing module is used to acquire and preprocess the image of the work area corresponding to the identification request when a recognition request is received from the power operation monitoring system, and generate an image matrix. The feature extraction module is used to extract target limb features from the image matrix through the target limb feature extraction network; The compliance judgment module is used to determine whether the behavior of the operator corresponding to the target limb feature is compliant based on the comparison results between the target limb feature and the preset feature range; The network training module includes: The target selection submodule is used to select one of the multiple historical image matrices as the target historical image matrix according to a preset rule. The feature extraction submodule is used to extract historical limb feature data from the target historical image matrix through the initial limb feature extraction network; The dual-ring optimization calculation submodule is used to calculate the comprehensive optimization value using the historical limb feature data, the actual feature data corresponding to the target historical image matrix, and the preset dual-ring optimization algorithm. The parameter adjustment submodule is used to adjust the parameters in the initial limb feature extraction network using the comprehensive optimization value, and generate an updated initial limb feature extraction network. The network training submodule is used to jump to the target selection submodule to perform corresponding steps until the comprehensive optimization value converges, thereby generating the target limb feature extraction network. The dual-loop optimization algorithm includes an automatic filter parameter adaptation function, an outer loop optimization function, an inner loop optimization function, and a comprehensive optimization function; the dual-loop optimization calculation submodule includes: An automatic adaptation filter parameter calculation unit is used to calculate the automatic adaptation filter parameters using the historical limb feature data and the actual feature data through the automatic adaptation filter parameter function. The outer loop optimization value calculation unit is used to calculate the outer loop optimization value using the automatic adaptation filter parameters and the historical limb feature data through the outer loop optimization function; The inner loop optimization value calculation unit is used to calculate the inner loop optimization value by using the automatic adaptation filter parameters, the historical limb feature data, and the actual feature data, through the inner loop optimization function. The comprehensive optimization value calculation unit is used to calculate the comprehensive optimization value using the outer loop optimization value and the inner loop optimization value through the comprehensive optimization function.

8. An electronic device, characterized in that, include: Memory and processor; The memory is used to store program code and transmit the program code to the processor; The processor is configured to execute the operator behavior compliance identification method according to any one of claims 1 to 6, based on the instructions in the program code.

Images